biondizzle

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:45:54 +00:00

8ee3f90e44 debug: handle flat_rank=8 for SF remap, add coordinate dump

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:34:48 +00:00

d2c1c76f5b debug: idx2crd+flatten approach with printf to determine flat_rank

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:32:13 +00:00

2ac3a7d631 fix: construct nested coordinate for CuTe layout shape ((32,4), K)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:31:06 +00:00

593ae998f8 fix: clean rewrite of cutlass_nvfp4_gemm.cu — no more file splicing

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:28:47 +00:00

196ee37fdb fix: rewrite SF remap kernel — source-iterating with layout_sf(m, k_elem)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:24:19 +00:00

fb390b24e2 debug: add printf to SF remap kernel to check flat_rank and layout shape

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:04:44 +00:00

8f5322ca31 fix: add missing extern "C" opening brace lost during file reconstruction

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 15:01:49 +00:00

a8bd962452 fix: SF remap — iterate dest indices, extract logical (m, k_sf) from nested coord

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 14:57:15 +00:00

395cc31883 fix: use layout_sf(m, k_elem) instead of make_coord for nested shapes

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 14:54:20 +00:00

d90967d6e9 fix: SF remap — element-space K coords + zero-init dest buffer

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 14:51:03 +00:00

5968ebad9f fix: SF remap was using idx2crd+flatten which gives atom sub-indices, not logical (m,k)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 14:19:41 +00:00

cf796e37cf debug: add weight_scale_2 shape/value logging in weight transform

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 14:17:44 +00:00

879adc324d fix: _fold_global_scale — remove broken logical_widths branch

biondizzle pushed to modelopt-nvfp4 at biondizzle/deepseek-v4-quant

2026-05-14 14:13:21 +00:00

f2656dcf6d sync B200 deployment files: Dockerfile, docker-compose, patches

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 13:44:43 +00:00

ef9cd023a9 fix: unpack_ue4m3_u32 — uint32 lacks CUDA bitwise ops, use int32

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 13:05:49 +00:00

1c39e21d87 fix: remove broken L1 weight interleave

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 13:04:33 +00:00

80495c0cd6 docs: clarify SF layout remap is in CUDA, not sf_layout.py

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 13:02:12 +00:00

16f91ff0e1 fix: rewrite stage_activation with proper E2M1 quantization

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 12:59:23 +00:00

3bcc0ac057 fix: unpack_ue4m3_u32 was value-casting instead of bit-reinterpreting

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel

2026-05-14 12:48:09 +00:00

8b7fa0c91e add README: pipeline diagram, file map, data formats, known issues