[XPU] enable XPU GRAPH by default (#51600)

Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>
This commit is contained in:
liuzhenwei
2026-09-24 19:17:25 +08:00
committed by GitHub
parent f5a78f2ad7
commit dcfc17e0b1
7 changed files with 38 additions and 29 deletions
-5
View File
@@ -317,7 +317,6 @@ if TYPE_CHECKING:
VLLM_ELASTIC_EP_SCALE_UP_LAUNCH: bool = False
VLLM_ELASTIC_EP_DRAIN_REQUESTS: bool = False
VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: bool = True
VLLM_XPU_ENABLE_XPU_GRAPH: bool = False
VLLM_XPU_FORCE_N_CONTIG_WEIGHT: bool = False
VLLM_XPU_USE_SAMPLER_KERNEL: bool = True
VLLM_XPU_INC_WNA16_BACKEND: Literal["auto", "ark", "w4a16", "w4a8"] = "auto"
@@ -2139,10 +2138,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS": lambda: bool(
int(os.getenv("VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS", "1"))
),
# Whether enable XPU graph on Intel GPU
"VLLM_XPU_ENABLE_XPU_GRAPH": lambda: bool(
int(os.getenv("VLLM_XPU_ENABLE_XPU_GRAPH", "0"))
),
# Force N-contiguous weight layout for all XPU unquantized linears.
"VLLM_XPU_FORCE_N_CONTIG_WEIGHT": lambda: bool(
int(os.getenv("VLLM_XPU_FORCE_N_CONTIG_WEIGHT", "0"))