Download t4_assets/compile_cache_note.json from tchbcb/qwenjev: direct link, hf CLI and curl.
- Browser
- Download file 2.17 kB
-
https://huggingface.co/tchbcb/qwenjev/resolve/main/t4_assets/compile_cache_note.json
- Command line
-
hf download hf://tchbcb/qwenjev/t4_assets/compile_cache_note.json
-
curl -L -o compile_cache_note.json https://huggingface.co/tchbcb/qwenjev/resolve/main/t4_assets/compile_cache_note.json
2.17 kB
| { | |
| "audit_version": "v2 (re-audited 2026-09-22 after BOTH runs: smoke + scale-up)", | |
| "conclusion": "No triton/torch_extensions JIT caches exist; exactly ONE NVRTC kernel cache was produced (/root/.cache/torch/kernels/, cc7.5 + cu128 prod-reduction) and is captured in compile_cache/torch_kernels_t4.tar.gz. The initial v1 note ('zero JIT caches') missed it because it only searched for torch_extensions/triton directory names.", | |
| "searched": [ | |
| "/root/.cache", | |
| "/content", | |
| "/root/.julia (ptx matches only)" | |
| ], | |
| "found": { | |
| "torch_extensions": "absent (nothing calls torch.compile; transformers 5.16 eager path)", | |
| "triton_cache": "absent (no triton kernels JIT-compiled; T4 cc7.5 used prebuilt kernels only)", | |
| "torch_kernels_nvrtc": "PRESENT: ~/.cache/torch/kernels/reduction_prod_arch7.5_nvrtc12.8_sass_30566_aa1ebe12288bb269e63981021cb5935eca5f1576 (64KB; captured as compile_cache/torch_kernels_t4.tar.gz, 12857 bytes; restore with: tar xzf torch_kernels_t4.tar.gz -C ~/.cache/)", | |
| "julia_ptx": "16 .ptx files under /root/.julia = VM-preinstalled Julia CUDA.jl examples, unrelated to this pipeline" | |
| }, | |
| "prebuilt_cuda_assets_reused_instead": { | |
| "torch": "2.11.0+cu128 (PyPI default wheel ships cu128; pip install torch==2.11.0)", | |
| "bitsandbytes": "0.50.2 prebuilt manylinux_2_24 cu12 wheel -> t4_assets/wheels/", | |
| "triton": "3.8.0 wheel -> t4_assets/wheels/triton-3.8.0-cp313-*.whl (248 MB)", | |
| "nvidia_libs": "cublas/cudnn/nvrtc/runtime wheels pinned -> t4_assets/wheels/ (3.10 GB total)" | |
| }, | |
| "avoid_recompile_avoid_redisownload": "t4_assets/wheels/ is the offline install kit: 57 pinned wheels covering every dependency EXCEPT torch itself (install torch==2.11.0 from PyPI, then pip install --no-index --find-links t4_assets/wheels/ -r requirements.txt). Wheels target CPython 3.13; other versions: re-run pip download.", | |
| "version_pins": { | |
| "torch": "2.11.0+cu128", | |
| "transformers": "5.16.1", | |
| "peft": "0.20.0", | |
| "accelerate": "1.14.0", | |
| "bitsandbytes": "0.50.2", | |
| "hub": "1.29.0", | |
| "cuda": "12.8", | |
| "capability": [ | |
| 7, | |
| 5 | |
| ], | |
| "gpu": "Tesla T4" | |
| }, | |
| "doc_pointer": "train.md v0.4.0 sections 12.6 (audit) + 13 (offline repro kit)" | |
| } |