[feat]解决EP开启pd_padding大输入卡住问题

7439b205 · 王敏 · b91ae72f · 7439b205 · 7439b205
Commit 7439b205 authored Dec 24, 2025 by 王敏
Show whitespace changes
Inline Side-by-side

Showing with 4 additions and 4 deletions

vllm/v1/spec_decode/eagle.py vllm/v1/spec_decode/eagle.py +2 -2

vllm/zero_overhead/v1/eagle.py vllm/zero_overhead/v1/eagle.py +2 -2

No files found.
--- a/vllm/v1/spec_decode/eagle.py
+++ b/vllm/v1/spec_decode/eagle.py
@@ -223,8 +223,8 @@ class EagleProposer:

        with set_forward_context(per_layer_attn_metadata,
                                 self.vllm_config,
-                                 num_tokens=num_input_tokens,
-                                 skip_cuda_graphs=not decoding):
+                                 num_tokens=num_input_tokens,):
+                                 #skip_cuda_graphs=not decoding):
            ret_hidden_states = self.model(
                self.input_ids[:num_input_tokens],
                self.positions[:num_input_tokens],

--- a/vllm/zero_overhead/v1/eagle.py
+++ b/vllm/zero_overhead/v1/eagle.py
@@ -146,8 +146,8 @@ class V1ZeroEagleProposer(EagleProposer):

        with set_forward_context(per_layer_attn_metadata,
                                 self.vllm_config,
-                                 num_tokens=num_input_tokens,
-                                 skip_cuda_graphs=not decoding):
+                                 num_tokens=num_input_tokens,):
+                                 #skip_cuda_graphs=not decoding):
            ret_hidden_states = self.model(
                self.input_ids[:num_input_tokens],
                self.positions[:num_input_tokens],