NVIDIA GeForce RTX 4090

### Name and Version $ ./llama-server --version ggml_cuda_init: found 2 CUDA devices (Total VRAM: 48161 MiB): Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes, VRAM: 24080 MiB

{"text":"Qwen3.6-27B-UD-Q4_K_XL.gguf using llama-server with openclaw. Failed to parse input."}

### Your current environment ``` uv is set ============================== System Info ============================== OS : Ubuntu 24.04.4 LTS (x86_64) GCC version

{"text":"Qwen 3.6 AWQ model, OOM error during loading."}

## Overview Adds native MTP speculative decoding support, including Qwen3.5/Qwen3.5-MoE MTP wiring and Gemma 4 MTP assistant support. This includes GGUF conversion/metadata handling for MTP

{"text":"Gemma 4 31B (q4km GGUF) with MTP speculative decoding, throughput approx 60 tok/s."}

# [Kernel][MLA] Triton-fused TurboQuant decode backend (#41803) > **Status:** Ready for **directional review**. Phase 3 benchmarks complete on 4×RTX 4090 + DeepSeek-V2-Lite, TP=4. DCO ✅. Phase-4 kern

{"text":"DeepSeek-V2-Lite, TP=4, TurboQuant 3/4bit_nc. 1.16-1.96x decode throughput improvement."}

### What is the issue? After updating to version `0.23.0` on Arch Linux, Ollama fails to discover my NVIDIA RTX 4090, reporting `total_vram="0 B"` and falling back to the CPU runner. Downgrading back

{"text":"Ollama v0.23.0 on Arch Linux; GPU discovery fails (total_vram=0 B) and falls back to CPU."}

### What is the issue? I was running granite4.1:30b (17 GB) and noticed it was running slow on my hardware, given the GPUs I have. When I ran `ollama ps`, I saw that the model was using 97 GB . I onl

{"text":"granite4.1:30b and 8b models. Context window size (131072) exceeds VRAM (48GB), causing CPU spillover and slow performance."}

### Name and Version $ ./llama-server --version ggml_cuda_init: found 2 CUDA devices (Total VRAM: 97020 MiB): Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes, VRAM: 48510 MiB

{"text":"Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf, llama-server, tool call validation error with anyOf schema"}

### Custom Node Testing - [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-wi

{"text":"Ernie-Image, Wan 2.2, LTX 2.3 等のモデルで、プロンプト変更後の再実行時にVRAM使用量が23.3/24GBに達し、システムがスタックするメモリ管理の問題が発生。"}

### Name and Version version: 8668 (5d3a4a7da) built with Clang 19.1.5 for Windows x86_64 ### Operating systems Windows ### GGML backends CUDA ### Hardware 4090 ### Models Qwen3.5-27B-Q4_K_M

{"text":"Qwen3.5-27B-Q4_K_M and Qwen3.5-35B-A3B-UD-IQ4_NL models using CUDA backend with openclaw, resulting in parse errors."}

### What is the issue? ### Description When using `gemma4:26b-a4b-it-q4_K_M` with the `format` parameter (JSON schema structured output), setting `think=false` causes the `format` constraint to be *

{"text":"gemma4:26b-a4b-it-q4_K_M モデルにおいて、think=false 設定時に format (JSON schema) 制約が無視され、プレーンテキストが出力される。"}

### Name and Version llama.cpp b8227-b8460 ### Operating systems Windows ### GGML backends CUDA ### Hardware intel core ultra 7 265K + RTX 4090 ### Models Qwen3.5-2B (I converted it myself) U

{"text":"Qwen3.5-2B (BF16) and Qwen3.5-9B (Q4_K_XL) on llama.cpp b8460; models think by default unexpectedly."}

### Custom Node Testing - [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-wi

{"text":"Switching from QWENEdit to SDXL causes slow sampling due to suspected VRAM swapping, despite 10GB available."}

### Is there an existing issue for this problem? - [X] I have searched the existing issues ### Operating system Windows ### GPU vendor Nvidia (CUDA) ### GPU model 4090 ### GPU

{"text":"OneTrainerで作成されたFlux LoRAのインポート時にInvalidModelConfigExceptionエラーが発生。"}