LoginSign Up
← Back to all sparks
V

vLLM

AI-ASSISTANTS
Velocity5.0

High-throughput inference and serving engine for large language models

docs.vllm.ai ↗

vLLM opens its 0.34 cycle by moving more models onto the Transformers backend

llm-inferencerelease-candidatesnvfp4rocmhardware-support
◆Current state
vLLM's feed shows release candidates rather than release notes. The 0.33 cycle ran four candidates in late October with ROCm, NVFP4 and test-checkpoint fixes. The 0.34 cycle opened on 8 November by moving Gemma and StableLM onto the Transformers backend; rc1 added a NaN fix for speculative decoding on Intel XPU, and rc2 on 12 November only touches docs and tests.
◆Where it's heading
Cycles are short and the visible work is hardware coverage and consolidation. Moving older model families onto the generic Transformers backend points to less per-model code to maintain as new architectures arrive.
◆Prediction
Expect one or two more 0.34 candidates before a final release; the feed does not show the cycle's full feature set, so its scope stays unclear.

◆Recent moves

  1. 2d ago

    v0.34.0rc2 trims GPU memory settings from docs and tests

    The third 0.34 candidate; its only visible change stops docs and tests from setting gpu_memory_utilization. Three candidates in five days suggests the 0.34 final is close, but its feature set still isn't visible in the feed.

    View source ↗
  2. 2d ago

    v0.34.0rc1 fixes NaN handling in XPU speculative decoding

    The second 0.34 candidate; its only visible change treats NaN as negative infinity in greedy argmax for speculative decoding on Intel XPU. Hardware-specific stabilisation, in line with the rest of the cycle.

    View source ↗
  3. 5d ago

    v0.34.0rc0 moves Gemma and StableLM to the Transformers backend

    The first 0.34 candidate, whose only visible change migrates Gemma and StableLM onto the Transformers backend. It continues the consolidation pattern rather than adding user-facing capability.

    View source ↗
  4. 14d ago

    v0.33.0rc3 pins a test checkpoint before release

    The fourth 0.33 candidate, whose only visible change pins a test checkpoint. Stabilisation ahead of the final release.

    View source ↗
  5. 15d ago

    v0.33.0rc2 fixes a large-GEMM overflow on AMD gfx950

    Skips AITER-tuned GEMM on gfx950 when the matrix exceeds 2^31 elements, a ROCm overflow fix cherry-picked into the 0.33 candidate.

    View source ↗
  6. 17d ago

    v0.33.0rc1 fixes NVFP4 decode tests on DGX Spark

    A CI fix for NVFP4 decode tests on DGX Spark, cherry-picked into the release branch. No user-facing change is visible.

    View source ↗