zzh/vllm-npu-plugin
1
0
Fork 0
mirror of https://github.com/handsomezhuzhu/vllm-npu-plugin.git synced 2026-02-20 11:42:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: Defer SOC version initialization to avoid NPU lazy init issues

Browse Source
This commit is contained in:
handsomezhuzhu
2026-02-10 23:19:34 +08:00
parent f49538ea8d
commit ae10ce68f0
2 changed files with 10 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm_npu/utils.py
View File

@@ -67,11 +67,13 @@ _IS_EAGLE_MODE = None
def is_310p(): def is_310p():
global _IS_310P global _IS_310P
if _IS_310P is None: if _IS_310P is None:
try: # Check if SOC version is already known from init_ascend_soc_version()
soc_version = torch_npu.npu.get_soc_version() if _ascend_soc_version is not None:
# 310P soc_version range: 200-209 _IS_310P = False # 310P is not A2 or A3
_IS_310P = 200 <= soc_version <= 209 else:
except Exception: # Avoid calling torch_npu.npu.get_soc_version() here as it
# triggers NPU lazy init which breaks forked subprocesses.
# Default to False; will be updated after init_device().
_IS_310P = False _IS_310P = False
return _IS_310P return _IS_310P

5
vllm_npu/worker/worker_v1.py
View File

@@ -87,9 +87,8 @@ class NPUWorker(WorkerBase):
ops.register_dummy_fusion_op() ops.register_dummy_fusion_op()
_register_atb_extensions() _register_atb_extensions()
register_ascend_customop(vllm_config) register_ascend_customop(vllm_config)
# init ascend config and soc version # init ascend config (soc version deferred to init_device)
init_ascend_config(vllm_config) init_ascend_config(vllm_config)
init_ascend_soc_version()
use_sparse = False use_sparse = False
if vllm_config.model_config is not None: if vllm_config.model_config is not None:
use_sparse = hasattr(vllm_config.model_config.hf_config, use_sparse = hasattr(vllm_config.model_config.hf_config,
@@ -213,6 +212,8 @@ class NPUWorker(WorkerBase):
def init_device(self): def init_device(self):
device = self._init_device() device = self._init_device()
# Now that NPU device is initialized, get soc version
init_ascend_soc_version()
# Init ModelRunner here, so that we have access to self.device. # Init ModelRunner here, so that we have access to self.device.
self.model_runner = NPUModelRunner(self.vllm_config, device) self.model_runner = NPUModelRunner(self.vllm_config, device)