File size: 2,336 Bytes
6abc190 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 | # Source Sync
- Upstream FlashRT source: `../official/FlashRT`
- Original SM110 sync commit: `132049d7c3a3534fb7d35676cd726f39408b1af6`
- GROOT N1.7 fused-epilogue sync commit:
`24df793f4fa2d50780aea03b644208c6e0cb4162`
- Initial package date: June 20, 2026
Copied source files:
- `csrc/gemm/fp4/cutlass_nvfp4_w4a16_gemm_sm120.cu/.cuh`
- `csrc/gemm/fp4/cutlass_nvfp4_w4a16_gemm_sm100.cu/.cuh`
- `csrc/gemm/fp4/cutlass_fp4_gemm_bias_bf16_sm100.cu/.cuh`
- `csrc/quantize/quantize_fp4_sfa.cu/.cuh`
- `csrc/quantize/quantize_fp4_sfa_bf16.cu/.cuh`
- `cutlass/util/packed_stride.hpp`, copied from CUTLASS tools util headers
into `csrc/cutlass/util/packed_stride.hpp` so the Hub package does not
depend on a local `third_party/cutlass/tools/util/include` path.
Packaging helper:
- `csrc/dequantize_fp4_sfa.cu/.cuh` derived from the SFA dequant validation
helper used in `fp4-fused-ops`; this package adds `is_sfb` support so tests
can dequant both A/SFA and B/SFB.
Local packaging edits:
- Added Tensor-facing PyTorch custom ops in `torch-ext/torch_binding.cpp`.
- Added Python wrappers and fake registrations in `torch-ext/fp4_gemm`.
- Added the BF16 direct SFA/SFB producer as an input-type specialization of
the existing FP16 producer. Its E2M1 encoding and CUTLASS scale layout are
unchanged; the additive entry removes a standalone activation cast.
- Public APIs accept CUDA tensors only; no raw pointers or stream arguments.
- CUTLASS SM100/SM120 block-scaled layout support is treated as package scope,
not as a test-only compiler define.
- The Tensor binding dispatches the canonical BF16-output GEMM by runtime
compute capability. CUDA 12.8 artifacts link only the SM120 implementation;
CUDA 13 artifacts link both SM110 and SM120 implementations.
- SM110 production auto-dispatch was tiled against PI0.5, GROOT, Cosmos Edge,
and LingBot VLA projection shapes. Explicit schedule IDs remain diagnostic.
Architecture limits:
- The canonical BF16-output GEMM and SFA/SFB helpers support SM110 and SM120.
- Bias, residual, and bias/GELU-to-FP4 epilogues have an independent SM110
backend copied from the production GROOT N1.7 path. Stream-K remains an
SM120-only API and rejects on SM110.
- SM110 requires CUDA 13 and the package's pinned CUTLASS 4.4 target; SM120
requires CUDA 12.8 and CUTLASS 4.0.
|