[CPU] Support int8 compute mode in CPU AWQ (#35697)

Signed-off-by: Yintong Lu <yintong.lu@intel.com>
This commit is contained in:
Yintong Lu
2026-03-31 15:27:37 +08:00
committed by GitHub
parent 42318c840b
commit f09daea261
10 changed files with 1197 additions and 11 deletions
+3
View File
@@ -52,6 +52,7 @@ if TYPE_CHECKING:
VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None
VLLM_CPU_SGL_KERNEL: bool = False
VLLM_ZENTORCH_WEIGHT_PREPACK: bool = True
VLLM_CPU_INT4_W4A8: bool = True
VLLM_XLA_CACHE_PATH: str = os.path.join(VLLM_CACHE_ROOT, "xla_cache")
VLLM_XLA_CHECK_RECOMPILATION: bool = False
VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE: Literal["auto", "nccl", "shm"] = "auto"
@@ -728,6 +729,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_ZENTORCH_WEIGHT_PREPACK": lambda: bool(
int(os.getenv("VLLM_ZENTORCH_WEIGHT_PREPACK", "1"))
),
# (CPU backend only) whether to use SGLang INT4 W4A8 kernels for AWQ.
"VLLM_CPU_INT4_W4A8": lambda: bool(int(os.getenv("VLLM_CPU_INT4_W4A8", "1"))),
# If the env var is set, Ray Compiled Graph uses the specified
# channel type to communicate between workers belonging to
# different pipeline-parallel stages.