ViperEkura
4244df2785
perf: pure FP8 fwd/bwd and lean non-transposed GEMM
...
- drop the fused kernel; forward/backward are quantize + a pre-quantized GEMM
- rename module fp8_mm -> fp8_ops (mm.cu -> ops.cu)
- kernels/launchers fp8_gemm_kernel / launch_fp8_gemm; drop PqTraits/gather_trans/pack_fp8x4_vector
- remove the in-kernel transposed-operand branches (TransA/TransB)
- backward: quantize g once (amax_g here), explicit fp8 transposes, fast non-transposed GEMMs (dX = g@w^T, dW = g^T@x^T)
- each pass uses a single FP8 format (E4M3 fwd / E5M2 bwd)
2026-08-23 15:38:30 +08:00
ViperEkura
16a55bb474
refactor: reorganize CUDA kernels into per-family directories
...
- move attention kernels to csrc/kernels/attention/ and rotary to rotary/
- add shared common/mma.cuh (mma_sync, ldmatrix) and device.cuh (sm checks)
- split fp8_mm into three-layer fp8/common.h, gemm.cuh, mm.cu
- fix fused FP8 GEMM ldmatrix lane indexing to fix OOB shared reads
- update extension ops, loader, and kernel tests
2026-08-22 20:40:31 +08:00
ViperEkura
6ac3b51496
refactor: separate extension ops and backends
2026-08-16 21:15:52 +08:00
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
ViperEkura
0378e62e17
refactor: split fp8 into fp8_ops adapter and fp8 policy module
...
- fp8_ops is the only module touching the pybind (kernel interface)
- fp8.py keeps scaling state, delayed amax and aten::linear dispatch
- remove circular imports between old fp8_ops/fp8_state/fp8_dispatch
2026-08-14 12:25:29 +08:00
ViperEkura