biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

VLLM_USE_TRITON_FLASH_ATTN V0 variable deprecation (#27611)

Browse Source 
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
This commit is contained in:
Andreas Karatzas
2025-11-11 20:34:36 -06:00
committed by GitHub
parent 7f829be7d3
commit 9f0247cfa4
15 changed files with 12 additions and 1588 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
vllm/platforms/rocm.py
View File

@@ -49,25 +49,8 @@ _ROCM_UNSUPPORTED_MODELS: list[str] = []
# Models partially supported by ROCm.
# Architecture -> Reason.
_ROCM_SWA_REASON = (
"Sliding window attention (SWA) is not yet supported in "
"Triton flash attention. For half-precision SWA support, "
"please use CK flash attention by setting "
"`VLLM_USE_TRITON_FLASH_ATTN=0`"
)
_ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {
"Qwen2ForCausalLM": _ROCM_SWA_REASON,
"MistralForCausalLM": _ROCM_SWA_REASON,
"MixtralForCausalLM": _ROCM_SWA_REASON,
"PaliGemmaForConditionalGeneration": (
"ROCm flash attention does not yet fully support 32-bit precision on PaliGemma"
),
"Phi3VForCausalLM": (
"ROCm Triton flash attention may run into compilation errors due to "
"excessive use of shared memory. If this happens, disable Triton FA "
"by setting `VLLM_USE_TRITON_FLASH_ATTN=0`"
),
}
_ROCM_SWA_REASON = ()
_ROCM_PARTIALLY_SUPPORTED_MODELS: dict[str, str] = {}
_ROCM_DEVICE_ID_NAME_MAP: dict[str, str] = {
"0x74a0": "AMD_Instinct_MI300A",
"0x74a1": "AMD_Instinct_MI300X",