vllm-project/vllm-metal v0.30.0.dev20261004205607

on GitHub

What's Changed

• [Perf] Project only the sampled logits rows on multimodal steps by @luke-edward in #1000
• [Model] Serve DiffusionGemma block diffusion on the paged Metal path by @marcomd in #994
• [KV Offload] KV cache offloading for the Metal backend by @RobbieJ in #998
• [Perf] Calibrate hd128 TurboQuant prefill admission by @suntp in #990
• GGUF: share the Q4 nibble split and accept remote :Q5_0/:Q5_1 tags by @begcdn in #997
• [Benchmarks] Compare DSpark HTTP serving with matched baselines by @StevenWang-CY in #999
• Add a KV offloading serving benchmark by @RobbieJ in #737

New Contributors

• @marcomd made their first contribution in #994
• @RobbieJ made their first contribution in #998

Full Changelog: v0.30.0.dev20261004100413...v0.30.0.dev20261004205607