xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 11:26:15 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc][Gaudi] Avoid torch.compile and enable lazy collectives (#10897)

Browse Source 
Signed-off-by: Konrad Zawora <kzawora@habana.ai>
This commit is contained in:
Konrad Zawora 2024-12-05 17:47:46 +01:00 committed by GitHub
parent 998eeafe58
commit a43065272f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 14 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/plugins/__init__.py
View File

@ -29,6 +29,20 @@ def load_general_plugins():
if current_platform.is_xpu():
# see https://github.com/pytorch/pytorch/blob/8cada5cbe5450e17c26fb8b358116785324537b2/torch/_dynamo/config.py#L158 # noqa
os.environ['TORCH_COMPILE_DISABLE'] = 'True'
if current_platform.is_hpu():
# NOTE(kzawora): PT HPU lazy backend (PT_HPU_LAZY_MODE = 1)
# does not support torch.compile
# Eager backend (PT_HPU_LAZY_MODE = 0) must be selected for
# torch.compile support
is_lazy = os.environ.get('PT_HPU_LAZY_MODE', '1') == '1'
if is_lazy:
# see https://github.com/pytorch/pytorch/blob/43c5f59/torch/_dynamo/config.py#L158
torch._dynamo.config.disable = True
# NOTE(kzawora) multi-HPU inference with HPUGraphs (lazy-only)
# requires enabling lazy collectives
# see https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html # noqa: E501
os.environ['PT_HPU_ENABLE_LAZY_COLLECTIVES'] = 'true'
global plugins_loaded
if plugins_loaded:
return