NVIDIA GeForce RTX 4060 Ti

## Overview **Testing WIP - will update status once ready for review** Intel B70 Arc Pro (Win) and Nvidia RTX 4060 TI (Win) ### B70 Arc Pro (Windows) (Before) b11370>llama-bench.exe -p 8192

{"text":"llama.cpp VulkanバックエンドでGemma-4-26B-A4B (Q4_K_M)モデルを動作させ、pp8192で約1544 t/sの性能を確認。"}

### Name and Version version: 0.5.0-dev (build 11369, commit 4ebdf2c74) built with MSVC 19.50.35727.0 for x64 ### Operating systems Windows ### GGML backends Vulkan ### Hardware Intel B70 Arc P

{"text":"Vulkanバックエンド使用時、Flash Attention有効化でKVキャッシュ復元エラーが発生したが、無効化時は動作した"}

I have a system with two GPUs: 7800XT 4060ti I run two ollama instances on different ports for each. Can you improve the installer to download both the cuda and rocm runtimes. Right now, it will

{"text":"Ubuntu 24.04環境でNVIDIA GPUとAMD GPUの混在時に、インストーラーがCUDAのみを優先的にダウンロードしROCmが欠落する問題が発生。"}

## Overview I added CUDA support for `MUL_MAT_HADAMARD` / FWHT dimensions (384, 640, 768, and 1280). They are not powers of two. This was done using Kronecker products with the fixed `H12`/`H20

{"text":"CUDA 13.3環境でMUL_MAT_HADAMARDのテストが16/16成功"}

### Name and Version /llama-cli --version ATTENTION: default value of option radv_enable_unified_heap_on_apu overridden by environment. version: 0.3.0-dev (build 10679, commit 50f068fff) built with G

{"text":"Qwen3.8-Flash-Next (IQ1_S) モデルを RPC 経由で Vulkan バックエンド (BC250) にオフロードした際、ggml-vulkan.cpp のアサーション失敗によりクラッシュ"}

### Name and Version .\llama-cli --version version: 0.3.0-dev (build 10672, commit 511f9c137) built with Clang 20.1.8 for Windows x86_64 ### Operating systems Windows ### GGML backends CUDA ###

{"text":"Ministral-3-8B-Reasoning-2512-GGUF:Q8_0モデル使用時にCLIPグラフのIM2COL演算子がCUDAバックエンドでサポートされていない警告が発生し、パフォーマンスが最適化されない状態。"}

# GitHub Issue Draft: MiniMax H3 — text conditioning has zero effect on output (same seed + different prompts → bit-identical videos) ## Summary When running MiniMax H3 locally via the built-in `Min

{"text":"MiniMax H3モデルでテキストプロンプトが出力に影響せず、異なるプロンプトでもピクセル同一の動画が生成される"}

# GitHub Issue Draft: MiniMax H3 — text conditioning has zero effect on output (same seed + different prompts → bit-identical videos) ## Summary When running MiniMax H3 locally via the built-in `Min

{"text":"MiniMax H3モデル使用時、テキストプロンプトが出力に影響せず、異なるプロンプトでもピクセル単位で同一の動画が生成される。"}

### What is the issue? I am using Gemma 4. When I submit an image, neither the GUI nor the CLI reads it. Even when I send an image and ask "What is this?", it appears the image is not being sent or r

{"text":"Ollama 0.32.15環境でGemma 4モデルを使用し、画像添付しても認識されない"}

### 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":"単一のGPU環境で複数検出されるバグ報告"}

### Name and Version Head: version: 0.1.0-dev (build 10433, commit 9b05354ec) built with MSVC 19.51.36256.0 for x64 Worker: version: 0.1.0-dev (build 10433, commit 9b05354ec) built with MSVC 19.51.3

{"text":"Qwen3.6-27B-UD-Q4_K_XL.ggufを使用。llama-cliではRPC接続に成功するが、llama-serverでは接続に失敗する。"}

Performance & Architectural Improvements: - Fuse RMSNorm and scale/shift modulation into comfy.quant_ops.ck.rms_adaln in DiTBlock.forward - Vectorize segment modulation in _mod_scale_shift and _mod_

{"text":"CUDA 12.4, FP16. 30-step generationで約0.921s短縮、スループットが1.56から1.64 steps/sに向上。"}

Performance & Architectural Improvements: - Integrate fused comfy.quant_ops.ck.rms_adaln GPU kernels for norm1 and norm2 in all DiT blocks - Vectorize sequence modulation using pre-gathered segm

{"text":"CUDA 12.4, FP16. 30-step generationで約0.806秒の短縮、サンプリング速度が1.56から1.63 steps/sに向上。"}

## Overview <!-- Describe what this PR does and why. Be concise but complete --> This PR is to fix a performance bug with FA set to auto. ### Context - The V cache has two possible layo

{"text":"CUDA バックエンド、Qwen3-4B-Q4_0 モデル、depth 4096"}

## Overview <!-- Describe what this PR does and why. Be concise but complete --> This PR is to fix a performance bug with FA set to auto. ### Context - The V cache has two possible layo

{"text":"Qwen3-4B-Q4_0, depth 4096. Performance restored to parity with -fa 0 after fix."}

## Overview <!-- Describe what this PR does and why. Be concise but complete --> This PR is to fix a performance bug with FA set to auto. ### Context - The V cache has two possible layo

{"text":"Qwen3-4B-Q4_0モデル、tg32、depth 4096環境でFlash Attentionのレイアウトバグ修正により、-fa 0と同等のパフォーマンス（-0.15%）を達成。"}

### What is the issue? I keep getting "Error: context deadline exceeded" for a lot of models that I download from hf. ollama run hf.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-

{"text":"Qwen3.6-27B-Fable-Fusion (Q4_K_M) の実行時に context deadline exceeded エラーが発生。"}

### What is the issue? I keep getting "Error: context deadline exceeded" for a lot of models that I download from hf. ollama run hf.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-

{"text":"Qwen3.6-27B (Q4_K_M) の実行時に context deadline exceeded エラーが発生。"}

### What is the issue? I keep getting "Error: context deadline exceeded" for a lot of models that I download from hf. ollama run hf.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-

{"text":"Qwen3.6-27B (Q4_K_M) モデルの読み込み時に 'context deadline exceeded' エラーが発生"}

After the latest ComfyUI update, the following error appears at the end of every audio generation (ACE-Step workflow). It did not occur before the update. [ERROR] Exception in callback _ProactorBase

{"text":"ACE-Step audio generation (T2A). Audio generates correctly, but a cosmetic ConnectionResetError occurs after completion."}