[TPU] Add TPU specific var VLLM_TPU_MOST_MODEL_LEN (#19919)

Signed-off-by: Chenyaaang <chenyangli@google.com>
This commit is contained in:
Chenyaaang
2025-06-25 15:51:02 -07:00
committed by GitHub
parent 55c65ab495
commit 2d7620c3eb
5 changed files with 184 additions and 76 deletions
+3
View File
@@ -119,6 +119,7 @@ if TYPE_CHECKING:
VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
VLLM_V0_USE_OUTLINES_CACHE: bool = False
VLLM_TPU_BUCKET_PADDING_GAP: int = 0
VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
VLLM_USE_DEEP_GEMM: bool = False
VLLM_XGRAMMAR_CACHE_MB: int = 0
VLLM_MSGPACK_ZERO_COPY_THRESHOLD: int = 256
@@ -833,6 +834,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_TPU_BUCKET_PADDING_GAP":
lambda: int(os.environ["VLLM_TPU_BUCKET_PADDING_GAP"])
if "VLLM_TPU_BUCKET_PADDING_GAP" in os.environ else 0,
"VLLM_TPU_MOST_MODEL_LEN":
lambda: maybe_convert_int(os.environ.get("VLLM_TPU_MOST_MODEL_LEN", None)),
# Allow use of DeepGemm kernels for fused moe ops.
"VLLM_USE_DEEP_GEMM":