[Bugfix] Refactor Flashinfer TRTLLM attention kernel selection logic (#24600)

Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
elvischenv
2025-09-17 15:36:29 -07:00
committed by GitHub
co-authored by Michael Goin
parent 9f882d8791
commit e67a79db03
3 changed files with 65 additions and 29 deletions
+5 -2
View File
@@ -1223,9 +1223,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_USE_CUDNN_PREFILL":
lambda: bool(int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))),
# If set to 1, use the TRTLLM attention backend in flashinfer.
# If set to 1/True, use the TRTLLM attention backend in flashinfer.
# If set to 0/False, use the default attention backend in flashinfer.
# If not set, auto-detect the attention backend in flashinfer.
"VLLM_USE_TRTLLM_ATTENTION":
lambda: os.getenv("VLLM_USE_TRTLLM_ATTENTION", None),
lambda: (None if "VLLM_USE_TRTLLM_ATTENTION" not in os.environ else
os.environ["VLLM_USE_TRTLLM_ATTENTION"].lower() in ("1", "true")),
# If set to 1, when we use fp8 kv, we do not quantize Q to fp8
"VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION":