diff --git a/cvs/core/orchestrators/container.py b/cvs/core/orchestrators/container.py
index 0d1d2026b..7f6c5c207 100644
--- a/cvs/core/orchestrators/container.py
+++ b/cvs/core/orchestrators/container.py
@@ -477,11 +477,16 @@ def setup_sshd(self):
             "chown -R root:root /root/.ssh",
             "bash -c 'chmod 700 /root/.ssh && chmod 600 /root/.ssh/*'",
             "mkdir -p /run/sshd",  # Create privilege separation directory for sshd
+            "bash -c 'which sshd > /dev/null 2>&1 || (apt-get update -qq && apt-get install -y -q openssh-server iproute2)'",
             "/usr/sbin/sshd -p2224",
         ]
 
+        timeouts = {
+            "bash -c 'which sshd > /dev/null 2>&1 || (apt-get update -qq && apt-get install -y -q openssh-server iproute2)'": 120,
+        }
+
         for cmd in ssh_setup_commands:
-            result = self.exec(cmd, timeout=10, detailed=True)
+            result = self.exec(cmd, timeout=timeouts.get(cmd, 10), detailed=True)
             # Check if command succeeded on all hosts
             for hostname, output in result.items():
                 if output['exit_code'] != 0:
diff --git a/cvs/input/cluster_file/mi300x_atom_multi.json b/cvs/input/cluster_file/mi300x_atom_multi.json
new file mode 100644
index 000000000..1ca3c32b2
--- /dev/null
+++ b/cvs/input/cluster_file/mi300x_atom_multi.json
@@ -0,0 +1,34 @@
+{
+  "_comment": "2-node MI300X example for inferencex_atom multinode (M5). Replace mgmt_ip / node_dict with lab IPs. head_node_dict.mgmt_ip should match the rank-0 (master) node VPC IP used in params.master_addr.",
+  "orchestrator": "container",
+  "username": "{user-id}",
+  "priv_key_file": "/home/{user-id}/.ssh/id_rsa",
+  "head_node_dict": {
+    "mgmt_ip": "{head-node-ip}"
+  },
+  "env_vars": {},
+  "node_dict": {
+    "{head-node-ip}": {
+      "bmc_ip": "NA",
+      "vpc_ip": "{head-node-ip}"
+    },
+    "{worker-node-ip}": {
+      "bmc_ip": "NA",
+      "vpc_ip": "{worker-node-ip}"
+    }
+  },
+  "container": {
+    "lifetime": "per_run",
+    "image": "rocm/atom-dev:latest",
+    "name": "inferencex_atom_mi300x_multi",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G"
+      }
+    }
+  }
+}
diff --git a/cvs/input/cluster_file/mi300x_atom_single.json b/cvs/input/cluster_file/mi300x_atom_single.json
index a57b158cd..f8b2cd835 100644
--- a/cvs/input/cluster_file/mi300x_atom_single.json
+++ b/cvs/input/cluster_file/mi300x_atom_single.json
@@ -1,5 +1,5 @@
 {
-  "_comment": "Single-node MI300X example for inferencex_atom_single (ATOM driver). Replace mgmt_ip / node_dict with lab IPs.",
+  "_comment": "Single-node MI300X example for inferencex_atom (ATOM driver). Replace mgmt_ip / node_dict with lab IPs.",
   "orchestrator": "container",
   "username": "{user-id}",
   "priv_key_file": "/home/{user-id}/.ssh/id_rsa",
diff --git a/cvs/input/cluster_file/mi355x_atom_multi.json b/cvs/input/cluster_file/mi355x_atom_multi.json
new file mode 100644
index 000000000..a577addce
--- /dev/null
+++ b/cvs/input/cluster_file/mi355x_atom_multi.json
@@ -0,0 +1,34 @@
+{
+  "_comment": "2-node MI355X example for inferencex_atom multinode (M5). Replace mgmt_ip / node_dict with lab IPs. head_node_dict.mgmt_ip should match the rank-0 (master) node VPC IP used in params.master_addr.",
+  "orchestrator": "container",
+  "username": "{user-id}",
+  "priv_key_file": "/home/{user-id}/.ssh/id_rsa",
+  "head_node_dict": {
+    "mgmt_ip": "{head-node-ip}"
+  },
+  "env_vars": {},
+  "node_dict": {
+    "{head-node-ip}": {
+      "bmc_ip": "NA",
+      "vpc_ip": "{head-node-ip}"
+    },
+    "{worker-node-ip}": {
+      "bmc_ip": "NA",
+      "vpc_ip": "{worker-node-ip}"
+    }
+  },
+  "container": {
+    "lifetime": "per_run",
+    "image": "rocm/atom-dev:nightly_202606211542",
+    "name": "inferencex_atom_mi355x_multi",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G"
+      }
+    }
+  }
+}
diff --git a/cvs/input/cluster_file/mi355x_atom_single.json b/cvs/input/cluster_file/mi355x_atom_single.json
index 0ea1e0eaf..3ca015048 100644
--- a/cvs/input/cluster_file/mi355x_atom_single.json
+++ b/cvs/input/cluster_file/mi355x_atom_single.json
@@ -1,5 +1,5 @@
 {
-  "_comment": "Single-node MI355X example for inferencex_atom_single (ATOM driver). Pin image to match variant config / ATOM CI run.",
+  "_comment": "Single-node MI355X example for inferencex_atom (ATOM driver). Pin image to match variant config / ATOM CI run.",
   "orchestrator": "container",
   "username": "{user-id}",
   "priv_key_file": "/home/{user-id}/.ssh/id_rsa",
diff --git a/cvs/input/config_file/inference/inferencex_atom/README.md b/cvs/input/config_file/inference/inferencex_atom/README.md
new file mode 100644
index 000000000..470c64838
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/README.md
@@ -0,0 +1,370 @@
+# InferenceX ATOM variants
+
+W1 **DeepSeek R1 FP8** on 8× GPU, ISL=OSL=1024, TP8.
+
+## Layout
+
+**In the CVS repo**, all variants live as flat sibling pairs in **this directory**:
+
+```text
+{gpu}_inferencex-atom-single_{model}_{precision}[_{mode}]_config.json
+{gpu}_inferencex-atom-single_{model}_{precision}[_{mode}]_threshold.json
+```
+
+**On your lab machine** (`~/input/config_file/inference/inferencex_atom/`), copy each variant into its **own subdirectory** so only one `*threshold.json` sits next to the config you pass to `--config_file`. `substitute_config` globs the config's parent directory; multiple `*threshold.json` files there raises `ValueError: multiple *threshold.json files … (ambiguous)`.
+
+```text
+~/input/.../inferencex_atom/smoke/   # smoke config + smoke threshold only
+~/input/.../inferencex_atom/perf/    # perf config + perf threshold only
+```
+
+Each shipped config sets `"threshold_json"` to the sibling threshold filename (resolved relative to the config directory). You may also use an absolute path (vLLM-style).
+
+Legacy nested layouts (`deepseek_r1_fp8_mi300x_atom_perf/`, `inferencemax/`, etc.) are **removed** from the repo tree. Use only the flat stems below.
+
+**Config filename example:** `mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json`
+
+| Variant | GPU | Notes |
+|---------|-----|-------|
+| `mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke` | MI300X | Quick path check (C=128, 128 prompts) |
+| `mi300x_inferencex-atom-single_deepseek-r1_fp8_perf` | MI300X | W1 perf, portable min-SLO thresholds, server reuse across sweep |
+| `mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf` | MI300X | **M4 vLLM parity** sibling of W1 perf (`framework: inferencex_atom_vllm_single`) |
+| `mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi` | MI300X | **M4 vLLM parity** 2-node (`nnodes=2`, `PP=2`); scaling + optional `compare.vllm.*` |
+| `mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi` | MI300X | W1 **2-node** scaling (`nnodes=2`, `PP=2`); enforced thresholds + scaling gate |
+| `mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3` | MI300X | W1 FP8+MTP3 |
+| `mi355x_inferencex-atom-single_deepseek-r1_fp8_perf` | MI355X | W1 perf (CI seeds, `enforce_thresholds: false`) |
+| `mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi` | MI355X | W1 **2-node** scaling (`nnodes=2`, `PP=2`); `enforce_thresholds: false` until lab confirm |
+| `mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi` | MI355X | **M4 vLLM parity** 2-node; `enforce_thresholds: false` until lab confirm |
+| `mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3` | MI355X | W1 FP8+MTP3 |
+| `mi300x_inferencex-atom-single_gpt-oss-120b_bf16` | MI300X | GPT-OSS uplift placeholder (`driver: vllm`, inline `serve_args`) |
+| `mi355x_inferencex-atom-single_gpt-oss-120b_bf16` | MI355X | GPT-OSS uplift placeholder |
+
+ATOM server CLI is inline in each config under `roles.server.atom_args` (vLLM-style, same as `roles.server.serve_args` on `vllm_single`). MTP3 variants also set `params.bench_extra_args`.
+
+## Cluster + container naming
+
+Use `cvs/input/cluster_file/mi300x_atom_single.json` or `mi355x_atom_single.json` for single-node runs. For **multinode**, use `mi300x_atom_multi.json` or `mi355x_atom_multi.json` with two entries in `node_dict` and set `params.master_addr` in the variant config to the head node's VPC IP. `params.nnodes` must match the cluster host count; `test_setup_sshd` starts in-container sshd on port 2224 when `len(node_dict) > 1`. The cluster `container.name` must match the variant (`inferencex_atom_mi300x` / `inferencex_atom_mi355x` / `inferencex_atom_mi300x_multi` / `inferencex_atom_mi355x_multi` / `inferencex_atom_vllm_mi300x` / `inferencex_atom_vllm_mi300x_multi` / `inferencex_atom_vllm_mi355x_multi`); the suite deep-merges variant container settings over the cluster file.
+
+## Shared suite helpers (reusable by other inference suites)
+
+| Module | Purpose |
+|--------|---------|
+| `cvs/lib/inference/inference_suite_lifecycle.py` | Lifecycle stage tests, `InferenceLifecycle`, pytest HTML hooks |
+| `cvs/lib/inference/inference_suite_results_table.py` | Configurable results table (`make_print_results_table`) |
+| `cvs/lib/inference/unittests/fake_orch.py` | `FakeOrch` for Job parse unit tests |
+
+`inferencex_atom` imports these today; `vllm_single` may adopt them in a follow-up without duplicating code.
+
+## Pytest layout
+
+1. `test_launch_container` → `test_setup_sshd` → `test_model_fetch`
+2. `test_inferencex_atom_inference` (per sweep cell; reuses server when `reuse_server_across_sweep: true`)
+3. `test_cell_metrics` (one HTML row per **metric tier** per cell: throughput, ttft, tpot, health, record)
+4. `test_print_results_table` → `test_teardown`
+
+W1 MI300X perf with two concurrency cells expects **~17** pytest rows (not one row per scalar metric).
+
+## Before the first lab run
+
+- On the **launcher** host after `git checkout` / `git pull`: run **`make install` first**, then **`source .cvs_venv/bin/activate`**. Do not activate `.cvs_venv` before `make install` — the Makefile manages that venv and install can fail if it is already active.
+
+```bash
+cd ~/cvs
+git fetch origin hnimrama/ix-atom-multinode
+git reset --hard origin/hnimrama/ix-atom-multinode
+make install
+source .cvs_venv/bin/activate
+```
+
+- Edit `~/input/cluster_file/mi300x_atom_single.json` (or `mi355x_atom_single.json`): node IPs, `username`, `priv_key_file`, `container.image`.
+- **Launcher vs GPU node:** CVS pytest runs on the launcher; `ContainerOrchestrator` SSHes to cluster nodes and runs `sudo docker` there. Local Docker on the launcher is not used. Prerequisites split by host:
+
+  | Item | Launcher | GPU node (cluster `mgmt_ip`) |
+  |------|----------|------------------------------|
+  | `cvs run`, venv, `~/input/`, `~/cvs_results/` | Yes | No |
+  | `priv_key_file`, `~/.hf_token` (read locally by pytest) | Yes | No |
+  | `~/models` (when `model.remote: 0`) | No | Yes |
+  | `rocm/atom-dev` image, `sudo docker` | No | Yes |
+  | `~/LOGS/` (server/bench logs via volume mount) | No | Yes |
+
+- Preflight from the launcher: `ssh -i ~/.ssh/<key> <user>@<mgmt_ip> 'sudo docker images | grep atom-dev; du -sh ~/models'`
+
+## Smoke (MI300X)
+
+One cell, 128 prompts — run this before the full perf matrix.
+
+```bash
+cd ~/cvs
+make install   # after git pull only; run before activating venv
+source .cvs_venv/bin/activate
+mkdir -p ~/cvs_results ~/input/cluster_file
+
+SMOKE_DIR=~/input/config_file/inference/inferencex_atom/smoke
+mkdir -p "$SMOKE_DIR"
+
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json \
+  --output "$SMOKE_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json"
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json \
+  --output "$SMOKE_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json"
+cvs copy-config mi300x_atom_single.json --output ~/input/cluster_file/mi300x_atom_single.json
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-smoke_mi300x.html
+LOG=~/cvs_results/${TS}_ix-atom-smoke_mi300x.log
+
+cvs run inferencex_atom \
+  --cluster_file ~/input/cluster_file/mi300x_atom_single.json \
+  --config_file "$SMOKE_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
+
+When `--html` is set, the **IX Run Deck** (`inferencex_atom_run_deck.html`, `.json`,
+`_viewer.html`) is generated at session end and bundled into the pytest zip.
+See `cvs/lib/report/README.md` for wiring other suites. Open the pytest HTML **Reports**
+section for links. Render-only; does not affect gates.
+
+## W1 perf (MI300X)
+
+Two concurrency cells (C=128, C=256), 1000 prompts. Second cell reuses the ATOM server when `reuse_server_across_sweep: true`.
+
+```bash
+cd ~/cvs
+make install   # after git pull only; run before activating venv
+source .cvs_venv/bin/activate
+
+PERF_DIR=~/input/config_file/inference/inferencex_atom/perf
+mkdir -p "$PERF_DIR"
+
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json \
+  --output "$PERF_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json \
+  --output "$PERF_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json"
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-w1-perf_mi300x.html
+LOG=~/cvs_results/${TS}_ix-atom-w1-perf_mi300x.log
+
+cvs run inferencex_atom \
+  --cluster_file ~/input/cluster_file/mi300x_atom_single.json \
+  --config_file "$PERF_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
+
+## W1 perf multinode (MI300X, 2-node)
+
+Requires a 2-node cluster file, ATOM image with distributed serve support, and fabric env in `roles.server.env` (NCCL socket ifnames, etc.). One cell matches the single-node W1 reference (ISL=OSL=1024, C=128).
+
+```bash
+cd ~/cvs
+make install   # after git pull only; run before activating venv
+source .cvs_venv/bin/activate
+
+MULTI_DIR=~/input/config_file/inference/inferencex_atom/multi
+mkdir -p "$MULTI_DIR"
+
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json \
+  --output "$MULTI_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json"
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json \
+  --output "$MULTI_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json"
+cvs copy-config mi300x_atom_multi.json --output ~/input/cluster_file/mi300x_atom_multi.json
+
+# Edit cluster + config: replace {head-node-ip} / {worker-node-ip} and set params.master_addr.
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-w1-perf-multi_mi300x.html
+LOG=~/cvs_results/${TS}_ix-atom-w1-perf-multi_mi300x.log
+
+cvs run inferencex_atom \
+  --cluster_file ~/input/cluster_file/mi300x_atom_multi.json \
+  --config_file "$MULTI_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
+
+## W1 perf multinode (MI355X, 2-node)
+
+Same sweep matrix as MI300X multinode. Thresholds are seeded from the MI355X single-node CI reference; `enforce_thresholds` stays `false` until a 2-node MI355X lab run confirms.
+
+```bash
+cd ~/cvs
+make install   # after git pull only; run before activating venv
+source .cvs_venv/bin/activate
+
+MULTI_DIR=~/input/config_file/inference/inferencex_atom/mi355x_multi
+mkdir -p "$MULTI_DIR"
+
+cvs copy-config inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json \
+  --output "$MULTI_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json"
+cvs copy-config inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json \
+  --output "$MULTI_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json"
+cvs copy-config mi355x_atom_multi.json --output ~/input/cluster_file/mi355x_atom_multi.json
+
+# Edit cluster + config: replace {head-node-ip} / {worker-node-ip} and set params.master_addr.
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-w1-perf-multi_mi355x.html
+LOG=~/cvs_results/${TS}_ix-atom-w1-perf-multi_mi355x.log
+
+cvs run inferencex_atom \
+  --cluster_file ~/input/cluster_file/mi355x_atom_multi.json \
+  --config_file "$MULTI_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
+
+## W1 vLLM parity (MI300X, M4)
+
+Same sweep cells as ATOM W1 perf. Uses ``cvs run inferencex_atom_vllm_single`` (not legacy ``vllm_single``). Set ``container.image`` to your ROCm vLLM image. Optional ``run_card.parity_reference_json`` enables ``compare.vllm.*`` ratio gates (see ``mi300x_inferencex-atom-single_deepseek-r1_fp8_atom_reference.example.json``).
+
+```bash
+cd ~/cvs && source .cvs_venv/bin/activate
+
+VLLM_DIR=~/input/config_file/inference/inferencex_atom/vllm_perf
+mkdir -p "$VLLM_DIR"
+
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_config.json \
+  --output "$VLLM_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_config.json"
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_threshold.json \
+  --output "$VLLM_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_threshold.json"
+cvs copy-config mi300x_atom_single.json --output ~/input/cluster_file/mi300x_atom_single.json
+
+# After a green ATOM W1 run, copy cell metrics into a reference JSON and set
+# run_card.parity_reference_json in the vLLM config.
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-vllm-parity_mi300x.html
+LOG=~/cvs_results/${TS}_ix-atom-vllm-parity_mi300x.log
+
+cvs run inferencex_atom_vllm_single \
+  --cluster_file ~/input/cluster_file/mi300x_atom_single.json \
+  --config_file "$VLLM_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
+
+## W1 vLLM parity multinode (MI300X)
+
+Same 15-cell sweep as ATOM `perf_multi`. Uses ``mi300x_atom_multi.json`` and vLLM pipeline parallelism across nodes (`PP=2`, `nnodes=2`). Populate ``run_card.parity_reference_json`` from a green ATOM multinode run for ``compare.vllm.*`` gates.
+
+```bash
+cd ~/cvs && source .cvs_venv/bin/activate
+
+VLLM_MULTI_DIR=~/input/config_file/inference/inferencex_atom/vllm_perf_multi
+mkdir -p "$VLLM_MULTI_DIR"
+
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json \
+  --output "$VLLM_MULTI_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json"
+cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json \
+  --output "$VLLM_MULTI_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json"
+cvs copy-config mi300x_atom_multi.json --output ~/input/cluster_file/mi300x_atom_multi.json
+
+# Edit cluster + config: replace {head-node-ip} / {worker-node-ip}, set container.image and params.master_addr.
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-vllm-parity-multi_mi300x.html
+LOG=~/cvs_results/${TS}_ix-atom-vllm-parity-multi_mi300x.log
+
+cvs run inferencex_atom_vllm_single \
+  --cluster_file ~/input/cluster_file/mi300x_atom_multi.json \
+  --config_file "$VLLM_MULTI_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
+
+## W1 vLLM parity multinode (MI355X)
+
+Same sweep matrix as MI355X ATOM multinode. Thresholds are seeded; ``enforce_thresholds`` stays ``false`` until a 2-node MI355X lab run confirms.
+
+```bash
+cd ~/cvs && source .cvs_venv/bin/activate
+
+VLLM_MULTI_DIR=~/input/config_file/inference/inferencex_atom/mi355x_vllm_multi
+mkdir -p "$VLLM_MULTI_DIR"
+
+cvs copy-config inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json \
+  --output "$VLLM_MULTI_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json"
+cvs copy-config inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json \
+  --output "$VLLM_MULTI_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json"
+cvs copy-config mi355x_atom_multi.json --output ~/input/cluster_file/mi355x_atom_multi.json
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-vllm-parity-multi_mi355x.html
+LOG=~/cvs_results/${TS}_ix-atom-vllm-parity-multi_mi355x.log
+
+cvs run inferencex_atom_vllm_single \
+  --cluster_file ~/input/cluster_file/mi355x_atom_multi.json \
+  --config_file "$VLLM_MULTI_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
+
+## W1 perf (MI355X)
+
+Thresholds are seeded from [ROCm/ATOM run 27912164002](https://github.com/ROCm/ATOM/actions/runs/27912164002). `enforce_thresholds` stays `false` until an MI355X lab run confirms.
+
+```bash
+cd ~/cvs
+make install   # after git pull only; run before activating venv
+source .cvs_venv/bin/activate
+
+PERF_DIR=~/input/config_file/inference/inferencex_atom/mi355x_perf
+mkdir -p "$PERF_DIR"
+
+cvs copy-config inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json \
+  --output "$PERF_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
+cvs copy-config inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json \
+  --output "$PERF_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json"
+cvs copy-config mi355x_atom_single.json --output ~/input/cluster_file/mi355x_atom_single.json
+
+TS=$(date +%Y%m%d_%H%M%S)
+HTML=~/cvs_results/${TS}_ix-atom-w1-perf_mi355x.html
+LOG=~/cvs_results/${TS}_ix-atom-w1-perf_mi355x.log
+
+cvs run inferencex_atom \
+  --cluster_file ~/input/cluster_file/mi355x_atom_single.json \
+  --config_file "$PERF_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json" \
+  --html="$HTML" \
+  --self-contained-html \
+  --log-file="$LOG" \
+  -vvv -s
+
+echo "HTML: $HTML"
+echo "LOG:  $LOG"
+```
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_atom_reference.example.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_atom_reference.example.json
new file mode 100644
index 000000000..b16672939
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_atom_reference.example.json
@@ -0,0 +1,11 @@
+{
+  "_comment": "Example ATOM reference metrics for M4 compare.* gates. Populate client.* scalars from a green inferencex_atom run per cell_key. Multinode cells include PP= and NNODES= segments (e.g. ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16).",
+  "ISL=1024,OSL=1024,TP=8,CONC=128": {
+    "client.output_throughput": 3000,
+    "client.mean_ttft_ms": 400
+  },
+  "ISL=1024,OSL=1024,TP=8,CONC=256": {
+    "client.output_throughput": 5000,
+    "client.mean_ttft_ms": 450
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
similarity index 98%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
index 5778609cc..dea198029 100644
--- a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
@@ -1,7 +1,7 @@
 {
   "_comment": "W1 DeepSeek R1 FP8+MTP3 on 8x MI300X. Recipe: dsr1-fp8-mi300x-atom-mtp3. Thresholds: plan Section 4.2.",
   "schema_version": 1,
-  "framework": "inferencex_atom_single",
+  "framework": "inferencex_atom",
   "gpu_arch": "mi300x",
   "enforce_thresholds": false,
   "threshold_json": "mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json",
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json
similarity index 100%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
similarity index 98%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
index a7ad3014a..333e4ac15 100644
--- a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
@@ -1,7 +1,7 @@
 {
   "_comment": "W1 DeepSeek R1 FP8 on 8x MI300X. Recipe: dsr1-fp8-mi300x-atom. Thresholds: portable minimum SLOs (throughput + health), not per-node calibration.",
   "schema_version": 1,
-  "framework": "inferencex_atom_single",
+  "framework": "inferencex_atom",
   "gpu_arch": "mi300x",
   "enforce_thresholds": true,
   "threshold_json": "mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json",
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json
new file mode 100644
index 000000000..5c63ecae0
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json
@@ -0,0 +1,101 @@
+{
+  "_comment": "W1 DeepSeek R1 FP8 multinode (2×8 GPU). Conc 16/64/128 per ISL/OSL shape (15 cells); runs grouped by shape for server reuse. max_model_length=4096.",
+  "schema_version": 1,
+  "framework": "inferencex_atom",
+  "gpu_arch": "mi300x",
+  "enforce_thresholds": true,
+  "threshold_json": "mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json",
+  "run_card": {
+    "atom_image_pin": "rocm/atom-dev:latest",
+    "notes": "MI300X 2-node scaling reference (M5); align master_addr with cluster head VPC IP"
+  },
+  "paths": {
+    "shared_fs": "/home/{user-id}",
+    "models_dir": "{shared_fs}/models",
+    "log_dir": "{shared_fs}/LOGS",
+    "hf_token_file": "{shared_fs}/.hf_token"
+  },
+  "model": {
+    "id": "deepseek-ai/DeepSeek-R1-0528",
+    "remote": 0,
+    "precision": "fp8"
+  },
+  "container": {
+    "lifetime": "per_run",
+    "name": "inferencex_atom_mi300x_multi",
+    "image": "rocm/atom-dev:latest",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G",
+        "volumes": [
+          "/home/{user-id}:/home/{user-id}"
+        ],
+        "devices": [
+          "/dev/dri",
+          "/dev/kfd"
+        ]
+      }
+    }
+  },
+  "roles": {
+    "server": {
+      "atom_args": [
+        "-tp",
+        "8",
+        "--kv_cache_dtype",
+        "fp8",
+        "--trust-remote-code"
+      ],
+      "env": {
+        "ATOM_DISABLE_MMAP": "true"
+      }
+    }
+  },
+  "params": {
+    "driver": "atom",
+    "port_no": "8000",
+    "tensor_parallelism": "8",
+    "pipeline_parallel_size": "2",
+    "nnodes": "2",
+    "master_addr": "{head-node-ip}",
+    "master_port": "29501",
+    "scaling_baseline_output_throughput": "1500",
+    "random_range_ratio": "0.8",
+    "num_prompts": "1000",
+    "max_model_length": "4096",
+    "metric_percentiles": "95,99",
+    "reuse_server_across_sweep": "true",
+    "client_poll_count": "80",
+    "client_poll_wait_time": "60"
+  },
+  "sweep": {
+    "sequence_combinations": [
+      {"name": "w1_512_512", "isl": "512", "osl": "512"},
+      {"name": "w1_1k_1k", "isl": "1024", "osl": "1024"},
+      {"name": "w1_2k_1k", "isl": "2048", "osl": "1024"},
+      {"name": "w1_1k_2k", "isl": "1024", "osl": "2048"},
+      {"name": "w1_2k_2k", "isl": "2048", "osl": "2048"}
+    ],
+    "runs": [
+      {"combo": "w1_512_512", "concurrency": 16},
+      {"combo": "w1_512_512", "concurrency": 64},
+      {"combo": "w1_512_512", "concurrency": 128},
+      {"combo": "w1_1k_1k", "concurrency": 16},
+      {"combo": "w1_1k_1k", "concurrency": 64},
+      {"combo": "w1_1k_1k", "concurrency": 128},
+      {"combo": "w1_2k_1k", "concurrency": 16},
+      {"combo": "w1_2k_1k", "concurrency": 64},
+      {"combo": "w1_2k_1k", "concurrency": 128},
+      {"combo": "w1_1k_2k", "concurrency": 16},
+      {"combo": "w1_1k_2k", "concurrency": 64},
+      {"combo": "w1_1k_2k", "concurrency": 128},
+      {"combo": "w1_2k_2k", "concurrency": 16},
+      {"combo": "w1_2k_2k", "concurrency": 64},
+      {"combo": "w1_2k_2k", "concurrency": 128}
+    ]
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json
new file mode 100644
index 000000000..4d57a2ac7
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json
@@ -0,0 +1,1593 @@
+{
+  "_comment": "MI300X W1 multinode seeded thresholds (scaling.efficiency_pct min 50% floor (lab seed)). Cell key includes PP and NNODES segments.",
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json
similarity index 100%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json
similarity index 98%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json
index 0abd57f8f..7755e17ef 100644
--- a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json
@@ -1,7 +1,7 @@
 {
   "_comment": "MI300X W1 smoke: one cell (C=128), num_prompts=128. Record-only; do not use for threshold calibration.",
   "schema_version": 1,
-  "framework": "inferencex_atom_single",
+  "framework": "inferencex_atom",
   "gpu_arch": "mi300x",
   "enforce_thresholds": false,
   "threshold_json": "mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json",
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json
similarity index 100%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_config.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_config.json
new file mode 100644
index 000000000..780964128
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_config.json
@@ -0,0 +1,88 @@
+{
+  "_comment": "W1 DeepSeek R1 FP8 vLLM parity on 8x MI300X. Same sweep cells as mi300x_inferencex-atom-single_deepseek-r1_fp8_perf. Set container.image to your ROCm vLLM image. Optional run_card.parity_reference_json points at ATOM reference cell metrics for compare.* gates.",
+  "schema_version": 1,
+  "framework": "inferencex_atom_vllm_single",
+  "gpu_arch": "mi300x",
+  "enforce_thresholds": false,
+  "threshold_json": "mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_threshold.json",
+  "run_card": {
+    "vllm_image_pin": "<changeme>",
+    "parity_reference_json": "",
+    "notes": "M4 vLLM parity sibling of ATOM W1 perf; calibrate thresholds independently"
+  },
+  "paths": {
+    "shared_fs": "/home/{user-id}",
+    "models_dir": "{shared_fs}/models",
+    "log_dir": "{shared_fs}/LOGS",
+    "hf_token_file": "{shared_fs}/.hf_token"
+  },
+  "model": {
+    "id": "deepseek-ai/DeepSeek-R1-0528",
+    "remote": 0,
+    "precision": "fp8"
+  },
+  "container": {
+    "lifetime": "per_run",
+    "name": "inferencex_atom_vllm_mi300x",
+    "image": "<changeme>",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G",
+        "volumes": [
+          "/home/{user-id}:/home/{user-id}"
+        ],
+        "devices": [
+          "/dev/dri",
+          "/dev/kfd"
+        ]
+      }
+    }
+  },
+  "roles": {
+    "server": {
+      "serve_args": {
+        "trust-remote-code": true,
+        "kv-cache-dtype": "fp8"
+      },
+      "env": {
+        "VLLM_ROCM_USE_AITER": "1",
+        "AMDGCN_USE_BUFFER_OPS": "0"
+      }
+    }
+  },
+  "params": {
+    "driver": "vllm",
+    "port_no": "8000",
+    "tensor_parallelism": "8",
+    "random_range_ratio": "0.8",
+    "num_prompts": "1000",
+    "max_model_length": "4096",
+    "metric_percentiles": "95,99",
+    "reuse_server_across_sweep": "true",
+    "client_poll_count": "80",
+    "client_poll_wait_time": "60"
+  },
+  "sweep": {
+    "sequence_combinations": [
+      {
+        "name": "w1_1k_1k",
+        "isl": "1024",
+        "osl": "1024"
+      }
+    ],
+    "runs": [
+      {
+        "combo": "w1_1k_1k",
+        "concurrency": 128
+      },
+      {
+        "combo": "w1_1k_1k",
+        "concurrency": 256
+      }
+    ]
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json
new file mode 100644
index 000000000..d7c6cc72a
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json
@@ -0,0 +1,100 @@
+{
+  "_comment": "W1 DeepSeek R1 FP8 vLLM parity multinode (2×8 GPU MI300X). Same sweep cells as mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi. Set container.image to your ROCm vLLM image. Optional run_card.parity_reference_json enables compare.* gates.",
+  "schema_version": 1,
+  "framework": "inferencex_atom_vllm_single",
+  "gpu_arch": "mi300x",
+  "enforce_thresholds": false,
+  "threshold_json": "mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json",
+  "run_card": {
+    "vllm_image_pin": "<changeme>",
+    "parity_reference_json": "",
+    "notes": "M4 vLLM parity multinode sibling of ATOM W1 perf_multi; calibrate thresholds independently"
+  },
+  "paths": {
+    "shared_fs": "/home/{user-id}",
+    "models_dir": "{shared_fs}/models",
+    "log_dir": "{shared_fs}/LOGS",
+    "hf_token_file": "{shared_fs}/.hf_token"
+  },
+  "model": {
+    "id": "deepseek-ai/DeepSeek-R1-0528",
+    "remote": 0,
+    "precision": "fp8"
+  },
+  "container": {
+    "lifetime": "per_run",
+    "name": "inferencex_atom_vllm_mi300x_multi",
+    "image": "<changeme>",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G",
+        "volumes": [
+          "/home/{user-id}:/home/{user-id}"
+        ],
+        "devices": [
+          "/dev/dri",
+          "/dev/kfd"
+        ]
+      }
+    }
+  },
+  "roles": {
+    "server": {
+      "serve_args": {
+        "trust-remote-code": true,
+        "kv-cache-dtype": "fp8"
+      },
+      "env": {
+        "VLLM_ROCM_USE_AITER": "1",
+        "AMDGCN_USE_BUFFER_OPS": "0"
+      }
+    }
+  },
+  "params": {
+    "driver": "vllm",
+    "port_no": "8000",
+    "tensor_parallelism": "8",
+    "pipeline_parallel_size": "2",
+    "nnodes": "2",
+    "master_addr": "{head-node-ip}",
+    "master_port": "29501",
+    "scaling_baseline_output_throughput": "1500",
+    "random_range_ratio": "0.8",
+    "num_prompts": "1000",
+    "max_model_length": "4096",
+    "metric_percentiles": "95,99",
+    "reuse_server_across_sweep": "true",
+    "client_poll_count": "80",
+    "client_poll_wait_time": "60"
+  },
+  "sweep": {
+    "sequence_combinations": [
+      {"name": "w1_512_512", "isl": "512", "osl": "512"},
+      {"name": "w1_1k_1k", "isl": "1024", "osl": "1024"},
+      {"name": "w1_2k_1k", "isl": "2048", "osl": "1024"},
+      {"name": "w1_1k_2k", "isl": "1024", "osl": "2048"},
+      {"name": "w1_2k_2k", "isl": "2048", "osl": "2048"}
+    ],
+    "runs": [
+      {"combo": "w1_512_512", "concurrency": 16},
+      {"combo": "w1_512_512", "concurrency": 64},
+      {"combo": "w1_512_512", "concurrency": 128},
+      {"combo": "w1_1k_1k", "concurrency": 16},
+      {"combo": "w1_1k_1k", "concurrency": 64},
+      {"combo": "w1_1k_1k", "concurrency": 128},
+      {"combo": "w1_2k_1k", "concurrency": 16},
+      {"combo": "w1_2k_1k", "concurrency": 64},
+      {"combo": "w1_2k_1k", "concurrency": 128},
+      {"combo": "w1_1k_2k", "concurrency": 16},
+      {"combo": "w1_1k_2k", "concurrency": 64},
+      {"combo": "w1_1k_2k", "concurrency": 128},
+      {"combo": "w1_2k_2k", "concurrency": 16},
+      {"combo": "w1_2k_2k", "concurrency": 64},
+      {"combo": "w1_2k_2k", "concurrency": 128}
+    ]
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json
new file mode 100644
index 000000000..d54446c34
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json
@@ -0,0 +1,1713 @@
+{
+  "_comment": "MI300X W1 vLLM parity multinode thresholds. Seeded from ATOM multinode scaffold; compare.* gates apply when run_card.parity_reference_json is set. Calibrate independently after lab run.",
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 1500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 188
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 94
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 3500
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 1750
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 438
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 219
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 5000
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 2500
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 625
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 313
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_threshold.json
new file mode 100644
index 000000000..7bc3c619b
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_threshold.json
@@ -0,0 +1,61 @@
+{
+  "_comment": "MI300X W1 vLLM parity thresholds (independent calibration from ATOM). compare.* gates apply when run_card.parity_reference_json is set.",
+  "ISL=1024,OSL=1024,TP=8,CONC=128": {
+    "client.total_token_throughput": {"kind": "min_tok_s", "value": 3000},
+    "client.output_throughput": {"kind": "min_tok_s", "value": 1500},
+    "client.per_gpu_throughput": {"kind": "min_tok_s", "value": 375},
+    "client.output_tput_per_gpu": {"kind": "min_tok_s", "value": 188},
+    "client.mean_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p90_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.mean_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p90_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.mean_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.mean_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p90_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.success_rate": {"kind": "min", "value": 1},
+    "client.failed": {"kind": "max", "value": 0},
+    "compare.vllm.output_throughput_ratio": {"kind": "min", "value": 0.85},
+    "compare.vllm.mean_ttft_ms_ratio": {"kind": "max", "value": 1.1}
+  },
+  "ISL=1024,OSL=1024,TP=8,CONC=256": {
+    "client.total_token_throughput": {"kind": "min_tok_s", "value": 5000},
+    "client.output_throughput": {"kind": "min_tok_s", "value": 2500},
+    "client.per_gpu_throughput": {"kind": "min_tok_s", "value": 625},
+    "client.output_tput_per_gpu": {"kind": "min_tok_s", "value": 313},
+    "client.mean_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p90_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_ttft_ms": {"kind": "max_ms", "value": 1000000},
+    "client.mean_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p90_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_tpot_ms": {"kind": "max_ms", "value": 1000000},
+    "client.mean_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_itl_ms": {"kind": "max_ms", "value": 1000000},
+    "client.mean_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.median_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p90_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p95_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.p99_e2el_ms": {"kind": "max_ms", "value": 1000000},
+    "client.success_rate": {"kind": "min", "value": 1},
+    "client.failed": {"kind": "max", "value": 0},
+    "compare.vllm.output_throughput_ratio": {"kind": "min", "value": 0.85},
+    "compare.vllm.mean_ttft_ms_ratio": {"kind": "max", "value": 1.1}
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
similarity index 98%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
index 6a7f8fa21..328bce4cb 100644
--- a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
+++ b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
@@ -1,7 +1,7 @@
 {
   "_comment": "Single-node InferenceX ATOM (schema_version 1). Set container.image and container.name. ISL+OSL must fit params.max_model_length. Volume bind home only; ContainerOrchestrator adds /home/<user>:/workspace.",
   "schema_version": 1,
-  "framework": "inferencex_atom_single",
+  "framework": "inferencex_atom",
   "gpu_arch": "mi300x",
   "enforce_thresholds": false,
   "threshold_json": "mi300x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json",
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json
similarity index 100%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json
rename to cvs/input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
similarity index 98%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
rename to cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
index bfcc75110..f928116fc 100644
--- a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
+++ b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json
@@ -1,7 +1,7 @@
 {
   "_comment": "W1 DeepSeek R1 FP8+MTP3 on 8x MI355X. Recipe: dsr1-fp8-mi355x-atom-mtp3. Thresholds: plan Section 4.3.2.",
   "schema_version": 1,
-  "framework": "inferencex_atom_single",
+  "framework": "inferencex_atom",
   "gpu_arch": "mi355x",
   "enforce_thresholds": false,
   "threshold_json": "mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json",
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json
similarity index 100%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json
rename to cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_threshold.json
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
similarity index 98%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
rename to cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
index 6d2413b86..4abeccda3 100644
--- a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
+++ b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json
@@ -1,7 +1,7 @@
 {
   "_comment": "W1 DeepSeek R1 FP8 on 8x MI355X. Recipe: dsr1-fp8-mi355x-atom. Thresholds: plan Section 4.3 (ATOM run 27912164002).",
   "schema_version": 1,
-  "framework": "inferencex_atom_single",
+  "framework": "inferencex_atom",
   "gpu_arch": "mi355x",
   "enforce_thresholds": false,
   "threshold_json": "mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json",
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json
new file mode 100644
index 000000000..bf170c467
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json
@@ -0,0 +1,102 @@
+{
+  "_comment": "W1 DeepSeek R1 FP8 multinode (2×8 GPU MI355X). Conc 16/64/128 per ISL/OSL shape (15 cells); runs grouped by shape for server reuse. max_model_length=4096.",
+  "schema_version": 1,
+  "framework": "inferencex_atom",
+  "gpu_arch": "mi355x",
+  "enforce_thresholds": false,
+  "threshold_json": "mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json",
+  "run_card": {
+    "upstream_run_url": "https://github.com/ROCm/ATOM/actions/runs/27912164002",
+    "atom_image_pin": "rocm/atom-dev:nightly_202606211542",
+    "notes": "MI355X 2-node scaling reference (M5); align master_addr with cluster head VPC IP. Lab re-run required before enforce_thresholds: true."
+  },
+  "paths": {
+    "shared_fs": "/home/{user-id}",
+    "models_dir": "{shared_fs}/models",
+    "log_dir": "{shared_fs}/LOGS",
+    "hf_token_file": "{shared_fs}/.hf_token"
+  },
+  "model": {
+    "id": "deepseek-ai/DeepSeek-R1-0528",
+    "remote": 0,
+    "precision": "fp8"
+  },
+  "container": {
+    "lifetime": "per_run",
+    "name": "inferencex_atom_mi355x_multi",
+    "image": "rocm/atom-dev:nightly_202606211542",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G",
+        "volumes": [
+          "/home/{user-id}:/home/{user-id}"
+        ],
+        "devices": [
+          "/dev/dri",
+          "/dev/kfd"
+        ]
+      }
+    }
+  },
+  "roles": {
+    "server": {
+      "atom_args": [
+        "-tp",
+        "8",
+        "--kv_cache_dtype",
+        "fp8",
+        "--trust-remote-code"
+      ],
+      "env": {
+        "ATOM_DISABLE_MMAP": "true"
+      }
+    }
+  },
+  "params": {
+    "driver": "atom",
+    "port_no": "8000",
+    "tensor_parallelism": "8",
+    "pipeline_parallel_size": "2",
+    "nnodes": "2",
+    "master_addr": "{head-node-ip}",
+    "master_port": "29501",
+    "scaling_baseline_output_throughput": "4000",
+    "random_range_ratio": "0.8",
+    "num_prompts": "1000",
+    "max_model_length": "4096",
+    "metric_percentiles": "95,99",
+    "reuse_server_across_sweep": "true",
+    "client_poll_count": "80",
+    "client_poll_wait_time": "60"
+  },
+  "sweep": {
+    "sequence_combinations": [
+      {"name": "w1_512_512", "isl": "512", "osl": "512"},
+      {"name": "w1_1k_1k", "isl": "1024", "osl": "1024"},
+      {"name": "w1_2k_1k", "isl": "2048", "osl": "1024"},
+      {"name": "w1_1k_2k", "isl": "1024", "osl": "2048"},
+      {"name": "w1_2k_2k", "isl": "2048", "osl": "2048"}
+    ],
+    "runs": [
+      {"combo": "w1_512_512", "concurrency": 16},
+      {"combo": "w1_512_512", "concurrency": 64},
+      {"combo": "w1_512_512", "concurrency": 128},
+      {"combo": "w1_1k_1k", "concurrency": 16},
+      {"combo": "w1_1k_1k", "concurrency": 64},
+      {"combo": "w1_1k_1k", "concurrency": 128},
+      {"combo": "w1_2k_1k", "concurrency": 16},
+      {"combo": "w1_2k_1k", "concurrency": 64},
+      {"combo": "w1_2k_1k", "concurrency": 128},
+      {"combo": "w1_1k_2k", "concurrency": 16},
+      {"combo": "w1_1k_2k", "concurrency": 64},
+      {"combo": "w1_1k_2k", "concurrency": 128},
+      {"combo": "w1_2k_2k", "concurrency": 16},
+      {"combo": "w1_2k_2k", "concurrency": 64},
+      {"combo": "w1_2k_2k", "concurrency": 128}
+    ]
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json
new file mode 100644
index 000000000..071b1ba74
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_threshold.json
@@ -0,0 +1,1593 @@
+{
+  "_comment": "MI355X W1 multinode seeded thresholds (scaling.efficiency_pct min 50% floor). Throughput mins scaled from MI300X multinode scaffold \u00d7 single-node MI355X/MI300X ratio. Lab re-run required before enforce_thresholds: true.",
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    }
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json
similarity index 100%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json
rename to cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json
new file mode 100644
index 000000000..005177b2f
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json
@@ -0,0 +1,100 @@
+{
+  "_comment": "W1 DeepSeek R1 FP8 vLLM parity multinode (2×8 GPU MI355X). Same sweep cells as mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi. Set container.image to your ROCm vLLM image.",
+  "schema_version": 1,
+  "framework": "inferencex_atom_vllm_single",
+  "gpu_arch": "mi355x",
+  "enforce_thresholds": false,
+  "threshold_json": "mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json",
+  "run_card": {
+    "vllm_image_pin": "<changeme>",
+    "parity_reference_json": "",
+    "notes": "MI355X vLLM parity multinode; align master_addr with cluster head VPC IP. Lab re-run required before enforce_thresholds: true."
+  },
+  "paths": {
+    "shared_fs": "/home/{user-id}",
+    "models_dir": "{shared_fs}/models",
+    "log_dir": "{shared_fs}/LOGS",
+    "hf_token_file": "{shared_fs}/.hf_token"
+  },
+  "model": {
+    "id": "deepseek-ai/DeepSeek-R1-0528",
+    "remote": 0,
+    "precision": "fp8"
+  },
+  "container": {
+    "lifetime": "per_run",
+    "name": "inferencex_atom_vllm_mi355x_multi",
+    "image": "<changeme>",
+    "runtime": {
+      "name": "docker",
+      "args": {
+        "network": "host",
+        "ipc": "host",
+        "privileged": true,
+        "shm_size": "128G",
+        "volumes": [
+          "/home/{user-id}:/home/{user-id}"
+        ],
+        "devices": [
+          "/dev/dri",
+          "/dev/kfd"
+        ]
+      }
+    }
+  },
+  "roles": {
+    "server": {
+      "serve_args": {
+        "trust-remote-code": true,
+        "kv-cache-dtype": "fp8"
+      },
+      "env": {
+        "VLLM_ROCM_USE_AITER": "1",
+        "AMDGCN_USE_BUFFER_OPS": "0"
+      }
+    }
+  },
+  "params": {
+    "driver": "vllm",
+    "port_no": "8000",
+    "tensor_parallelism": "8",
+    "pipeline_parallel_size": "2",
+    "nnodes": "2",
+    "master_addr": "{head-node-ip}",
+    "master_port": "29501",
+    "scaling_baseline_output_throughput": "4000",
+    "random_range_ratio": "0.8",
+    "num_prompts": "1000",
+    "max_model_length": "4096",
+    "metric_percentiles": "95,99",
+    "reuse_server_across_sweep": "true",
+    "client_poll_count": "80",
+    "client_poll_wait_time": "60"
+  },
+  "sweep": {
+    "sequence_combinations": [
+      {"name": "w1_512_512", "isl": "512", "osl": "512"},
+      {"name": "w1_1k_1k", "isl": "1024", "osl": "1024"},
+      {"name": "w1_2k_1k", "isl": "2048", "osl": "1024"},
+      {"name": "w1_1k_2k", "isl": "1024", "osl": "2048"},
+      {"name": "w1_2k_2k", "isl": "2048", "osl": "2048"}
+    ],
+    "runs": [
+      {"combo": "w1_512_512", "concurrency": 16},
+      {"combo": "w1_512_512", "concurrency": 64},
+      {"combo": "w1_512_512", "concurrency": 128},
+      {"combo": "w1_1k_1k", "concurrency": 16},
+      {"combo": "w1_1k_1k", "concurrency": 64},
+      {"combo": "w1_1k_1k", "concurrency": 128},
+      {"combo": "w1_2k_1k", "concurrency": 16},
+      {"combo": "w1_2k_1k", "concurrency": 64},
+      {"combo": "w1_2k_1k", "concurrency": 128},
+      {"combo": "w1_1k_2k", "concurrency": 16},
+      {"combo": "w1_1k_2k", "concurrency": 64},
+      {"combo": "w1_1k_2k", "concurrency": 128},
+      {"combo": "w1_2k_2k", "concurrency": 16},
+      {"combo": "w1_2k_2k", "concurrency": 64},
+      {"combo": "w1_2k_2k", "concurrency": 128}
+    ]
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json
new file mode 100644
index 000000000..ca708966b
--- /dev/null
+++ b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_threshold.json
@@ -0,0 +1,1713 @@
+{
+  "_comment": "MI355X W1 vLLM parity multinode thresholds. Seeded from ATOM multinode scaffold; compare.* when parity_reference_json is set. Lab re-run required before enforce_thresholds: true.",
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=1024,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=1024,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=16": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 8009.32
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 4004.66
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 1003.83
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 501.92
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=64": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 18688.41
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 9344.21
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 2338.72
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1169.36
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  },
+  "ISL=2048,OSL=2048,TP=8,PP=2,NNODES=2,CONC=128": {
+    "client.total_token_throughput": {
+      "kind": "min_tok_s",
+      "value": 26697.73
+    },
+    "client.output_throughput": {
+      "kind": "min_tok_s",
+      "value": 13348.87
+    },
+    "client.per_gpu_throughput": {
+      "kind": "min_tok_s",
+      "value": 3337.22
+    },
+    "client.output_tput_per_gpu": {
+      "kind": "min_tok_s",
+      "value": 1671.28
+    },
+    "client.mean_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_ttft_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_tpot_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_itl_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.mean_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.median_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p90_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p95_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.p99_e2el_ms": {
+      "kind": "max_ms",
+      "value": 1000000
+    },
+    "client.success_rate": {
+      "kind": "min",
+      "value": 1
+    },
+    "client.failed": {
+      "kind": "max",
+      "value": 0
+    },
+    "scaling.efficiency_pct": {
+      "kind": "min",
+      "value": 50
+    },
+    "compare.vllm.output_throughput_ratio": {
+      "kind": "min",
+      "value": 0.85
+    },
+    "compare.vllm.mean_ttft_ms_ratio": {
+      "kind": "max",
+      "value": 1.1
+    }
+  }
+}
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_config.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
similarity index 98%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
rename to cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
index 2b7537697..a0b574ef7 100644
--- a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
+++ b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_config.json
@@ -1,7 +1,7 @@
 {
   "_comment": "Single-node InferenceX ATOM sample (schema_version 1). Set container.image and container.name. Verify serve_args against your InferenceX revision.",
   "schema_version": 1,
-  "framework": "inferencex_atom_single",
+  "framework": "inferencex_atom",
   "gpu_arch": "mi355x",
   "enforce_thresholds": false,
   "threshold_json": "mi355x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json",
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json b/cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json
similarity index 100%
rename from cvs/input/config_file/inference/inferencex_atom_single/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json
rename to cvs/input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json
diff --git a/cvs/input/config_file/inference/inferencex_atom_single/README.md b/cvs/input/config_file/inference/inferencex_atom_single/README.md
deleted file mode 100644
index 63915906e..000000000
--- a/cvs/input/config_file/inference/inferencex_atom_single/README.md
+++ /dev/null
@@ -1,177 +0,0 @@
-# InferenceX ATOM single-node variants
-
-W1 **DeepSeek R1 FP8** on 8× GPU, ISL=OSL=1024, TP8.
-
-## Layout
-
-**In the CVS repo**, all variants live as flat sibling pairs in **this directory**:
-
-```text
-{gpu}_inferencex-atom-single_{model}_{precision}[_{mode}]_config.json
-{gpu}_inferencex-atom-single_{model}_{precision}[_{mode}]_threshold.json
-```
-
-**On your lab machine** (`~/input/config_file/inference/inferencex_atom_single/`), copy each variant into its **own subdirectory** so only one `*threshold.json` sits next to the config you pass to `--config_file`. `substitute_config` globs the config's parent directory; multiple `*threshold.json` files there raises `ValueError: multiple *threshold.json files … (ambiguous)`.
-
-```text
-~/input/.../inferencex_atom_single/smoke/   # smoke config + smoke threshold only
-~/input/.../inferencex_atom_single/perf/    # perf config + perf threshold only
-```
-
-Each shipped config sets `"threshold_json"` to the sibling threshold filename (resolved relative to the config directory). You may also use an absolute path (vLLM-style).
-
-Legacy nested layouts (`deepseek_r1_fp8_mi300x_atom_perf/`, `inferencemax/`, etc.) are **removed** from the repo tree. Use only the flat stems below.
-
-**Config filename example:** `mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json`
-
-| Variant | GPU | Notes |
-|---------|-----|-------|
-| `mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke` | MI300X | Quick path check (C=128, 128 prompts) |
-| `mi300x_inferencex-atom-single_deepseek-r1_fp8_perf` | MI300X | W1 perf, portable min-SLO thresholds, server reuse across sweep |
-| `mi300x_inferencex-atom-single_deepseek-r1_fp8_mtp3` | MI300X | W1 FP8+MTP3 |
-| `mi355x_inferencex-atom-single_deepseek-r1_fp8_perf` | MI355X | W1 perf (CI seeds, `enforce_thresholds: false`) |
-| `mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3` | MI355X | W1 FP8+MTP3 |
-| `mi300x_inferencex-atom-single_gpt-oss-120b_bf16` | MI300X | GPT-OSS uplift placeholder (`driver: vllm`, inline `serve_args`) |
-| `mi355x_inferencex-atom-single_gpt-oss-120b_bf16` | MI355X | GPT-OSS uplift placeholder |
-
-ATOM server CLI is inline in each config under `roles.server.atom_args` (vLLM-style, same as `roles.server.serve_args` on `vllm_single`). MTP3 variants also set `params.bench_extra_args`.
-
-## Cluster + container naming
-
-Use `cvs/input/cluster_file/mi300x_atom_single.json` or `mi355x_atom_single.json`. The cluster `container.name` must match the variant (`inferencex_atom_mi300x` / `inferencex_atom_mi355x`); the suite deep-merges variant container settings over the cluster file.
-
-## Shared suite helpers (reusable by other inference suites)
-
-| Module | Purpose |
-|--------|---------|
-| `cvs/lib/inference/inference_suite_lifecycle.py` | Lifecycle stage tests, `InferenceLifecycle`, pytest HTML hooks |
-| `cvs/lib/inference/inference_suite_results_table.py` | Configurable results table (`make_print_results_table`) |
-| `cvs/lib/inference/unittests/fake_orch.py` | `FakeOrch` for Job parse unit tests |
-
-`inferencex_atom_single` imports these today; `vllm_single` may adopt them in a follow-up without duplicating code.
-
-## Pytest layout
-
-1. `test_launch_container` → `test_setup_sshd` → `test_model_fetch`
-2. `test_inferencex_atom_inference` (per sweep cell; reuses server when `reuse_server_across_sweep: true`)
-3. `test_cell_metrics` (one HTML row per **metric tier** per cell: throughput, ttft, tpot, health, record)
-4. `test_print_results_table` → `test_teardown`
-
-W1 MI300X perf with two concurrency cells expects **~17** pytest rows (not one row per scalar metric).
-
-## Before the first lab run
-
-- `git checkout` the branch under test, then `make install` and `source .cvs_venv/bin/activate` on the **launcher** host.
-- Edit `~/input/cluster_file/mi300x_atom_single.json` (or `mi355x_atom_single.json`): node IPs, `username`, `priv_key_file`, `container.image`.
-- **Launcher vs GPU node:** CVS pytest runs on the launcher; `ContainerOrchestrator` SSHes to cluster nodes and runs `sudo docker` there. Local Docker on the launcher is not used. Prerequisites split by host:
-
-  | Item | Launcher | GPU node (cluster `mgmt_ip`) |
-  |------|----------|------------------------------|
-  | `cvs run`, venv, `~/input/`, `~/cvs_results/` | Yes | No |
-  | `priv_key_file`, `~/.hf_token` (read locally by pytest) | Yes | No |
-  | `~/models` (when `model.remote: 0`) | No | Yes |
-  | `rocm/atom-dev` image, `sudo docker` | No | Yes |
-  | `~/LOGS/` (server/bench logs via volume mount) | No | Yes |
-
-- Preflight from the launcher: `ssh -i ~/.ssh/<key> <user>@<mgmt_ip> 'sudo docker images | grep atom-dev; du -sh ~/models'`
-
-## Smoke (MI300X)
-
-One cell, 128 prompts — run this before the full perf matrix.
-
-```bash
-cd ~/cvs && source .cvs_venv/bin/activate
-mkdir -p ~/cvs_results ~/input/cluster_file
-
-SMOKE_DIR=~/input/config_file/inference/inferencex_atom_single/smoke
-mkdir -p "$SMOKE_DIR"
-
-cvs copy-config inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json \
-  --output "$SMOKE_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json"
-cvs copy-config inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json \
-  --output "$SMOKE_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_threshold.json"
-cvs copy-config mi300x_atom_single.json --output ~/input/cluster_file/mi300x_atom_single.json
-
-TS=$(date +%Y%m%d_%H%M%S)
-HTML=~/cvs_results/${TS}_ix-atom-smoke_mi300x.html
-LOG=~/cvs_results/${TS}_ix-atom-smoke_mi300x.log
-
-cvs run inferencex_atom_single \
-  --cluster_file ~/input/cluster_file/mi300x_atom_single.json \
-  --config_file "$SMOKE_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json" \
-  --html="$HTML" \
-  --self-contained-html \
-  --log-file="$LOG" \
-  -vvv -s
-
-echo "HTML: $HTML"
-echo "LOG:  $LOG"
-```
-
-When `--html` is set, the **IX Run Deck** (`inferencex_atom_run_deck.html`, `.json`,
-`_viewer.html`) is generated at session end and bundled into the pytest zip.
-See `cvs/lib/report/README.md` for wiring other suites. Open the pytest HTML **Reports**
-section for links. Render-only; does not affect gates.
-
-## W1 perf (MI300X)
-
-Two concurrency cells (C=128, C=256), 1000 prompts. Second cell reuses the ATOM server when `reuse_server_across_sweep: true`.
-
-```bash
-cd ~/cvs && source .cvs_venv/bin/activate
-
-PERF_DIR=~/input/config_file/inference/inferencex_atom_single/perf
-mkdir -p "$PERF_DIR"
-
-cvs copy-config inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json \
-  --output "$PERF_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
-cvs copy-config inference/inferencex_atom_single/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json \
-  --output "$PERF_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json"
-
-TS=$(date +%Y%m%d_%H%M%S)
-HTML=~/cvs_results/${TS}_ix-atom-w1-perf_mi300x.html
-LOG=~/cvs_results/${TS}_ix-atom-w1-perf_mi300x.log
-
-cvs run inferencex_atom_single \
-  --cluster_file ~/input/cluster_file/mi300x_atom_single.json \
-  --config_file "$PERF_DIR/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json" \
-  --html="$HTML" \
-  --self-contained-html \
-  --log-file="$LOG" \
-  -vvv -s
-
-echo "HTML: $HTML"
-echo "LOG:  $LOG"
-```
-
-## W1 perf (MI355X)
-
-Thresholds are seeded from [ROCm/ATOM run 27912164002](https://github.com/ROCm/ATOM/actions/runs/27912164002). `enforce_thresholds` stays `false` until an MI355X lab run confirms.
-
-```bash
-cd ~/cvs && source .cvs_venv/bin/activate
-
-PERF_DIR=~/input/config_file/inference/inferencex_atom_single/mi355x_perf
-mkdir -p "$PERF_DIR"
-
-cvs copy-config inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json \
-  --output "$PERF_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
-cvs copy-config inference/inferencex_atom_single/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json \
-  --output "$PERF_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json"
-cvs copy-config mi355x_atom_single.json --output ~/input/cluster_file/mi355x_atom_single.json
-
-TS=$(date +%Y%m%d_%H%M%S)
-HTML=~/cvs_results/${TS}_ix-atom-w1-perf_mi355x.html
-LOG=~/cvs_results/${TS}_ix-atom-w1-perf_mi355x.log
-
-cvs run inferencex_atom_single \
-  --cluster_file ~/input/cluster_file/mi355x_atom_single.json \
-  --config_file "$PERF_DIR/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json" \
-  --html="$HTML" \
-  --self-contained-html \
-  --log-file="$LOG" \
-  -vvv -s
-
-echo "HTML: $HTML"
-echo "LOG:  $LOG"
-```
diff --git a/cvs/lib/inference/ADDING_A_SUITE.md b/cvs/lib/inference/ADDING_A_SUITE.md
index f1ab5327a..94e90a35c 100644
--- a/cvs/lib/inference/ADDING_A_SUITE.md
+++ b/cvs/lib/inference/ADDING_A_SUITE.md
@@ -593,7 +593,7 @@ auto-loads it when the stem matches `cvs run …` and writes HTML/JSON/viewer at
 end when `--html` is set. Reports are render-only and do not change pass/fail.
 
 Do **not** wire reports in suite `conftest.py`. **IX-atom reference:**
-`presets/inferencex_atom_single.py` + `presets/inferencex_atom.py`.
+`presets/inferencex_atom.py` + `presets/inferencex_atom.py`.
 
 ---
 
diff --git a/cvs/lib/inference/inference_suite_lifecycle.py b/cvs/lib/inference/inference_suite_lifecycle.py
index 6d2135426..1aeee3084 100644
--- a/cvs/lib/inference/inference_suite_lifecycle.py
+++ b/cvs/lib/inference/inference_suite_lifecycle.py
@@ -4,7 +4,7 @@
 
 Reusable **lifecycle-as-tests** helpers for DTNI inference suites.
 
-``inferencex_atom_single`` imports the stage tests from here today; other suites
+``inferencex_atom`` imports the stage tests from here today; other suites
 (``vllm_single``, future IX parity frameworks) can reuse the same module instead
 of copying launch / sshd / model-fetch / teardown blocks.
 
@@ -46,6 +46,7 @@
     pytest_html = None
 
 from cvs.lib import globals
+from cvs.lib.inference.cache_probe import du_bytes
 
 log = globals.log
 
@@ -78,9 +79,6 @@ def sweep_cell_result_key(variant_config, seq_combo, isl, osl, concurrency):
     )
 
 
-from cvs.lib.inference.cache_probe import du_bytes
-
-
 def test_launch_container(orch, variant_config, lifecycle, request):
     """Stage 1: launch the container."""
     t = time.monotonic()
@@ -106,7 +104,11 @@ def test_setup_sshd(orch, lifecycle, request):
         lifecycle.failed = True
         pytest.fail("setup_sshd() returned False")
     if len(orch.hosts) > 1:
-        probe = orch.exec("bash -c 'ss -ltn 2>/dev/null | grep -q :2224 && echo OK || echo NO'")
+        probe = orch.exec(
+            "bash -c '"
+            "(ss -ltn 2>/dev/null || netstat -lnt 2>/dev/null) | grep -q :2224 && echo OK || "
+            "(pgrep -f \"sshd.*2224\" >/dev/null && echo OK) || echo NO'"
+        )
         if not any("OK" in (v or "") for v in (probe or {}).values()):
             lifecycle.failed = True
             pytest.fail("sshd not listening on 2224 after setup_sshd()")
@@ -194,10 +196,7 @@ def attach_lifecycle_html_table(item, report):
         return
     if pytest_html is None:
         return
-    body = "".join(
-        f"<tr><td>{label}</td><td>{value:.1f}</td><td>{unit}</td></tr>"
-        for label, value, unit in rows
-    )
+    body = "".join(f"<tr><td>{label}</td><td>{value:.1f}</td><td>{unit}</td></tr>" for label, value, unit in rows)
     html = f"<table><tr><th>stage</th><th>value</th><th>unit</th></tr>{body}</table>"
     extras = getattr(report, "extras", [])
     extras.append(pytest_html.extras.html(html))
@@ -210,7 +209,7 @@ def html_metric_table_header(cells):
 
 
 def html_metric_table_row(report, cells):
-    props = dict(report.user_properties)
+    props = dict(getattr(report, "user_properties", ()))
     has = "metric_value" in props
     val = props.get("metric_value")
     unit = props.get("metric_unit", "") if has else ""
diff --git a/cvs/lib/inference/inference_suite_results_table.py b/cvs/lib/inference/inference_suite_results_table.py
index b98049694..39adf0f2f 100644
--- a/cvs/lib/inference/inference_suite_results_table.py
+++ b/cvs/lib/inference/inference_suite_results_table.py
@@ -36,6 +36,7 @@
     ("Mean TTFT (ms)", "client.mean_ttft_ms"),
     ("Mean TPOT (ms)", "client.mean_tpot_ms"),
     ("P99 ITL (ms)", "client.p99_itl_ms"),
+    ("Scaling eff. (%)", "scaling.efficiency_pct"),
 )
 
 # Optional preset for suites that want vLLM-style columns (not wired in vllm_single yet).
diff --git a/cvs/lib/inference/inferencex_atom_orch.py b/cvs/lib/inference/inferencex_atom_orch.py
index 234adcdad..8bc8382f2 100644
--- a/cvs/lib/inference/inferencex_atom_orch.py
+++ b/cvs/lib/inference/inferencex_atom_orch.py
@@ -2,13 +2,22 @@
 Copyright 2025 Advanced Micro Devices, Inc.
 All rights reserved.
 
-Standalone InferenceX ATOM single-node job driven by a ContainerOrchestrator.
+InferenceX ATOM job driven by a ContainerOrchestrator (single- or multi-node).
 
 ``params.driver=atom`` (target): ``atom.entrypoints.openai_server`` +
 ``atom.benchmarks.benchmark_serving`` with ATOM JSON artifacts.
 
 ``params.driver=vllm`` (interim uplift): ``vllm serve`` + ``vllm bench serve``.
 
+When ``params.nnodes`` > 1 the job launches one ATOM server per cluster host.
+For ``driver=atom``, multinode uses ATOM SPMD data-parallel env vars
+(``ATOM_DP_*``) plus ``-dp N`` when each node has enough GPUs for local DP
+replicas (``tensor_parallelism * nnodes <= 8``). With TP8 across two 8-GPU
+nodes, each host runs an independent ``-tp 8`` replica (no ``-dp``) because
+ATOM cannot place two local DP ranks when TP already consumes all eight GPUs;
+the benchmark client still runs on the head node only. For ``driver=vllm``,
+multinode keeps vLLM's ``--node-rank`` / ``--master-addr`` flags.
+
 Does NOT subclass :class:`cvs.lib.inference.base.InferenceBaseJob`.
 '''
 
@@ -26,7 +35,7 @@
 
 
 class InferenceXAtomJob:
-    """Single-node InferenceX ATOM benchmark job driven by an injected ContainerOrchestrator."""
+    """InferenceX ATOM benchmark job driven by an injected ContainerOrchestrator."""
 
     READINESS_RE = re.compile(r"Application startup complete|Uvicorn running|Started server", re.I)
     COMPLETION_RE = re.compile(r"Serving Benchmark Result", re.I)
@@ -37,7 +46,10 @@ class InferenceXAtomJob:
         re.I,
     )
     EARLY_FAILURE_RE = re.compile(
-        r"no such file or directory|command not found|cannot access|failed to start",
+        r"no such file or directory|command not found|cannot access|failed to start"
+        r"|Free memory on device.*less than desired"
+        r"|Engine core initialization failed"
+        r"|WorkerProc failed to start",
         re.I,
     )
 
@@ -46,6 +58,18 @@ class InferenceXAtomJob:
         "no-enable-prefix-caching": True,
     }
 
+    # vLLM multinode flags; ATOM openai_server rejects these (use ATOM_DP_* / -dp instead).
+    _VLLM_DISTRIBUTED_FLAGS = frozenset(
+        {
+            "--node-rank",
+            "--master-addr",
+            "--master-port",
+            "--nnodes",
+            "--pipeline-parallel-size",
+            "--distributed-executor-backend",
+        }
+    )
+
     def __init__(
         self,
         orch,
@@ -75,8 +99,14 @@ def __init__(
         self.log_subdir = log_subdir
 
         p = variant.params
-        self.driver = p.driver
+        self.driver = str(p.driver or "atom").strip().lower()
         self.tp = p.tensor_parallelism
+        self.pp = p.pipeline_parallel_size
+        self.nnodes = int(p.nnodes)
+        self.distributed = self.nnodes > 1
+        raw_master = (p.master_addr or "").strip()
+        self.master_addr = raw_master or (orch.hosts[0] if getattr(orch, "hosts", None) else "localhost")
+        self.master_port = p.master_port
         self.port_no = p.port_no
         self.random_range_ratio = p.random_range_ratio
         self.random_prefix_len = p.random_prefix_len
@@ -92,6 +122,8 @@ def __init__(
         self.max_model_length = str(p.max_model_length)
         self.bench_extra_args = (p.bench_extra_args or "").strip()
         self.result_stem = (p.result_filename or "results").removesuffix(".json")
+        raw_baseline = (p.scaling_baseline_output_throughput or "").strip()
+        self._scaling_baseline = float(raw_baseline) if raw_baseline else None
 
         self.model_id = variant.model.id
         self.log_dir = variant.paths.log_dir
@@ -100,20 +132,11 @@ def __init__(
         self.atom_server_args = list(variant.roles.server.atom_args)
         self.server_env = dict(variant.roles.server.env)
 
-        self.out_dir = (
-            f"{self.log_dir}/{self.log_subdir}/out-node0/"
-            f"isl{self.isl}_osl{self.osl}_conc{self.concurrency}"
-        )
-        self.server_log = (
-            f"{self.out_dir}/atom_server.log"
-            if self.driver == "atom"
-            else f"{self.out_dir}/vllm_serve_server.log"
-        )
+        self.out_dir = self._node_out_dir(0)
+        self.server_log = self._rank_server_log(0)
         self.client_log = f"{self.out_dir}/client.log"
         self._result_artifact = (
-            f"{self.out_dir}/{self.result_stem}.json"
-            if self.driver == "atom"
-            else f"{self.out_dir}/{self.result_stem}"
+            f"{self.out_dir}/{self.result_stem}.json" if self.driver == "atom" else f"{self.out_dir}/{self.result_stem}"
         )
 
         self._precheck_wait = server_precheck_wait_s
@@ -159,9 +182,29 @@ def _int_attr(name, default):
         kw.update(overrides)
         return cls(**kw)
 
+    def _node_out_dir(self, rank):
+        return f"{self.log_dir}/{self.log_subdir}/out-node{rank}/isl{self.isl}_osl{self.osl}_conc{self.concurrency}"
+
+    def _rank_server_log(self, rank):
+        base = self._node_out_dir(rank)
+        name = "atom_server.log" if self.driver == "atom" else "vllm_serve_server.log"
+        return f"{base}/{name}"
+
+    def _exec_all(self, cmd, **kwargs):
+        return self.orch.exec(cmd, **kwargs)
+
+    def _exec_head(self, cmd, **kwargs):
+        if self.distributed:
+            return self.orch.exec_on_head(cmd, **kwargs)
+        return self.orch.exec(cmd, **kwargs)
+
     def prepare_cell_out_dir(self):
         """Create per-cell output directory without touching server env or cache."""
-        self.orch.exec(f"mkdir -p {shlex.quote(self.out_dir)}")
+        if self.distributed:
+            for rank in range(self.nnodes):
+                self._exec_all(f"mkdir -p {shlex.quote(self._node_out_dir(rank))}")
+        else:
+            self._exec_all(f"mkdir -p {shlex.quote(self.out_dir)}")
 
     @classmethod
     def _merged_serve_args(cls, variant):
@@ -189,6 +232,72 @@ def _flatten_serve_args(mapping):
                 argv.extend([opt, str(value)])
         return argv
 
+    @staticmethod
+    def _argv_has_flag(argv, *names):
+        for tok in argv:
+            if tok in names:
+                return True
+        return False
+
+    def _without_vllm_distributed_flags(self, argv):
+        """Strip vLLM multinode tokens from ``roles.server.atom_args`` if present."""
+        out = []
+        skip_next = False
+        for tok in argv:
+            if skip_next:
+                skip_next = False
+                continue
+            if tok in self._VLLM_DISTRIBUTED_FLAGS:
+                skip_next = True
+                continue
+            out.append(tok)
+        return out
+
+    def _atom_spmd_dp_cli(self):
+        """Return ``-dp N`` argv when local GPUs can host *N* TP shards on one node."""
+        if self.driver != "atom" or not self.distributed:
+            return []
+        tp = int(self.tp)
+        if tp * self.nnodes > 8:
+            return []
+        atom_argv = self._without_vllm_distributed_flags(self.atom_server_args)
+        if self._argv_has_flag(atom_argv, "-dp", "--data-parallel-size"):
+            return []
+        return ["-dp", str(self.nnodes)]
+
+    def _atom_multinode_argv(self):
+        """ATOM-only multinode CLI tokens (never vLLM ``--node-rank`` / ``--pipeline-parallel-size``)."""
+        return self._atom_spmd_dp_cli()
+
+    def _atom_spmd_env_exports(self, rank):
+        if self.driver != "atom" or not self.distributed or not self._atom_spmd_dp_cli():
+            return []
+        return [
+            f"export ATOM_DP_RANK={rank}",
+            f"export ATOM_DP_SIZE={self.nnodes}",
+            "export ATOM_DP_RANK_LOCAL=0",
+            f"export ATOM_DP_MASTER_IP={shlex.quote(self.master_addr)}",
+            f"export ATOM_DP_MASTER_PORT={self.master_port}",
+        ]
+
+    def _vllm_distributed_argv(self, rank):
+        if not self.distributed:
+            return []
+        return [
+            "--node-rank",
+            str(rank),
+            "--master-addr",
+            str(self.master_addr),
+            "--master-port",
+            str(self.master_port),
+            "--nnodes",
+            str(self.nnodes),
+            "--pipeline-parallel-size",
+            str(self.pp),
+            "--distributed-executor-backend",
+            "mp",
+        ]
+
     def build_server_cmd(self, *, clear_atom_cache=True):
         env_lines = [
             f"export HF_TOKEN={shlex.quote(self.hf_token)}",
@@ -207,12 +316,16 @@ def build_server_cmd(self, *, clear_atom_cache=True):
                 continue
             env_lines.append(f"export {k}={shlex.quote(str(v))}")
         env_script = "\n".join(env_lines) + "\n"
-        self.orch.exec("bash -c " + shlex.quote(f"printf '%s' {shlex.quote(env_script)} > /tmp/server_env_script.sh"))
-        self.orch.exec(f"mkdir -p {shlex.quote(self.out_dir)}")
+        self._exec_all("bash -c " + shlex.quote(f"printf '%s' {shlex.quote(env_script)} > /tmp/server_env_script.sh"))
+        if self.distributed:
+            for rank in range(self.nnodes):
+                self._exec_all(f"mkdir -p {shlex.quote(self._node_out_dir(rank))}")
+        else:
+            self._exec_all(f"mkdir -p {shlex.quote(self.out_dir)}")
         if self.driver == "atom" and clear_atom_cache:
-            self.orch.exec("bash -c 'rm -rf ~/.cache/atom/* 2>/dev/null || true'")
+            self._exec_all("bash -c 'rm -rf ~/.cache/atom/* 2>/dev/null || true'")
 
-    def _server_argv(self):
+    def _server_argv(self, rank=0):
         argv = [
             "vllm",
             "serve",
@@ -226,10 +339,11 @@ def _server_argv(self):
             "--port",
             str(self.port_no),
         ]
+        argv.extend(self._vllm_distributed_argv(rank))
         argv.extend(self._flatten_serve_args(self.serve_args))
         return argv
 
-    def _atom_server_argv(self):
+    def _atom_server_argv(self, rank=0):
         argv = [
             "python",
             "-m",
@@ -239,29 +353,48 @@ def _atom_server_argv(self):
             "--server-port",
             str(self.port_no),
         ]
-        argv.extend(self.atom_server_args)
+        argv.extend(self._without_vllm_distributed_flags(self.atom_server_args))
+        argv.extend(self._atom_multinode_argv())
         return argv
 
     def start_server(self):
-        if self.driver == "atom":
-            serve_cmd = " ".join(shlex.quote(str(a)) for a in self._atom_server_argv())
-        else:
-            serve_cmd = " ".join(shlex.quote(str(a)) for a in self._server_argv())
-        inner = (
-            f"source /tmp/server_env_script.sh && "
-            f"nohup {serve_cmd} > {shlex.quote(self.server_log)} 2>&1 &"
-        )
-        out = self.orch.exec("bash -c " + shlex.quote(inner))
+        hosts = list(getattr(self.orch, "hosts", []) or ["node0"])
+        if self.distributed and len(hosts) != self.nnodes:
+            raise RuntimeError(
+                f"params.nnodes={self.nnodes} but cluster has {len(hosts)} host(s); "
+                "align cluster node_dict with params.nnodes"
+            )
         label = "atom" if self.driver == "atom" else "vllm"
-        for host, output in out.items():
-            if self.EARLY_FAILURE_RE.search(output or ""):
-                raise RuntimeError(f"{label} server failed to launch on {host}: {output[-500:]}")
+        launch_hosts = enumerate(hosts) if self.distributed else [(0, hosts[0])]
+        for rank, host in launch_hosts:
+            if self.driver == "atom":
+                argv = self._atom_server_argv(rank)
+            elif self.driver == "vllm":
+                argv = self._server_argv(rank)
+            else:
+                raise RuntimeError(f"unsupported params.driver={self.driver!r}; expected 'atom' or 'vllm'")
+            serve_cmd = " ".join(shlex.quote(str(a)) for a in argv)
+            rank_log = self._rank_server_log(rank)
+            rank_env = " && ".join(self._atom_spmd_env_exports(rank))
+            env_prefix = f"{rank_env} && " if rank_env else ""
+            inner = (
+                f"source /tmp/server_env_script.sh && {env_prefix}nohup {serve_cmd} > {shlex.quote(rank_log)} 2>&1 &"
+            )
+            if self.distributed:
+                out = self._exec_all("bash -c " + shlex.quote(inner), hosts=[host])
+            else:
+                out = self._exec_all("bash -c " + shlex.quote(inner))
+            for h, output in out.items():
+                if self.EARLY_FAILURE_RE.search(output or ""):
+                    raise RuntimeError(f"{label} server failed to launch on {h} (rank {rank}): {output[-500:]}")
 
     def _atom_health_ok(self):
         url = f"http://localhost:{self.port_no}/health"
-        out = self.orch.exec(
-            f"curl -sf {shlex.quote(url)} -o /dev/null && echo OK || echo NO"
-        )
+        probe = f"curl -sf {shlex.quote(url)} -o /dev/null && echo OK || echo NO"
+        if self.distributed:
+            out = self._exec_all("bash -c " + shlex.quote(probe))
+        else:
+            out = self._exec_head("bash -c " + shlex.quote(probe))
         return bool(out) and all("OK" in (v or "") for v in out.values())
 
     def _atom_warmup_ok(self):
@@ -274,24 +407,46 @@ def _atom_warmup_ok(self):
             f"curl -sf {shlex.quote(url)} -H 'Content-Type: application/json' "
             f"-d {shlex.quote(payload)} -o /dev/null --max-time 120 && echo OK || echo NO"
         )
-        out = self.orch.exec("bash -c " + shlex.quote(inner))
+        out = self._exec_head("bash -c " + shlex.quote(inner))
         return bool(out) and all("OK" in (v or "") for v in out.values())
 
     def is_ready(self):
         if self.driver == "atom":
             return self._atom_health_ok()
+        if self.distributed:
+            pattern = self.READINESS_RE.pattern
+            ready = True
+            for rank in range(self.nnodes):
+                log_path = self._rank_server_log(rank)
+                out = self._exec_all(
+                    f"grep -qiE {shlex.quote(pattern)} {shlex.quote(log_path)}",
+                    detailed=True,
+                )
+                if not out or not all(r["exit_code"] == 0 for r in out.values()):
+                    ready = False
+                    break
+            return ready
         pattern = self.READINESS_RE.pattern
-        out = self.orch.exec(
+        out = self._exec_all(
             f"grep -qiE {shlex.quote(pattern)} {shlex.quote(self.server_log)}",
             detailed=True,
         )
         return bool(out) and all(r["exit_code"] == 0 for r in out.values())
 
+    def _tail_server_logs(self, lines=30):
+        if self.distributed:
+            out = {}
+            for rank in range(self.nnodes):
+                chunk = self._exec_all(f"tail -{lines} {shlex.quote(self._rank_server_log(rank))}")
+                out.update(chunk or {})
+            return out
+        return self._exec_all(f"tail -{lines} {shlex.quote(self.server_log)}")
+
     def wait_ready(self):
         log.info("waiting %ds for server log to materialise", self._precheck_wait)
         time.sleep(self._precheck_wait)
 
-        out = self.orch.exec(f"tail -30 {shlex.quote(self.server_log)}")
+        out = self._tail_server_logs(30)
         for host, output in out.items():
             if self.EARLY_FAILURE_RE.search(output or ""):
                 label = "atom" if self.driver == "atom" else "vllm"
@@ -303,12 +458,10 @@ def wait_ready(self):
         for it in range(self._server_poll_count):
             if not self.is_ready():
                 if self.driver == "atom":
-                    poll_out = self.orch.exec(f"tail -30 {shlex.quote(self.server_log)}")
+                    poll_out = self._tail_server_logs(30)
                     for host, output in poll_out.items():
                         if self.EARLY_FAILURE_RE.search(output or ""):
-                            raise RuntimeError(
-                                f"atom server early failure on {host}: {output[-500:]}"
-                            )
+                            raise RuntimeError(f"atom server early failure on {host}: {output[-500:]}")
                 time.sleep(self._server_poll_wait)
                 continue
             log.info("server health ready (iter=%d)", it)
@@ -327,16 +480,13 @@ def wait_ready(self):
     def stop_server(self):
         if self.driver == "atom":
             log.info("stopping atom server")
-            self.orch.exec(
+            self._exec_all(
                 "bash -c "
-                + shlex.quote(
-                    "pkill -f 'atom.entrypoints.openai_server' || "
-                    "pkill -f 'openai_server' || true"
-                )
+                + shlex.quote("pkill -f 'atom.entrypoints.openai_server' || pkill -f 'openai_server' || true")
             )
         else:
             log.info("stopping vllm server")
-            self.orch.exec("bash -c 'pkill -f \"vllm serve\" || true'")
+            self._exec_all("bash -c 'pkill -f \"vllm serve\" || true'")
         time.sleep(5)
 
     def _atom_client_argv(self):
@@ -429,21 +579,21 @@ def _vllm_client_argv(self):
     def _clear_stale_result_artifact(self):
         """Remove a prior run's result file so poll logic cannot treat it as complete."""
         artifact = shlex.quote(self._result_artifact)
-        self.orch.exec(f"rm -f {artifact}")
+        self._exec_head(f"rm -f {artifact}")
 
     def run_client(self):
         self._clear_stale_result_artifact()
         args = self._atom_client_argv() if self.driver == "atom" else self._vllm_client_argv()
         bench_cmd = " ".join(shlex.quote(str(a)) for a in args)
         client_cmd = f"source /tmp/server_env_script.sh && {bench_cmd} > {shlex.quote(self.client_log)} 2>&1 &"
-        self.orch.exec("bash -c " + shlex.quote(client_cmd))
+        self._exec_head("bash -c " + shlex.quote(client_cmd))
 
     def _atom_result_ready(self):
-        out = self.orch.exec(f"test -s {shlex.quote(self._result_artifact)} && echo OK || echo NO")
+        out = self._exec_head(f"test -s {shlex.quote(self._result_artifact)} && echo OK || echo NO")
         return bool(out) and all("OK" in (v or "") for v in out.values())
 
     def _client_log_failures(self, tail_lines=2000):
-        out = self.orch.exec(f"tail -{tail_lines} {shlex.quote(self.client_log)}")
+        out = self._exec_head(f"tail -{tail_lines} {shlex.quote(self.client_log)}")
         failed = []
         for host, output in out.items():
             txt = output or ""
@@ -497,7 +647,7 @@ def wait_client_complete(self):
                     log.info("client complete (iter=%d)", it)
                     return
             else:
-                out = self.orch.exec(f"tail -2000 {shlex.quote(self.client_log)}")
+                out = self._exec_head(f"tail -2000 {shlex.quote(self.client_log)}")
                 done = [bool(self.COMPLETION_RE.search(txt or "")) for txt in out.values()]
                 if done and all(done):
                     log.info("client complete (iter=%d)", it)
@@ -506,7 +656,7 @@ def wait_client_complete(self):
         raise RuntimeError("client did not complete before poll cap")
 
     def parse_results(self):
-        out = self.orch.exec(f"cat {shlex.quote(self._result_artifact)}")
+        out = self._exec_head(f"cat {shlex.quote(self._result_artifact)}")
         results = {}
         for host, text in out.items():
             text = (text or "").strip()
@@ -515,11 +665,15 @@ def parse_results(self):
             try:
                 raw = json.loads(text)
             except (json.JSONDecodeError, ValueError) as e:
-                raise RuntimeError(
-                    f"unparseable results artifact on {host}: {self._result_artifact}: {e}"
-                ) from e
+                raise RuntimeError(f"unparseable results artifact on {host}: {self._result_artifact}: {e}") from e
             if self.driver == "atom":
                 raw.setdefault("random_input_len", int(self.isl))
                 raw.setdefault("random_output_len", int(self.osl))
-            results[host] = to_client_metrics(raw, tp=self.tp, isl=self.isl)
+            results[host] = to_client_metrics(
+                raw,
+                tp=self.tp,
+                isl=self.isl,
+                scaling_baseline_output_throughput=self._scaling_baseline,
+                nnodes=self.nnodes,
+            )
         return results
diff --git a/cvs/lib/inference/unittests/fake_orch.py b/cvs/lib/inference/unittests/fake_orch.py
index 2edc75488..22e60b401 100644
--- a/cvs/lib/inference/unittests/fake_orch.py
+++ b/cvs/lib/inference/unittests/fake_orch.py
@@ -10,10 +10,17 @@
 
 
 class FakeOrch:
-    def __init__(self, exec_return=None):
+    def __init__(self, exec_return=None, hosts=None, exec_on_head_return=None):
+        self.hosts = list(hosts or ["node0"])
         self.exec_return = exec_return if exec_return is not None else {}
+        self.exec_on_head_return = exec_on_head_return if exec_on_head_return is not None else self.exec_return
         self.commands = []
+        self.exec_on_head_commands = []
 
-    def exec(self, cmd, **kwargs):
-        self.commands.append(cmd)
+    def exec(self, cmd, hosts=None, **kwargs):
+        self.commands.append((cmd, hosts))
         return self.exec_return
+
+    def exec_on_head(self, cmd, **kwargs):
+        self.exec_on_head_commands.append(cmd)
+        return self.exec_on_head_return
diff --git a/cvs/lib/inference/unittests/test_inferencex_atom_config_loader.py b/cvs/lib/inference/unittests/test_inferencex_atom_config_loader.py
index 142fe8aed..68a5344ea 100644
--- a/cvs/lib/inference/unittests/test_inferencex_atom_config_loader.py
+++ b/cvs/lib/inference/unittests/test_inferencex_atom_config_loader.py
@@ -29,20 +29,50 @@ class TestInferenceXAtomConfigLoader(unittest.TestCase):
     def test_load_mi300x_sample_config(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
-            "mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json"
+            "input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json"
         )
         variant = load_variant(config, _cluster_dict())
-        self.assertEqual(variant.framework, "inferencex_atom_single")
+        self.assertEqual(variant.framework, "inferencex_atom")
         self.assertEqual(variant.params.driver, "vllm")
         self.assertEqual(variant.expected_cells(), ["ISL=7168,OSL=1024,TP=8,CONC=64"])
         self.assertIn("enforce-eager", variant.roles.server.serve_args)
 
+    def test_deprecated_framework_alias_normalizes(self):
+        sweep = Sweep(
+            sequence_combinations=[SeqCombo(name="w1", isl="1024", osl="1024")],
+            runs=[Run(combo="w1", concurrency=128)],
+        )
+        variant = InferenceXAtomVariantConfig(
+            schema_version=1,
+            framework="inferencex_atom_single",
+            gpu_arch="mi300x",
+            enforce_thresholds=False,
+            paths={
+                "shared_fs": "/home/x",
+                "models_dir": "/home/x/models",
+                "log_dir": "/home/x/LOGS",
+                "hf_token_file": "/home/x/.hf",
+            },
+            model={"id": "deepseek-ai/DeepSeek-R1-0528", "remote": 0, "precision": "fp8"},
+            container={
+                "name": "c",
+                "image": "img",
+                "runtime": {"name": "docker", "args": {"volumes": ["/home/x:/home/x"]}},
+            },
+            params={
+                "driver": "atom",
+                "tensor_parallelism": "8",
+                "num_prompts": "100",
+            },
+            sweep=sweep,
+            thresholds={},
+        )
+        self.assertEqual(variant.framework, "inferencex_atom")
+
     def test_load_w1_mi300x_atom_variant(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
-            "mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
+            "input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
         )
         variant = load_variant(config, _cluster_dict())
         self.assertEqual(variant.threshold_json, "mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_threshold.json")
@@ -67,10 +97,49 @@ def test_load_w1_mi300x_atom_variant(self):
         ):
             self.assertIn(key, variant.thresholds[cell])
 
+    def test_load_w1_mi300x_multinode_variant(self):
+        root = Path(__file__).resolve().parents[3]
+        config = root / (
+            "input/config_file/inference/inferencex_atom/"
+            "mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json"
+        )
+        variant = load_variant(config, _cluster_dict())
+        self.assertEqual(variant.params.nnodes, "2")
+        self.assertEqual(variant.params.pipeline_parallel_size, "2")
+        self.assertEqual(variant.params.scaling_baseline_output_throughput, "1500")
+        self.assertTrue(variant.enforce_thresholds)
+        self.assertEqual(len(variant.expected_cells()), 15)
+        cell = "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16"
+        self.assertIn(cell, variant.expected_cells())
+        self.assertEqual(
+            variant.thresholds[cell]["scaling.efficiency_pct"],
+            {"kind": "min", "value": 50},
+        )
+
+    def test_load_w1_mi355x_multinode_variant(self):
+        root = Path(__file__).resolve().parents[3]
+        config = root / (
+            "input/config_file/inference/inferencex_atom/"
+            "mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_multi_config.json"
+        )
+        variant = load_variant(config, _cluster_dict())
+        self.assertEqual(variant.gpu_arch, "mi355x")
+        self.assertEqual(variant.params.nnodes, "2")
+        self.assertEqual(variant.params.pipeline_parallel_size, "2")
+        self.assertEqual(variant.params.scaling_baseline_output_throughput, "4000")
+        self.assertFalse(variant.enforce_thresholds)
+        self.assertEqual(len(variant.expected_cells()), 15)
+        cell = "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16"
+        self.assertIn(cell, variant.expected_cells())
+        self.assertEqual(
+            variant.thresholds[cell]["scaling.efficiency_pct"],
+            {"kind": "min", "value": 50},
+        )
+
     def test_load_w1_mi300x_smoke_variant(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
+            "input/config_file/inference/inferencex_atom/"
             "mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json"
         )
         variant = load_variant(config, _cluster_dict())
@@ -80,8 +149,7 @@ def test_load_w1_mi300x_smoke_variant(self):
     def test_load_w1_mi355x_atom_perf_variant_and_thresholds(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
-            "mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
+            "input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
         )
         variant = load_variant(config, _cluster_dict())
         self.assertEqual(variant.gpu_arch, "mi355x")
@@ -103,8 +171,7 @@ def test_load_w1_mi355x_atom_perf_variant_and_thresholds(self):
     def test_load_w1_mi355x_atom_mtp3_inline_bench_args(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
-            "mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json"
+            "input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json"
         )
         variant = load_variant(config, _cluster_dict())
         self.assertIn("--method", variant.roles.server.atom_args)
@@ -113,8 +180,7 @@ def test_load_w1_mi355x_atom_mtp3_inline_bench_args(self):
     def test_load_w1_mi355x_atom_mtp3_thresholds(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
-            "mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json"
+            "input/config_file/inference/inferencex_atom/mi355x_inferencex-atom-single_deepseek-r1_fp8_mtp3_config.json"
         )
         variant = load_variant(config, _cluster_dict())
         cell = "ISL=1024,OSL=1024,TP=8,CONC=256"
@@ -133,7 +199,7 @@ def test_orchestrator_container_includes_server_env(self):
         }
         variant = InferenceXAtomVariantConfig(
             schema_version=1,
-            framework="inferencex_atom_single",
+            framework="inferencex_atom",
             gpu_arch="mi300x",
             enforce_thresholds=False,
             paths={
@@ -159,8 +225,7 @@ def test_orchestrator_container_includes_server_env(self):
     def test_expand_sweep_matches_w1_perf(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
-            "mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
+            "input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
         )
         import json
 
@@ -174,8 +239,7 @@ def test_expand_sweep_matches_w1_perf(self):
     def test_w1_perf_threshold_health_gates_tight_when_enforcing(self):
         root = Path(__file__).resolve().parents[3]
         config = root / (
-            "input/config_file/inference/inferencex_atom_single/"
-            "mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
+            "input/config_file/inference/inferencex_atom/mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json"
         )
         variant = load_variant(config, _cluster_dict())
         self.assertTrue(variant.enforce_thresholds)
@@ -199,7 +263,7 @@ def test_atom_driver_requires_inline_atom_args(self):
         with self.assertRaises(ValueError):
             InferenceXAtomVariantConfig(
                 schema_version=1,
-                framework="inferencex_atom_single",
+                framework="inferencex_atom",
                 gpu_arch="mi300x",
                 enforce_thresholds=False,
                 paths={
@@ -239,6 +303,124 @@ def test_expand_sweep_parametrize_tier_ids(self):
         _, _, ids = expand_sweep_parametrize(sweep, ("metric_tier",))
         self.assertIn("w1-conc128-throughput", ids)
 
+    def test_load_w1_mi300x_vllm_parity_variant(self):
+        root = Path(__file__).resolve().parents[3]
+        config = root / (
+            "input/config_file/inference/inferencex_atom/"
+            "mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_config.json"
+        )
+        variant = load_variant(config, _cluster_dict())
+        self.assertEqual(variant.framework, "inferencex_atom_vllm_single")
+        self.assertEqual(variant.params.driver, "vllm")
+        self.assertIn("trust-remote-code", variant.roles.server.serve_args)
+        self.assertEqual(
+            variant.expected_cells(),
+            ["ISL=1024,OSL=1024,TP=8,CONC=128", "ISL=1024,OSL=1024,TP=8,CONC=256"],
+        )
+
+    def test_load_w1_mi300x_vllm_multinode_variant(self):
+        root = Path(__file__).resolve().parents[3]
+        config = root / (
+            "input/config_file/inference/inferencex_atom/"
+            "mi300x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json"
+        )
+        variant = load_variant(config, _cluster_dict())
+        self.assertEqual(variant.framework, "inferencex_atom_vllm_single")
+        self.assertEqual(variant.params.driver, "vllm")
+        self.assertEqual(variant.params.nnodes, "2")
+        self.assertEqual(variant.params.pipeline_parallel_size, "2")
+        self.assertEqual(variant.params.scaling_baseline_output_throughput, "1500")
+        self.assertFalse(variant.enforce_thresholds)
+        self.assertEqual(len(variant.expected_cells()), 15)
+        cell = "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16"
+        self.assertIn(cell, variant.expected_cells())
+        self.assertEqual(
+            variant.thresholds[cell]["scaling.efficiency_pct"],
+            {"kind": "min", "value": 50},
+        )
+        self.assertEqual(
+            variant.thresholds[cell]["compare.vllm.output_throughput_ratio"],
+            {"kind": "min", "value": 0.85},
+        )
+
+    def test_load_w1_mi355x_vllm_multinode_variant(self):
+        root = Path(__file__).resolve().parents[3]
+        config = root / (
+            "input/config_file/inference/inferencex_atom/"
+            "mi355x_inferencex-atom-single_deepseek-r1_fp8_vllm_perf_multi_config.json"
+        )
+        variant = load_variant(config, _cluster_dict())
+        self.assertEqual(variant.gpu_arch, "mi355x")
+        self.assertEqual(variant.framework, "inferencex_atom_vllm_single")
+        self.assertEqual(variant.params.driver, "vllm")
+        self.assertEqual(variant.params.nnodes, "2")
+        self.assertEqual(variant.params.scaling_baseline_output_throughput, "4000")
+        self.assertFalse(variant.enforce_thresholds)
+        cell = "ISL=512,OSL=512,TP=8,PP=2,NNODES=2,CONC=16"
+        self.assertIn(cell, variant.expected_cells())
+
+    def test_vllm_framework_rejects_atom_driver(self):
+        sweep = Sweep(
+            sequence_combinations=[SeqCombo(name="w1", isl="1024", osl="1024")],
+            runs=[Run(combo="w1", concurrency=128)],
+        )
+        thresholds = {"ISL=1024,OSL=1024,TP=8,CONC=128": placeholder_gated_threshold_cell()}
+        with self.assertRaises(ValueError):
+            InferenceXAtomVariantConfig(
+                schema_version=1,
+                framework="inferencex_atom_vllm_single",
+                gpu_arch="mi300x",
+                enforce_thresholds=False,
+                paths={
+                    "shared_fs": "/home/x",
+                    "models_dir": "/home/x/models",
+                    "log_dir": "/home/x/LOGS",
+                    "hf_token_file": "/home/x/.hf",
+                },
+                model={"id": "deepseek-ai/DeepSeek-R1-0528", "remote": 0, "precision": "fp8"},
+                container={
+                    "name": "c",
+                    "image": "img",
+                    "runtime": {"name": "docker", "args": {"volumes": ["/home/x:/home/x"]}},
+                },
+                roles={"server": {"serve_args": {"trust-remote-code": True}}},
+                params={"driver": "atom", "tensor_parallelism": "8"},
+                sweep=sweep,
+                thresholds=thresholds,
+            )
+
+    def test_server_session_key_differs_for_serve_args(self):
+        from types import SimpleNamespace
+
+        base = SimpleNamespace(
+            model=SimpleNamespace(id="m"),
+            params=SimpleNamespace(
+                driver="vllm",
+                tensor_parallelism="8",
+                nnodes="1",
+                pipeline_parallel_size="1",
+                master_addr="",
+                master_port="29501",
+            ),
+            roles=SimpleNamespace(server=SimpleNamespace(serve_args={"kv-cache-dtype": "fp8"})),
+        )
+        other = SimpleNamespace(
+            model=SimpleNamespace(id="m"),
+            params=SimpleNamespace(
+                driver="vllm",
+                tensor_parallelism="8",
+                nnodes="1",
+                pipeline_parallel_size="1",
+                master_addr="",
+                master_port="29501",
+            ),
+            roles=SimpleNamespace(server=SimpleNamespace(serve_args={"kv-cache-dtype": "bf16"})),
+        )
+        self.assertNotEqual(
+            server_session_key(base, "1024", "1024"),
+            server_session_key(other, "1024", "1024"),
+        )
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/cvs/lib/inference/unittests/test_inferencex_atom_orch_parse.py b/cvs/lib/inference/unittests/test_inferencex_atom_orch_parse.py
index 2911b183f..868cdfc22 100644
--- a/cvs/lib/inference/unittests/test_inferencex_atom_orch_parse.py
+++ b/cvs/lib/inference/unittests/test_inferencex_atom_orch_parse.py
@@ -21,10 +21,17 @@
 _TP = 8
 
 
-def _fake_variant(*, driver="vllm"):
+def _fake_variant(
+    *, driver="vllm", nnodes="1", pipeline_parallel_size="1", master_addr="", scaling_baseline_output_throughput=""
+):
     params = SimpleNamespace(
         driver=driver,
         tensor_parallelism=str(_TP),
+        pipeline_parallel_size=pipeline_parallel_size,
+        nnodes=nnodes,
+        master_addr=master_addr,
+        master_port="29501",
+        scaling_baseline_output_throughput=scaling_baseline_output_throughput,
         port_no="8000",
         random_range_ratio="0.8",
         random_prefix_len="0",
@@ -64,12 +71,8 @@ def test_parse_results_maps_client_metrics(self):
         w = raw
         self.assertIn("client.output_throughput", metrics)
         self.assertIn("client.mean_ttft_ms", metrics)
-        self.assertAlmostEqual(
-            metrics["client.per_gpu_throughput"], w["total_token_throughput"] / _TP
-        )
-        self.assertAlmostEqual(
-            metrics["client.output_tput_per_gpu"], w["output_throughput"] / _TP
-        )
+        self.assertAlmostEqual(metrics["client.per_gpu_throughput"], w["total_token_throughput"] / _TP)
+        self.assertAlmostEqual(metrics["client.output_tput_per_gpu"], w["output_throughput"] / _TP)
         self.assertEqual(metrics["client.p99_ttft_ms"], w["p99_ttft_ms"])
 
     def test_parse_results_w1_tail_metrics_from_widened_fixture(self):
@@ -114,10 +117,10 @@ def test_run_client_clears_stale_result_artifact(self):
             num_prompts=1000,
         )
         job.run_client()
-        rm_cmds = [c for c in orch.commands if c.startswith("rm -f ")]
+        rm_cmds = [c for c, _ in orch.commands if c.startswith("rm -f ")]
         self.assertEqual(len(rm_cmds), 1)
         self.assertIn(job._result_artifact, rm_cmds[0])
-        self.assertTrue(any("benchmark_serving" in c for c in orch.commands))
+        self.assertTrue(any("benchmark_serving" in c for c, _ in orch.commands))
 
     def test_parse_results_empty_artifact_raises(self):
         job = InferenceXAtomJob(
@@ -177,7 +180,7 @@ def test_build_server_cmd_suppresses_gpu_memory_env_vars(self):
             num_prompts=100,
         )
         job.build_server_cmd()
-        env_cmd = orch.commands[0]
+        env_cmd = orch.commands[0][0]
         self.assertNotIn("CVS_GPU_MEMORY_UTIL", env_cmd)
         self.assertNotIn("VLLM_GPU_MEMORY_UTIL", env_cmd)
         self.assertNotIn("VLLM_ENFORCE_EAGER", env_cmd)
@@ -185,9 +188,7 @@ def test_build_server_cmd_suppresses_gpu_memory_env_vars(self):
 
     def test_client_log_failures_traceback(self):
         job = InferenceXAtomJob(
-            orch=FakeOrch(
-                exec_return={"node0": "Traceback (most recent call last):\n  boom"}
-            ),
+            orch=FakeOrch(exec_return={"node0": "Traceback (most recent call last):\n  boom"}),
             variant=_fake_variant(driver="atom"),
             hf_token="tok",
             isl="1024",
@@ -255,6 +256,165 @@ def test_early_failure_regexes(self):
         self.assertTrue(job.CLIENT_LAUNCH_FAIL_RE.search("unrecognized arguments: --bad"))
         self.assertTrue(job.EARLY_FAILURE_RE.search("No such file or directory"))
 
+    def test_distributed_start_server_targets_each_host(self):
+        orch = FakeOrch(hosts=["10.0.0.1", "10.0.0.2"])
+        job = InferenceXAtomJob(
+            orch=orch,
+            variant=_fake_variant(driver="atom", nnodes="2", pipeline_parallel_size="2", master_addr="10.0.0.1"),
+            hf_token="tok",
+            isl="1024",
+            osl="1024",
+            concurrency=128,
+            num_prompts=100,
+        )
+        job.build_server_cmd(clear_atom_cache=False)
+        job.start_server()
+        launch_cmds = [c for c, hosts in orch.commands if hosts]
+        self.assertEqual(len(launch_cmds), 2)
+        self.assertNotIn("--node-rank", launch_cmds[0])
+        self.assertNotIn("--distributed-executor-backend", launch_cmds[0])
+        self.assertIn("openai_server", launch_cmds[0])
+
+    def test_distributed_vllm_start_server_uses_node_rank_and_pp(self):
+        orch = FakeOrch(hosts=["10.0.0.1", "10.0.0.2"])
+        job = InferenceXAtomJob(
+            orch=orch,
+            variant=_fake_variant(driver="vllm", nnodes="2", pipeline_parallel_size="2", master_addr="10.0.0.1"),
+            hf_token="tok",
+            isl="1024",
+            osl="1024",
+            concurrency=128,
+            num_prompts=100,
+        )
+        job.build_server_cmd()
+        job.start_server()
+        launch_cmds = [c for c, hosts in orch.commands if hosts]
+        self.assertEqual(len(launch_cmds), 2)
+        self.assertIn("vllm serve", launch_cmds[0])
+        self.assertNotIn("openai_server", launch_cmds[0])
+        self.assertIn("--node-rank 0", launch_cmds[0])
+        self.assertIn("--node-rank 1", launch_cmds[1])
+        self.assertIn("--nnodes 2", launch_cmds[0])
+        self.assertIn("--pipeline-parallel-size 2", launch_cmds[0])
+        self.assertIn("--master-addr 10.0.0.1", launch_cmds[0])
+        self.assertIn("--distributed-executor-backend mp", launch_cmds[0])
+
+    def test_vllm_distributed_argv(self):
+        job = InferenceXAtomJob(
+            orch=FakeOrch(hosts=["10.0.0.1", "10.0.0.2"]),
+            variant=_fake_variant(driver="vllm", nnodes="2", pipeline_parallel_size="2", master_addr="10.0.0.1"),
+            hf_token="tok",
+            isl="1024",
+            osl="1024",
+            concurrency=128,
+            num_prompts=100,
+        )
+        argv = job._vllm_distributed_argv(rank=1)
+        self.assertEqual(
+            argv,
+            [
+                "--node-rank",
+                "1",
+                "--master-addr",
+                "10.0.0.1",
+                "--master-port",
+                "29501",
+                "--nnodes",
+                "2",
+                "--pipeline-parallel-size",
+                "2",
+                "--distributed-executor-backend",
+                "mp",
+            ],
+        )
+
+    def test_distributed_atom_spmd_env_and_dp_when_tp_allows(self):
+        orch = FakeOrch(hosts=["10.0.0.1", "10.0.0.2"])
+        variant = _fake_variant(driver="atom", nnodes="2", pipeline_parallel_size="2", master_addr="10.0.0.1")
+        variant.params.tensor_parallelism = "4"
+        variant.roles.server.atom_args = ["-tp", "4"]
+        job = InferenceXAtomJob(
+            orch=orch,
+            variant=variant,
+            hf_token="tok",
+            isl="1024",
+            osl="1024",
+            concurrency=128,
+            num_prompts=100,
+        )
+        job.start_server()
+        launch_cmds = [c for c, hosts in orch.commands if hosts]
+        self.assertIn("-dp 2", launch_cmds[0])
+        self.assertIn("ATOM_DP_RANK=0", launch_cmds[0])
+        self.assertIn("ATOM_DP_RANK=1", launch_cmds[1])
+        self.assertIn("ATOM_DP_MASTER_IP=10.0.0.1", launch_cmds[0])
+
+    def test_distributed_atom_tp8_multinode_never_passes_vllm_flags(self):
+        orch = FakeOrch(hosts=["10.0.0.1", "10.0.0.2"])
+        variant = _fake_variant(driver="atom", nnodes="2", pipeline_parallel_size="2", master_addr="10.0.0.1")
+        variant.roles.server.atom_args = [
+            "-tp",
+            "8",
+            "--node-rank",
+            "1",
+            "--pipeline-parallel-size",
+            "2",
+        ]
+        job = InferenceXAtomJob(
+            orch=orch,
+            variant=variant,
+            hf_token="tok",
+            isl="1024",
+            osl="1024",
+            concurrency=128,
+            num_prompts=100,
+        )
+        argv = job._atom_server_argv(rank=1)
+        joined = " ".join(argv)
+        self.assertNotIn("--node-rank", joined)
+        self.assertNotIn("--pipeline-parallel-size", joined)
+        self.assertNotIn("--master-addr", joined)
+        self.assertIn("-tp 8", joined)
+        job.start_server()
+        launch_cmds = [c for c, hosts in orch.commands if hosts]
+        self.assertNotIn("--node-rank", launch_cmds[1])
+        self.assertNotIn("--pipeline-parallel-size", launch_cmds[1])
+
+    def test_distributed_client_uses_exec_on_head(self):
+        orch = FakeOrch(hosts=["10.0.0.1", "10.0.0.2"])
+        job = InferenceXAtomJob(
+            orch=orch,
+            variant=_fake_variant(driver="atom", nnodes="2", pipeline_parallel_size="2"),
+            hf_token="tok",
+            isl="1024",
+            osl="1024",
+            concurrency=128,
+            num_prompts=100,
+        )
+        job.run_client()
+        self.assertEqual(len(orch.exec_on_head_commands), 2)
+        self.assertTrue(any("benchmark_serving" in c for c in orch.exec_on_head_commands))
+
+    def test_parse_results_scaling_efficiency(self):
+        raw = json.loads((_FIXTURES / "vllm_results_sample.json").read_text())
+        job = InferenceXAtomJob(
+            orch=FakeOrch(exec_on_head_return={"head": json.dumps(raw)}),
+            variant=_fake_variant(
+                driver="atom",
+                nnodes="2",
+                scaling_baseline_output_throughput="100",
+            ),
+            hf_token="tok",
+            isl="1024",
+            osl="1024",
+            concurrency=128,
+            num_prompts=100,
+        )
+        metrics = job.parse_results()["head"]
+        self.assertIn("scaling.efficiency_pct", metrics)
+        expected = (raw["output_throughput"] / (100.0 * 2)) * 100.0
+        self.assertAlmostEqual(metrics["scaling.efficiency_pct"], expected)
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/cvs/lib/inference/unittests/test_inferencex_atom_parsing.py b/cvs/lib/inference/unittests/test_inferencex_atom_parsing.py
index db2447f1d..4d8c2e0e0 100644
--- a/cvs/lib/inference/unittests/test_inferencex_atom_parsing.py
+++ b/cvs/lib/inference/unittests/test_inferencex_atom_parsing.py
@@ -51,6 +51,34 @@ def test_tier_metric_specs_record_includes_non_tiered(self):
         self.assertIn("client.median_ttft_ms", specs)
         self.assertNotIn("client.output_throughput", specs)
 
+    def test_tier_metric_specs_scaling(self):
+        cell = {
+            "scaling.efficiency_pct": {"kind": "min", "value": 50},
+            "client.output_throughput": {"kind": "min_tok_s", "value": 1},
+        }
+        specs = tier_metric_specs(cell, "scaling")
+        self.assertEqual(specs, {"scaling.efficiency_pct": {"kind": "min", "value": 50}})
+
+    def test_tier_metric_specs_compare_vllm(self):
+        cell = {
+            "compare.vllm.output_throughput_ratio": {"kind": "min", "value": 0.85},
+            "client.output_throughput": {"kind": "min_tok_s", "value": 1},
+        }
+        specs = tier_metric_specs(cell, "compare")
+        self.assertEqual(
+            specs,
+            {"compare.vllm.output_throughput_ratio": {"kind": "min", "value": 0.85}},
+        )
+
+    def test_compare_vllm_metrics(self):
+        from cvs.lib.inference.utils.inferencex_atom_parsing import compare_vllm_metrics
+
+        actual = {"client.output_throughput": 2000, "client.mean_ttft_ms": 440}
+        reference = {"client.output_throughput": 2500, "client.mean_ttft_ms": 400}
+        out = compare_vllm_metrics(actual, reference)
+        self.assertAlmostEqual(out["compare.vllm.output_throughput_ratio"], 0.8)
+        self.assertAlmostEqual(out["compare.vllm.mean_ttft_ms_ratio"], 1.1)
+
     def test_gated_metrics_subset_of_client_metrics(self):
         client_short = {short for short, _unit in CLIENT_METRICS}
         missing = GATED_METRICS - client_short
diff --git a/cvs/lib/inference/utils/docs/inferencex-atom-parsing.md b/cvs/lib/inference/utils/docs/inferencex-atom-parsing.md
index 2ef1c2ded..3e68cb0a1 100644
--- a/cvs/lib/inference/utils/docs/inferencex-atom-parsing.md
+++ b/cvs/lib/inference/utils/docs/inferencex-atom-parsing.md
@@ -24,4 +24,4 @@ same artifact keys.
 
 - `inferencex_atom_orch.InferenceXAtomJob.parse_results`
 - `inferencex_atom_config_loader` threshold coverage (`gated_metrics=GATED_METRICS`)
-- `cvs.tests.inference.inferencex_atom.inferencex_atom_single` — `test_cell_metrics` tiers (throughput, ttft, tpot, health, record)
+- `cvs.tests.inference.inferencex_atom.inferencex_atom` — `test_cell_metrics` tiers (throughput, ttft, tpot, health, record)
diff --git a/cvs/lib/inference/utils/inferencex_atom_config_loader.py b/cvs/lib/inference/utils/inferencex_atom_config_loader.py
index 535bc7c0e..4f7723558 100644
--- a/cvs/lib/inference/utils/inferencex_atom_config_loader.py
+++ b/cvs/lib/inference/utils/inferencex_atom_config_loader.py
@@ -2,7 +2,7 @@
 Copyright 2025 Advanced Micro Devices, Inc.
 All rights reserved.
 
-InferenceX ATOM suite config schema (``framework: inferencex_atom_single``).
+InferenceX ATOM suite config schema (``inferencex_atom`` / ``inferencex_atom_vllm_single``).
 
 Generic paths/model/container/threshold plumbing lives in
 :mod:`cvs.lib.utils.config_loader`. Sweep selector types are shared with
@@ -13,7 +13,7 @@
 
 from typing import Any, Dict, List
 
-from pydantic import model_validator
+from pydantic import field_validator, model_validator
 from typing_extensions import Literal
 
 from cvs.lib.inference.utils.inferencing_config_loader import (
@@ -66,16 +66,38 @@ class InferenceXAtomParams(_Forbid):
     bench_max_failed_requests: str = "0"
     bench_extra_args: str = ""
     result_filename: str = "results"
+    # Multinode (M5): omit or set nnodes=1 for single-node runs. When nnodes>1,
+    # cluster node_dict must list the same number of hosts and test_setup_sshd runs.
+    nnodes: str = "1"
+    pipeline_parallel_size: str = "1"
+    master_addr: str = ""
+    master_port: str = "29501"
+    # Optional single-node reference output_throughput for scaling.efficiency_pct.
+    scaling_baseline_output_throughput: str = ""
 
 
 class InferenceXAtomRunCard(_Forbid):
     upstream_run_url: str = ""
     atom_image_pin: str = ""
+    vllm_image_pin: str = ""
     notes: str = ""
+    # Optional ATOM reference metrics JSON (cell_key -> client.* scalars) for M4 parity gates.
+    parity_reference_json: str = ""
+
+
+INFERENCEX_ATOM_FRAMEWORKS = ("inferencex_atom", "inferencex_atom_vllm_single")
 
 
 class InferenceXAtomVariantConfig(BaseVariantConfig):
-    framework: Literal["inferencex_atom_single"]
+    framework: Literal["inferencex_atom", "inferencex_atom_vllm_single"]
+
+    @field_validator("framework", mode="before")
+    @classmethod
+    def _normalize_deprecated_framework(cls, value):
+        if value == "inferencex_atom_single":
+            return "inferencex_atom"
+        return value
+
     gpu_arch: str
     run_card: InferenceXAtomRunCard = InferenceXAtomRunCard()
     roles: InferenceXAtomRoles = InferenceXAtomRoles()
@@ -83,14 +105,17 @@ class InferenceXAtomVariantConfig(BaseVariantConfig):
     sweep: Sweep
 
     def cell_key(self, isl, osl, concurrency):
-        return f"ISL={isl},OSL={osl},TP={self.params.tensor_parallelism},CONC={concurrency}"
+        p = self.params
+        key = f"ISL={isl},OSL={osl},TP={p.tensor_parallelism}"
+        if int(p.pipeline_parallel_size) > 1 or int(p.nnodes) > 1:
+            key += f",PP={p.pipeline_parallel_size}"
+        if int(p.nnodes) > 1:
+            key += f",NNODES={p.nnodes}"
+        return f"{key},CONC={concurrency}"
 
     def expected_cells(self) -> List[str]:
         by_name = {c.name: c for c in self.sweep.sequence_combinations}
-        return [
-            self.cell_key(by_name[r.combo].isl, by_name[r.combo].osl, r.concurrency)
-            for r in self.sweep.runs
-        ]
+        return [self.cell_key(by_name[r.combo].isl, by_name[r.combo].osl, r.concurrency) for r in self.sweep.runs]
 
     @model_validator(mode="after")
     def _check_thresholds_cover_sweep(self):
@@ -100,6 +125,22 @@ def _check_thresholds_cover_sweep(self):
             enforce_thresholds=self.enforce_thresholds,
             gated_metrics=GATED_METRICS,
         )
+        if int(self.params.nnodes) > 1 and (self.params.scaling_baseline_output_throughput or "").strip():
+            missing = []
+            for cell in self.expected_cells():
+                specs = self.thresholds.get(cell) or {}
+                if "scaling.efficiency_pct" not in specs:
+                    missing.append(cell)
+            if missing:
+                msg = (
+                    "multinode variant with scaling_baseline_output_throughput requires "
+                    f"scaling.efficiency_pct in every cell; missing: {missing}"
+                )
+                if self.enforce_thresholds:
+                    raise ValueError(msg)
+                import warnings
+
+                warnings.warn(f"{msg} (enforce_thresholds=false -> record-only)", stacklevel=2)
         return self
 
     @model_validator(mode="after")
@@ -111,6 +152,20 @@ def _atom_driver_requires_inline_server_args(self):
             )
         return self
 
+    @model_validator(mode="after")
+    def _framework_driver_consistency(self):
+        if self.framework == "inferencex_atom_vllm_single":
+            if self.params.driver != "vllm":
+                raise ValueError(
+                    "framework='inferencex_atom_vllm_single' requires params.driver='vllm'"
+                )
+            if not self.roles.server.serve_args:
+                raise ValueError(
+                    "framework='inferencex_atom_vllm_single' requires "
+                    "roles.server.serve_args (inline vLLM serve flags)"
+                )
+        return self
+
 
 def expand_sweep(sweep):
     """Expand a sweep into ``(cases, ids)`` for pytest parametrization."""
@@ -138,23 +193,60 @@ def reuse_server_flag(params) -> bool:
     return raw in ("true", "1", "yes")
 
 
+def _serve_args_session_key(serve_args: dict) -> tuple:
+    return tuple(sorted((str(k), str(v)) for k, v in (serve_args or {}).items()))
+
+
 def server_session_key(variant_config, isl, osl):
     """Stable key for server reuse across sweep cells with identical model/shape."""
     p = variant_config.params
+    server = variant_config.roles.server
+    if p.driver == "vllm":
+        server_tokens = _serve_args_session_key(server.serve_args)
+    else:
+        server_tokens = tuple(server.atom_args)
     return (
         variant_config.model.id,
         p.driver,
         str(isl),
         str(osl),
-        tuple(variant_config.roles.server.atom_args),
+        server_tokens,
         p.tensor_parallelism,
+        p.nnodes,
+        p.pipeline_parallel_size,
+        p.master_addr,
+        p.master_port,
     )
 
 
-def expand_sweep_parametrize(sweep, fixturenames):
+def load_parity_reference_cells(path: str) -> dict[str, dict]:
+    """Load cell_key -> metric dict from a parity reference JSON file."""
+    if not (path or "").strip():
+        return {}
+    from pathlib import Path
+    import json
+
+    ref_path = Path(path)
+    if not ref_path.is_file():
+        raise FileNotFoundError(f"parity reference not found: {path}")
+    data = json.loads(ref_path.read_text(encoding="utf-8"))
+    if not isinstance(data, dict):
+        raise ValueError(f"parity reference must be a JSON object: {path}")
+    out: dict[str, dict] = {}
+    for key, value in data.items():
+        if str(key).startswith("_"):
+            continue
+        if not isinstance(value, dict):
+            raise ValueError(f"parity reference cell {key!r} must be an object")
+        out[str(key)] = value
+    return out
+
+
+def expand_sweep_parametrize(sweep, fixturenames, *, metric_tier_order=None):
     """Build pytest parametrize args for inference or metric-tier collection."""
     from cvs.lib.inference.utils.inferencex_atom_parsing import METRIC_TIER_ORDER
 
+    tiers = metric_tier_order or METRIC_TIER_ORDER
     cases, ids = expand_sweep(sweep)
     if "metric_tier" in fixturenames:
         if not cases:
@@ -162,7 +254,7 @@ def expand_sweep_parametrize(sweep, fixturenames):
         tier_cases = []
         tier_ids = []
         for (combo, c), cid in zip(cases, ids):
-            for tier in METRIC_TIER_ORDER:
+            for tier in tiers:
                 tier_cases.append((combo, c, tier))
                 tier_ids.append(f"{cid}-{tier}")
         return ("seq_combo,concurrency,metric_tier", tier_cases, tier_ids)
@@ -192,7 +284,6 @@ def placeholder_gated_threshold_cell(
 ) -> Dict[str, Any]:
     """Return one sweep cell's ``client.*`` specs covering every ``GATED_METRICS`` member."""
     loose_ms = {"kind": "max_ms", "value": 1_000_000}
-    loose_tok = {"kind": "min_tok_s", "value": 0}
     return {
         "client.total_token_throughput": {"kind": "min_tok_s", "value": total_token_throughput_min},
         "client.output_throughput": {"kind": "min_tok_s", "value": output_throughput_min},
diff --git a/cvs/lib/inference/utils/inferencex_atom_parsing.py b/cvs/lib/inference/utils/inferencex_atom_parsing.py
index e31f930b5..e6f18170a 100644
--- a/cvs/lib/inference/utils/inferencex_atom_parsing.py
+++ b/cvs/lib/inference/utils/inferencex_atom_parsing.py
@@ -57,34 +57,89 @@
         "success_rate",
         "failed",
     ),
+    "scaling": ("efficiency_pct",),
+}
+
+SCALING_METRICS: tuple[str, ...] = METRIC_TIERS["scaling"]
+SCALING_METRIC_UNITS: dict[str, str] = {"efficiency_pct": "%"}
+
+COMPARE_VLLM_METRICS: tuple[str, ...] = (
+    "output_throughput_ratio",
+    "mean_ttft_ms_ratio",
+)
+COMPARE_VLLM_METRIC_UNITS: dict[str, str] = {
+    "output_throughput_ratio": "ratio",
+    "mean_ttft_ms_ratio": "ratio",
 }
 
 METRIC_TIER_ORDER: tuple[str, ...] = tuple(METRIC_TIERS.keys()) + ("record",)
+VLLM_METRIC_TIER_ORDER: tuple[str, ...] = tuple(METRIC_TIERS.keys()) + ("compare", "record")
 
 _tiered = {m for names in METRIC_TIERS.values() for m in names}
-RECORD_METRICS: tuple[str, ...] = tuple(
-    short for short, _unit in CLIENT_METRICS if short not in _tiered
-)
+RECORD_METRICS: tuple[str, ...] = tuple(short for short, _unit in CLIENT_METRICS if short not in _tiered)
 
 ENFORCED_METRICS = frozenset(_tiered)
 
 
-def to_client_metrics(raw, *, tp, isl):
+def scaling_efficiency_pct(actual_output_throughput, *, baseline_single_node, nnodes):
+    """Linear scaling efficiency: actual / (single-node baseline × nnodes)."""
+    denom = _safe_div(baseline_single_node, 1)
+    if denom is None or int(nnodes) < 1:
+        return None
+    ideal = denom * int(nnodes)
+    if ideal <= 0:
+        return None
+    return _safe_div(actual_output_throughput, ideal)
+
+
+def to_client_metrics(raw, *, tp, isl, scaling_baseline_output_throughput=None, nnodes=1):
     """Map an ATOM ``results.json`` dict to the ``client.*`` namespace for IX."""
     m = _vllm_to_client_metrics(raw, tp=tp, isl=isl)
     m["client.output_tput_per_gpu"] = _safe_div(raw.get("output_throughput"), tp)
+    if scaling_baseline_output_throughput is not None:
+        eff = scaling_efficiency_pct(
+            raw.get("output_throughput"),
+            baseline_single_node=scaling_baseline_output_throughput,
+            nnodes=nnodes,
+        )
+        if eff is not None:
+            m["scaling.efficiency_pct"] = eff * 100.0
     return m
 
 
+def compare_vllm_metrics(actual: dict, reference: dict) -> dict[str, float]:
+    """M4 parity ratios: vLLM cell metrics relative to ATOM reference cell."""
+    out: dict[str, float] = {}
+    vllm_tput = actual.get("client.output_throughput")
+    atom_tput = reference.get("client.output_throughput")
+    ratio = _safe_div(vllm_tput, atom_tput)
+    if ratio is not None:
+        out["compare.vllm.output_throughput_ratio"] = ratio
+    vllm_ttft = actual.get("client.mean_ttft_ms")
+    atom_ttft = reference.get("client.mean_ttft_ms")
+    ratio = _safe_div(vllm_ttft, atom_ttft)
+    if ratio is not None:
+        out["compare.vllm.mean_ttft_ms_ratio"] = ratio
+    return out
+
+
 def tier_metric_specs(thresholds_cell: dict, tier: str) -> dict[str, dict]:
-    """Return ``client.*`` threshold specs for one tier in a sweep cell."""
+    """Return threshold specs for one tier in a sweep cell."""
     if tier == "record":
         names = RECORD_METRICS
+        prefix = "client."
+    elif tier == "scaling":
+        names = SCALING_METRICS
+        prefix = "scaling."
+    elif tier == "compare":
+        names = COMPARE_VLLM_METRICS
+        prefix = "compare.vllm."
     else:
         names = METRIC_TIERS.get(tier, ())
+        prefix = "client."
     specs = {}
     for short in names:
-        full = f"client.{short}"
+        full = f"{prefix}{short}"
         spec = thresholds_cell.get(full)
         if spec is not None:
             specs[full] = spec
diff --git a/cvs/lib/inference_lib.py b/cvs/lib/inference_lib.py
index 217fea5fb..173cc3a0b 100644
--- a/cvs/lib/inference_lib.py
+++ b/cvs/lib/inference_lib.py
@@ -28,7 +28,7 @@ def __init__(self, *args, **kwargs):
         raise NotImplementedError(
             "InferenceMax is deprecated. Use ``cvs.lib.inference.inferencex_atom_orch.InferenceXAtomJob`` "
             "with ``ContainerOrchestrator`` from the tests under ``cvs.tests.inference.inferencex_atom`` "
-            "(``inferencex_atom_single`` suite and schema_version 1 configs)."
+            "(``inferencex_atom`` suite and schema_version 1 configs)."
         )
 
 
@@ -60,8 +60,7 @@ def _detect_framework(cls, inference_config_dict):
         """
         if 'inferencemax_repo' in inference_config_dict:
             log.warning(
-                "inferencemax_repo detected; InferenceMax is deprecated — "
-                "use the inferencex_atom_single suite instead"
+                "inferencemax_repo detected; InferenceMax is deprecated — use the inferencex_atom suite instead"
             )
             return 'inferencex_atom'
         elif 'vllm_script_path' in inference_config_dict:
@@ -110,7 +109,7 @@ def create_job(
 
         framework_lower = framework.lower()
         if framework_lower == 'inferencemax':
-            log.warning("framework='inferencemax' is deprecated; use inferencex_atom_single")
+            log.warning("framework='inferencemax' is deprecated; use inferencex_atom")
             framework_lower = 'inferencex_atom'
 
         if framework_lower not in cls._FRAMEWORK_CLASSES:
diff --git a/cvs/lib/report/README.md b/cvs/lib/report/README.md
index a57fc1fef..f2e59429b 100644
--- a/cvs/lib/report/README.md
+++ b/cvs/lib/report/README.md
@@ -4,9 +4,8 @@ CVS can attach an HTML/JSON **suite report** to pytest runs that use `--html`. R
 generated at **session end** and bundled into the same results zip as the pytest HTML report.
 They are **render-only** — they do not change pass/fail or threshold enforcement.
 
-**IX-atom reference:** `presets/inferencex_atom.py` + shim `presets/inferencex_atom_single.py`
-(auto-loaded for `cvs run inferencex_atom_single`). Suite owners also see Step 8 in
-`cvs/lib/inference/ADDING_A_SUITE.md`.
+**IX-atom reference:** `presets/inferencex_atom.py` (auto-loaded for `cvs run inferencex_atom`).
+Suite owners also see Step 8 in `cvs/lib/inference/ADDING_A_SUITE.md`.
 
 ## Quick start
 
@@ -25,8 +24,10 @@ Your suite must already collect:
 | `variant_config` | Thresholds, `enforce_thresholds`, `cell_key(isl, osl, conc)` |
 | `lifecycle` | `.record(nodeid, label, seconds)` on server/client stages |
 
+## Unit tests
+
 ```bash
-cvs run inferencex_atom_single --cluster_file ... --config_file ... --html=~/cvs_results/run.html
+cvs run inferencex_atom --cluster_file ... --config_file ... --html=~/cvs_results/run.html
 python -m pytest cvs/lib/report/unittests/ -q
 ```
 
diff --git a/cvs/lib/report/cell_build.py b/cvs/lib/report/cell_build.py
index 38e7e742f..a34313d7d 100644
--- a/cvs/lib/report/cell_build.py
+++ b/cvs/lib/report/cell_build.py
@@ -141,20 +141,15 @@ def build_cell_record(
                 "label": label,
                 "metric": full,
                 "actual": actual,
-                "unit": config.metric_units.get(short, ""),
+                "unit": config.metric_units.get(short.rsplit(".", 1)[-1], ""),
                 "spec": spec,
                 "status": metric_pass(full, actual, spec) if enforce and spec else "record",
-                "bar_pct": bar_pct(float(actual), spec)
-                if spec is not None and actual is not None
-                else None,
+                "bar_pct": bar_pct(float(actual), spec) if spec is not None and actual is not None else None,
                 "margin": margin_text(actual, spec) if spec else None,
             }
         )
 
-    tiers = {
-        tier: tier_status(config, actuals, thresholds_cell, tier, enforce)
-        for tier in config.metric_tier_order
-    }
+    tiers = {tier: tier_status(config, actuals, thresholds_cell, tier, enforce) for tier in config.metric_tier_order}
     pytest_links = resolve_pytest_nodeids_for_cell(config, lifecycle_report, conc)
 
     return {
@@ -183,7 +178,10 @@ def build_all_cells(
     lifecycle_report: Mapping[str, list],
 ) -> List[dict]:
     cells: List[dict] = []
-    for key, host_dict in sorted(inf_res_dict.items(), key=lambda kv: (kv[0][4], kv[0][5]) if isinstance(kv[0], tuple) and len(kv[0]) >= 6 else (0, 0)):
+    for key, host_dict in sorted(
+        inf_res_dict.items(),
+        key=lambda kv: (kv[0][4], kv[0][5]) if isinstance(kv[0], tuple) and len(kv[0]) >= 6 else (0, 0),
+    ):
         if not isinstance(key, tuple) or len(key) != 6:
             continue
         if not isinstance(host_dict, dict) or not host_dict:
diff --git a/cvs/lib/report/inference_payload.py b/cvs/lib/report/inference_payload.py
index 9a1f8e509..3b91d8ccb 100644
--- a/cvs/lib/report/inference_payload.py
+++ b/cvs/lib/report/inference_payload.py
@@ -68,9 +68,7 @@ def sweep_has_multi_shape_comparison(cells: List[dict]) -> bool:
     return len(shapes) >= 2 and len(concurrencies) >= 2
 
 
-def build_chart_series(
-    config: InferenceReportConfig, cells: List[dict]
-) -> Dict[str, List[dict]]:
+def build_chart_series(config: InferenceReportConfig, cells: List[dict]) -> Dict[str, List[dict]]:
     """Per-metric sweep charts grouped by ISL/OSL shape.
 
     Each metric maps to a list of ``{isl, osl, label, points}`` entries so
@@ -80,7 +78,7 @@ def build_chart_series(
     groups = group_cells_by_shape(cells)
     series: Dict[str, List[dict]] = {}
     for chart in config.chart_series:
-        full = config.full_metric(chart.metric_suffix)
+        full = chart.metric_key or config.full_metric(chart.metric_suffix)
         group_entries: List[dict] = []
         for (isl, osl), group_cells in sorted(groups.items()):
             values_by_conc = metric_values_by_concurrency(group_cells, full)
@@ -180,11 +178,7 @@ def _build_run_card_display(
         run_card_rows.append((label, value, is_link))
 
     run_card_display = extend_run_card_display(run_card_rows, prov)
-    run_card_display = [
-        (label, value, is_link)
-        for label, value, is_link in run_card_display
-        if label != "CVS version"
-    ]
+    run_card_display = [(label, value, is_link) for label, value, is_link in run_card_display if label != "CVS version"]
     display_labels = {label for label, _value, _link in run_card_display}
     if prov.get("cvs_version") and "CVS" not in display_labels:
         run_card_display.append(("CVS", str(prov["cvs_version"]), False))
@@ -246,9 +240,7 @@ def build_inference_report_payload(
     if cvs_version:
         prov.setdefault("cvs_version", cvs_version)
 
-    run_card_display, run_card_notes, generated_at = _build_run_card_display(
-        config, variant_config, prov
-    )
+    run_card_display, run_card_notes, generated_at = _build_run_card_display(config, variant_config, prov)
 
     chart_series = build_chart_series(config, cells)
     panels = _build_panels(config, cells, report_dir)
diff --git a/cvs/lib/report/presets/_inference_suite_template.py b/cvs/lib/report/presets/_inference_suite_template.py
index f7ba8b32e..dfb24aa6e 100644
--- a/cvs/lib/report/presets/_inference_suite_template.py
+++ b/cvs/lib/report/presets/_inference_suite_template.py
@@ -5,9 +5,9 @@
 **Copy this file** to ``cvs/lib/report/presets/<cvs_run_stem>.py``.
 
 The filename must match the pytest module stem from ``cvs run <stem>`` (e.g.
-``inferencex_atom_single`` → ``presets/inferencex_atom_single.py``).
+``inferencex_atom`` → ``presets/inferencex_atom.py``).
 
-**Reference:** ``inferencex_atom.py`` (full preset) + ``inferencex_atom_single.py`` (auto-load shim).
+**Reference:** ``inferencex_atom.py`` (IX-atom preset; auto-loaded when ``cvs run inferencex_atom``).
 See ``cvs/lib/report/README.md`` for the IX-atom end-to-end example.
 
 Suite owners fill in the TODOs below, keep collecting ``inf_res_dict`` during tests,
diff --git a/cvs/lib/report/presets/inferencex_atom.py b/cvs/lib/report/presets/inferencex_atom.py
index e6aa1dc04..c8f013a02 100644
--- a/cvs/lib/report/presets/inferencex_atom.py
+++ b/cvs/lib/report/presets/inferencex_atom.py
@@ -15,6 +15,7 @@
 from cvs.lib.inference.utils.inferencex_atom_parsing import (
     CLIENT_METRIC_UNITS,
     METRIC_TIER_ORDER,
+    SCALING_METRIC_UNITS,
     tier_metric_specs,
 )
 from cvs.lib.report.chart_presets import DEFAULT_PERF_CHART_SERIES
@@ -23,6 +24,17 @@
     provenance_link_rows,
     thresholds_run_card_row,
 )
+from cvs.lib.report.types import ReportChartSeries
+
+_INFERENCEX_ATOM_CHART_SERIES = (
+    *DEFAULT_PERF_CHART_SERIES,
+    ReportChartSeries(
+        "efficiency_pct",
+        "Scaling efficiency",
+        "%",
+        metric_key="scaling.efficiency_pct",
+    ),
+)
 
 
 def _atom_run_card_display(variant: Any, provenance: dict) -> List[Tuple[str, str, bool]]:
@@ -46,12 +58,12 @@ def _atom_run_card_display(variant: Any, provenance: dict) -> List[Tuple[str, st
     report_basename="inferencex_atom_run_deck",
     title="IX Run Deck",
     subtitle="InferenceX ATOM \u00b7 lab performance summary",
-    footer="CVS inferencex_atom_single \u00b7 render-only \u00b7 does not affect gates",
+    footer="CVS inferencex_atom \u00b7 render-only \u00b7 does not affect gates",
     link_name="IX Run Deck",
     results_columns=INFERENCEX_ATOM_RESULTS_COLUMNS,
     metric_tier_order=METRIC_TIER_ORDER,
     tier_metric_specs=tier_metric_specs,
-    metric_units=CLIENT_METRIC_UNITS,
+    metric_units={**CLIENT_METRIC_UNITS, **SCALING_METRIC_UNITS},
     metric_prefix="client.",
     cell_highlights=(
         ("output_throughput", "Output tok/s"),
@@ -59,8 +71,9 @@ def _atom_run_card_display(variant: Any, provenance: dict) -> List[Tuple[str, st
         ("mean_tpot_ms", "Mean TPOT (ms)"),
         ("p99_ttft_ms", "P99 TTFT (ms)"),
         ("p95_tpot_ms", "P95 TPOT (ms)"),
+        ("scaling.efficiency_pct", "Scaling eff. (%)"),
     ),
-    chart_series=DEFAULT_PERF_CHART_SERIES,
+    chart_series=_INFERENCEX_ATOM_CHART_SERIES,
     inference_test_substring="test_inferencex_atom_inference",
     row_card_extras=False,
     row_card_test_names=("test_cell_metrics",),
diff --git a/cvs/lib/report/presets/inferencex_atom_single.py b/cvs/lib/report/presets/inferencex_atom_single.py
deleted file mode 100644
index c6f4642e3..000000000
--- a/cvs/lib/report/presets/inferencex_atom_single.py
+++ /dev/null
@@ -1,10 +0,0 @@
-'''

-Copyright 2025 Advanced Micro Devices, Inc.

-All rights reserved.

-

-Auto-loaded when running ``cvs run inferencex_atom_single`` (stem matches filename).

-'''

-

-from cvs.lib.report.presets.inferencex_atom import (

-    INFERENCEX_ATOM_REPORT_CONFIG as INFERENCEX_ATOM_SINGLE_REPORT_CONFIG,

-)

diff --git a/cvs/lib/report/presets/inferencex_atom_vllm.py b/cvs/lib/report/presets/inferencex_atom_vllm.py
new file mode 100644
index 000000000..91bd5f7c6
--- /dev/null
+++ b/cvs/lib/report/presets/inferencex_atom_vllm.py
@@ -0,0 +1,67 @@
+'''
+Copyright 2025 Advanced Micro Devices, Inc.
+All rights reserved.
+
+Auto-loaded when running ``cvs run inferencex_atom_vllm_single``.
+'''
+
+from cvs.lib.inference.inference_suite_results_table import INFERENCEX_ATOM_RESULTS_COLUMNS
+from cvs.lib.inference.utils.inferencex_atom_parsing import (
+    CLIENT_METRIC_UNITS,
+    COMPARE_VLLM_METRIC_UNITS,
+    SCALING_METRIC_UNITS,
+    VLLM_METRIC_TIER_ORDER,
+    tier_metric_specs,
+)
+from cvs.lib.report.presets.inferencex_atom import (
+    _INFERENCEX_ATOM_CHART_SERIES,
+    _atom_run_card_display,
+    _CELL_LIFECYCLE,
+    _SESSION_LIFECYCLE,
+)
+from cvs.lib.report.types import InferenceReportConfig, ReportChartSeries
+
+_VLLM_PARITY_CHART_SERIES = (
+    *_INFERENCEX_ATOM_CHART_SERIES,
+    ReportChartSeries(
+        "output_throughput_ratio",
+        "vLLM / ATOM output tput",
+        "ratio",
+        metric_key="compare.vllm.output_throughput_ratio",
+    ),
+)
+
+_VLLM_RESULTS_COLUMNS = INFERENCEX_ATOM_RESULTS_COLUMNS + (
+    ("vLLM/ATOM tput", "compare.vllm.output_throughput_ratio"),
+    ("vLLM/ATOM TTFT", "compare.vllm.mean_ttft_ms_ratio"),
+)
+
+INFERENCEX_ATOM_VLLM_SINGLE_REPORT_CONFIG = InferenceReportConfig(
+    suite_id="inferencex_atom_vllm",
+    report_basename="inferencex_atom_vllm_run_deck",
+    title="IX Run Deck",
+    subtitle="InferenceX vLLM parity \u00b7 W1 reference sweep",
+    footer="CVS inferencex_atom_vllm_single \u00b7 render-only \u00b7 does not affect gates",
+    link_name="IX vLLM Run Deck",
+    results_columns=_VLLM_RESULTS_COLUMNS,
+    metric_tier_order=VLLM_METRIC_TIER_ORDER,
+    tier_metric_specs=tier_metric_specs,
+    metric_units={**CLIENT_METRIC_UNITS, **SCALING_METRIC_UNITS, **COMPARE_VLLM_METRIC_UNITS},
+    metric_prefix="client.",
+    cell_highlights=(
+        ("output_throughput", "Output tok/s"),
+        ("mean_ttft_ms", "Mean TTFT (ms)"),
+        ("mean_tpot_ms", "Mean TPOT (ms)"),
+        ("compare.vllm.output_throughput_ratio", "vLLM/ATOM tput"),
+        ("compare.vllm.mean_ttft_ms_ratio", "vLLM/ATOM TTFT"),
+    ),
+    chart_series=_VLLM_PARITY_CHART_SERIES,
+    inference_test_substring="test_inferencex_atom_vllm_inference",
+    session_lifecycle_labels=_SESSION_LIFECYCLE,
+    cell_lifecycle_labels=_CELL_LIFECYCLE,
+    row_card_extras=False,
+    row_card_test_names=("test_cell_metrics",),
+    interactive_viewer=True,
+    viewer_cell_threshold=16,
+    run_card_display_builder=_atom_run_card_display,
+)
diff --git a/cvs/lib/report/types.py b/cvs/lib/report/types.py
index 56bf0ae2a..4c3458a13 100644
--- a/cvs/lib/report/types.py
+++ b/cvs/lib/report/types.py
@@ -32,6 +32,7 @@ class ReportChartSeries:
     title: str
     unit: str
     invert: bool = False
+    metric_key: str = ""
 
 
 @dataclass(frozen=True)
@@ -63,15 +64,15 @@ class InferenceReportConfig:
     interactive_viewer: bool = True
     viewer_cell_threshold: int = 24
     prev_run_json: str = ""
-    run_card_display_builder: RunCardDisplayFn = field(
-        default=lambda _variant, _prov: [("Suite", "inference", False)]
-    )
+    run_card_display_builder: RunCardDisplayFn = field(default=lambda _variant, _prov: [("Suite", "inference", False)])
 
     @property
     def gated_tiers(self) -> tuple[str, ...]:
         return tuple(t for t in self.metric_tier_order if t != self.record_tier)
 
     def full_metric(self, short: str) -> str:
+        if "." in short:
+            return short
         if short.startswith(f"{self.metric_prefix}"):
             return short
         return f"{self.metric_prefix}{short}"
diff --git a/cvs/tests/inference/inferencex_atom/conftest.py b/cvs/tests/inference/inferencex_atom/conftest.py
index 08d77b795..b3dffb361 100644
--- a/cvs/tests/inference/inferencex_atom/conftest.py
+++ b/cvs/tests/inference/inferencex_atom/conftest.py
@@ -33,14 +33,22 @@ def _log_variant_run_card(variant_config):
         f"model={variant_config.model.id}",
     ]
     atom_args = variant_config.roles.server.atom_args
+    serve_args = variant_config.roles.server.serve_args
     if atom_args:
         parts.append(f"atom_args={len(atom_args)} tokens")
+    if serve_args:
+        parts.append(f"serve_args={len(serve_args)} keys")
     if rc.atom_image_pin:
         parts.append(f"image_pin={rc.atom_image_pin}")
+    if rc.vllm_image_pin:
+        parts.append(f"vllm_image_pin={rc.vllm_image_pin}")
     if rc.upstream_run_url:
         parts.append(f"upstream_run={rc.upstream_run_url}")
     if rc.notes:
         parts.append(f"notes={rc.notes}")
+    if int(variant_config.params.nnodes) > 1:
+        parts.append(f"nnodes={variant_config.params.nnodes}")
+        parts.append(f"pp={variant_config.params.pipeline_parallel_size}")
     log.info("InferenceX ATOM run card: %s", "; ".join(parts))
 
 
@@ -55,6 +63,7 @@ def _emit_variant_run_card(variant_config):
     "test_setup_sshd": 1,
     "test_model_fetch": 2,
     "test_inferencex_atom_inference": 3,
+    "test_inferencex_atom_vllm_inference": 3,
     "test_cell_metrics": 4,
     "test_print_results_table": 5,
     "test_teardown": 6,
diff --git a/cvs/tests/inference/inferencex_atom/inferencex_atom_single.py b/cvs/tests/inference/inferencex_atom/inferencex_atom.py
similarity index 86%
rename from cvs/tests/inference/inferencex_atom/inferencex_atom_single.py
rename to cvs/tests/inference/inferencex_atom/inferencex_atom.py
index 3f1d92d2d..90bf4b9a7 100644
--- a/cvs/tests/inference/inferencex_atom/inferencex_atom_single.py
+++ b/cvs/tests/inference/inferencex_atom/inferencex_atom.py
@@ -12,10 +12,10 @@
 from cvs.lib import globals
 from cvs.lib.inference.inference_suite_lifecycle import (
     sweep_cell_result_key,
-    test_launch_container,
-    test_model_fetch,
-    test_setup_sshd,
-    test_teardown,
+    test_launch_container,  # noqa: F401
+    test_model_fetch,  # noqa: F401
+    test_setup_sshd,  # noqa: F401
+    test_teardown,  # noqa: F401
 )
 from cvs.lib.inference.inferencex_atom_orch import InferenceXAtomJob
 from cvs.lib.inference.utils.inferencex_atom_config_loader import (
@@ -27,6 +27,7 @@
     CLIENT_METRIC_UNITS as _METRIC_UNITS,
     METRIC_TIERS,
     RECORD_METRICS,
+    SCALING_METRIC_UNITS,
     tier_metric_specs,
 )
 from cvs.lib.utils.verdict import evaluate_all
@@ -45,9 +46,7 @@ def _tier_display_metric(tier):
 def pytest_generate_tests(metafunc):
     config_file = metafunc.config.getoption("config_file")
     if not config_file or not os.path.isfile(config_file):
-        raise pytest.UsageError(
-            f"--config_file not found or not specified: {config_file!r}"
-        )
+        raise pytest.UsageError(f"--config_file not found or not specified: {config_file!r}")
     with open(config_file) as fp:
         raw = json.load(fp)
     spec = expand_sweep_parametrize(raw.get("sweep", {}), metafunc.fixturenames)
@@ -140,21 +139,25 @@ def test_cell_metrics(
     specs = tier_metric_specs(thresholds_cell, metric_tier)
 
     display = _tier_display_metric(metric_tier)
-    full = f"client.{display}"
+    if metric_tier == "scaling":
+        full = f"scaling.{display}"
+        unit = SCALING_METRIC_UNITS.get(display, "%")
+    else:
+        full = f"client.{display}"
+        unit = _METRIC_UNITS.get(display, metric_tier)
     value = actuals.get(full)
-    unit = _METRIC_UNITS.get(display, metric_tier)
     request.node.user_properties.append(("metric_value", value))
     request.node.user_properties.append(("metric_unit", unit))
 
     if not variant_config.enforce_thresholds or metric_tier == "record":
         return
     if not specs:
+        if metric_tier == "scaling" and int(variant_config.params.nnodes) <= 1:
+            pytest.skip("scaling tier not configured for single-node runs")
         pytest.fail(f"no threshold specs for tier {metric_tier!r} in cell {cell!r}")
     # ATOM benchmark_serving may omit some tail percentiles even when
     # metric_percentiles requests them; only gate metrics present in actuals.
-    specs = {
-        k: v for k, v in specs.items() if k in actuals and actuals[k] is not None
-    }
+    specs = {k: v for k, v in specs.items() if k in actuals and actuals[k] is not None}
     if not specs:
         pytest.fail(
             f"no assertable threshold specs for tier {metric_tier!r} in cell {cell!r} "
diff --git a/cvs/tests/inference/inferencex_atom/inferencex_atom_vllm.py b/cvs/tests/inference/inferencex_atom/inferencex_atom_vllm.py
new file mode 100644
index 000000000..3be020a99
--- /dev/null
+++ b/cvs/tests/inference/inferencex_atom/inferencex_atom_vllm.py
@@ -0,0 +1,196 @@
+'''
+Copyright 2025 Advanced Micro Devices, Inc.
+All rights reserved.
+
+InferenceX vLLM parity suite (M4): same sweep layout as ``inferencex_atom``
+with ``framework: inferencex_atom_vllm_single`` and ``params.driver: vllm``.
+'''
+
+import json
+import os
+import time
+
+import pytest
+
+from cvs.lib import globals
+from cvs.lib.inference.inference_suite_lifecycle import (
+    sweep_cell_result_key,
+    test_launch_container,
+    test_model_fetch,
+    test_setup_sshd,
+    test_teardown,
+)
+from cvs.lib.inference.inferencex_atom_orch import InferenceXAtomJob
+from cvs.lib.inference.utils.inferencex_atom_config_loader import (
+    expand_sweep_parametrize,
+    load_parity_reference_cells,
+    reuse_server_flag,
+    server_session_key,
+)
+from cvs.lib.inference.utils.inferencex_atom_parsing import (
+    CLIENT_METRIC_UNITS as _METRIC_UNITS,
+    COMPARE_VLLM_METRIC_UNITS,
+    METRIC_TIERS,
+    RECORD_METRICS,
+    SCALING_METRIC_UNITS,
+    VLLM_METRIC_TIER_ORDER,
+    compare_vllm_metrics,
+    tier_metric_specs,
+)
+from cvs.lib.utils.verdict import evaluate_all
+from cvs.tests.inference.inferencex_atom._shared import test_print_results_table  # noqa: F401
+
+log = globals.log
+
+
+def _tier_display_metric(tier):
+    if tier == "record":
+        return RECORD_METRICS[0] if RECORD_METRICS else "output_throughput"
+    if tier == "compare":
+        from cvs.lib.inference.utils.inferencex_atom_parsing import COMPARE_VLLM_METRICS
+
+        return COMPARE_VLLM_METRICS[0]
+    names = METRIC_TIERS.get(tier, ())
+    return names[0] if names else tier
+
+
+@pytest.fixture(scope="module")
+def parity_reference_cells(variant_config):
+    path = (variant_config.run_card.parity_reference_json or "").strip()
+    if not path:
+        return {}
+    return load_parity_reference_cells(path)
+
+
+def pytest_generate_tests(metafunc):
+    config_file = metafunc.config.getoption("config_file")
+    if not config_file or not os.path.isfile(config_file):
+        raise pytest.UsageError(
+            f"--config_file not found or not specified: {config_file!r}"
+        )
+    with open(config_file) as fp:
+        raw = json.load(fp)
+    spec = expand_sweep_parametrize(
+        raw.get("sweep", {}),
+        metafunc.fixturenames,
+        metric_tier_order=VLLM_METRIC_TIER_ORDER,
+    )
+    if spec:
+        argnames, argvalues, ids = spec
+        metafunc.parametrize(argnames, argvalues, ids=ids)
+
+
+def test_inferencex_atom_vllm_inference(
+    orch,
+    variant_config,
+    hf_token,
+    seq_combo,
+    concurrency,
+    inf_res_dict,
+    server_session,
+    lifecycle,
+    parity_reference_cells,
+    request,
+):
+    if lifecycle.failed:
+        pytest.skip("a prior lifecycle stage failed")
+
+    isl = seq_combo["isl"]
+    osl = seq_combo["osl"]
+    p = variant_config.params
+    job = InferenceXAtomJob.from_variant(
+        orch=orch,
+        variant=variant_config,
+        hf_token=hf_token,
+        isl=isl,
+        osl=osl,
+        concurrency=concurrency,
+    )
+
+    session_key = server_session_key(variant_config, isl, osl)
+    reuse = reuse_server_flag(p) and server_session.get("key") == session_key
+
+    try:
+        if not reuse:
+            job.stop_server()
+            job.build_server_cmd()
+            t = time.monotonic()
+            job.start_server()
+            job.wait_ready()
+            lifecycle.record(request.node.nodeid, "server_ready", time.monotonic() - t)
+            if reuse_server_flag(p):
+                server_session["key"] = session_key
+        else:
+            log.info("reusing vLLM server across sweep cell (key=%s)", session_key)
+            job.prepare_cell_out_dir()
+        t_client = time.monotonic()
+        job.run_client()
+        job.wait_client_complete()
+        results = job.parse_results()
+    except Exception:
+        lifecycle.failed = True
+        raise
+
+    cell = variant_config.cell_key(isl, osl, concurrency)
+    ref_cell = parity_reference_cells.get(cell)
+    if ref_cell:
+        for host, metrics in results.items():
+            metrics.update(compare_vllm_metrics(metrics, ref_cell))
+
+    inf_res_dict[sweep_cell_result_key(variant_config, seq_combo, isl, osl, concurrency)] = results
+    lifecycle.record(request.node.nodeid, "client_complete", time.monotonic() - t_client)
+
+
+def test_cell_metrics(
+    seq_combo,
+    concurrency,
+    metric_tier,
+    inf_res_dict,
+    variant_config,
+    lifecycle,
+    request,
+):
+    if lifecycle.failed:
+        pytest.skip("a prior lifecycle stage failed")
+    isl = seq_combo["isl"]
+    osl = seq_combo["osl"]
+    key = sweep_cell_result_key(variant_config, seq_combo, isl, osl, concurrency)
+    if key not in inf_res_dict:
+        pytest.skip(f"no recorded results for cell {key!r} (inference did not run)")
+    host_dict = inf_res_dict[key]
+    _host, actuals = next(iter(host_dict.items()))
+    cell = variant_config.cell_key(isl, osl, concurrency)
+    thresholds_cell = variant_config.thresholds.get(cell) or {}
+    specs = tier_metric_specs(thresholds_cell, metric_tier)
+
+    display = _tier_display_metric(metric_tier)
+    if metric_tier == "scaling":
+        full = f"scaling.{display}"
+        unit = SCALING_METRIC_UNITS.get(display, "%")
+    elif metric_tier == "compare":
+        full = f"compare.vllm.{display}"
+        unit = COMPARE_VLLM_METRIC_UNITS.get(display, "ratio")
+    else:
+        full = f"client.{display}"
+        unit = _METRIC_UNITS.get(display, metric_tier)
+    value = actuals.get(full)
+    request.node.user_properties.append(("metric_value", value))
+    request.node.user_properties.append(("metric_unit", unit))
+
+    if not variant_config.enforce_thresholds or metric_tier == "record":
+        return
+    if not specs:
+        if metric_tier == "scaling" and int(variant_config.params.nnodes) <= 1:
+            pytest.skip("scaling tier not configured for single-node runs")
+        if metric_tier == "compare" and not (variant_config.run_card.parity_reference_json or "").strip():
+            pytest.skip("compare tier requires run_card.parity_reference_json")
+        pytest.fail(f"no threshold specs for tier {metric_tier!r} in cell {cell!r}")
+    specs = {
+        k: v for k, v in specs.items() if k in actuals and actuals[k] is not None
+    }
+    if not specs:
+        pytest.fail(
+            f"no assertable threshold specs for tier {metric_tier!r} in cell {cell!r} "
+            f"(metrics missing from benchmark artifact)"
+        )
+    evaluate_all(actuals, specs)
diff --git a/docs/how-to/run-cvs-tests.rst b/docs/how-to/run-cvs-tests.rst
index c48ff7e21..5443f2941 100644
--- a/docs/how-to/run-cvs-tests.rst
+++ b/docs/how-to/run-cvs-tests.rst
@@ -42,7 +42,7 @@ You can list available tests using either `cvs run` (with no arguments) or `cvs
       • install_ibperf_tools
   
     cvs.tests.inference.inferencex_atom (1 test suite)
-      • inferencex_atom_single
+      • inferencex_atom
   
     cvs.tests.inference.pytorch_xdit (2 test suites)
       • pytorch_xdit_flux1_dev_single
@@ -631,11 +631,11 @@ You can list all available InferenceX ATOM test cases using the CLI:
 
 .. code:: bash
 
-  cvs list inferencex_atom_single
+  cvs list inferencex_atom
 
 .. code:: text
 
-  Available tests in inferencex_atom_single:
+  Available tests in inferencex_atom:
     - test_launch_container
     - test_inferencex_atom_inference
     - test_print_results_table
@@ -644,15 +644,15 @@ You can list all available InferenceX ATOM test cases using the CLI:
 Use these scripts to run the InferenceX ATOM tests. Supply your own suite JSON
 (``schema_version: 1`` variant config); see :doc:`../reference/configuration-files/inferencex_atom`.
 After ``cvs copy-config``, keep **one** ``*threshold.json`` in the same directory as the
-``--config_file`` you pass (per-variant subdirs under ``~/input/.../inferencex_atom_single/``).
-Copy-paste lab commands: ``cvs/input/config_file/inference/inferencex_atom_single/README.md``.
+``--config_file`` you pass (per-variant subdirs under ``~/input/.../inferencex_atom/``).
+Copy-paste lab commands: ``cvs/input/config_file/inference/inferencex_atom/README.md``.
 
 .. code:: bash
 
   TS=$(date +%Y%m%d_%H%M%S)
-  cvs run inferencex_atom_single \
+  cvs run inferencex_atom \
     --cluster_file ~/input/cluster_file/mi300x_atom_single.json \
-    --config_file ~/input/config_file/inference/inferencex_atom_single/smoke/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json \
+    --config_file ~/input/config_file/inference/inferencex_atom/smoke/mi300x_inferencex-atom-single_deepseek-r1_fp8_smoke_config.json \
     --html=~/cvs_results/${TS}_ix-atom-smoke_mi300x.html \
     --self-contained-html \
     --log-file=~/cvs_results/${TS}_ix-atom-smoke_mi300x.log \
diff --git a/docs/install/cvs-install.rst b/docs/install/cvs-install.rst
index a08de631f..f2d46ee0e 100644
--- a/docs/install/cvs-install.rst
+++ b/docs/install/cvs-install.rst
@@ -312,8 +312,8 @@ CVS provides comprehensive inference testing configurations for various LLM serv
 
 .. code:: bash
 
-     cvs copy-config inference/inferencex_atom_single/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json --output ~/my_inferencex_atom_config.json
-     cvs copy-config inference/inferencex_atom_single/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json --output ~/my_inferencex_atom_threshold.json
+     cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_config.json --output ~/my_inferencex_atom_config.json
+     cvs copy-config inference/inferencex_atom/mi300x_inferencex-atom-single_gpt-oss-120b_bf16_threshold.json --output ~/my_inferencex_atom_threshold.json
 
 2. Edit the files and modify these parameters:
 
diff --git a/docs/reference/configuration-files/inferencex_atom.rst b/docs/reference/configuration-files/inferencex_atom.rst
index de64b4fcc..8d1ffa759 100644
--- a/docs/reference/configuration-files/inferencex_atom.rst
+++ b/docs/reference/configuration-files/inferencex_atom.rst
@@ -19,7 +19,7 @@ The suite checks:
 - **Result verification**: Tiered ``client.*`` thresholds when ``enforce_thresholds`` is true
 
 Configs use flat ``*_config.json`` + sibling ``*_threshold.json`` pairs under
-``cvs/input/config_file/inference/inferencex_atom_single/``. Filename pattern:
+``cvs/input/config_file/inference/inferencex_atom/``. Filename pattern:
 ``{gpu}_inferencex-atom-single_{model}_{precision}[_{mode}]_config.json``.
 Pass ``--config_file`` to the ``*_config.json``; :func:`cvs.lib.utils.config_loader.substitute_config`
 discovers the sole sibling ``*threshold.json`` in the **config file's parent directory** when
@@ -28,10 +28,10 @@ loading fails with an ambiguous-threshold ``ValueError``.
 
 **Lab ``~/input`` layout:** the repo keeps every variant flat in one tree, but after
 ``cvs copy-config`` you should place each run's config + threshold pair in a dedicated
-subdirectory (for example ``~/input/.../inferencex_atom_single/smoke/``) so only one
+subdirectory (for example ``~/input/.../inferencex_atom/smoke/``) so only one
 threshold file sits beside the config you pass to ``--config_file``. Alternatively set
 ``"threshold_json"`` in the config to an explicit path. See the in-tree README at
-``cvs/input/config_file/inference/inferencex_atom_single/README.md`` for copy-paste commands.
+``cvs/input/config_file/inference/inferencex_atom/README.md`` for copy-paste commands.
 
 **Cluster file:** use ``cvs/input/cluster_file/mi300x_atom_single.json`` (or ``mi355x_atom_single.json``).
 Container ``name`` must match the variant (``inferencex_atom_mi300x`` / ``inferencex_atom_mi355x``);
@@ -47,7 +47,7 @@ cluster nodes (``cluster_file`` ``mgmt_ip`` / ``node_dict``) and runs ``sudo doc
 ``roles.server.serve_args`` on ``vllm_single``). When ``params.driver`` is ``atom``, ``atom_args``
 is required. MTP3 variants also set ``params.bench_extra_args`` (for example ``--use-chat-template``).
 
-Pytest and HTML layout (inferencex_atom_single)
+Pytest and HTML layout (inferencex_atom)
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. list-table::
@@ -82,7 +82,7 @@ Pytest and HTML layout (inferencex_atom_single)
 Example variant layout
 ======================
 
-Each stem has ``<stem>_config.json`` (``schema_version: 1``, ``framework: inferencex_atom_single``)
+Each stem has ``<stem>_config.json`` (``schema_version: 1``, ``framework: inferencex_atom``)
 and sibling ``<stem>_threshold.json``. In the CVS source tree many stems share one directory;
 on a lab machine, copy only the pair you need into a per-variant subdirectory (or set
 ``threshold_json``). See ``mi300x_inferencex-atom-single_deepseek-r1_fp8_perf_config.json``
@@ -120,7 +120,7 @@ Top-level blocks follow the DTNI variant schema. InferenceX ATOM-specific keys:
      - Example
      - Description
    * - ``framework``
-     - ``inferencex_atom_single``
+     - ``inferencex_atom``
      - Suite identifier for :func:`load_variant`.
    * - ``gpu_arch``
      - ``mi300x``
@@ -158,6 +158,15 @@ Top-level blocks follow the DTNI variant schema. InferenceX ATOM-specific keys:
    * - ``params.reuse_server_across_sweep``
      - ``true``
      - Skip server restart when only concurrency changes between sweep cells.
+   * - ``params.nnodes`` / ``params.pipeline_parallel_size``
+     - ``2`` / ``2``
+     - Multinode distributed serve (M5). Omit or set ``nnodes: "1"`` for single-node.
+   * - ``params.master_addr`` / ``params.master_port``
+     - head VPC IP / ``29501``
+     - Rendezvous for distributed ATOM/vLLM executor; defaults to cluster head when empty.
+   * - ``params.scaling_baseline_output_throughput``
+     - ``1500``
+     - Single-node reference ``output_throughput`` for ``scaling.efficiency_pct`` (record-only).
    * - ``params.server_warmup_wait_s`` / ``client_initial_wait_s``
      - ``330`` / ``120``
      - Config-driven server warmup and client poll floor (shorter on smoke configs).