Skip to content

[V0][Sampler] Use raw logits for greedy argmax #13312

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

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

[V0][Sampler] Use raw logits for greedy argmax #13312

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ddd7620
[V0][Sampler] Use raw logits for greedy argmax
njhill Feb 14, 2025
c882516
Merge remote-tracking branch 'origin/main' into greedy-logit-argmax
njhill Mar 19, 2025
491d923
Merge remote-tracking branch 'refs/remotes/origin/main' into greedy-l…
njhill Jun 19, 2025
6ef7fb7
Merge remote-tracking branch 'refs/remotes/origin/main' into greedy-l…
njhill Jul 7, 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
42 changes: 9 additions & 33 deletions vllm/model_executor/layers/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -296,6 +296,7 @@ def forward(

# Sample the next tokens.
maybe_deferred_sample_results, maybe_sampled_tokens_tensor = _sample(
logits,
probs,
logprobs,
sampling_metadata,
Expand Down Expand Up @@ -611,7 +612,8 @@ def get_pythonized_sample_results(
]


def _sample_with_torch(
def _sample(
logits: torch.Tensor,
probs: torch.Tensor,
logprobs: torch.Tensor,
sampling_metadata: SamplingMetadata,
Expand All @@ -629,6 +631,11 @@ def _sample_with_torch(
* Perform GPU-side sampling computation
* Defer Pythonization & preserve GPU-side
tensors required for Pythonization

Returns:
(next_token_ids, parent_seq_ids) for each seq group in a batch.
If sampling is skipped, it returns ([], [])
sampled_token_ids_tensor: A tensor of sampled token ids.
'''

categorized_seq_group_ids: dict[SamplingType, list[int]] = {
Expand Down Expand Up @@ -668,8 +675,7 @@ def _sample_with_torch(
sample_metadata[sampling_type] = (seq_group_id, seq_groups)
long_sample_indices = sample_indices.long()
if sampling_type == SamplingType.GREEDY:
greedy_samples = torch.argmax(logprobs[long_sample_indices],
dim=-1)
greedy_samples = torch.argmax(logits[long_sample_indices], dim=-1)

if sampled_token_ids_tensor is not None:
# Store sampled tokens in output tensor.
Expand Down Expand Up @@ -735,36 +741,6 @@ def _sample_with_torch(
)


def _sample(
probs: torch.Tensor,
logprobs: torch.Tensor,
sampling_metadata: SamplingMetadata,
sampling_tensors: SamplingTensors,
include_gpu_probs_tensor: bool,
modify_greedy_probs: bool,
) -> SampleReturnType:
"""
Args:
probs: (num_query_tokens_in_batch, num_vocab)
logprobs: (num_query_tokens_in_batch, num_vocab)
sampling_metadata: The metadata for a batch for sampling.
sampling_tensors: Tensors that include sampling related metadata.

Returns:
(next_token_ids, parent_seq_ids) for each seq group in a batch.
If sampling is skipped, it returns ([], [])
sampled_token_ids_tensor: A tensor of sampled token ids.
"""
return _sample_with_torch(
probs,
logprobs,
sampling_metadata,
sampling_tensors,
include_gpu_probs_tensor=include_gpu_probs_tensor,
modify_greedy_probs=modify_greedy_probs,
)


def _get_ranks(x: torch.Tensor, indices: torch.Tensor) -> torch.Tensor:
"""
This function calculates the ranks of the chosen tokens in a logprob tensor.
Expand Down