[Feature] support prompt repetition_penalty (#2806)

fastdeploy/model_executor/layers/sample/meta_data.py

@@ -43,3 +43,5 @@ class SamplingMetadata:
     top_p: paddle.Tensor
     top_k: Optional[paddle.Tensor] = None
     max_num_logprobs: Optional[int] = None
+    prompt_ids: Optional[paddle.Tensor] = None
+    prompt_lens: Optional[paddle.Tensor] = None

fastdeploy/model_executor/layers/sample/ops/apply_penalty_multi_scores.py

@@ -21,6 +21,8 @@ from fastdeploy.platforms import current_platform
 
 def apply_penalty_multi_scores(
     pre_token_ids: paddle.Tensor,
+    prompt_ids: paddle.Tensor,
+    prompt_lens: paddle.Tensor,
     logits: paddle.Tensor,
     repetition_penalties: paddle.Tensor,
     frequency_penalties: paddle.Tensor,
@@ -39,6 +41,8 @@ def apply_penalty_multi_scores(
             get_token_penalty_multi_scores
         logits = get_token_penalty_multi_scores(
             pre_token_ids,
+            prompt_ids,
+            prompt_lens,
             logits,
             repetition_penalties,
             frequency_penalties,
@@ -69,6 +73,8 @@ def apply_penalty_multi_scores(
             get_token_penalty_multi_scores
         logits = get_token_penalty_multi_scores(
             pre_token_ids,
+            prompt_ids,
+            prompt_lens,
             logits,
             repetition_penalties,
             frequency_penalties,

fastdeploy/model_executor/layers/sample/sampler.py

@@ -253,6 +253,8 @@ class Sampler(nn.Layer):
 
         logits = apply_penalty_multi_scores(
             sampling_metadata.pre_token_ids,
+            sampling_metadata.prompt_ids,
+            sampling_metadata.prompt_lens,
             logits,
             sampling_metadata.repetition_penalties,
             sampling_metadata.frequency_penalties,