fix: Complete speculators Eagle support fixes

- Updated llama_eagle.py to skip transformer weights (loaded separately)
- Added num_lookahead_tokens to speculators config (required for Eagle)
- Together these fixes allow speculators Eagle models to work with V1 engine

Signed-off-by: rtuli@redhat.com

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>
Signed-off-by: Rahul Tuli <rtuli@redhat.com>

Author: rahul-tuli

vllm/model_executor/models/llama_eagle.py

@@ -117,9 +117,8 @@ def load_weights(self, weights: Iterable[tuple[str,
             if name in speculators_name_map:
                 name = speculators_name_map[name]
             elif name.startswith("transformer."):
-                # transformer.* -> model.layers.0.*
-                suffix = name[len("transformer."):]
-                name = f"model.layers.0.{suffix}"
+                # Skip transformer weights - they're loaded separately
+                continue
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:

vllm/transformers_utils/configs/speculators_eagle.py

@@ -89,6 +89,7 @@ def _convert_speculators_to_vllm(cls, speculators_config: dict) -> dict:
             "eagle_fc_bias": speculators_config.get("fusion_bias", False),
             "truncated_vocab_size": transformer_config.get("vocab_size"),
             "method": speculators_config.get("speculators_model_type", "eagle"),  # Use speculators_model_type
+            "num_lookahead_tokens": 5,  # Default number of speculative tokens for Eagle
         }
 
         # Preserve any additional fields that might be needed