diff --git a/scripts/01-rocm-env-for-triton.sh b/scripts/01-rocm-env-for-triton.sh index c1e1ec8..b91c282 100644 --- a/scripts/01-rocm-env-for-triton.sh +++ b/scripts/01-rocm-env-for-triton.sh @@ -1,4 +1,8 @@ # Required for Strix Halo / RDNA3.5 on vLLM export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" -export VLLM_TARGET_DEVICE=rocm \ No newline at end of file +export VLLM_TARGET_DEVICE=rocm +export VLLM_USE_MMAP=0 +export VLLM_ROCM_USE_AITER=0 +export VLLM_ROCM_USE_AITER_MOE=0 +export VLLM_USE_TRITON_AWQ=1 \ No newline at end of file