VoxCPM2 inductor compile cache for Kaggle T4
torch.compiler.save_cache_artifacts() output for openbmb/VoxCPM2 with
torch.compile(mode="reduce-overhead", fullgraph=True) applied to the fixed-shape feat_encoder (LocEnc) and
feat_decoder.estimator (LocDiT) in float16, as done by the wfly-spring Kaggle VoxCPM2 service.
Loading it with torch.compiler.load_cache_artifacts() before compiling cuts warm-up compilation from ~31s to ~7s.
| Directory | torch | triton | GPU | voxcpm |
|---|---|---|---|---|
torch-2.11.0-cu128_triton-3.6.0_sm75/ |
2.11.0+cu128 | 3.6.0 | NVIDIA Tesla T4 (sm_75) | 2.0.3 |
One file per GPU index (compile_cache_cuda0.bin, compile_cache_cuda1.bin): inductor cache keys include the
device index, so a cache produced on cuda:0 misses on cuda:1.
The cache is only valid for the exact torch / triton / GPU architecture above. A mismatched environment should fall back
to compiling from scratch (the service looks up torch-{torch.__version__}_triton-{triton.__version__}_sm{major}{minor}
with + replaced by -, file compile_cache_cuda{index}.bin, and compiles normally when it is missing).
Generated 2026-10-23 on Kaggle (Python 3.13). Contains only compiled Triton kernels / FX graphs, no weights.