Add RDNA4 GPU support (RX 9070 XT, RX 9070, RX 9060 XT, RX 9000 XT) f…

docs/compatibility.rst

@@ -65,24 +65,67 @@ The following table lists the validated Hyperloom version and component combinat
 
 .. role:: version-start
 
-.. table::
-   :widths: 6 27 10 10 14 30 3
-   :align: left
+.. list-table::
+   :header-rows: 1
+   :widths: 15 30 45 35 20 15 5
    :class: compat-matrix format-big-table
 
-+-------------------+---------------------------+------------------------+----------------------------+---------------+-------------+-----------------------------+
-| Hyperloom version | Component                 | GPU                    | ROCm version               | Ubuntu        | Python      | GitHub                      |
-+===================+===========================+========================+============================+===============+=============+=============================+
-| 1.0.0a3           | `TraceLens 0.1.0`_        | Hardware-agnostic      | No dependency              | OS-independent| >= 3.6      | |tracelens-github|          |
-+                   +---------------------------+------------------------+----------------------------+---------------+-------------+-----------------------------+
-|                   | `GEAK 4.0.0`_             | MI300X, MI325X, MI355X | 6.4.x, 7.0.x, 7.1.x, 7.2.x | 22.04, 24.04  | 3.8, 3.12   | |geak-github|               |
-+                   +---------------------------+------------------------+----------------------------+---------------+-------------+-----------------------------+
-|                   | `IntelliKit 0.1.0`_       | MI300X, MI325X, MI355X | 7.2.x                      | 22.04, 24.04  | >= 3.10     | |intellikit-github|         |
-+                   +---------------------------+------------------------+----------------------------+---------------+-------------+-----------------------------+
-|                   | `AgentKernelArena 0.2.0`_ | MI300X, MI325X, MI355X | 7.2.x                      | 22.04, 24.04  | >= 3.10     | |agent-kernel-arena-github| |
-+                   +---------------------------+------------------------+----------------------------+---------------+-------------+-----------------------------+
-|                   | `Magpie 0.2.0`_           | MI300X, MI325X, MI355X | 7.0.x, 7.1.x, 7.2.x        | 22.04, 24.04  | >= 3.10     | |magpie-github|             |
-+-------------------+---------------------------+------------------------+----------------------------+---------------+-------------+-----------------------------+
+   * - Hyperloom version
+     - Component
+     - GPU
+     - ROCm version
+     - Ubuntu
+     - Python
+     - GitHub
+   * - 1.0.0a1
+     - `TraceLens 0.1.0`_
+     - Hardware-agnostic
+     - No dependency
+     - OS-independent
+     - >= 3.6
+     - |tracelens-github|
+   * - 1.0.0a2
+     - `TraceLens 0.1.0`_
+     - Hardware-agnostic
+     - No dependency
+     - OS-independent
+     - >= 3.6
+     - |tracelens-github|
+   * - 1.0.0a3
+     - `TraceLens 0.1.0`_
+     - Hardware-agnostic
+     - No dependency
+     - OS-independent
+     - >= 3.6
+     - |tracelens-github|
+   * - 1.0.0a2
+     - `GEAK 4.0.0`_
+     - MI300X, MI325X, MI355X, RX 9070 XT, RX 9070, RX 9060 XT, R9000
+     - 6.4.x, 7.0.x, 7.1.x, 7.2.x
+     - 22.04, 24.04
+     - 3.8, 3.12
+     - |geak-github|
+   * - 1.0.0a2
+     - `IntelliKit 0.1.0`_
+     - MI300X, MI325X, MI355X
+     - 7.2.x
+     - 22.04, 24.04
+     - >= 3.10
+     - |intellikit-github|
+   * - 1.0.0a2
+     - `AgentKernelArena 0.2.0`_
+     - MI300X, MI325X, MI355X
+     - 7.2.x
+     - 22.04, 24.04
+     - >= 3.10
+     - |agent-kernel-arena-github|
+   * - 1.0.0a2
+     - `Magpie 0.2.0`_
+     - MI300X, MI325X, MI355X, RX 9070 XT, RX 9070, RX 9060 XT, R9000
+     - 7.0.x, 7.1.x, 7.2.x (Linux), 7.3+ (Windows for RDNA4)
+     - 22.04, 24.04, Win 11
+     - >= 3.10
+     - |magpie-github|
 
 .. _TraceLens 0.1.0: https://rocm.docs.amd.com/projects/tracelens/en/docs-0.1.0/
 .. _GEAK 4.0.0: https://rocm.docs.amd.com/projects/geak/en/docs-4.0.0/
@@ -196,3 +239,4 @@ rocm720, prefer ``docker`` mode with
 These are recommended defaults, not hard pins. Framework and ROCm versions are
 overridable via env (``SGLANG_REF``, ``SGLANG_ROCM_EXTRA``, ``VLLM_VERSION``,
 ``VLLM_ROCM_VARIANT``) for hosts that need a different pinned stack.
+

src/hyperloom/inference_optimizer/cli/parser.py

@@ -272,7 +272,7 @@ def _build_parser() -> argparse.ArgumentParser:
     opt.add_argument(
         "--gpu-type",
         type=str.lower,
-        choices=["mi300x", "mi308x", "mi325x", "mi355x"],
+        choices=["mi300x", "mi308x", "mi325x", "mi355x", "rx9070xt", "rx9070", "rx9060xt", "r9000"],
         default=None,
         help="Hint for the real target GPU. The rocm-smi probe always "
         "wins when both are present and disagree; a WARN is "

src/hyperloom/inference_optimizer/gpu_types.py

@@ -6,29 +6,52 @@
 from __future__ import annotations
 
 import os
+import sys
 
 
-_AMD_GPU_TYPES = frozenset({"mi300x", "mi308x", "mi325x", "mi355x"})
+_AMD_GPU_TYPES = frozenset({
+    "mi300x", "mi308x", "mi325x", "mi355x",
+    "rx9070xt", "rx9070", "rx9060xt", "r9000"
+})
 
 _GFX_TO_RUNNER: dict[str, str] = {
-    # gfx arch -> Magpie runner label, so launchers and runtime materializers
-    # agree on the selected benchmark script.
     "gfx942": "mi300x",
     "gfx950": "mi355x",
+    "gfx1201": "rx9070xt",
+    "gfx1203": "r9000",
+    "gfx1206": "rx9060xt",
+    "gfx1207": "rx9070",
 }
 
 _AMD_GPU_DISPATCH_IDENTITIES: dict[str, tuple[str, int]] = {
     "mi300x": ("gfx942", 304),
     "mi308x": ("gfx942", 304),
     "mi325x": ("gfx942", 304),
     "mi355x": ("gfx950", 256),
+    "rx9070xt": ("gfx1201", 64),
+    "r9000": ("gfx1203", 64),
+    "rx9070": ("gfx1207", 56),
+    "rx9060xt": ("gfx1206", 32),
 }
 
+# GPU types for which Magpie ships a benchmark runner script (sglang_<runner>.sh).
+# MI308X/MI325X are mapped to the MI300X runner by _gpu_runner_type(); every
+# RDNA4 SKU resolves to itself. The matching rx9xxx runner scripts are shipped
+# by the AMD-AGI/Magpie package (see docs/components/magpie.md).
+_SHIPPED_MAGPIE_RUNNERS: frozenset[str] = frozenset({
+    "mi300x", "mi355x",
+    "rx9070xt", "rx9070", "rx9060xt", "r9000",
+})
+
 
 def _gpu_runner_type(gpu_type: str) -> str:
-    """Return the Magpie runner label for a resolved real GPU type."""
+    """Return the Magpie runner label for a resolved real GPU type.
+
+    MI308X/MI325X collapse to the MI300X runner (shipped as
+    ``sglang_mi300x.sh``); every other supported AMD type resolves to itself.
+    """
     normalized = str(gpu_type or "").strip().lower()
-    if normalized in ("mi325x", "mi308x"):
+    if normalized in ("mi300x", "mi308x", "mi325x"):
         return "mi300x"
     return normalized
 
@@ -51,22 +74,44 @@ def _resolve_gpu_type(
 
 
 def _autodetect_gpu_type() -> str | None:
-    """Return mi300x|mi308x|mi325x|mi355x or None if undetectable."""
+    """Return mi300x|mi308x|mi325x|mi355x|rx9070xt|rx9070|rx9060xt|r9000 or None."""
     import subprocess
 
     try:
-        out = subprocess.run(
-            ["rocm-smi", "--showproductname"],
-            capture_output=True,
-            text=True,
-            timeout=5,
-        ).stdout.upper()
-        for tag in ("MI355X", "MI325X", "MI308X", "MI300X"):
-            if tag in out:
-                return tag.lower()
+        if sys.platform == "win32":
+            out = subprocess.run(
+                ["hipConfig", "--show-device"],
+                capture_output=True,
+                text=True,
+                timeout=10,
+            ).stdout.upper()
+            # Match whole device names; bare "RX 9070" (no " XT") must still
+            # resolve to rx9070 rather than collapsing into the empty set.
+            for tag in (
+                "RX 9070 XT", "R9000", "RX 9060 XT",
+                "RX 9070 ", "RX 9070", "RX 9060",
+                "MI355X", "MI300X",
+            ):
+                if tag in out:
+                    return tag.replace("MI", "mi").replace("RX ", "rx").replace(" XT", "xt").replace(" ", "").lower()
+            # Fallbacks for "AMD Radeon RX 9070" / "AMD Radeon Pro R9000" (full
+            # product strings) so bare non-XT SKUs are detected.
+            import re as _re
+            bare = _re.search(r"AMD\s+Radeon\s+(Pro\s+R9000|RX\s+\d{3,4}\s*XT|RX\s+\d{3,4})\b", out)
+            if bare:
+                tok = bare.group(1).replace("Pro ", "Pro_").replace("XT", "xt").replace(" ", "").replace("_", "").replace("RX", "rx").lower()
+                return "r9000" if "pro" in tok or "r9000" in tok else tok
+        else:
+            out = subprocess.run(
+                ["rocm-smi", "--showproductname"],
+                capture_output=True,
+                text=True,
+                timeout=5,
+            ).stdout.upper()
+            for tag in ("MI355X", "MI325X", "MI308X", "MI300X", "RX9070XT", "RX9070", "RX9060XT", "R9000"):
+                if tag in out:
+                    return tag.lower()
     except (FileNotFoundError, subprocess.TimeoutExpired, PermissionError, OSError):
-        # rocm-smi missing / slow / not permitted; fall through to the torch
-        # gcnArchName probe below (autodetect is best-effort).
         pass
     try:
         import torch

src/hyperloom/orchestrator/kernel/roofline_ceiling.py

@@ -58,6 +58,40 @@
     "fp4": 10066.4,
     "float4": 10066.4,
 }
+#: RDNA4 vendor-dense peak TFLOPS — per SKU. The four RX 9000-series SKUs do
+#: *not* share a single table: AMD's published CU count and game clock differ
+#: per SKU (see wiki/RDNA_4). rx9070xt (64 CU @ 2.4 GHz) is the baseline; the
+#: other rows are scaled by (CUs * game_clock) so FP32/FP16/FP8 throughput is
+#: distinct and correctly ordered: 9070xt > 9000xt > 9070 > 9060xt. Absolute
+#: magnitude is preserved for rx9070xt so existing KB recipes stay calibrated.
+_RDNA4_PEAK_TFLOPS_RX9070XT: dict[str, float] = {
+    "bf16": 194.8,
+    "bfloat16": 194.8,
+    "fp16": 194.8,
+    "float16": 194.8,
+    "fp8": 389.0,
+    "float8_e4m3fn": 389.0,
+    "float8_e5m2": 389.0,
+    "fp32": 14.6,
+    "float32": 14.6,
+    "fp64": 0.46,
+    "float64": 0.46,
+    "int4": 1557.0,
+}
+
+
+def _rdna4_peak_tflops(cu_count: int, boost_ghz: float, base: dict[str, float]) -> dict[str, float]:
+    """Scale a reference RDNA4 (RX 9070 XT, 64 CU @ 2.4 GHz) peak table by
+    the CU-count × boost-clock ratio for the target SKU."""
+    ref_cu = 64
+    ref_ghz = 2.4
+    scale = (cu_count * boost_ghz) / (ref_cu * ref_ghz)
+    return {dtype: round(v * scale, 1) for dtype, v in base.items()}
+
+
+_RDNA4_PEAK_TFLOPS_R9000: dict[str, float] = _RDNA4_PEAK_TFLOPS_RX9070XT
+_RDNA4_PEAK_TFLOPS_RX9070: dict[str, float] = _rdna4_peak_tflops(56, 2.07, _RDNA4_PEAK_TFLOPS_RX9070XT)
+_RDNA4_PEAK_TFLOPS_RX9060XT: dict[str, float] = _rdna4_peak_tflops(32, 2.53, _RDNA4_PEAK_TFLOPS_RX9070XT)
 HW_SPECS: dict[str, dict[str, Any]] = {
     "mi300x": {
         "hbm_gb": 192.0,
@@ -79,6 +113,26 @@
         "hbm_bw_gbps": 8000.0,
         "peak_tflops": _MI355X_PEAK_TFLOPS,
     },
+    "rx9070xt": {
+        "gddr_gb": 16.0,
+        "hbm_bw_gbps": 640.0,
+        "peak_tflops": _RDNA4_PEAK_TFLOPS_RX9070XT,
+    },
+    "rx9070": {
+        "gddr_gb": 16.0,
+        "hbm_bw_gbps": 640.0,
+        "peak_tflops": _RDNA4_PEAK_TFLOPS_RX9070,
+    },
+    "rx9060xt": {
+        "gddr_gb": 16.0,
+        "hbm_bw_gbps": 640.0,
+        "peak_tflops": _RDNA4_PEAK_TFLOPS_RX9060XT,
+    },
+    "r9000": {
+        "gddr_gb": 32.0,
+        "hbm_bw_gbps": 640.0,
+        "peak_tflops": _RDNA4_PEAK_TFLOPS_R9000,
+    },
 }
 
 
@@ -1574,6 +1628,29 @@ def read_baseline_server_args(state: Any) -> str:
     "float32": 137.0,
 }
 
+
+#: RDNA4 sustained (achievable) TFLOPS — reference is the RX 9070 XT
+#: (64 CU, Navi 48). Other SKUs scaled by (CUs × game_clock) ratio.
+#: The R9000 Pro is the workstation twin of the RX 9070 XT (same die,
+#: 64 CU) and therefore shares its achievable table.
+_RDNA4_ACHIEVABLE_TFLOPS_RX9070XT: dict[str, float] = {
+    "bf16": 140.0,
+    "bfloat16": 140.0,
+    "fp16": 135.0,
+    "float16": 135.0,
+    "fp8": 300.0,
+    "float8_e4m3fn": 300.0,
+    "float8_e5m2": 300.0,
+    "fp64": 0.035,
+    "float64": 0.035,
+    "int4": 1100.0,
+    "fp32": 11.0,
+    "float32": 11.0,
+}
+_RDNA4_ACHIEVABLE_TFLOPS_R9000: dict[str, float] = _RDNA4_ACHIEVABLE_TFLOPS_RX9070XT
+_RDNA4_ACHIEVABLE_TFLOPS_RX9070: dict[str, float] = _rdna4_peak_tflops(56, 2.07, _RDNA4_ACHIEVABLE_TFLOPS_RX9070XT)
+_RDNA4_ACHIEVABLE_TFLOPS_RX9060XT: dict[str, float] = _rdna4_peak_tflops(32, 2.53, _RDNA4_ACHIEVABLE_TFLOPS_RX9070XT)
+
 HW_SPECS_ACHIEVABLE: dict[str, dict[str, Any]] = {
     "mi300x": {
         "hbm_bw_gbps": 5300.0,
@@ -1590,6 +1667,26 @@ def read_baseline_server_args(state: Any) -> str:
         "hbm_gb": 288.0,
         "peak_tflops": _MI355X_ACHIEVABLE_TFLOPS,
     },
+    "rx9070xt": {
+        "hbm_bw_gbps": 640.0,
+        "hbm_gb": 16.0,
+        "peak_tflops": _RDNA4_ACHIEVABLE_TFLOPS_RX9070XT,
+    },
+    "rx9070": {
+        "hbm_bw_gbps": 640.0,
+        "hbm_gb": 16.0,
+        "peak_tflops": _RDNA4_ACHIEVABLE_TFLOPS_RX9070,
+    },
+    "rx9060xt": {
+        "hbm_bw_gbps": 640.0,
+        "hbm_gb": 16.0,
+        "peak_tflops": _RDNA4_ACHIEVABLE_TFLOPS_RX9060XT,
+    },
+    "r9000": {
+        "hbm_bw_gbps": 640.0,
+        "hbm_gb": 32.0,
+        "peak_tflops": _RDNA4_ACHIEVABLE_TFLOPS_R9000,
+    },
 }