[CI][Spec Decode] fix: broken test for EAGLE model (vllm-project#11972)

Signed-off-by: Sungjae Lee <[email protected]>

Author: llsj14

.buildkite/test-pipeline.yaml

@@ -231,13 +231,15 @@ steps:
     - pytest -v -s test_logits_processor.py
     - pytest -v -s model_executor/test_guided_processors.py
 
-- label: Speculative decoding tests # 30min
+- label: Speculative decoding tests # 40min
   source_file_dependencies:
   - vllm/spec_decode
   - tests/spec_decode
+  - vllm/model_executor/models/eagle.py
   commands:
     - pytest -v -s spec_decode/e2e/test_multistep_correctness.py
     - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py
+    - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
   mirror_hardwares: [amd]

vllm/model_executor/models/eagle.py

@@ -19,6 +19,11 @@
 
 class DummyInputLayerNorm(nn.Module):
 
+    def __init__(self, weight=None, bias=None):
+        super().__init__()
+        self.weight = nn.Parameter(weight) if weight is not None else None
+        self.bias = nn.Parameter(bias) if bias is not None else None
+
     def forward(self, x):
         return x
 
@@ -69,7 +74,11 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
 
         # Modify layer normalization and residual connections as suggested
         # in the EAGLE framework: https://github.com/SafeAILab/EAGLE
-        self.model.model.layers[0].input_layernorm = DummyInputLayerNorm()
+        # While weights and biases are generally not needed,
+        # they are retained here to support certain unit tests
+        # (e.g., spec_decode/e2e/test_eagle_correctness.py).
+        self.model.model.layers[0].input_layernorm = DummyInputLayerNorm(
+            weight=self.model.model.layers[0].input_layernorm.weight)
         self.model.model.norm = DummyOutputNorm()
 
         self.orig_vocab_size = config.vocab_size