remove relaxed thinking

Signed-off-by: qgai <[email protected]>

Author: sunnyqgg

examples/llm-api/quickstart_advanced.py

@@ -176,12 +176,7 @@ def setup_llm(args, **kwargs):
                 max_draft_len=args.spec_decode_max_draft_len,
                 speculative_model_dir=args.model_dir,
                 eagle3_one_model=args.use_one_model,
-                is_mtp_eagle=True,
-                use_relaxed_acceptance_for_thinking=args.
-                use_relaxed_acceptance_for_thinking,
-                relaxed_topk=args.relaxed_topk,
-                relaxed_delta=args.relaxed_delta,
-            )
+                is_mtp_eagle=True)
         else:
             spec_config = MTPDecodingConfig(
                 num_nextn_predict_layers=args.spec_decode_max_draft_len,

tensorrt_llm/_torch/models/modeling_auto.py

@@ -20,7 +20,7 @@ def from_config(
                                         "")  # Strip the appended EAGLE3
         if hasattr(config.pretrained_config, "draft_vocab_size"):
             model_arch = "EAGLE3" + model_arch
-        if model_arch == "DeepseekV3ForCausalLM" and config.spec_config.max_draft_len == 0:
+        if model_arch == "DeepseekV3ForCausalLM" and config.spec_config is not None and config.spec_config.max_draft_len == 0:
             model_arch = "MTPDraftModelForCausalLM"
 
         cls = MODEL_CLASS_MAPPING.get(model_arch)

tensorrt_llm/llmapi/llm_args.py

@@ -427,10 +427,6 @@ class EagleDecodingConfig(DecodingBaseConfig):
     eagle3_one_model: Optional[bool] = True
     eagle3_layers_to_capture: Optional[Tuple[int, ...]] = None
     is_mtp_eagle: Optional[bool] = False  # which also means eagle2
-    #TODO:add the support of relaxed acceptance for thinking
-    use_relaxed_acceptance_for_thinking: Optional[bool] = False
-    relaxed_topk: Optional[int] = 10
-    relaxed_delta: Optional[float] = 0.01
 
     @classmethod
     def from_dict(cls, data: dict):