[Model Runner V2] Support num NaNs in logits (#30187)

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-12-09 10:00:49 -08:00
committed by GitHub
parent 9e6562a3f6
commit d471b2aff0
7 changed files with 89 additions and 26 deletions
+10 -2
View File
@@ -3,13 +3,15 @@
import torch
import vllm.envs as envs
from vllm.config.model import LogprobsMode
from vllm.v1.outputs import SamplerOutput
from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p
from vllm.v1.worker.gpu.metrics.logits import get_num_nans
from vllm.v1.worker.gpu.sample.gumbel import gumbel_sample
from vllm.v1.worker.gpu.sample.logprob import compute_topk_logprobs
from vllm.v1.worker.gpu.sample.metadata import SamplingMetadata
from vllm.v1.worker.gpu.sample.min_p import apply_min_p
from vllm.v1.worker.gpu.sample.output import SamplerOutput
from vllm.v1.worker.gpu.sample.penalties import apply_penalties_and_temperature
@@ -21,12 +23,16 @@ class Sampler:
if logprobs_mode not in ["processed_logprobs", "raw_logprobs"]:
raise NotImplementedError(f"Unsupported logprobs_mode: {logprobs_mode}")
self.logprobs_mode = logprobs_mode
self.compute_nans = envs.VLLM_COMPUTE_NANS_IN_LOGITS # False by default.
def __call__(
self,
logits: torch.Tensor,
sampling_metadata: SamplingMetadata,
) -> SamplerOutput:
# NOTE(woosuk): We intentionally compute num_nans before sampling to make clear
# that num_nans is computed before applying penalties and temperature.
num_nans = get_num_nans(logits) if self.compute_nans else None
sampled, processed_logits = self.sample(logits, sampling_metadata)
if sampling_metadata.max_num_logprobs is not None:
logits = (
@@ -49,6 +55,7 @@ class Sampler:
# token per request.
sampled_token_ids=sampled.view(-1, 1),
logprobs_tensors=logprobs_tensors,
num_nans=num_nans,
)
return sampler_output
@@ -63,7 +70,8 @@ class Sampler:
# Apply penalties and temperature in place.
apply_penalties_and_temperature(logits, sampling_metadata)
# Apply min_p in place.
apply_min_p(logits, sampling_metadata.min_p)
if sampling_metadata.min_p is not None:
apply_min_p(logits, sampling_metadata.min_p)
# Apply top_k and/or top_p. This might return a new tensor.
logits = apply_top_k_top_p(
logits, sampling_metadata.top_k, sampling_metadata.top_p