●Change8

Migrating to vLLM v0.9.2rc1

Version v0.9.2rc1 introduces 3 breaking changes. This guide details how to update your code.

Released: 7/3/2025

3
Breaking Changes
4
Migration Steps
7
Affected Symbols

⚠️ Check Your Code

If you use any of these symbols, you need to read this guide:

MultiModalHasher.hash_prompt_mm_dataFusedMoEModularKernelAutoWeightsLoaderTorchAOConfigw8a8_block_fp8_matmul_deepgemmQwen3-RerankerVLLM_USE_V1

Breaking Changes

●Issue #1

ROCm platforms now use V1 engine by default, which may change performance characteristics or feature availability.

●Issue #2

Removed MultiModalHasher.hash_prompt_mm_data; code relying on this internal method will fail.

●Issue #3

New security policy prevents new imports of (cloud)pickle to mitigate deserialization vulnerabilities.

Migration Steps

  1. 1
    Update FlashInfer to 0.2.6.post1 if using FlashInfer backend.
  2. 2
    Ensure inputs are contiguous when using dynamic_per_token FP8/INT8 quantization.
  3. 3
    If using ROCm, review the V1 User Guide as it is now the default engine.
  4. 4
    Replace any custom usage of MultiModalHasher.hash_prompt_mm_data with standard hashing logic.

Release Summary

This release introduces support for Qwen3 Embedding/Reranker models, enables ROCm V1 by default, and adds several performance optimizations including deep_gemm support and vectorized INT8 kernels. It also includes critical bug fixes for structured outputs and CUDAGraph stability.

Need More Details?

View the full release notes and all changes for vLLM v0.9.2rc1.

View Full Changelog