Remove COMPILE_CUSTOM_KERNELS env (#4864)

With more and more custom ops merged, disable `COMPILE_CUSTOM_KERNELS `
for vllm ascend seems useless now. Let's enable csrc compile by default.

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm_ascend/envs.py

@@ -36,12 +36,6 @@ env_variables: Dict[str, Callable[[], Any]] = {
     # Release, Debug, RelWithDebugInfo. If not set, the default value is Release.
     "CMAKE_BUILD_TYPE":
     lambda: os.getenv("CMAKE_BUILD_TYPE"),
-    # Whether to compile custom kernels. If not set, the default value is True.
-    # If set to False, the custom kernels will not be compiled. Please note that
-    # the sleep mode feature will be disabled as well if custom kernels are not
-    # compiled.
-    "COMPILE_CUSTOM_KERNELS":
-    lambda: bool(int(os.getenv("COMPILE_CUSTOM_KERNELS", "1"))),
     # The CXX compiler used for compiling the package. If not set, the default
     # value is None, which means the system default CXX compiler will be used.
     "CXX_COMPILER":

vllm_ascend/utils.py

@@ -49,7 +49,6 @@ ACL_FORMAT_FRACTAL_ND = 2
 ACL_FORMAT_FRACTAL_NZ = 29
 
 _CUSTOM_OP_ENABLED = None
-_SLEEP_MODE_ENABLED = None
 _CURRENT_STREAM = None
 _PREFETCH_STREAM = None
 _SHARED_EXPERTS_CALCULATION_STREAM = None
@@ -125,14 +124,6 @@ def is_enable_nz():
     return envs_ascend.VLLM_ASCEND_ENABLE_NZ
 
 
-def sleep_mode_enabled():
-    global _SLEEP_MODE_ENABLED
-    if _SLEEP_MODE_ENABLED is None:
-        from vllm_ascend import _build_info  # type: ignore
-        _SLEEP_MODE_ENABLED = _build_info.__sleep_mode_enabled__
-    return _SLEEP_MODE_ENABLED
-
-
 def _round_up(x: int, align: int):
     # round up x to align, for example, if align is 16, x will be rounded up to 16, 32, 48, etc.
     # input: 15, 16 -> output: 16

vllm_ascend/worker/worker_v1.py

@@ -54,7 +54,7 @@ from vllm_ascend.ops.triton.triton_utils import init_device_properties_triton
 from vllm_ascend.platform import NPUPlatform
 from vllm_ascend.utils import (check_ascend_device_type, enable_sp,
                                is_enable_nz, register_ascend_customop,
-                               sleep_mode_enabled, try_register_lib)
+                               try_register_lib)
 from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
 
 torch._dynamo.trace_rules.clear_lru_cache()  # noqa: E402
@@ -129,7 +129,7 @@ class NPUWorker(WorkerBase):
             init_cached_hf_modules()
 
         self.profiler = self._init_profiler()
-        if sleep_mode_enabled():
+        if vllm_config.model_config and vllm_config.model_config.enable_sleep_mode:
             # Buffers saved before sleep
             self._sleep_saved_buffers: dict[str, torch.Tensor] = {}
 
@@ -140,10 +140,6 @@ class NPUWorker(WorkerBase):
             WEIGHT_LOADER_V2_SUPPORTED.remove("UnquantizedLinearMethod")
 
     def sleep(self, level: int = 1) -> None:
-        if not sleep_mode_enabled():
-            raise ValueError(
-                "Sleep mode is not enabled. Please compile vllm-ascend with COMPILE_CUSTOM_KERNELS=1."
-            )
         free_bytes_before_sleep = NPUPlatform.mem_get_info()[0]
         # Save the buffers before level 2 sleep
         if level == 2:
@@ -164,11 +160,6 @@ class NPUWorker(WorkerBase):
             used_bytes / GiB_bytes)
 
     def wake_up(self, tags: Optional[list[str]] = None) -> None:
-        if not sleep_mode_enabled():
-            raise ValueError(
-                "Sleep mode is not enabled. Please compile vllm-ascend with COMPILE_CUSTOM_KERNELS=1."
-            )
-
         if is_enable_nz():
             raise ValueError(
                 "FRACTAL_NZ mode is enabled. This may cause model parameter precision issues "