flashrt-vocab-ce-train
Streaming huge-vocab CE training API skeleton with eager autograd reference.
Available functions
vocab_ce_lossvocab_ce_fwd
Acceptance status
- Reference/eager autograd path: available.
- CUDA optimized implementation: pending acceptance.
- Precision mode: fp32/bf16 training reference only, no FP8/FP4.
Use this package to lock API integration and run correctness harnesses before replacing the internal reference path with optimized CUDA kernels.
- Downloads last month
- 12
Supported hardwares new
CUDA
- OS
- linux
- Arch
- x86_64





