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

[Bugfix] Remove VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE (#2969)

Browse Source 
### What this PR does / why we need it?
This PR prepares for deleting this enviroment variable,
`VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE`, as vllm requires `fullgraph=True`
to run

- Fixes https://github.com/vllm-project/vllm/issues/21834

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
See CI 

- vLLM version: v0.10.2
- vLLM main:
99cc41ad50

---------

Signed-off-by: Lucas Kabela <lucaskabela@meta.com>
This commit is contained in:
Lucas Kabela
2025-09-19 17:22:30 -07:00
committed by GitHub
parent e26fe1caf1
commit 53ecd89e8f
2 changed files with 10 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm_ascend/spec_decode/mtp_proposer.py
View File

@@ -3,7 +3,6 @@ import types
import torch import torch
import torch.nn as nn import torch.nn as nn
import torchair import torchair
import vllm.envs as envs_vllm
from torchair import patch_for_hcom from torchair import patch_for_hcom
from vllm.attention.layer import Attention from vllm.attention.layer import Attention
from vllm.config import (VllmConfig, get_layers_from_vllm_config, from vllm.config import (VllmConfig, get_layers_from_vllm_config,
@@ -596,11 +595,10 @@ class MtpProposer(Proposer):
torch.npu.set_compile_mode(jit_compile=False) torch.npu.set_compile_mode(jit_compile=False)
if not self.runner.use_cached_npu_graph: if not self.runner.use_cached_npu_graph:
npu_backend = torchair.get_npu_backend(compiler_config=config) npu_backend = torchair.get_npu_backend(compiler_config=config)
self.torchair_compiled_model = torch.compile( self.torchair_compiled_model = torch.compile(self.model,
self.model, dynamic=True,
dynamic=True, fullgraph=True,
fullgraph=envs_vllm.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE, backend=npu_backend)
backend=npu_backend)
return self.torchair_compiled_model return self.torchair_compiled_model
else: else:
# Generate a new forward proxy code object to prevent the invalidation of # Generate a new forward proxy code object to prevent the invalidation of
@@ -622,7 +620,7 @@ class MtpProposer(Proposer):
batch_size] = torchair.inference.cache_compile( batch_size] = torchair.inference.cache_compile(
self.model.__dict__[forward_proxy_name], self.model.__dict__[forward_proxy_name],
dynamic=True, dynamic=True,
fullgraph=envs_vllm.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE, fullgraph=True,
cache_dir=TORCHAIR_CACHE_DIR, cache_dir=TORCHAIR_CACHE_DIR,
config=config, config=config,
ge_cache=False) ge_cache=False)

12
vllm_ascend/torchair/torchair_model_runner.py
View File

@@ -25,7 +25,6 @@ import torch
import torch.distributed as dist import torch.distributed as dist
import torch.nn as nn import torch.nn as nn
import torch_npu import torch_npu
import vllm.envs as envs_vllm
from vllm.config import VllmConfig from vllm.config import VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.parallel_state import get_dp_group from vllm.distributed.parallel_state import get_dp_group
@@ -373,11 +372,10 @@ class NPUTorchairModelRunner(NPUModelRunner):
torch.npu.set_compile_mode(jit_compile=False) torch.npu.set_compile_mode(jit_compile=False)
if not self.use_cached_npu_graph: if not self.use_cached_npu_graph:
npu_backend = torchair.get_npu_backend(compiler_config=config) npu_backend = torchair.get_npu_backend(compiler_config=config)
self.torchair_compiled_model = torch.compile( self.torchair_compiled_model = torch.compile(self.model,
self.model, dynamic=True,
dynamic=True, fullgraph=True,
fullgraph=envs_vllm.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE, backend=npu_backend)
backend=npu_backend)
return self.torchair_compiled_model return self.torchair_compiled_model
else: else:
# Generate a new forward proxy code object to prevent the invalidation of # Generate a new forward proxy code object to prevent the invalidation of
@@ -399,7 +397,7 @@ class NPUTorchairModelRunner(NPUModelRunner):
batch_size] = torchair.inference.cache_compile( batch_size] = torchair.inference.cache_compile(
self.model.__dict__[forward_proxy_name], self.model.__dict__[forward_proxy_name],
dynamic=True, dynamic=True,
fullgraph=envs_vllm.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE, fullgraph=True,
cache_dir=TORCHAIR_CACHE_DIR, cache_dir=TORCHAIR_CACHE_DIR,
config=config, config=config,
ge_cache=False) ge_cache=False)