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

remove useless patch (#4699)

Browse Source 
patach_config is useless now. Let's remove it


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>
This commit is contained in:
wangxiyuan
2025-12-08 11:02:42 +08:00
committed by GitHub
parent 866347a621
commit 0b65ac6c4b
21 changed files with 30 additions and 277 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/torchair/torchair_mla.py
View File

@@ -317,7 +317,7 @@ class AscendMLATorchairMetadataBuilder:
dtype=self.model_config.dtype,
device=device)
if self.vllm_config.speculative_config is not None and\
self.vllm_config.speculative_config.method == 'deepseek_mtp':
self.vllm_config.speculative_config.method == 'mtp':
attn_state = AscendAttentionState.SpecDecoding
num_decode_tokens = 2
else:

2
vllm_ascend/torchair/torchair_model_runner.py
View File

@@ -501,7 +501,7 @@ class NPUTorchairModelRunner(NPUModelRunner):
def update_torchair_graph_batch_sizes(self):
# return graph_batch_sizes according to the max number of tokens
# first pad according to the number of requests
if self.is_kv_consumer and self.speculative_config and self.speculative_config.method == 'deepseek_mtp':
if self.is_kv_consumer and self.speculative_config and self.speculative_config.method == 'mtp':
# pd disaggregation scenario may incorrectly calculate the batch in mtp scenario, so we force set it to max_num_reqs
self.torchair_graph_batch_sizes = [self.max_num_reqs]
logger.warning(

2
vllm_ascend/torchair/torchair_sfa.py
View File

@@ -319,7 +319,7 @@ class AscendSFATorchairMetadataBuilder:
device=device)
if self.vllm_config.speculative_config is not None and\
self.vllm_config.speculative_config.method == 'deepseek_mtp':
self.vllm_config.speculative_config.method == 'mtp':
attn_state = AscendAttentionState.SpecDecoding
num_decode_tokens = 2
else: