NVIDIA GeForce RTX 4090

### 🚀 The feature, motivation and pitch ### Motivation `vllm_gguf_plugin` is popular for running quantized GGUF weights. During autoregressive decoding ($N=1$), memory bandwidth is the bottleneck, an

{"text":"Llama-3.1-8B-Q8_0モデルで8,192トークンのプロンプト処理時、cuBLAS tensor cores利用により12.7倍高速化を確認"}

### Proposal to improve performance ### Summary When scoring prompt logprobs on long sequences ($T=8,192, K=32$), vLLM v1's Triton kernels (`_topk_log_softmax_kernel` and `_ranks_kernel` in `vllm/v1/

{"text":"Llama-3.1-8Bモデル、8192トークン長、prompt_logprobs=32の条件下で、TritonカーネルよりPyTorchネイティブ実装が約100倍高速化"}

The official library artifacts for Qwen2.5-Coder-3B-Instruct at four low-bit quantization levels produce fluent-looking but entirely non-functional output. On a 15-task functional smoke suite (easy H

{"text":"Qwen2.5-Coder-3B-Instruct の低ビット量子化モデル（q2_K, q3_K_S, q3_K_M, q3_K_L）が機能不全（コード生成成功率0%）"}

## Purpose BLIP-2 Q-Former attention currently materializes the attention scores, applies scaling and softmax separately, and launches a second matrix multiplication: ```text before: {matmul: 2, mul

{"text":"BLIP-2 Q-Former attentionのSDPA使用によるベンチマークで、FP16/BF16環境下で1.4x〜1.5xの高速化を確認。"}

### Your current environment <details> <summary>The output of <code>python collect_env.py</code></summary> ```text Collecting environment information... uv is set ==============================

{"text":"vLLM環境でFourierRotaryEmbeddingのバグにより非回転チャネルが破損する"}

## Summary This draft adds initial Mamba2 support for `VLLM_BATCH_INVARIANT=1` for the validated Triton SSU configuration. The same Mamba2 prefill can currently be factored into different physical S

{"text":"Granite 4.0 H 350M (bf16)モデルで、Triton SSU設定下におけるバッチ不変性(bitwise-equal)を検証し、25個のスケジューラテストが成功した。"}

### Is there an existing issue for this problem? - [x] I have searched the existing issues ### Install method Invoke's Launcher ### Operating system Windows ### GPU vendor Nvidia (CUDA) ### GP

{"text":"FP4 Mistral 3 Text Encoder (FLUX.2 DEV) の設定誤りによるバグ報告"}

## What happened Adding the top-level `logprobs`/`top_logprobs` fields to an `/api/chat` request changes the *generated text itself* under fully pinned greedy decoding — not just the metadata returne

{"text":"Ollama 0.30.5, gpt-oss:20b, greedy decoding (temperature 0) で logprobs 有効時に長文生成が非決定的になる"}

### Environment - vLLM 0.28.0, V1 engine, TP=2 - 2 × NVIDIA RTX 4090 (24,564 MiB each), driver 580.173.02, CUDA 13.0 - Qwen/Qwen3-32B-AWQ, `--max-model-len 8192` - Ubuntu 24.04, Python 3.12.3, `VLLM_

{"text":"vLLM 0.28.0, Qwen3-32B-AWQ, TP=2, 初回起動時にKVキャッシュが約5%少なくなる"}

## Purpose Add tuned Triton fused-MoE block configs for NVIDIA GeForce RTX 4090 (BF16), for Qwen1.5-MoE-A2.7B: - `E=30,N=1408,device_name=NVIDIA_GeForce_RTX_4090.json` — TP=2 + expert parallel - `E=

{"text":"Qwen1.5-MoE-A2.7B (BF16) の推論において、Triton fused-MoE カーネルのチューニングによりデコード(M=1)で約8%、バッチプリフィル(M>=128)で約4%の性能向上を確認。"}

## Purpose Related to #48485. When prefix caching is enabled, ordinary LRU eviction can discard blocks that ready waiting requests are about to reuse. Those requests then recompute a prefix even tho

{"text":"Qwen3-0.6Bモデルを使用し、prefix caching有効環境でのA/Bテストにおいて、提案された実装がベースラインより有益（BENEFICIAL）であることが確認された。"}

EDIT: it may be related to https://github.com/LostRuins/koboldcpp/issues/2005 I'm getting CUDA errors occasionally when processing prompts (never when actually generating). This is not a new issue bu

{"text":"Qwen3.6-35B-A3B-UD-IQ4_NLモデル使用時、プロンプト処理中にCUDA error: invalid argumentが発生し、特定コンテキストで再現する"}

## Overview <!-- Describe what this PR does and why. Be concise but complete --> DSV4, GLM use a new variant of GLU which clamps the SWIGLU. Currently in the graph it is not represented well for

{"text":"DSV4/GLMモデルでSWIGLU_CLAMPを使用し、4x4090環境でTG 2%、PP 1-2%の性能向上を確認"}

### Motivation. When EngineCore dies, an operator can often see that serving stopped but cannot recover the short sequence of scheduler, KV-cache, and worker state that led to the failure. Existing m

{"text":"vLLM 0.26.1rc1.dev1103+g7ca49fbe4環境で、外部レコーダーによるEngineCore失敗時の状態監視実験を実施。"}

### Name and Version [b10664](https://github.com/ggml-org/llama.cpp/releases/tag/b10664), [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/b10664/llama-b10664-bin-win-c

{"text":"Windows 10環境でmmap/tensor-read-lazy使用時にQwen3.8-Flash-Next-Q8_0モデル実行中、ディスクへの大量書き込み(~2GB/s)が発生するバグ報告"}

### Environment - Hardware: 8x NVIDIA GeForce RTX 4090 (Ada, compute capability 8.9, 48 GiB each) - Image: `vllm/vllm-openai:glm53-flash-cu129` (vLLM `0.1.dev20051+g487ecf187`, built from PR #53906)

{"text":"GLM-5.3-Flash (FP8) の sparse-MLA attention path が sm_89 (Ada) でサポートされていないため、モデルロード時にエラーが発生。"}

## Environment - vLLM 0.27.1, `vllm/vllm-openai` image digest `sha256:0a51ea5b4ae2dc5d81890e5173f54203d2a3ae0cfffe51b8fd2afd4391bfd967` (created 2026-08-11) - Model: `nvidia/NVIDIA-Nemotron-3.5-Ligh

{"text":"vLLM 0.27.1, Nemotron-3.5-Lightning-30B-A3B-NVFP4 (Marlin), TP=2, WSL2環境でのSleep/Wake操作によりLoRA状態が破損し、固定シード生成が分岐する。"}

## Summary Two independent defects prevent `incoai/Qwen3.8-27B-DFlash2` (the DFlash2 reference drafter) from loading on current main / nightly (`nightly-f94666b60d4c58ec0807d22c837cfae322a1dde9`): *

{"text":"RTX 4090 (24GB) 上で Qwen3.8-27B W4A16 と DFlash2 drafter の組み合わせ時に、初期化時のメモリ割り当てにより OOM エラーが発生。"}

### Your current environment <details> <summary>The output of <code>python collect_env.py</code></summary> ```text OS: Ubuntu 22.04.5 LTS Kernel: Linux 5.15.0-173-generic Machine: x86_64 Python: 3.1

{"text":"vLLM 0.27.1環境でQwen1.5-MoE-A2.7B-Chatモデルを使用し、同一入力に対するgreedy top-1トークンが実行ごとに異なるバグを確認"}

Fixes #15781 `MiniMaxH3.memory_usage_factor` (`comfy/supported_models.py`) was `0.114`, which underestimates the real sampling working set by roughly 1.45x on a 24GB GPU. Since `MIN_WEIGHT_MEMORY_RAT

{"text":"MiniMaxH3モデルで480x720x107フレームおよび736x1152x192フレームの生成がOOMなしに完了"}