Skip to content

[Metrics] Read scheduled prefill KV tokens from batch prefix_lens #42797

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hnyls2002 merged 2 commits into from
Oct 6, 2026
Merged

[Metrics] Read scheduled prefill KV tokens from batch prefix_lens #42797

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a1e4460
fpm: read prefill kv tokens from batch prefix_lens
hnyls2002 Oct 6, 2026
47ad375
trim comments
hnyls2002 Oct 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion python/sglang/srt/managers/scheduler_components/metrics_reporter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -418,7 +418,9 @@ def _build_scheduled_request_metrics(self, batch: ScheduleBatch):
prefill_lengths.add(len(req.origin_input_ids))
num_prefill_requests = stats.num_new_seqs if stats else len(prefill_reqs)
sum_prefill_tokens = stats.log_input_tokens if stats else 0
sum_prefill_kv_tokens = sum(len(req.prefix_indices) for req in prefill_reqs)
# Prefill reqs lead batch.reqs, so they own the head of prefix_lens;
# each req's own prefix was already advanced by result processing.
sum_prefill_kv_tokens = sum(batch.prefix_lens[: len(prefill_reqs)])

decode_kv = WelfordAccumulator()
if batch.forward_mode.is_mixed():
Expand Down
8 changes: 4 additions & 4 deletions test/registered/unit/observability/test_forward_pass_metrics.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,11 +29,9 @@ def __init__(
self,
prompt_len: int,
output_len: int = 0,
prefix_len: int = 0,
):
self.origin_input_ids = list(range(prompt_len))
self.output_ids = list(range(output_len))
self.prefix_indices = list(range(prefix_len))
self.seqlen = prompt_len + output_len


Expand Down Expand Up @@ -153,6 +151,7 @@ def _make_batch(self, **overrides):
forward_mode=_FakeForwardMode(),
reqs=[],
decoding_reqs=[],
prefix_lens=[],
prefill_stats=None,
seq_lens_cpu=[],
fpm_start_time=100.0,
Expand All @@ -164,13 +163,14 @@ def test_emit_mixed_batch_separates_prefill_and_decode(self):
self.scheduler._fpm_dp_rank = 3
self.scheduler.waiting_queue = [_FakeReq(6), _FakeReq(4, output_len=2)]

prefill_a = _FakeReq(10, prefix_len=2)
prefill_b = _FakeReq(14, prefix_len=3)
prefill_a = _FakeReq(10)
prefill_b = _FakeReq(14)
decode_req = _FakeReq(8, output_len=3)
batch = self._make_batch(
forward_mode=_FakeForwardMode(is_mixed=True, is_extend=True),
reqs=[prefill_a, prefill_b, decode_req],
decoding_reqs=[decode_req],
prefix_lens=[2, 3, decode_req.seqlen - 1],
prefill_stats=PrefillStats(
log_input_tokens=12,
log_hit_tokens=5,
Expand Down
Loading