Skip to content

[V1][Sampler] Don't apply temp for greedy-only #13311

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 3 commits into from
Feb 15, 2025
Merged

[V1][Sampler] Don't apply temp for greedy-only #13311

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9a1ceb4
[V1][Sampler] Don't apply temp for greedy-only
njhill Feb 14, 2025
2a5392e
Merge remote-tracking branch 'origin/main' into no-temp-greedy
njhill Feb 14, 2025
131250d
fix min_p application
njhill Feb 15, 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
24 changes: 15 additions & 9 deletions vllm/v1/sample/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,8 +41,6 @@ def forward(
logits = self.apply_logits_bias(logits, sampling_metadata)
# Apply penalties (e.g., min_tokens, freq_penalties).
logits = self.apply_penalties(logits, sampling_metadata)
# Apply temperature.
logits = self.apply_temperature(logits, sampling_metadata.temperature)
# Sample the next token.
sampled = self.sample(logits, sampling_metadata)

Expand Down Expand Up @@ -82,9 +80,21 @@ def sample(
) -> torch.Tensor:
assert not (sampling_metadata.all_greedy
and sampling_metadata.all_random)
if sampling_metadata.all_greedy:
return self.greedy_sample(logits)
if sampling_metadata.all_random:
greedy_sampled = None
else:
greedy_sampled = self.greedy_sample(logits)
if sampling_metadata.all_greedy:
return greedy_sampled

# Apply temperature.
logits = self.apply_temperature(logits, sampling_metadata.temperature)

# Apply min_p.
if not sampling_metadata.no_min_p:
logits = self.apply_min_p(logits, sampling_metadata.min_p)

# Apply top_k and/or top_p.
random_sampled = self.topk_topp_sampler(
logits,
sampling_metadata.generators,
Expand All @@ -94,13 +104,9 @@ def sample(
sampling_metadata.top_p,
)

if not sampling_metadata.no_min_p:
logits = self.apply_min_p(logits, sampling_metadata.min_p)

if sampling_metadata.all_random:
if greedy_sampled is None:
return random_sampled

greedy_sampled = self.greedy_sample(logits)
sampled = torch.where(
sampling_metadata.temperature < _SAMPLING_EPS,
greedy_sampled,
Expand Down