[{"data":1,"prerenderedAt":242},["ShallowReactive",2],{"finding:c33-cuda-setup-notes":3,"finding-runs:c33-cuda-setup-notes":205,"finding-related:c33-cuda-setup-notes":206},{"meta":4,"impact":30,"sections":35},{"id":5,"title":6,"subtitle":7,"eyebrow":8,"date":9,"status":10,"category":11,"polarity":12,"axes":13,"tags":15,"task_code":26,"related_findings":27},"c33-cuda-setup-notes","c33 CUDA env setup — gsplat \u002F orig 3DGS \u002F brush の 3 env を A6000 + NFS 共有ホームで構築","Phase 2 (A.3) の準備。c33 (A6000 × 3, sm_86, CUDA 12.1) に gsplat-env \u002F orig3dgs-env \u002F Rust + brush の 3 env を build。NFS 共有ホーム (\u002Frda5、11TB 空き) なので c32 \u002F c34 へ継承可能、累計 ~18.6 GB。実 training は Phase 2 で実施。","Phase 2 prep · CUDA env setup","2026-05-22","stable","setup","neutral",[14],2,[16,11,17,18,19,20,21,22,23,24,25],"phase-2","cuda","a6000","c33","gsplat","original-3dgs","brush","conda","rust","nfs","A.3",[28,29],"c32-gsplat-smoke","c32-orig3dgs-bench",{"summary":31,"rank":32,"verdict":33,"delta_wallclock":34},"c33 (A6000, sm_86) に gsplat-env \u002F orig3dgs-env \u002F Rust + brush の 3 env を build。NFS 共有ホーム経由で c32 \u002F c34 にも継承可能。3 env とも import \u002F --help \u002F --version レベルで動作確認 OK。実 training (Lego 30k) は Phase 2 で。","mid","accepted","~18.6 GB disk",[36,39,56,59,89,95,97,101,104,106,108,111,114,116,118,120,122,124,126,128,129,131,133,135,136,138,143,144,146,148,150,154,155,157,159,160,162,164,166,168,170,172,174,176,178,180,182,189,191,197,199,201,203],{"type":37,"text":38},"lead","卒研「第 2 軸 (wgpu 抽象コスト) 三層対比表」用に、c33 (compute-0-33, RTX A6000 48GB × 3, driver 530.30.02, CUDA 12.1) で \u003Cstrong>gsplat-env \u002F orig3dgs-env \u002F Rust + brush\u003C\u002Fstrong> の 3 env を構築。NFS 共有ホーム (\u003Ccode>\u002Frda5\u003C\u002Fcode>、11TB 空き) なので c32 \u002F c34 にも継承される。3 env とも \u003Ccode>import\u003C\u002Fcode> \u002F \u003Ccode>--help\u003C\u002Fcode> \u002F \u003Ccode>--version\u003C\u002Fcode> レベルで動作確認 OK。\u003Cstrong>実 training (Lego 30k) は Phase 2 で実施\u003C\u002Fstrong>、本タスクは環境構築 + 動作確認のみ。",{"type":40,"items":41},"kv",[42,44,47,50,53],{"key":43,"value":9},"実施日",{"key":45,"value":46},"対象機","matsudalab-c33 (compute-0-33, RTX A6000 48GB × 3, driver 530.30.02, CUDA 12.1)",{"key":48,"value":49},"ホーム","NFS 共有 \u002Frda5、11TB 空き — c32 \u002F c34 と共有",{"key":51,"value":52},"累計 disk","~18.6 GB (miniconda 12G + brush target 3.9G + .rustup 1.4G + .cargo 954M + gaussian-splatting 308M)",{"key":54,"value":55},"GPU 占有 (setup 中)","GPU0 48.5\u002F49.1 GiB、GPU1\u002F2 16.1\u002F49.1 GiB、Util 81-93% — 他ユーザ稼働中",{"type":57,"text":58},"heading","TL;DR",{"type":60,"columns":61,"align":68,"rows":70},"table",[62,63,64,65,66,67],"env","用途","python","torch \u002F rust","CUDA","動作確認",[69,69,69,69,69,69],"left",[71,78,82],[72,73,74,75,76,77],"\u003Ccode>gsplat-env\u003C\u002Fcode>","gsplat (CUDA native via PyTorch extension)","3.11","torch 2.4.1+cu121","12.1","\u003Ccode>gsplat 1.5.3\u003C\u002Fcode> import OK、\u003Ccode>cuda.is_available()\u003C\u002Fcode> True、device \u003Ccode>NVIDIA RTX A6000\u003C\u002Fcode>",[79,80,74,75,76,81],"\u003Ccode>orig3dgs-env\u003C\u002Fcode>","graphdeco-inria\u002Fgaussian-splatting","\u003Ccode>diff_gaussian_rasterization\u003C\u002Fcode> + \u003Ccode>simple_knn\u003C\u002Fcode> + \u003Ccode>fused_ssim\u003C\u002Fcode> 全 import OK、\u003Ccode>train.py --help\u003C\u002Fcode> OK",[83,84,85,86,87,88],"Rust (system-wide, conda 外)","brush (wgpu → Vulkan)","—","rustc stable","— (Vulkan 1.2.131 loader)","\u003Ccode>~\u002Frepos\u002Fbrush\u002Ftarget\u002Frelease\u002Fbrush --version\u003C\u002Fcode> → \u003Ccode>brush-cli 0.3.0\u003C\u002Fcode>、HEAD \u003Ccode>ce6ef3f\u003C\u002Fcode>",{"type":90,"items":91},"list",[92,93,94],"\u003Cstrong>NFS 共有ホーム\u003C\u002Fstrong> (\u003Ccode>\u002Frda5\u003C\u002Fcode>、11TB 空き) なので c32 \u002F c34 にもそのまま継承","\u003Cstrong>disk 使用\u003C\u002Fstrong>: 累計 ~18.6 GB (miniconda 12G + brush target 3.9G + .rustup 1.4G + .cargo 954M + gaussian-splatting 308M)","\u003Cstrong>GPU 占有状況\u003C\u002Fstrong> (setup 中スナップショット): GPU0 48.5\u002F49.1 GiB、GPU1\u002F2 16.1\u002F49.1 GiB、Util 81-93% — \u003Cstrong>他ユーザ稼働中\u003C\u002Fstrong>、Phase 2 実 train は要調整",{"type":57,"text":96},"1. 状態点検 (setup 前)",{"type":98,"lang":99,"text":100},"code","text","hostname: compute-0-33\nuser: otake_26 (Domain Users, NFS home \u002Frda5\u002Fusers\u002Fotake_26)\nOS: Ubuntu 20.04 系 (Linux 5.4.0-148-generic)\nshell: \u002Fbin\u002Fbash\nhome: 空 (.bashrc \u002F .profile のみ、conda \u002F cargo \u002F python venv なし)\ndisk: NFS 192.168.2.226:\u002Fdata2 → \u002Frda5、Size 55T \u002F Avail 11T \u002F Use 80%\nnvcc: \u002Fusr\u002Flocal\u002Fcuda\u002Fbin\u002Fnvcc → 12.1.105 (cuda_12.1.r12.1)\nCUDA 候補: cuda-11.4, 11.5, 11.6, 12, 12.1, cuda (default → 12.1)\nGPU driver: 530.30.02, RTX A6000 48GB × 3\nVulkan loader: \u002Fusr\u002Flib\u002Fx86_64-linux-gnu\u002Flibvulkan.so.1.2.131 (brush の wgpu→Vulkan path 用)\ntmux: 3.0a (長時間 install は tmux session 分離)\ngcc: 既存 (toolchain 確認略)\n",{"type":102,"text":103},"paragraph","\u003Cstrong>判断\u003C\u002Fstrong>: driver 530.30.02 が CUDA 12.1 までサポート、system nvcc が 12.1 なので \u003Cstrong>PyTorch も cu121 で統一\u003C\u002Fstrong>。advisor 推奨どおり cu118\u002F12.1 ミスマッチを避ける。",{"type":102,"text":105},"\u003Ccode>tiny-cuda-nn\u003C\u002Fcode> は \u003Cstrong>原著 3DGS には不要\u003C\u002Fstrong> (advisor 指摘、\u003Ccode>graphdeco-inria\u002Fgaussian-splatting\u003C\u002Fcode> 本来の依存は \u003Ccode>diff-gaussian-rasterization\u003C\u002Fcode> + \u003Ccode>simple-knn\u003C\u002Fcode> + 任意 \u003Ccode>fused-ssim\u003C\u002Fcode> のみ)。タスク仕様の記述を訂正済。",{"type":57,"text":107},"2. 共通 setup",{"type":57,"level":109,"text":110},3,"2.1 Miniconda",{"type":98,"lang":112,"text":113},"bash","mkdir -p ~\u002Fminiconda3 && cd ~\u002Fminiconda3\nwget -q https:\u002F\u002Frepo.anaconda.com\u002Fminiconda\u002FMiniconda3-latest-Linux-x86_64.sh -O install.sh\nbash install.sh -b -u -p ~\u002Fminiconda3\n~\u002Fminiconda3\u002Fbin\u002Fconda init bash\n# Anaconda TOS (conda 26 系で必須):\n~\u002Fminiconda3\u002Fbin\u002Fconda tos accept --override-channels --channel https:\u002F\u002Frepo.anaconda.com\u002Fpkgs\u002Fmain\n~\u002Fminiconda3\u002Fbin\u002Fconda tos accept --override-channels --channel https:\u002F\u002Frepo.anaconda.com\u002Fpkgs\u002Fr\n",{"type":102,"text":115},"\u003Ccode>conda 26.3.2\u003C\u002Fcode> がインストールされた。\u003Ccode>.bashrc\u003C\u002Fcode> に conda init block が追加された。",{"type":57,"level":109,"text":117},"2.2 CUDA path (.bashrc)",{"type":98,"lang":112,"text":119},"# >>> CUDA toolkit (added 2026-05-22 for 3DGS env setup) >>>\nexport CUDA_HOME=\u002Fusr\u002Flocal\u002Fcuda\nexport PATH=$CUDA_HOME\u002Fbin:$PATH\nexport LD_LIBRARY_PATH=$CUDA_HOME\u002Flib64:${LD_LIBRARY_PATH:-}\n# \u003C\u003C\u003C CUDA toolkit \u003C\u003C\u003C\n",{"type":102,"text":121},"NFS 共有ホームで c32 \u002F c34 にもこの block が反映される。c32 \u002F c34 で別 CUDA 系を使う場合は per-env で \u003Ccode>CUDA_HOME\u003C\u002Fcode> を上書きする (conda env config vars set) 方が安全だが、現状 3 機とも \u003Ccode>\u002Fusr\u002Flocal\u002Fcuda\u003C\u002Fcode> 経由で十分。",{"type":57,"text":123},"3. gsplat-env",{"type":57,"level":109,"text":125},"手順",{"type":98,"lang":112,"text":127},"conda create -y -n gsplat-env python=3.11\nconda activate gsplat-env\npip install --upgrade pip\npip install torch==2.4.1 torchvision --index-url https:\u002F\u002Fdownload.pytorch.org\u002Fwhl\u002Fcu121\npip install ninja gsplat\n",{"type":57,"level":109,"text":67},{"type":98,"lang":99,"text":130},"$ python -c \"import gsplat, torch; print(gsplat.__version__, torch.__version__, torch.version.cuda, torch.cuda.is_available(), torch.cuda.get_device_name(0))\"\nGSPLAT_VER         1.5.3\nTORCH_VER          2.4.1+cu121\nTORCH_CUDA_BUILT   12.1\nTORCH_CUDA_AVAIL   True\nDEVICE_NAME        NVIDIA RTX A6000\n",{"type":102,"text":132},"\u003Cstrong>OK\u003C\u002Fstrong>。gsplat 1.5.3 は JIT で CUDA extension をビルド (初回 import 時)、本タスクでは \u003Ccode>cuda.is_available()\u003C\u002Fcode> のみで alloc は走らせていない。",{"type":57,"text":134},"4. orig3dgs-env (graphdeco-inria\u002Fgaussian-splatting)",{"type":57,"level":109,"text":125},{"type":98,"lang":112,"text":137},"conda create -y -n orig3dgs-env python=3.11\nconda activate orig3dgs-env\nexport TORCH_CUDA_ARCH_LIST=\"8.6\"  # sm_86 = A6000 \u002F 3090 \u002F A40\n\npip install torch==2.4.1 torchvision --index-url https:\u002F\u002Fdownload.pytorch.org\u002Fwhl\u002Fcu121\npip install ninja plyfile tqdm opencv-python joblib\n\nmkdir -p ~\u002Frepos && cd ~\u002Frepos\ngit clone --recursive https:\u002F\u002Fgithub.com\u002Fgraphdeco-inria\u002Fgaussian-splatting.git\ncd gaussian-splatting\n\n# build isolation を切らないと subprocess で torch を find できない\npip install --no-build-isolation submodules\u002Fdiff-gaussian-rasterization\npip install --no-build-isolation submodules\u002Fsimple-knn\npip install --no-build-isolation submodules\u002Ffused-ssim\n",{"type":139,"label":140,"variant":141,"text":142},"callout","詰まりポイント","warn","\u003Ccode>pip install submodules\u002Fdiff-gaussian-rasterization\u003C\u002Fcode> がデフォルトで build-isolated subprocess を作るので \u003Ccode>setup.py\u003C\u002Fcode> から \u003Ccode>import torch\u003C\u002Fcode> が \u003Ccode>ModuleNotFoundError\u003C\u002Fcode>。\u003Ccode>--no-build-isolation\u003C\u002Fcode> が必須。",{"type":57,"level":109,"text":67},{"type":98,"lang":99,"text":145},"$ python -c \"import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available(), torch.cuda.get_device_name(0))\"\n2.4.1+cu121 12.1 True NVIDIA RTX A6000\n\n$ python -c \"import diff_gaussian_rasterization, simple_knn, fused_ssim\"\n(no error)\n\n$ cd ~\u002Frepos\u002Fgaussian-splatting && python train.py --help | head -10\nusage: train.py [-h] [--sh_degree SH_DEGREE] [--source_path SOURCE_PATH] ...\n",{"type":102,"text":147},"\u003Cstrong>OK\u003C\u002Fstrong>。\u003Ccode>diff_gaussian_rasterization\u003C\u002Fcode> 0.0.0、\u003Ccode>simple_knn\u003C\u002Fcode> 0.0.0、\u003Ccode>fused_ssim\u003C\u002Fcode> 0.0.0 (Kerbl 公式 commit)、いずれも cu121 + sm_86 でビルド成功 (advisor 懸念点 \"diff-gaussian-rasterization HEAD が cu121 で建つか\" → OK)。",{"type":57,"text":149},"5. brush (Rust + wgpu → Vulkan)",{"type":139,"label":151,"variant":152,"text":153},"Note","info","タスク仕様の \"wgpu cuda backend\" は誤り (advisor 指摘)。wgpu は Linux 上では Vulkan、Apple は Metal、Windows は DX12 を採用。c33 では \u003Ccode>libvulkan.so.1.2.131\u003C\u002Fcode> を loader として wgpu→Vulkan で NVIDIA driver にディスパッチする。",{"type":57,"level":109,"text":125},{"type":98,"lang":112,"text":156},"curl --proto '=https' --tlsv1.2 -sSf https:\u002F\u002Fsh.rustup.rs | sh -s -- -y --default-toolchain stable\nsource ~\u002F.cargo\u002Fenv\n\nmkdir -p ~\u002Frepos && cd ~\u002Frepos\ngit clone https:\u002F\u002Fgithub.com\u002FArthurBrussee\u002Fbrush.git\ncd ~\u002Frepos\u002Fbrush\ncargo build --release   # 5m 54s\n",{"type":102,"text":158},"target は \u003Ccode>target\u002Frelease\u002Fbrush\u003C\u002Fcode> (217 MB ELF、ldd 動的リンク)。\u003Ccode>--bin brush_app\u003C\u002Fcode> 指定は不要 (workspace default で main binary \u003Ccode>brush\u003C\u002Fcode> が出る)。",{"type":57,"level":109,"text":67},{"type":98,"lang":99,"text":161},"$ ~\u002Frepos\u002Fbrush\u002Ftarget\u002Frelease\u002Fbrush --version\nbrush-cli 0.3.0\n\n$ ~\u002Frepos\u002Fbrush\u002Ftarget\u002Frelease\u002Fbrush --help | head\nBrush - universal splats\n\nUsage: brush [OPTIONS] [PATH_OR_URL]\n\nArguments:\n  [PATH_OR_URL]  Source to load from (path or URL)\n\nOptions:\n      --with-viewer  Spawn a viewer to visualize the training\n  -h, --help         Print help\n  -V, --version      Print version\n...\n\n$ cd ~\u002Frepos\u002Fbrush && git rev-parse HEAD\nce6ef3f8e4c03c231020ebf8049e5c19259a2923\n",{"type":102,"text":163},"\u003Cstrong>OK\u003C\u002Fstrong>。HEAD \u003Ccode>ce6ef3f\u003C\u002Fcode> (2026-05-xx 時点 main)。\u003Ccode>--with-viewer\u003C\u002Fcode> 省略すれば server (CLI training) mode。",{"type":57,"level":109,"text":165},"GPU 選択 (Phase 2 で要設定)",{"type":102,"text":167},"wgpu は \u003Ccode>WGPU_BACKEND\u003C\u002Fcode> env (vulkan \u002F dx12 \u002F metal \u002F gl) と \u003Ccode>WGPU_ADAPTER_NAME\u003C\u002Fcode> \u002F \u003Ccode>WGPU_POWER_PREF\u003C\u002Fcode> でバックエンド・物理 GPU を選ぶ。c33 で 3 枚ある A6000 から特定の GPU を指定する場合、Phase 2 で \u003Ccode>WGPU_BACKEND=vulkan WGPU_ADAPTER_NAME=\"NVIDIA RTX A6000\"\u003C\u002Fcode> あたりを試す (詳細は Phase 2 で確認)。",{"type":57,"text":169},"6. ディスク使用量 (setup 完了時)",{"type":98,"lang":99,"text":171},"~\u002Fminiconda3            12 G   ← gsplat-env + orig3dgs-env (それぞれ torch + cu121 wheels)\n~\u002Frepos\u002Fbrush          3.9 G   ← incl. target\u002Frelease (debug 含まず)\n~\u002Frepos\u002Fgaussian-splatting  308 M\n~\u002F.cargo                954 M   ← cargo registry\n~\u002F.rustup               1.4 G   ← rustc toolchain\n合計                     ~18.6 G\n",{"type":102,"text":173},"NFS 全体は 55T \u002F 11T 空き (Use 80%)。当面問題なし。",{"type":57,"text":175},"7. リモートファイル状態 (Phase 2 へ引き継ぐ参照)",{"type":98,"lang":99,"text":177},"\u002Fhome\u002Fotake_26\u002F\n├── .bashrc                       (conda init + CUDA toolkit block 追加済)\n├── miniconda3\u002F                   (base + gsplat-env + orig3dgs-env)\n├── .cargo\u002F, .rustup\u002F             (Rust stable toolchain)\n├── repos\u002F\n│   ├── brush\u002F                    (HEAD ce6ef3f8e4c03c231020ebf8049e5c19259a2923)\n│   │   └── target\u002Frelease\u002Fbrush  (217 MB, brush-cli 0.3.0)\n│   └── gaussian-splatting\u002F       (graphdeco-inria, --recursive で submodule clone 済)\n│       └── submodules\u002F{diff-gaussian-rasterization, simple-knn, fused-ssim}\u002F  (build 済、pip にも install 済)\n├── setup_gsplat.sh, setup_orig3dgs_resume.sh, setup_brush.sh   (再現用 script)\n├── gsplat_verify.log, orig_verify.log, brush_verify.log         (動作確認 output)\n└── *_done.flag                   (setup 完了マーカー、すべて POST_RC=0)\n",{"type":57,"text":179},"8. Phase 2 への残課題",{"type":57,"level":109,"text":181},"8.1 必須",{"type":90,"ordered":183,"items":184},true,[185,186,187,188],"\u003Cstrong>NeRF Synthetic Lego データセット c33 配置\u003C\u002Fstrong>: 開発機 m3 mac (\u003Ccode>~\u002Fdev\u002F3dgs-workspace\u002Fdatasets\u002Fnerf_synthetic\u002Flego\u003C\u002Fcode>) にしか無いので、\u003Ccode>rsync\u003C\u002Fcode> で c33 に。サイズ ~300 MB なので NFS で問題なし。","\u003Cstrong>GPU 占有予約\u003C\u002Fstrong>: setup 時点で 3 枚とも 80-93% Util の他ユーザ job 走行中。Phase 2 の training run は他ユーザの状況を見つつ、空いた GPU を \u003Ccode>CUDA_VISIBLE_DEVICES\u003C\u002Fcode> で固定する。","\u003Cstrong>wgpu バックエンド確認\u003C\u002Fstrong>: brush の Linux + NVIDIA 環境で \u003Ccode>WGPU_BACKEND=vulkan\u003C\u002Fcode> がデフォルトで通るか、\u003Ccode>vulkaninfo\u003C\u002Fcode> の有無を含めて初回 run で確認。","\u003Cstrong>再現性条件の整理\u003C\u002Fstrong>: 原著 3DGS は paper の Lego baseline (~35.78 PSNR、Kerbl 2023) と一致するかは hyperparameter (iter=30000 既定、\u003Ccode>--white_background\u003C\u002Fcode> 必要) に注意。gsplat は既定 strategy 設定を gsplat docs で参照、本リポの brush との直接比較は loss\u002Frefine 仕様の差を別途記録。brush は c33 (Vulkan + A6000) と m3 mac (Metal + M4 Max) で同 dataset \u002F iter \u002F config を回す。",{"type":57,"level":109,"text":190},"8.2 任意 \u002F 推奨",{"type":90,"items":192},[193,194,195,196],"\u003Cstrong>conda env config vars\u003C\u002Fstrong>: per-env で \u003Ccode>CUDA_HOME\u003C\u002Fcode> を pin しておくと c32 \u002F c34 への移行時に CUDA path 衝突を避けられる (例: \u003Ccode>conda env config vars set CUDA_HOME=\u002Fusr\u002Flocal\u002Fcuda-12.1 -n orig3dgs-env\u003C\u002Fcode>)。","\u003Cstrong>\u003Ccode>tmux\u003C\u002Fcode> の常用\u003C\u002Fstrong>: setup でも有効だったが、Phase 2 の 30k run (推定 ~15-30 min\u002Frun) も tmux session で分離して、SSH 切断耐性を確保。","\u003Cstrong>GPU メモリ予算\u003C\u002Fstrong>: A6000 48GB、Lego 30k は 2-4GB なので 1 GPU を他ユーザと共有しつつも余裕あり。ただし他ユーザ job との衝突に注意 (\u003Ccode>nvidia-smi\u003C\u002Fcode> で確認後 \u003Ccode>CUDA_VISIBLE_DEVICES\u003C\u002Fcode> 設定)。","\u003Cstrong>brush の \u003Ccode>--no-default-features\u003C\u002Fcode> 等の調査\u003C\u002Fstrong>: 今回 default features で build 通った (UI deps fontconfig\u002Fxkbcommon は X11 系なしでも crate level fallback で成立)。training-only バイナリ最小化は Phase 2 でも必要なら検討。",{"type":57,"level":109,"text":198},"8.3 (将来) wgpu バックエンド切替実験",{"type":102,"text":200},"第 2 軸 (wgpu 抽象コスト) の本丸として、c33 で \u003Cstrong>wgpu→Vulkan\u003C\u002Fstrong> と \u003Cstrong>CUDA native (原著 \u002F gsplat)\u003C\u002Fstrong> を同 hardware で比較する。Phase 2 のメイン deliverable は同じ Lego 30k を 3 実装で回して \u003Ccode>wallclock \u002F PSNR \u002F VRAM\u003C\u002Fcode> を表に揃えること。",{"type":57,"text":202},"9. setup スクリプト保管",{"type":102,"text":204},"リモート (\u003Ccode>~\u002Fsetup_gsplat.sh\u003C\u002Fcode>、\u003Ccode>~\u002Fsetup_orig3dgs_resume.sh\u003C\u002Fcode>、\u003Ccode>~\u002Fsetup_brush.sh\u003C\u002Fcode>) に残してあるので、c32 \u002F c34 で再実行する際はそのまま使える。ただし NFS 共有ホームで conda env も共有されるため、c32 \u002F c34 でも同 conda env を \u003Ccode>conda activate\u003C\u002Fcode> するだけで OK (再 install 不要)。",[],[207,224],{"id":28,"title":208,"date":209,"status":10,"polarity":210,"category":11,"axes":211,"tags":212,"task_code":26,"related_runs":217,"delta_psnr":220,"delta_wallclock":221,"rank":32,"verdict":33,"impact_summary":222,"detail_path":223},"c32 V100 gsplat smoke — NFS 共有 env を異 sm 機へ持ち込み JIT 再 build 1 回で動作確認","2026-05-23","positive",[14],[16,20,213,214,17,215,25,216],"v100","c32","smoke","jit",[218,219],"gsplat-lego-1k-smoke","gsplat-lego-50-dryrun",19.81,"10.5s \u002F 1k step","NFS 共有 gsplat-env を異 sm 機 (c33 sm_86 → c32 sm_70) に持ち込み、TORCH_CUDA_ARCH_LIST=7.0 で JIT 再 build 1 回 (93s) → 即動作。Lego 1k iter で wallclock 10.5s \u002F val PSNR 19.81 dB。30k full は Phase 2b。","\u002Ffindings\u002Fc32-gsplat-smoke\u002F",{"id":29,"title":225,"date":209,"status":10,"polarity":226,"category":227,"axes":228,"tags":229,"task_code":26,"related_runs":233,"delta_psnr":236,"delta_wallclock":237,"rank":238,"verdict":239,"impact_summary":240,"detail_path":241},"c32 V100 原著 3DGS 30k bench — A.5 三層対比表の最終 row & eval convention 乖離 finding","mixed","experiment",[14],[16,21,213,214,17,230,231,232],"bench","eval-convention","abstraction-cost",[234,235],"orig3dgs-lego-1k-smoke","orig3dgs-lego-30k",28.384,"10m37s","high","investigative","原著 3DGS を V100 で 30k 学習 (PSNR 28.38 dB \u002F 10m37s \u002F 237k splats)。同 V100・同 30k で brush (wgpu→Vulkan) 8m24s \u002F 37.46 dB を上回れず、抽象コスト ≪ 実装最適化レベル を CUDA 機でも再確認。さらに codebase eval と paper-standard eval で 12 dB 乖離 (28.4 vs 14.6) を発見、A.5 表は eval convention 注記必須。","\u002Ffindings\u002Fc32-orig3dgs-bench\u002F",1782449788627]