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

[Refactor] move the metadata from attention_v1 to util(ready for extract common_cp) & realize Ascendmetadata inherit from the parent class. (#5203)

Browse Source 
RFC: https://github.com/vllm-project/vllm-ascend/issues/4629

1. Remove the pcp-related code from attention_v1.
2. Establish the inheritance relationship of CommonAttentionMetadata.

TODO
1. extract common_cp
2. move cp metadata to common_cp.
3. remove commonAttentionMetadata for aclgraph.

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

---------

Signed-off-by: weijinqian_v1 <weijinqian@huawei.com>
Co-authored-by: weijinqian_v1 <weijinqian@huawei.com>
This commit is contained in:
weijinqian0
2025-12-23 00:10:52 +08:00
committed by GitHub
parent 3d9954eff0
commit 95e8a52156
9 changed files with 83 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1045,7 +1045,7 @@ class NPUModelRunner(GPUModelRunner):
max_query_len=max_num_scheduled_tokens,
decode_token_per_req=self.decode_token_per_req,
prefill_context_parallel_metadata=long_seq_metadata,
)
max_seq_len=0)
if self.speculative_config and self.pcp_size * self.dcp_size > 1:
# For pcp + spec decode, we flatten block_table
@@ -1874,7 +1874,7 @@ class NPUModelRunner(GPUModelRunner):
max_query_len=max_query_len,
decode_token_per_req=self.decode_token_per_req,
prefill_context_parallel_metadata=long_seq_metadata,
)
max_seq_len=0)
if self.pcp_size * self.dcp_size > 1:
common_attn_metadata.block_table_tensor = \
block_table_tensor[:num_reqs * self.decode_threshold]