[Core] Add VLLM_GPU_SYNC_CHECK env var (#44800)

This commit is contained in:
Nick Hill
2026-06-26 08:24:33 -07:00
committed by GitHub
parent e71bc6da85
commit 4e07ca2c92
5 changed files with 279 additions and 0 deletions
+8
View File
@@ -87,6 +87,7 @@ if TYPE_CHECKING:
VLLM_FLOAT32_MATMUL_PRECISION: Literal["highest", "high", "medium"] = "highest"
VLLM_BATCH_INVARIANT: bool = False
VLLM_TRITON_ATTN_USE_TD: bool | None = None
VLLM_GPU_SYNC_CHECK: Literal["warn", "error"] | None = None
MAX_JOBS: str | None = None
NVCC_THREADS: str | None = None
VLLM_USE_PRECOMPILED: bool = False
@@ -584,6 +585,13 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_TRITON_ATTN_USE_TD": lambda: {"1": True, "0": False}.get(
os.getenv("VLLM_TRITON_ATTN_USE_TD", "").strip()
),
# If set, enable PyTorch's GPU<->CPU synchronization debug mode around
# the worker's `execute_model` and `sample_tokens` calls. Valid values
# are "warn" (print a warning on each sync) or "error" (raise on sync).
# Unset disables the check. See `torch.cuda.set_sync_debug_mode`.
"VLLM_GPU_SYNC_CHECK": env_with_choices(
"VLLM_GPU_SYNC_CHECK", None, ["warn", "error"]
),
# Maximum number of compilation jobs to run in parallel.
# By default this is the number of CPUs
"MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),