# vLLM 0.33.0rc0 pre-release fixes CUDA verification cost profiling bug

> **Key takeaway:** The vLLM inference engine's first release candidate for v0.33.0 patches a CUDA-specific bug in the MRV2 scheduler's adaptive verification cost profiling that could affect serving throughput.

**Published:** 2026-10-26T17:59:24+00:00 
**Source:** vLLM High-Throughput Serving Engine 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/vllm-0330rc0-pre-release-fixes-cuda-verification-cost-profil.html 

## Summary
vLLM, the Apache-2.0 licensed high-throughput serving engine for large language models, has published v0.33.0rc0, the initial release candidate for its next minor version. This pre-release contains a targeted bugfix for the MRV2 (Multi-Request Version 2) scheduling subsystem that skips adaptive verification cost profiling on the CUDA compute backend. The change was tagged by contributor LucasWilkinson on 26 October.

The fix addresses a specific performance or correctness issue in how vLLM profiles verification overhead during speculative decoding or similar adaptive execution paths. vLLM supports multiple backends including CUDA, ROCm, and CPU, though this particular change only affects the NVIDIA CUDA path. The project remains one of the most widely deployed open-source inference servers for production LLM serving.

## What changes in vLLM 0.33.0rc0
- **MRV2 scheduler:** The bugfix targets the Multi-Request Version 2 scheduling system, vLLM's newer request orchestration layer designed for higher throughput and lower latency in batched inference.
- **CUDA backend only:** The adaptive verification cost profiling skip applies specifically to the CUDA compute path, not to ROCm, CPU, or other supported backends.
- **Verification cost profiling:** This refers to runtime measurement of overhead for verifying speculative decoding outputs or other correctness checks; the fix prevents this profiling from executing in a problematic configuration.
- **Release candidate stage:** v0.33.0rc0 indicates the project is in final testing before the stable v0.33.0 release, with this commit (4135590) signed via GitHub's verified signature infrastructure.

## Why it matters
For organizations running vLLM on NVIDIA GPUs in production, this pre-release signals that the maintainers have identified and resolved a throughput-affecting edge case in their latest scheduler generation before final tagging. The Apache-2.0 licensing and multi-backend support continue to make vLLM a vendor-neutral alternative to proprietary serving stacks from NVIDIA Triton or cloud APIs, though users should await the stable release before deploying to production environments.