Repository Issue Activity (beta)

#53748 [Bug]: MLA decode fails on GB10 — shared memory 102400 > 101376; num_stages guard in _decode_grouped_att_m_fwd only fires for BLOCK_DMODEL >= 1024

#53413 [Bug]: GLM-5.2 FP8 on 8×H200 dies with runtime CUDA OOM in sparse_decode_fwd

#49238 [Bug]: Decode instance segfaults on NIXL `loadRemoteMD` after prefill pod restarts in P/D disaggregation

#53137 [Bug]: tools/recipes/recipe_json_to_vllm_config.py silently mis-parses the -cc (compilation-config) short alias

#53745 [Bug]: Strict tool calling attaches no structural tag when the reasoning and tool parsers share a parser engine

#50587 [Feature]: Kimi K3 Performance Optimization

#50877 [Bug]: DSpark speculative decoding triggers FlashInfer MNNVL allreduce "buffer size insufficient" via draft model's embed_input_ids (TP8, GB200 NVL72)

#51581 [Bug][Spec Decode]: DFlash fused-KV projection calls F.linear on a sliced qkv_proj weight — breaks (and can silently corrupt) any weight-quantized drafter

#53718 [Bug]: Named shared memory may be closed before peer ranks open it

#53749 [Bug]: enable_prefix_caching=True but zero hits for any shared prefix shorter than one attention block on hybrid checkpoints (2096 / 1920 tokens), with nothing in the logs or metrics naming the minimum

#51472 [RFC] Multimodal RL inputs for /inference/v1/generate

#53469 [Bug]: structured outputs never enforced for Muse Glimmer with --reasoning-parser muse_glimmer (json_object too; duplicate of #52594)

#53742 [Feature]: Rust frontend: /v1/embeddings endpoint

#51873 [Feature][DSpark]: Enable logprobs with adaptive verification

#53737 [Bug]: DeepSeek V3.2 encoder renders a trailing system message inside the assistant think block

#44494 [Bug]: Gemma 4 12B is not working

#53726 Silent CUDA IMA (exit 0) in hybrid GDN + MTP k=3 + async scheduling on RTX 3090; persists through #50021/#45100/#53613-class fixes

#15081 [Installation]: how to install v0.8.0

#53620 [Bug]: test_spec_decode_logprobs never runs on NVIDIA CI, and is ill-defined at top-k ties when it does

#52544 [Bug]: VLLM_USE_PRECOMPILED can fall back across CUDA variants without compatibility validation

#53642 [Bug]: Gemma4 parser swallows the rest of the turn when the model emits call:name(...) instead of call:name{...}

#35603 [Bug]: vllm: error: unrecognized arguments: --task embedding

#53665 [Bug]: min_tokens + structured output can empty the token mask and return HTTP 500

#52803 [ROCm][AMD] Kimi-K3 gfx942 / MI325X Gap and Roadmap

#53679 [Bug] DSpark/DFlash speculator CUDA graph capture crashes with 'CachingHostAllocator use_count > 0 INTERNAL ASSERT FAILED' after memory profiling