fix

Signed-off-by: wwl2755 <wangwenlong2755@gmail.com>

Author: wwl2755

tests/spec_decode/e2e/test_eagle_correctness.py

@@ -179,7 +179,7 @@ def test_eagle_e2e_greedy_correctness_cuda_graph(
 
 
 # TRACKING: https://github.com/vllm-project/vllm/issues/18166
-@pytest.mark.skip(reason="RE-ENABLE: Failing on main.")
+# @pytest.mark.skip(reason="RE-ENABLE: Failing on main.")
 @pytest.mark.parametrize(
     "common_llm_kwargs",
     [{

vllm/model_executor/models/eagle.py

@@ -145,6 +145,17 @@ def forward(
         if inputs_embeds is None:
             inputs_embeds = self.get_input_embeddings(input_ids)
 
+        # Handle both empty previous_hidden_states
+        # and mismatched batch size
+        batch_size = inputs_embeds.size(0)
+        if previous_hidden_states.size(0) == 0 or \
+           previous_hidden_states.size(0) != batch_size:
+            hidden_dim = self.config.model.hidden_size
+            device = inputs_embeds.device
+            # Create zero tensor with matching batch size
+            previous_hidden_states = \
+                torch.zeros(batch_size, hidden_dim, device=device)
+
         if self.add_para_norm:
             inputs_embeds = torch.cat([
                 self.enorm(inputs_embeds),

vllm/sequence.py

@@ -1330,6 +1330,8 @@ def prune(self,
         # may be "paused" then "resumed" later. This should only prune sequences
         # which are confirmed to be aborted.
         seq_ids = get_all_seq_ids(seq_group_metadata_list)
+        # Only keep sequence IDs that exist in self._seq_ids
+        seq_ids = [seq_id for seq_id in seq_ids if seq_id in self._seq_ids]
         if seq_ids != self._seq_ids:
             # Batch contents changed - prune removed sequences.
             index = [self._seq_ids.index(seq_id) for seq_id in seq_ids]