NVIDIA GeForce RTX 4070

## 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":"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: 0.2.0-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 Ollama reuses a processed prompt prefix only while the runner is alive. For llama-server, the slot KV cache disappears with the subprocess, so a keep-alive expiry or model switch makes t

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

### 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"}

### 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":"MiniMaxH3モデルを使用。DynamicVRAM有効時に無限ハング、またはCUDA illegal memory accessが発生。"}

### 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":"MiniMaxH3モデル（int8_convrot/NVFP4）でDynamicVRAM有効時にハングまたはCUDA illegal memory accessが発生"}

### Name and Version version: 0.1.0-dev (build 10398, commit 8e7f22b67) built with GNU 13.3.0 for Linux x86_64 ### Operating systems Linux ### GGML backends CUDA CPU ### Hardware Intel Core i7

{"text":"Gemma 3 4B IT Q4_K_Mを使用。スロット復元の失敗後にK/Vデータが破損し、後続の推論が正常に行えなくなるバグ報告。"}

Follow-up to #16835, @jessegross replied: > Yes, if this was handled in a natural and Ollama native way that would be > useful. Ideally, support would be added for both runners. This issue proposes

{"text":"llama3.2:3bを使用。プリフィルキャッシュの再利用により、31Kトークン時のTTFTが約7秒から19-26msまで短縮された。"}

## Summary <!--A description of the changes in this PR. Include the kind of change (fix, feature, docs, etc), the "why" and the "how". Screenshots or videos are useful for frontend changes.--> U

{"text":"Triposplatモデル（約4GB）を使用し、2D画像から3Dへの変換およびキャンバス上でのレンダリングをRTX 4070 Laptop GPUで実行"}

## Purpose The target of this PR is to address issue #38175 related to the Idefics3 / SmolVLM ViT CUDA Graph. This PR adds ViT encoder CUDA Graph support for Idefics3 / SmolVLM, so that the encoder f

{"text":"Model: SmolVLM-256M-Instruct. Encoder CUDA Graph enabled, showing slight latency improvement (mean: -0.56%)."}

### Problem A single-user local agent typically prepends the same large system prompt (rules, tool definitions) on every turn. Ollama recomputes that full prefill each request, so every turn pays a f

{"text":"llama3.2:3bモデルでKVキャッシュの保存/復元機能を実装し、33KトークンプロンプトでTTFTを417倍高速化"}

### Name and Version > .\llama-cli.exe --version ggml_cuda_init: found 1 CUDA devices (Total VRAM: 12281 MiB): Device 0: NVIDIA GeForce RTX 4070, compute capability 8.9, VMM: yes, VRAM: 12281 MiB l

{"text":"Gemma 4 (Q8_0, BF16) でツール呼び出しがコンテンツとして出力されるバグ報告。"}

### What is the issue? CPU-only is available when starting ollama nvidia-smi works flawlessly from within a container, other containers work with nvidia 6.18.15+deb13-amd64 Driver Version: 550.163.0

{"text":"Ollama 20.3 in Docker container, Driver 550.163.01. GPU is not initialized and falls back to CPU-only."}

### 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":"Flux.2 Klein Img2Img でポスタリゼーション（階調飛び）が発生。"}

### 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":"Flux.2 Klein Img2Imgでのポスター化バグ報告"}

### 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":"ComfyUI Launcher 2.8.14.415 環境下で 'TypeError: Failed to fetch' エラーが発生"}

### 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":"ComfyUI Launcher 2.8.14.415 環境下で 'TypeError: Failed to fetch' エラーが発生"}