[https://nvbugs/5451426][fix] Avoid torch compile on full eagle3 worker

Signed-off-by: Jin Li <[email protected]>

Author: liji-nv

tensorrt_llm/_torch/speculative/eagle3.py

@@ -266,7 +266,6 @@ def __init__(self, spec_config: "EagleDecodingConfig", mapping: Mapping):
         self.max_draft_len = self.spec_config.max_draft_len
         self.mapping = mapping
 
-    @torch.compile(options={"max-autotune": True})
     def forward(self, input_ids, position_ids, hidden_states, logits,
                 attn_metadata, spec_metadata, draft_model):
         batch_size = attn_metadata.num_seqs