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 2, 20:45 UTC
[XPU] Ensure RNG offset alignment with PyTorch requirements in XPU sampler (#43028)
Signed-off-by: chaojun-zhang <chaojun.zhang@intel.com> Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Chaojun Zhang committed
May 26, 2026 at 10:01 UTC
71d810bbf44b34f3a019730a6878fbcbf2480499
1 file changed
+2
-1
vllm/v1/sample/ops/topk_topp_sampler.py
+2
-1
@@ -291,7 +291,8 @@ class TopKTopPSampler(nn.Module):
291
)
292
# The custom XPU sampler kernel consumes RNG values internally, so advance
293
# the default generator's offset to keep future draws deterministic.
294
- offset += logits.numel()
294
+ # pytorch: offset must be multiple of 4
295
+ offset = (offset + logits.numel() + 3) // 4 * 4
296
state.view(torch.int64)[1] = offset
297
generator.set_state(state)
298
return random_sampled, logits_to_return