- Fix interleave_l1_weights: remove //2 bug (g=granularity_bf16 for N-axis)
- Apply L1 weight+SF interleave in runner._ensure_stacked() and moe_pipeline
- De-interleave L1 GEMM output before gate/up split
- Fused SwiGLU kernel: epi_tile=(128,8) for subtile-level pairing
- Even subtiles = gate: SiLU in FP32 registers, save to register buffer
- Odd subtiles = up: silu(gate)*up from buffer
- Both branches produce same BF16 tensor type (CuTeDSL constraint)
- run_nvfp4_moe_fused() pipeline: fused L1 + PyTorch L2
- Runner: fused_swiglu=True option for CuTeDSLMoERunner
- Layertest: both fused and non-fused paths PASS (cosine 0.988)
- README.md updated with current status and lessons learned
SiLU in registers: PASS (0.034% error, Step 1 stable)
Gate/up subtile detection: blocked by CuTeDSL type system
CuTeDSL compiles the kernel for ALL subtile iterations at once.
Runtime conditionals (if is_gate_subtile) that affect:
- Register tensor assignment → DSLRuntimeError (type structure mismatch)
- TMA store skipping → corrupted output
- Mask blending → wrong results
Path forward: use const_expr debug flag for the BF16 side output,
or process gate/up in a separate post-GEMM kernel.
Step 1 VALIDATED:
- cute.exp works on register tensors in the epilogue
- SiLU (x / (1+exp(-x))) produces correct results
- Relative error vs PyTorch: 0.034%, max abs: 0.0625 (BF16 precision)
Step 2 (gate/up pairing) approach:
- Register-level pairing requires understanding acc_vec layout from tiled_copy_r2s
- DeepGEMM pattern: (values[0], values[2]) pairs for tcgen05.ld
- CuTeDSL retile may produce different layout than direct PTX loads
- SMEM-level SiLU is a valid intermediate: avoids GMEM round-trip while
working in logical (M, N) coordinate space
- Non-interleaved weights + SMEM SiLU is simplest starting point
Stage 1 of the fused epilogue: applies SiLU (x * sigmoid(x)) to the
full accumulator register tensor before writing BF16 to C.
This validates that cute.exp and element-wise FP32 operations work
on CuTe register tensors in the epilogue. The gate/up pairing is
not yet implemented (Stage 2).
The fused_swiglu flag is const_expr(0) by default, so the standard
epilogue path is unchanged unless the flag is enabled.
Copied from CUTLASS examples (no more runtime dependency on
/root/cutlass/examples/). Fixed all imports to use cutedsl.kernel.*
instead of blackwell.kernel.*.
Structure:
cutedsl/__init__.py
cutedsl/kernel/__init__.py
cutedsl/kernel/moe/ (the MoE scaled grouped GEMM)
cutedsl/kernel/blockscaled_gemm/ (dense blockscaled GEMM)
test_cutedsl.py updated to import from our local copy.
