Merge branch 'v0.9.2-dev-ds_auto_21' into 'v0.9.2-dev-ds_auto_12.29'

[feat]解决deepep auto模式卡住问题

See merge request dcutoolkit/deeplearing/vllm!324

vllm/zero_overhead/v1/eagle.py

@@ -108,6 +108,9 @@ class V1ZeroEagleProposer(EagleProposer):
         else:
             num_input_tokens = num_tokens
 
+        num_pad, num_tokens_across_dp = self.get_dp_padding(num_input_tokens)
+        num_input_tokens += num_pad
+
         # copy inputs to buffer for cudagraph
         self.positions[:num_tokens] = target_positions
         self.hidden_states[:num_tokens] = target_hidden_states