From be3f614ff158ccecc99546958bf5e4893f8e0581 Mon Sep 17 00:00:00 2001 From: Taneem Ibrahim Date: Tue, 18 Aug 2026 07:28:00 -0500 Subject: [PATCH] [CI] Register CPU CI "VLLM_CPU_CI_ENV" environment variable (#52633) Signed-off-by: Taneem Ibrahim --- vllm/envs.py | 3 +++ vllm/platforms/cpu.py | 6 ++---- 2 files changed, 5 insertions(+), 4 deletions(-) diff --git a/vllm/envs.py b/vllm/envs.py index 38b64f8c38..1a7458802a 100755 --- a/vllm/envs.py +++ b/vllm/envs.py @@ -51,6 +51,7 @@ if TYPE_CHECKING: VLLM_CPU_KVCACHE_SPACE: int | None = 0 VLLM_CPU_OMP_THREADS_BIND: str = "auto" VLLM_CPU_NUM_OF_RESERVED_CPU: int | None = None + VLLM_CPU_CI_ENV: bool = False VLLM_CPU_ATTN_SPLIT_KV: bool = True VLLM_ZENTORCH_WEIGHT_PREPACK: bool = True VLLM_CPU_INT4_W4A8: bool = True @@ -885,6 +886,8 @@ environment_variables: dict[str, Callable[[], Any]] = { if "VLLM_CPU_NUM_OF_RESERVED_CPU" in os.environ else None ), + # (CPU backend only) whether vLLM is running in a CI environment. + "VLLM_CPU_CI_ENV": lambda: bool(int(os.getenv("VLLM_CPU_CI_ENV", "0"))), # (CPU backend only) whether to enable attention spilt KV. "VLLM_CPU_ATTN_SPLIT_KV": lambda: bool( int(os.getenv("VLLM_CPU_ATTN_SPLIT_KV", "1")) diff --git a/vllm/platforms/cpu.py b/vllm/platforms/cpu.py index 5633e160cc..dd18363e0d 100644 --- a/vllm/platforms/cpu.py +++ b/vllm/platforms/cpu.py @@ -10,6 +10,7 @@ from typing import TYPE_CHECKING import torch +from vllm import envs from vllm.logger import init_logger from vllm.utils.cpu_resource_utils import ( DEVICE_CONTROL_ENV_VAR, @@ -203,10 +204,7 @@ class CpuPlatform(Platform): # cache. So use VLLM_CPU_CI_ENV to indicate the CI environment, # and just execute model with dynamo + eager mode to save time. # VLLM_CPU_CI_ENV is only used as an internal variable. - if os.environ.get("VLLM_CPU_CI_ENV", "0") != "0": - backend = "eager" - else: - backend = "inductor" + backend = "eager" if envs.VLLM_CPU_CI_ENV else "inductor" compilation_config.mode = CompilationMode.DYNAMO_TRACE_ONCE compilation_config.backend = backend