[Feature] support prompt repetition_penalty (#2806)

2025-10-05 16:48:03 +08:00 · 2025-07-17 12:05:52 +08:00
parent 7dfd2ea052
commit 1f15ca21e4
8 changed files with 305 additions and 64 deletions
--- a/fastdeploy/model_executor/layers/sample/sampler.py
+++ b/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,