A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 05:45 UTC
[Misc][Gaudi] Avoid torch.compile and enable lazy collectives (#10897)
Signed-off-by: Konrad Zawora <kzawora@habana.ai>
Konrad Zawora committed
Dec 5, 2024 at 17:47 UTC
a43065272f73a7468b1a35dd44fb5b0ed80f88c7
1 file changed
+14
vllm/plugins/__init__.py
+14
@@ -29,6 +29,20 @@ def load_general_plugins():
29
if current_platform.is_xpu():
30
# see https://github.com/pytorch/pytorch/blob/8cada5cbe5450e17c26fb8b358116785324537b2/torch/_dynamo/config.py#L158 # noqa
31
os.environ['TORCH_COMPILE_DISABLE'] = 'True'
32
+ if current_platform.is_hpu():
33
+ # NOTE(kzawora): PT HPU lazy backend (PT_HPU_LAZY_MODE = 1)
34
+ # does not support torch.compile
35
+ # Eager backend (PT_HPU_LAZY_MODE = 0) must be selected for
36
+ # torch.compile support
37
+ is_lazy = os.environ.get('PT_HPU_LAZY_MODE', '1') == '1'
38
+ if is_lazy:
39
+ # see https://github.com/pytorch/pytorch/blob/43c5f59/torch/_dynamo/config.py#L158
40
+ torch._dynamo.config.disable = True
41
+ # NOTE(kzawora) multi-HPU inference with HPUGraphs (lazy-only)
42
+ # requires enabling lazy collectives
43
+ # see https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html # noqa: E501
44
+ os.environ['PT_HPU_ENABLE_LAZY_COLLECTIVES'] = 'true'
45
+
46
global plugins_loaded
47
if plugins_loaded:
48
return