EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[Bugfix] Disable torch.compile() (#370)

Browse Source 
### What this PR does / why we need it?
To resolve this
[patch](https://github.com/vllm-project/vllm-ascend/pull/236/files#diff-43b96b39b5a52fe209d86449ad703a7ff5e1349ebaf1aa12ece8d82163ee5b61R24-R49)
, we need to set `torch.compile()` backend to `eager` to disable
compile, using default pytorch way.


---------

Signed-off-by: shen-shanshan <467638484@qq.com>
This commit is contained in:
Shanshan Shen
2025-03-21 15:55:51 +08:00
committed by GitHub
parent 9a175ca0fc
commit 89ca63a2c2
4 changed files with 30 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm_ascend/platform.py
View File

@@ -40,7 +40,7 @@ class NPUPlatform(Platform):
_enum = PlatformEnum.OOT
device_name: str = "npu"
device_type: str = "npu"
simple_compile_backend: str = "npu"
simple_compile_backend: str = "eager" # Disable torch.compile()
ray_device_key: str = "NPU"
device_control_env_var: str = "ASCEND_RT_VISIBLE_DEVICES"
dispatch_key: str = "PrivateUse1"
@@ -99,11 +99,13 @@ class NPUPlatform(Platform):
if parallel_config.worker_cls == "auto":
if envs.VLLM_USE_V1:
parallel_config.worker_cls = "vllm_ascend.worker.worker_v1.NPUWorker"
elif vllm_config.speculative_config:
parallel_config.worker_cls = "vllm.spec_decode.spec_decode_worker.create_spec_worker"
parallel_config.sd_worker_cls = "vllm_ascend.worker.worker.NPUWorker"
elif vllm_config.scheduler_config.is_multi_step:
parallel_config.worker_cls = "vllm_ascend.worker.multi_step_worker.MultiStepWorker"
else:
if vllm_config.scheduler_config.is_multi_step:
parallel_config.worker_cls = "vllm_ascend.worker.multi_step_worker.MultiStepWorker"
else:
parallel_config.worker_cls = "vllm_ascend.worker.worker.NPUWorker"
parallel_config.worker_cls = "vllm_ascend.worker.worker.NPUWorker"
cache_config = vllm_config.cache_config
if cache_config and cache_config.block_size is None: