[Speculative Decoding] Refactor Eagle MTP hidden states copy (#6812)

* reformat eagle_get_hidden_states & eagle_get_self_hidden_states

* readibility

* fix xpu bug

* fix coverage failure

* change luanch params & parallelize position_map compute

* Fix MTP-related bugs in FastDeploy centralized inference

* fix

* refactor mtp hidden_states process

* fix

* add unittest & optimize kernel

* remove useless code

* fix

fastdeploy/model_executor/layers/sample/sampler.py

@@ -1290,35 +1290,7 @@ class MTPSampler(nn.Layer):
             elif self.logprobs_mode == "raw_logits":
                 raw_logprobs = share_inputs["draft_logits"][:real_token_num, :].clone()
 
-        if sampling_metadata.token_ids_all is not None:
-            token_ids_all = sampling_metadata.token_ids_all
-            prompt_lens = sampling_metadata.prompt_lens
-        else:
-            token_ids_all = sampling_metadata.pre_token_ids
-            prompt_lens = sampling_metadata.fake_prompt_lens
-
-        logits = apply_speculative_penalty_multi_scores(
-            token_ids_all,
-            prompt_lens,
-            logits,
-            sampling_metadata.repetition_penalties,
-            sampling_metadata.frequency_penalties,
-            sampling_metadata.presence_penalties,
-            sampling_metadata.temperature,
-            sampling_metadata.bad_words_token_ids,
-            sampling_metadata.bad_words_token_len,
-            sampling_metadata.step_idx,
-            sampling_metadata.min_dec_lens,
-            sampling_metadata.eos_token_ids,
-            share_inputs["seq_lens_this_time"],
-            share_inputs["batch_id_per_token_output"],
-            share_inputs["cu_seqlens_q_output"],
-            max_model_len,
-            sampling_metadata.pre_token_ids,
-        )
-        probs = F.softmax(logits)
-
-        next_tokens = paddle.argmax(probs, axis=-1)
+        next_tokens = paddle.argmax(logits, axis=-1)
 
         token_ids = None
         logprobs_tensors = None