[ROCm] Add env to enable/disable aiter triton gemm (#28321)

Signed-off-by: Yong Hoon Shin <yhshin@meta.com>
2025-11-08 20:27:00 -10:00
parent e5e9067e61
commit de2b78305f
2 changed files with 8 additions and 0 deletions
--- a/vllm/model_executor/layers/utils.py
+++ b/vllm/model_executor/layers/utils.py
@@ -106,6 +106,7 @@ def default_unquantized_gemm(
 def use_aiter_triton_gemm(n, m, k, dtype):
    if (
        envs.VLLM_ROCM_USE_AITER == 0
+        or envs.VLLM_ROCM_USE_AITER_TRITON_GEMM == 0
        # MI300's - fp8nuz=True
        or current_platform.is_fp8_fnuz()
        or dtype not in [torch.float16, torch.bfloat16]