From a5ea699e29c28c72cacf1e7a55348691d313f380 Mon Sep 17 00:00:00 2001
From: lilinsiman <lilinsiman@gmail.com>
Date: Tue, 10 Mar 2026 16:32:49 +0800
Subject: [PATCH] [eagle][cp] fix eagle_cp enable bug2 (#7079)

### 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:
https://github.com/vllm-project/vllm/commit/4034c3d32e30d01639459edd3ab486f56993876d

---------

Signed-off-by: lilinsiman <lilinsiman@gmail.com>
---
 vllm_ascend/attention/context_parallel/attention_cp.py | 2 +-
 vllm_ascend/worker/model_runner_v1.py                  | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/vllm_ascend/attention/context_parallel/attention_cp.py b/vllm_ascend/attention/context_parallel/attention_cp.py
index eadd4fd7..655fec1d 100644
--- a/vllm_ascend/attention/context_parallel/attention_cp.py
+++ b/vllm_ascend/attention/context_parallel/attention_cp.py
@@ -143,7 +143,7 @@ class AscendAttentionCPMetadataBuilder(AscendAttentionMetadataBuilder):
         chunked_context_metadata = None
         attn_mask_seqlens = common_long_seq_metadata.attn_mask_seqlens
         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]
             max_context_len_cpu = context_lens_cpu.max().item()
             if self.chunked_prefill_enabled and max_context_len_cpu > 0:
diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index 1bb40291..6ed5caa6 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -1016,7 +1016,7 @@ class NPUModelRunner(GPUModelRunner):
                     target_positions = self._get_positions(num_scheduled_tokens)
                     target_hidden_states = hidden_states
                     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:
                     token_indices_to_sample = None
                     # input_ids can be None for multimodal models.