ViperEkura 69f35c46e0 fix: quantize amax from raw values, not scaled fp8 values ...

- amax for delayed scale was the quantized max (always ~448), so scale collapsed to 1
- this made fp8 gradients diverge (cosine 0.05) and training stall
- stop w/x transpose-quantize amax from polluting the grad scale

2026-08-14 14:26:16 +08:00

..

kernels

fix: quantize amax from raw values, not scaled fp8 values

2026-08-14 14:26:16 +08:00

tests

perf: flatten paged prefill tile dispatch

2026-08-09 23:12:53 +08:00

__init__.py

feat: add optional CUDA kernel system (csrc/) + fused GQA decode attention

2026-07-06 12:09:58 +08:00

CMakeLists.txt

feat: add fp8 training via cublasLt dispatch

2026-08-14 00:39:49 +08:00