@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 07064cb1d4
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 03:15 UTC

[Bugfix] Check chain_speculative_sampling before calling it (#11673)

Signed-off-by: Lu Fang <lufang@fb.com>

Lu Fang committed Jan 2, 2025 at 16:58 UTC 07064cb1d49d2b04ec58d8876bee2cd8281eedf5
1 file changed +1 -1
vllm/model_executor/layers/rejection_sampler.py
+1 -1
@@ -118,7 +118,7 @@ class RejectionSampler(SpecDecodeStochasticBaseSampler):
118
119 # If use Flashinfer chain_speculative_sampling kernel
120 # for rejection sampling
121 - if self.use_flashinfer:
121 + if self.use_flashinfer and chain_speculative_sampling is not None:
122 batch_size, k, _ = draft_probs.shape
123 uniform_samples = self._create_uniform_samples(
124 seeded_seqs, batch_size, k, draft_probs.device)