platform · inferred from evidence
CUDA
NVIDIA's parallel computing platform and API for GPU acceleration.
- 运行失败：python -m pip install git+https://github.com/Dao-AILab/flash-attention.git@v2.5.9.post1
flash-attention failing to build from git: skip the source build and install a prebuilt wheel matching your setup. Check your Python version, CUDA version (nvidia-smi), and CPython tag, then download the matching wheel from the flash-attent
- AssertionError: Torch not compiled with CUDA enabled - macOS Sequoia 15.1.1 (intel, AMD)
How to assertionError: Torch not compiled with CUDA enabled - macOS Sequoia 15.1.1 (intel, AMD). Verified in gh:Tencent-Hunyuan/HunyuanVideo#22.
- RuntimeError: Deserialization Fails on CPU-Only Systems Due to CUDA Mapping in torch.load
Teaches how to runtimeError: Deserialization Fails on CPU-Only Systems Due to CUDA Mapping in torch.load. Based on a real issue report and its verified fix.
- Diagnose and fix PyTorch CUDA OOM and allocator fragmentation
Diagnose a PyTorch CUDA out-of-memory crash: classify genuine OOM vs allocator fragmentation vs outside-allocator memory using allocated/reserved/free counters, capture peak-memory snapshots, and apply the right fix (batch/checkpointing/amp
- Diagnose and fix NaN or stalled loss with fastai mixed precision (learn.to_fp16())
Diagnose fastai mixed-precision training failures: classify early-NaN (fp16 forward overflow) vs mid-training NaN (chronic gradient overflow) vs plateau above the fp32 baseline (gradient underflow / skipped steps) using the learn.scales los