NVIDIA GeForce RTX 4070

## Overview <!-- Describe what this PR does and why. Be concise but complete --> This PR is to address performance degredation due to unexpected consequences to PR #29622. It resulted in small ope

{"text":"CUDA, Q4_K_M models, 8/32 threads"}

## Overview <!-- Describe what this PR does and why. Be concise but complete --> This PR is to address performance degredation due to unexpected consequences to PR #29622. It results in moving mov

{"text":"CUDA, Q4_K_M models, 8/32 threads, 512 tokens"}

### 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":"InvokeAI v6.14.2環境で、カスタムノードがcontext.boards.get_all()を呼び出した際にTypeErrorが発生"}

### Name and Version version commit 76a5bc86d1bdfae96feccdc7a41fea535e792e6e ### Operating systems Linux ### Which llama.cpp modules do you know to be affected? Other (Please specify in the next

{"text":"RPCサーバー起動時にシンボリックリンクされたキャッシュディレクトリを作成できないエラーが発生"}

### Name and Version version: b11194-5-g86a24a182 (master 86a24a182) ### Operating systems Linux ### GGML backends CUDA, HIP, Vulkan ### Hardware RTX 4070 (local). ggml-ci: T4, ROCm, Intel PTL,

{"text":"llama.cpp b11194-5-g86a24a182, hrm_text-dense.ggufモデル, GGML_SCHED_NO_REALLOC有効時のtest-save-load-stateテストでabort発生"}

## Overview For every new prompt the server creates context checkpoints (by default: start of the last user message, `4 + n_ubatch` tokens before the end, and 4 tokens before the end). For recurrent

{"text":"Qwen3.8-27B UD-IQ3_Sモデル、KV q4_0、64Kコンテキスト、MTP draft使用、2ノードRPCクラスタ環境"}

## Summary `ggml/src/ggml-cuda/argsort.cu` gates `STRIDED_ITERATOR_AVAILABLE` on a CCCL version check that compares the major and minor components independently: ```c #if (CCCL_MAJOR_VERSION >= 3 &&

{"text":"CUDA 13.4.92, CCCL 3.4.3, Ubuntu 26.04環境でビルド成功およびARGSORTテスト98/98合格"}

## Summary `ggml/src/ggml-cuda/argsort.cu` gates `STRIDED_ITERATOR_AVAILABLE` on a CCCL version check that compares the major and minor components independently: ```c #if (CCCL_MAJOR_VERSION >= 3 &&

{"text":"CUDA 13.4.92, CCCL 3.4.3"}

The MLX runner and Vulkan driver with Nvidia devices don't work in the docker container. Add missing libraries and configuration items. Pre: ```console $ docker run --gpus all --rm -it -e OLLAMA

{"text":"Dockerコンテナ内でNvidia GPU (RTX 4070) に対してVulkanおよびMLXランナーが正常に動作することを確認。"}

### What is the issue? i tried downloading a ollama model in the newest 0.34.2 and it didnt work it only worked after i switched to a 0.12 version i am also testing on the 0.34.1 version and its work

{"text":"Ollama 0.34.2環境でLlama-3.2-9Bモデルのダウンロード時にリダイレクトエラーが発生"}

### What is the issue? Windows build 10.0.26200 Ollama 0.34.1, clean reinstall RTX 4070 + RTX 3060, CUDA v13, driver 13.1 Recent security updates: KB5129195 (15 Sep), KB5124007 + KB5126052 (09 Sep

{"text":"Windows 11 26200環境でtrayアプリ経由のサーバー起動がタイムアウトするが、手動でollama serveを実行するとGPU検出され正常に動作する。"}

### What is the issue? Windows build 10.0.26200 Ollama 0.34.1, clean reinstall RTX 4070 + RTX 3060, CUDA v13, driver 13.1 Recent security updates: KB5129195 (15 Sep), KB5124007 + KB5126052 (09 Sep

{"text":"Windows 11 26200環境でtrayアプリ経由のサーバー起動がタイムアウトするが、手動でのollama serve実行ではGPU検出とAPI応答が正常に動作する。"}

## Bug description Both `pixal3d_bf16.safetensors` and `pixal3d_multiview_int8_convrot.safetensors` crash during structure/shape sampling with a tensor shape mismatch inside Trellis2's cross-attentio

{"text":"Pixal3D bf16およびmultiview checkpointsでtensor shape mismatchエラーが発生（int8_convrotは正常動作）"}

### Custom Node Testing - [ ] 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":"Trellis2 PixelArtistry_Pixal3D_native ワークフローで、PostProcessing > Remesh Mesh ステップ時にVRAM不足（Out of Memory）によりクラッシュ。"}

### Name and Version version: 0.3.0-dev (build 10636, commit 4d19b2876) built with GNU 16.2.1 for Linux x86_64 ### Operating systems Linux ### GGML backends CUDA ### Hardware i5 13400f + RTX 40

{"text":"CUDA backend, tensor parallelism, kv offload to cpu"}

### Name and Version version: 0.3.0-dev (build 10636, commit 4d19b2876) built with GNU 16.2.1 for Linux x86_64 ### Operating systems Linux ### GGML backends CUDA ### Hardware i5 13400f + RTX 40

{"text":"Tensor parallelismとKV cache offloadを併用した際にクラッシュが発生"}

### Name and Version version: 0.3.0-dev (build 10636, commit 4d19b2876) built with GNU 16.2.1 for Linux x86_64 ### Operating systems Linux ### GGML backends CUDA ### Hardware i5 13400f + RTX 40

{"text":"RTX 4070とRTX 3060の2枚構成でTensor Parallelismを使用し、KVキャッシュをCPUオフロードした際にクラッシュが発生。"}

### Name and Version version: 0.2.0-dev (build 10615, commit f280b2698) built with GNU 16.2.1 for Linux x86_64 ### Operating systems Linux ### Which llama.cpp modules do you know to be affected?

{"text":"CUDA FlashAttention (MMA) 実装において compute-sanitizer synccheck により divergent __syncthreads エラーが検出された"}

### What This PR adds experimental prefill/KV cache persistence across llama-server runner reloads. Ollama currently loses the processed prompt state when a runner unloads, so the next request m

{"text":"llama3.2:3b, CUDA 12.8, f16 KV, num_ctx=34000, OLLAMA_NUM_PARALLEL=1"}

### What This PR adds experimental prefill/KV cache persistence across llama-server runner reloads. Ollama currently loses the processed prompt state when a runner unloads, so the next request m

{"text":"llama3.2:3b, CUDA 12.8, f16 KV, num_ctx=34000"}