ViperEkura 76aa4edc9f perf: skip grad-bias reduce for bias-free linears ...

- _LinearFp8.backward computed g2.sum(0) unconditionally and dropped it when needs_input_grad[2] was false; now the column-sum only runs when the bias actually requires grad
- saves ~327 reduce kernels per train step on bias-free LLMs (215M GQA: end-to-end 1.08x -> 1.13x vs bf16)

2026-08-26 18:51:18 +08:00

..

backend

refactor: rework attention backend resolution

2026-08-23 14:47:02 +08:00

lib

refactor: move compiled kernel .so files into extension/lib

2026-07-31 15:36:32 +08:00

ops

perf: fast interior loop on the big cta and fused epilogue bias

2026-08-26 14:52:06 +08:00

__init__.py

refactor: separate extension ops and backends

2026-08-16 21:15:52 +08:00

fp8.py

perf: skip grad-bias reduce for bias-free linears

2026-08-26 18:51:18 +08:00

loader.py

refactor: unify kernel module loading and packaging

2026-08-23 15:57:04 +08:00