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

[eagle][cp] fix eagle_cp enable bug2 (#7079)

Browse Source 
### What this PR does / why we need it?
Fix acceptance and high-concurrency bug in eagle3 and cp enabled

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

### How was this patch tested?
tests and ut

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e

---------

Signed-off-by: lilinsiman <lilinsiman@gmail.com>
This commit is contained in:
lilinsiman
2026-03-10 16:32:49 +08:00
committed by GitHub
parent 67d40f23fd
commit a5ea699e29
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/attention/context_parallel/attention_cp.py
View File

@@ -143,7 +143,7 @@ class AscendAttentionCPMetadataBuilder(AscendAttentionMetadataBuilder):
chunked_context_metadata = None chunked_context_metadata = None
attn_mask_seqlens = common_long_seq_metadata.attn_mask_seqlens attn_mask_seqlens = common_long_seq_metadata.attn_mask_seqlens
if num_prefills > 0: if num_prefills > 0:
query_lens = query_lens[num_decode_tokens:] query_lens = query_lens[num_decodes:]
context_lens_cpu = num_computed_tokens_cpu[num_decodes:num_reqs] context_lens_cpu = num_computed_tokens_cpu[num_decodes:num_reqs]
max_context_len_cpu = context_lens_cpu.max().item() max_context_len_cpu = context_lens_cpu.max().item()
if self.chunked_prefill_enabled and max_context_len_cpu > 0: if self.chunked_prefill_enabled and max_context_len_cpu > 0:

2
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1016,7 +1016,7 @@ class NPUModelRunner(GPUModelRunner):
target_positions = self._get_positions(num_scheduled_tokens) target_positions = self._get_positions(num_scheduled_tokens)
target_hidden_states = hidden_states target_hidden_states = hidden_states
if self.use_aux_hidden_state_outputs: if self.use_aux_hidden_state_outputs:
target_hidden_states = torch.cat([h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1) target_hidden_states = torch.cat([h for h in aux_hidden_states], dim=-1)
else: else:
token_indices_to_sample = None token_indices_to_sample = None
# input_ids can be None for multimodal models. # input_ids can be None for multimodal models.