[cherry-pick pr-4254] bugfix for mtp>1 when lm_head_tp>1 (#4360)

### What this PR does / why we need it?
Previously, the dummy run executed compute_logits only once, regardless
of num_speculative_tokens. This caused execute_model to hang on
compute_logits when lm head tensor parallelism exceeded 1. The fix
ensures compute_logits executes correctly during dummy run, matching
num_speculative_tokens.

Signed-off-by: zouyida2052 <zouyida2002@gmail.com>

vllm_ascend/spec_decode/eagle_proposer.py

@@ -116,7 +116,8 @@ class EagleProposer(Proposer):
                   num_reqs: int = 0,
                   num_tokens_across_dp: Optional[torch.Tensor] = None,
                   aclgraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
-                  batch_descriptor=None):
+                  batch_descriptor=None,
+                  dummy_compute_logits=lambda hidden_states: None):
         moe_comm_type = self.runner._select_moe_comm_method(
             num_tokens, with_prefill)
         with set_ascend_forward_context(None,
@@ -128,6 +129,7 @@ class EagleProposer(Proposer):
                 positions=self.positions[:num_tokens],
                 hidden_states=self.hidden_states[:num_tokens],
             )
+            dummy_compute_logits(self.hidden_states)
 
     def generate_token_ids(self,
                            valid_sampled_token_ids: list[list[int]],

vllm_ascend/spec_decode/mtp_proposer.py

@@ -114,7 +114,8 @@ class MtpProposer(Proposer):
                   num_reqs: int = 0,
                   num_tokens_across_dp=None,
                   aclgraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
-                  batch_descriptor=None) -> None:
+                  batch_descriptor=None,
+                  dummy_compute_logits=lambda hidden_states: None) -> None:
         if not self.torchair_graph_enabled:
             # TODO: adapt enable_dbo later
             (num_tokens, num_tokens_across_dp, with_prefill,
@@ -188,6 +189,7 @@ class MtpProposer(Proposer):
                     self.model(input_ids=input_ids,
                                positions=positions,
                                hidden_states=previous_hidden_states)
+                dummy_compute_logits(previous_hidden_states)
             if with_prefill:
                 break
 
@@ -490,6 +492,7 @@ class MtpProposer(Proposer):
             logits = self.model.compute_logits(sample_hidden_states)
             if lmhead_tp_enable() and num_indices < logits.shape[0]:
                 logits = logits[:num_indices]
+                last_token_indices = last_token_indices[:num_indices]
             draft_token_ids = logits.argmax(dim=-1)
 
             if self.num_speculative_tokens == 1:
@@ -554,7 +557,7 @@ class MtpProposer(Proposer):
             # For the requests that exceed the max model length, we set the
             # sequence length to 1 to minimize their overheads in attention.
             exceeds_max_model_len_cpu = exceeds_max_model_len.to(
-                attn_metadata_i.seq_lens.device, non_blocking=True)
+                attn_metadata_i.seq_lens.device, non_blocking=False)
             attn_metadata_i.seq_lens[:batch_size].masked_fill_(
                 exceeds_max_model_len_cpu, 1)
             # Mask out the slot mappings that exceed the max model length.

vllm_ascend/worker/model_runner_v1.py

@@ -2465,13 +2465,21 @@ class NPUModelRunner(LoRAModelRunnerMixin):
             need_dummy_logits = (not self.in_profile_run
                                  and lmhead_tp_enable())
 
-            if need_dummy_logits:
+            max_num_reqs_across_dp = num_tokens if not with_prefill else max_num_reqs
-                max_num_reqs_across_dp = num_tokens if not with_prefill else max_num_reqs
+            dummy_indices = torch.zeros(max_num_reqs_across_dp,
-                dummy_indices = torch.zeros(max_num_reqs_across_dp,
+                                        dtype=torch.int32)
-                                            dtype=torch.int32)
 
-                def dummy_compute_logits(hidden_states):
+            def dummy_compute_logits(hidden_states):
-                    return self.model.compute_logits(
+                if not need_dummy_logits:
+                    return None
+                return self.model.compute_logits(hidden_states[dummy_indices])
+
+            def dummy_drafter_compute_logits(hidden_states):
+                if not need_dummy_logits or self.drafter is None:
+                    return
+                if hasattr(self.drafter, "model") and hasattr(
+                        self.drafter.model, "compute_logits"):
+                    return self.drafter.model.compute_logits(
                         hidden_states[dummy_indices])
 
             with set_ascend_forward_context(
@@ -2493,8 +2501,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                     with_prefill, is_torchair_compile, input_ids, positions,
                     attn_metadata, num_tokens, intermediate_tensors,
                     inputs_embeds)
-                if need_dummy_logits:
+                dummy_compute_logits(hidden_states)
-                    dummy_compute_logits(hidden_states)
 
             if self.drafter:
                 self.drafter.dummy_run(
@@ -2504,10 +2511,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                     num_reqs=num_reqs,
                     num_tokens_across_dp=num_tokens_across_dp,
                     aclgraph_runtime_mode=aclgraph_runtime_mode,
-                    batch_descriptor=batch_descriptor)
+                    batch_descriptor=batch_descriptor,
-                if need_dummy_logits:
+                    dummy_compute_logits=dummy_drafter_compute_logits)
-                    self.drafter.model.compute_logits(
-                        hidden_states[dummy_indices])
             if self.in_profile_run and self.dynamic_eplb:
                 self.model.clear_all_moe_loads()
             if not self.in_profile_run and self.dynamic_eplb: