[V1][TPU] Enable Top K (#15489)

Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Hyesoo Yang <hyeygit@gmail.com>
Co-authored-by: Hyesoo Yang <hyeygit@gmail.com>
This commit is contained in:
Nicolò Lucchesi
2025-04-17 18:18:11 +00:00
committed by GitHub
co-authored by Hyesoo Yang
parent 5989f4684d
commit eb5819b2d9
6 changed files with 69 additions and 26 deletions
-6
View File
@@ -103,7 +103,6 @@ if TYPE_CHECKING:
VLLM_DP_MASTER_PORT: int = 0
VLLM_MARLIN_USE_ATOMIC_ADD: bool = False
VLLM_V0_USE_OUTLINES_CACHE: bool = False
VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION: bool = False
VLLM_TPU_BUCKET_PADDING_GAP: int = 0
VLLM_USE_DEEP_GEMM: bool = False
VLLM_XGRAMMAR_CACHE_MB: int = 0
@@ -685,11 +684,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_V0_USE_OUTLINES_CACHE":
lambda: os.environ.get("VLLM_V0_USE_OUTLINES_CACHE", "0") == "1",
# If set, disables TPU-specific optimization for top-k & top-p sampling
"VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION":
lambda: bool(int(os.environ["VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION"]))
if "VLLM_TPU_DISABLE_TOPK_TOPP_OPTIMIZATION" in os.environ else None,
# Gap between padding buckets for the forward pass. So we have
# 8, we will run forward pass with [16, 24, 32, ...].
"VLLM_TPU_BUCKET_PADDING_GAP":