File size: 2,168 Bytes
52e11de
5d63911
 
52e11de
 
5d63911
 
52e11de
 
5d63911
 
 
 
52e11de
 
 
 
 
 
 
5d63911
52e11de
 
 
 
 
 
 
 
 
 
 
 
 
5d63911
 
52e11de
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
{
 "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)"
}