Fix input layer norm mismatch for Eagle Speculative Decoding compatib… (ROCm#548)

* Fix input layer norm mismatch for Eagle Speculative Decoding compatibility

The LLaMA decoder layer applies input layer normalization at every layer, whereas Eagle omits it for the initial layer, using a dummy InputLayerNorm class instead.

Recently, LLaMA's input layer norm implementation (https://github.com/ROCm/vllm/blob/262ed1e16c5bd71f0612b700186854b8c932565d/vllm/model_executor/models/llama.py#L326) was updated to accept at most 3 inputs. To maintain compatibility and prevent Eagle Speculative Decoding from failing, this dummy class needs to be updated accordingly.

* Update eagle.py

Author: Mohammad Mahdi Kamani

vllm/model_executor/models/eagle.py

@@ -29,7 +29,7 @@ def __init__(self, weight=None, bias=None):
         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):
+    def forward(self, x, residual=None, scale=None):
         return x