A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 16:00 UTC
[V1][Minor] Enhance SpecDecoding Metrics Log in V1 (#15902)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Woosuk Kwon committed
Apr 1, 2025 at 23:38 UTC
274d8e8818b85084c8d2e994b42dd77cb12d23af
1 file changed
+10
-7
vllm/v1/spec_decode/metrics.py
+10
-7
@@ -47,13 +47,16 @@ class SpecDecodingMetrics:
47
num_draft_tokens = np.sum(self.num_draft_tokens)
48
num_accepted_tokens = np.sum(self.num_accepted_tokens)
49
50
- draft_acceptance_rate = (num_accepted_tokens / num_draft_tokens
51
- if num_draft_tokens > 0 else float("nan"))
50
+ draft_acceptance_rate = (num_accepted_tokens / num_draft_tokens *
51
+ 100 if num_draft_tokens > 0 else float("nan"))
52
53
logger.info(
54
- "Speculative metrics: "
55
- "Draft acceptance rate: %.3f, "
56
- "Number of accepted tokens: %d, "
57
- "Number of draft tokens: %d, ", draft_acceptance_rate,
58
- num_accepted_tokens, num_draft_tokens)
54
+ "SpecDecoding metrics: "
55
+ "Draft acceptance rate: %.1f%%, "
56
+ "Accepted: %d tokens, "
57
+ "Drafted: %d tokens",
58
+ draft_acceptance_rate,
59
+ num_accepted_tokens,
60
+ num_draft_tokens,
61
+ )
62
self.reset()