[Bug] mhc_pre "CUDA error: invalid configuration argument" on MI300X (gfx942) during DeepSeek-V4-Flash-0731 profile_run

Environment

• GPU: 8x AMD Instinct MI300X (gfx942)
• ROCm: 7.1.1, PyTorch 2.14.0.dev20260803+rocm7.1
• aiter: reproduced on both v0.1.19 and main @ 17f24ec6e (2026-08-07), built with GPU_ARCHS=gfx942
• vLLM: commit 4719a9b8f, VLLM_ROCM_USE_AITER=1
• Model: deepseek-ai/DeepSeek-V4-Flash-0731

Command

vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
  --tensor-parallel-size 8 --gpu-memory-utilization 0.9 --max-model-len 8192 \
  --kv-cache-dtype fp8 --tokenizer-mode deepseek_v4 --enforce-eager

Error

All 8 TP workers crash identically during profile_run's dummy forward pass:

File ".../aiter/ops/mhc.py", line 350, in mhc_pre
    comb_mix = torch.empty(m, hc_mult, hc_mult, dtype=dtypes.fp32, device=device)
torch.AcceleratorError: CUDA error: invalid configuration argument

hc_mult=4 (from the model's config.json). Call path: vllm/models/deepseek_v4/amd/model.py -> hc_pre -> MHCPreOp.forward_hip -> torch.ops.vllm.mhc_pre_aiter -> aiter/ops/mhc.py:mhc_pre.

Reproduced identically (same line, same error) under both HIP_LAUNCH_BLOCKING=1 CUDA_LAUNCH_BLOCKING=1 and AMD_SERIALIZE_KERNEL=3, ruling out simple async-error misattribution.

Workaround

Forcing HAS_AITER_MHC = False in vllm/model_executor/layers/mhc.py routes MHCPreOp/MHCPostOp.forward_hip to forward_native instead, and the model loads and proceeds past this point.

Related

• vllm-project/vllm#42876 -- DeepSeek-V4 on MI350X: different MHC-path error (TileLang PDL), same op family
• vllm-project/vllm#46796 -- DeepSeek-V4 on NVIDIA B300: same error text ("invalid configuration argument") in the same MHC/hc_prenorm kernel family, different backend (DeepGEMM, not AITER)

Reproduce the DeepSeek-V4-Flash-0731 profile_run with the documented vLLM command, then start at aiter/ops/mhc.py:350 and trace the call path through vllm/models/deepseek_v4/amd/model.py and vllm/model_executor/layers/mhc.py. Compare the AITER path with the native fallback; done means the MI300X run proceeds past MHCPreOp without the invalid configuration error.