Skip to content

[V1][Spec Decode] Avoid logging useless nan metrics #16023

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
merged 2 commits into from
Apr 4, 2025
Merged

[V1][Spec Decode] Avoid logging useless nan metrics #16023

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
5ffca64
[V1][Spec Decode] Avoid logging useless nan metrics
markmc Apr 3, 2025
7943506
[V1][Spec Decode] Fix scheduler stats test
markmc Apr 4, 2025
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
17 changes: 9 additions & 8 deletions tests/v1/core/test_scheduler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -671,10 +671,7 @@ def test_schedule_spec_decoding_stats(spec_tokens, output_tokens, expected):
assert running_req.num_tokens_with_spec == 2 + len(spec_tokens[i])

# No draft or accepted tokens counted yet
assert engine_core_outputs.scheduler_stats.spec_decoding_stats is not None
stats = engine_core_outputs.scheduler_stats.spec_decoding_stats
assert stats.num_draft_tokens == 0
assert stats.num_accepted_tokens == 0
assert engine_core_outputs.scheduler_stats.spec_decoding_stats is None

# Schedule the speculated tokens for validation
output = scheduler.schedule()
Expand Down Expand Up @@ -702,7 +699,11 @@ def test_schedule_spec_decoding_stats(spec_tokens, output_tokens, expected):
engine_core_outputs = scheduler.update_from_output(output,
model_runner_output)

assert engine_core_outputs.scheduler_stats.spec_decoding_stats is not None
stats = engine_core_outputs.scheduler_stats.spec_decoding_stats
assert stats.num_draft_tokens == expected[0]
assert stats.num_accepted_tokens == expected[1]
scheduler_stats = engine_core_outputs.scheduler_stats
if expected[0] == 0:
assert scheduler_stats.spec_decoding_stats is None
else:
assert scheduler_stats.spec_decoding_stats is not None
stats = scheduler_stats.spec_decoding_stats
assert stats.num_draft_tokens == expected[0]
assert stats.num_accepted_tokens == expected[1]
25 changes: 19 additions & 6 deletions vllm/v1/core/sched/scheduler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -553,11 +553,11 @@ def update_from_output(
spec_token_ids = model_runner_output.spec_token_ids
logprobs = model_runner_output.logprobs
prompt_logprobs_dict = model_runner_output.prompt_logprobs_dict
spec_decoding_stats = SpecDecodingStats() if self.log_stats else None
num_scheduled_tokens = scheduler_output.num_scheduled_tokens

new_running: list[Request] = []
outputs: list[EngineCoreOutput] = []
spec_decoding_stats: Optional[SpecDecodingStats] = None

# NOTE(woosuk): As len(self.running) can be up to 1K or more, the below
# loop can be a performance bottleneck. We should do our best to avoid
Expand Down Expand Up @@ -585,11 +585,10 @@ def update_from_output(
num_tokens_rejected = (len(scheduled_spec_token_ids) + 1 -
len(generated_token_ids))
request.num_computed_tokens -= num_tokens_rejected

if spec_decoding_stats is not None:
spec_decoding_stats.observe(
num_draft_tokens=len(scheduled_spec_token_ids),
num_accepted_tokens=len(generated_token_ids) - 1)
spec_decoding_stats = self.make_spec_decoding_stats(
spec_decoding_stats,
num_draft_tokens=len(scheduled_spec_token_ids),
num_accepted_tokens=len(generated_token_ids) - 1)

cached_encoder_input_ids = (
self.encoder_cache_manager.get_cached_input_ids(request))
Expand Down Expand Up @@ -744,3 +743,17 @@ def make_stats(
prefix_cache_stats=self.kv_cache_manager.make_prefix_cache_stats(),
spec_decoding_stats=spec_decoding_stats,
)

def make_spec_decoding_stats(
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

could this be a separate util function instead of a method? I think it would be nice if clutter was kept to a very bare minimum on scheduler.py

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, I considered it, but it pairs nicely with Scheduler make_stats() and it needs access to Scheduler log_stats

self,
spec_decoding_stats: Optional[SpecDecodingStats],
num_draft_tokens: int,
num_accepted_tokens: int,
) -> Optional[SpecDecodingStats]:
if not self.log_stats:
return None
if spec_decoding_stats is None:
spec_decoding_stats = SpecDecodingStats()
spec_decoding_stats.observe(num_draft_tokens=num_draft_tokens,
num_accepted_tokens=num_accepted_tokens)
return spec_decoding_stats