Swap the cache config for the default cases of gfx1250-GEMM-A8W8_BLOCKSCALE.json

[jpvillam-amd]

Motivation

Speed up vLLM DeepseekV4-Flash

Technical Details

ca shows much higher decode throughput

Test Plan

tested on lm_eval with around an 8x improvement e2e and micro benchmark

│  #!/usr/bin/env python3
│  """Micro-benchmark: cache_modifier ".cg" vs ".ca" in aiter's gfx1250 block-scale
│  FP8 GEMM, at decode-sized M. Self-contained (torch + aiter only)."""
│  import time, copy, torch
│  from aiter.ops.triton.gemm.basic.gemm_a8w8_blockscale import gemm_a8w8_blockscale
│
│  N, K, M = 8192, 4096, 8                       # a DeepSeek-V4-Flash wo_a shape, decode M
│  dev = torch.device("cuda")
│
│  A  = torch.randn(M, K, device=dev).to(torch.float8_e4m3fn)
│  B  = torch.randn(N, K, device=dev).to(torch.float8_e4m3fn)
│  As = torch.ones(M, K // 128, device=dev)      # fp32 block scales
│  Bs = torch.ones(N // 128, K // 128, device=dev)
│
│  cfg = dict(BLOCK_SIZE_M=16, BLOCK_SIZE_N=32, BLOCK_SIZE_K=128, GROUP_SIZE_M=1,
│             num_warps=2, num_stages=2, waves_per_eu=1, matrix_instr_nonkdim=16, NUM_KSPLIT=1)
│
│  def bench(cache_modifier, iters=100):
│      c = dict(cfg, cache_modifier=cache_modifier)
│      run = lambda: gemm_a8w8_blockscale(A, B, As, Bs, dtype=torch.bfloat16,
│                                         config=copy.deepcopy(c), backend="gluon")
│      run(); torch.cuda.synchronize()           # warm up / JIT compile
│      t = time.perf_counter()
│      for _ in range(iters): run()
│      torch.cuda.synchronize()
│      return (time.perf_counter() - t) / iters * 1e3   # ms
│
│  cg = bench(".cg")
│  ca = bench(".ca")
│  print(f"shape (M={M}, N={N}, K={K})")
│  print(f".cg : {cg:.3f} ms")
│  print(f".ca : {ca:.3f} ms")
│  print(f"speedup: {cg/ca:.1f}x")

Test Result

.cg : 2.321 ms
.ca : 0.048 ms
speedup: 48.5x

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4359 --add-label <label>