Add metrics for speculative decoding (acceptance rate, average acceptance length) (#11441)

2025-10-13 11:24:27 -07:00
parent f4aa78801e
commit b6fb5d7666
9 changed files with 70 additions and 1 deletions
--- a/python/sglang/srt/managers/io_struct.py
+++ b/python/sglang/srt/managers/io_struct.py
@@ -823,6 +823,7 @@ class BatchTokenIDOutput(BaseBatchReq):
    completion_tokens: List[int]
    cached_tokens: List[int]
    spec_verify_ct: List[int]
+    spec_accepted_tokens: List[int]

    # Logprobs
    input_token_logprobs_val: List[float]
@@ -896,6 +897,7 @@ class BatchStrOutput(BaseBatchReq):
    completion_tokens: List[int]
    cached_tokens: List[int]
    spec_verify_ct: List[int]
+    spec_accepted_tokens: List[int]

    # Logprobs
    input_token_logprobs_val: List[float]