Merge branch 'v0.9.2-dev' into 'v0.9.2-dev-add_connector'

# Conflicts:
#   vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_engine.py

vllm/zero_overhead/v1/gpu_model_runner.py

@@ -465,7 +465,7 @@ class V1ZeroModelRunner(GPUModelRunner):
         num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
         
         # make sure that the padded length is divisible by attn_tp_size because we may need reduce-scatter across attn_tp dim.
-        if self.ep_sp:
+        if self.ep_sp or self.enable_dp_attention:
             num_input_tokens = round_up(num_scheduled_tokens, tp_size)
             if (self.use_cuda_graph
                     and num_input_tokens <= self.cudagraph_batch_sizes[-1]):
@@ -796,6 +796,7 @@ class V1ZeroModelRunner(GPUModelRunner):
                         req_state = self.requests[req_id]
                         token_idx = self.last_sampled_token_lens[req_idx]
                         if token_idx == -1:
+                            self.fix_sampled_token_ids[req_idx].clear()
                             continue
                         fix_len = len(self.fix_sampled_token_ids[req_idx])
                         req_state.output_token_ids[token_idx:token_idx + fix_len] = self.fix_sampled_token_ids[req_idx]