float4_e2m1fn_x2 packs 2 values per byte along K, not N.
The GEMM output N dimension is the logical N from mat_b.shape[2],
not 2x packed. Previous n_dim*2 was wrong — it accidentally worked
in the test because intermediate_size*2 == 2*intermediate_size.
Real model with N=9216 exposed the bug.
- assemble_activation_scales_gpu: builds padded+swizzled scale tensor
without .item() or .tolist() CPU syncs. Uses GPU index arange + cat
+ single scatter instead of per-expert Python slicing.
- Still has a for e in range(num_experts) loop but num_experts is
compile-time constant so torch.compile unrolls it.
- Added tests/cudagraph_test.py: attempts CUDA graph capture on the
MoE runner, diagnoses sync violations with patched torch functions.
- Removed the if total_slots == 0 early return (Python control flow
on GPU data)
The bridge's assemble_scales_3d_side expects (K_sf, N) input and
transposes to (N, K_sf) internally before swizzling. The checkpoint
stores scales as (N, K_sf). Without this transpose, the kernel was
reading completely wrong scale data — cosine dropped to 0.713.
Also fixed dual global scale normalization: after transpose, gate/up
are along dim 1 (columns), not dim 0 (rows).
finalize_weights() now view-casts checkpoint uint8 → float4_e2m1fn_x2
directly. Block scales (float8_e4m3fn) and global scales (float32)
pass through unchanged. Zero precision loss on the weights themselves.
L1 dual global scale handling: gate and up have different global scales.
Normalize to max(gate_gs, up_gs) and fold the ratio into block scales
via float32 (one multiply + float8 round-trip on the RATIO only —
much better than dequantizing the entire weight matrix).
layertest.py: updated to test direct path. Expect cosine improvement
from 0.989 → 0.995+ (matching the L1-only result).
The layertest dequantizes checkpoint NVFP4→BF16 then re-quantizes
BF16→NVFP4. This double quantization costs ~1% cosine. The kernel
itself is correct — the 0.989 cosine is expected quantization noise.
cutedsl/moe_pipeline.py: complete pipeline
- stage_activation: BF16 → NVFP4 (keeps data in FP4)
- L1 GEMM: NVFP4 × NVFP4 → BF16 (gate+up)
- SiLU(gate) * up: BF16 (only nonlinear, can't avoid)
- Re-quantize: BF16 → NVFP4 (back to native)
- L2 GEMM: NVFP4 × NVFP4 → BF16 (down_proj)
- Scatter with routing weights → BF16 output
layertest.py: now tests the FULL MoE pipeline against BF16 reference.
NVFP4-native: both GEMMs use float4_e2m1fn_x2 for A and B,
float8_e4m3fn for block scales, float32 for global scales.
BF16 only for SiLU activation and final scatter.
Tokens must be laid out as [expert0_tokens | expert1_tokens | ...]
for the 2Dx3D grouped GEMM. Each expert gets its own contiguous
block of tokens. Scale factors split by expert offsets.
