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

bugfix for mtp in fullgraph (#3878)

Browse Source 
### What this PR does / why we need it?
bugfix for mtp in fullgraph

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

---------

Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
This commit is contained in:
zouyida2052
2025-10-29 23:52:20 +08:00
committed by GitHub
parent 19f49ecb5f
commit d9249c968e
6 changed files with 58 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm_ascend/platform.py
View File

@@ -263,6 +263,7 @@ class NPUPlatform(Platform):
**********************************************************************************\033[0m **********************************************************************************\033[0m
""" """
logger.warning(warning_message) logger.warning(warning_message)
update_aclgraph_sizes(vllm_config)
else: else:
logger.info( logger.info(
"%s cudagraph_mode is not support on NPU. falling back to NONE", "%s cudagraph_mode is not support on NPU. falling back to NONE",

57
vllm_ascend/utils.py
View File

@@ -314,6 +314,13 @@ def get_max_hidden_layers(hf_config) -> int:
def update_aclgraph_sizes(vllm_config: VllmConfig) -> None: def update_aclgraph_sizes(vllm_config: VllmConfig) -> None:
"""Update ACL graph capture sizes based on hardware limitations""" """Update ACL graph capture sizes based on hardware limitations"""
from vllm.config.compilation import CUDAGraphMode
if vllm_config.compilation_config.cudagraph_mode == CUDAGraphMode.FULL_DECODE_ONLY:
if vllm_config.speculative_config is not None and \
vllm_config.speculative_config.num_speculative_tokens > 1:
_update_spec_aclgraph_sizes(vllm_config)
return
# NOTE: Currently, we can only capture 1800 graphs at most, # NOTE: Currently, we can only capture 1800 graphs at most,
# due to the limitation of ACL graph. This number is bounded by # due to the limitation of ACL graph. This number is bounded by
# the number of streams, which is 2048, we save 248 streams # the number of streams, which is 2048, we save 248 streams
@@ -421,25 +428,43 @@ def update_aclgraph_sizes(vllm_config: VllmConfig) -> None:
vllm_config.model_config.architectures[0], num_hidden_layers, vllm_config.model_config.architectures[0], num_hidden_layers,
len(original_sizes)) len(original_sizes))
if vllm_config.speculative_config is not None and \
vllm_config.speculative_config.num_speculative_tokens > 1:
_update_spec_aclgraph_sizes(vllm_config)
def _update_spec_aclgraph_sizes(vllm_config: VllmConfig) -> None:
# default or defined cudagraph_capture_sizes may not consider num_speculative_tokens>1 scenario # default or defined cudagraph_capture_sizes may not consider num_speculative_tokens>1 scenario
# the maximum size cudagraph_capture_sizes[0] should be greater or equal than # the maximum size cudagraph_capture_sizes[0] should be greater or equal than
# (num_speculative_tokens+1)*max_num_seqs, otherwise draft model will run in eager mode # (num_speculative_tokens+1)*max_num_seqs, otherwise draft model will run in eager mode
if vllm_config.speculative_config is not None and \ from vllm.config.compilation import CUDAGraphMode
vllm_config.speculative_config.num_speculative_tokens > 1: compilation_config = vllm_config.compilation_config
num_speculative_tokens = vllm_config.speculative_config.num_speculative_tokens num_speculative_tokens = vllm_config.speculative_config.num_speculative_tokens
max_num_seqs = vllm_config.scheduler_config.max_num_seqs uniform_decode_query_len = num_speculative_tokens + 1
original_sizes, compilation_config.cudagraph_capture_sizes = \ max_num_seqs = vllm_config.scheduler_config.max_num_seqs
compilation_config.cudagraph_capture_sizes, None max_num_tokens = max_num_seqs * uniform_decode_query_len
assert len(original_sizes) > 0 original_sizes, compilation_config.cudagraph_capture_sizes = \
if original_sizes[0] < (num_speculative_tokens + 1) * max_num_seqs: compilation_config.cudagraph_capture_sizes, None
enlarged_sizes = [(num_speculative_tokens + 1) * size assert len(original_sizes) > 0
for size in original_sizes]
compilation_config.init_with_cudagraph_sizes(enlarged_sizes) if vllm_config.compilation_config.cudagraph_mode == CUDAGraphMode.FULL_DECODE_ONLY and \
logger.info( not all(size % uniform_decode_query_len == 0 for size in original_sizes):
"Adjusted ACL graphs: %s → %s for speculative decoding", enlarged_sizes = [
original_sizes, enlarged_sizes) size * uniform_decode_query_len for size in original_sizes
else: if max_num_tokens >= size >= uniform_decode_query_len
compilation_config.cudagraph_capture_sizes = original_sizes ]
compilation_config.init_with_cudagraph_sizes(enlarged_sizes)
logger.info("Adjusted ACL graphs: %s → %s for speculative decoding",
original_sizes, enlarged_sizes)
elif original_sizes[0] < max_num_tokens:
enlarged_sizes = [
size * uniform_decode_query_len for size in original_sizes
]
compilation_config.init_with_cudagraph_sizes(enlarged_sizes)
logger.info("Adjusted ACL graphs: %s → %s for speculative decoding",
original_sizes, enlarged_sizes)
else:
compilation_config.cudagraph_capture_sizes = original_sizes
# TODO(wxy): Move to ops module # TODO(wxy): Move to ops module

8
vllm_ascend/worker/model_runner_v1.py
View File

@@ -3529,14 +3529,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
if aclgraph_mode.decode_mode() == CUDAGraphMode.FULL and \ if aclgraph_mode.decode_mode() == CUDAGraphMode.FULL and \
aclgraph_mode.separate_routine(): aclgraph_mode.separate_routine():
max_num_tokens = self.scheduler_config.max_num_seqs * \
self.uniform_decode_query_len
decode_cudagraph_batch_sizes = [
x for x in self.aclgraph_batch_sizes if x <= max_num_tokens
and x >= self.uniform_decode_query_len
]
compilation_cases_decode = list( compilation_cases_decode = list(
reversed(decode_cudagraph_batch_sizes)) reversed(self.aclgraph_batch_sizes))
self._capture_aclgraphs( self._capture_aclgraphs(
compilation_cases=compilation_cases_decode, compilation_cases=compilation_cases_decode,
aclgraph_runtime_mode=CUDAGraphMode.FULL, aclgraph_runtime_mode=CUDAGraphMode.FULL,