biondizzle

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

2026-05-16 20:19:23 +00:00

647c03b2ee fix: make_b_k_major must preserve shape — use double-permute trick

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

2026-05-16 20:18:20 +00:00

ed4f501bba fix: make_b_k_major stride check — K-major means stride[1]==1, not stride[2]==1

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

2026-05-16 20:16:31 +00:00

2162cee4ad fix: restore proper quantize_weight_to_nvfp4 — K is the packed dim, not N

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

2026-05-16 20:09:03 +00:00

10f1dca982 fix: import ceil_div from correct module

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

2026-05-16 20:08:24 +00:00

81632e2f21 fix: correct cutlass_torch import (cutlass.torch, not top-level)

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

2026-05-16 20:06:39 +00:00

16c4fad025 fix: remove cutlass.cute.backend import

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

2026-05-16 20:06:00 +00:00

44b40d41fe fix: compile CuTeDSL kernel with real tensors, not dummy shapes

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

2026-05-16 20:00:37 +00:00

79281b6fda fix: compute K_packed/N_packed before passing to _get_compiled_kernel

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

2026-05-16 19:59:45 +00:00

caf93d6c45 fix: pass K_packed/N_packed to _get_compiled_kernel

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

2026-05-16 19:58:14 +00:00

ecc7b83334 fix: compile CuTeDSL kernel with actual tensor shapes, not dummy 256x256

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

2026-05-16 19:55:21 +00:00

cc75a55bd9 restore: new bridge/moe_pipeline/layertest

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

2026-05-16 19:54:06 +00:00

0c878b3a9e temp: restore old layertest+bridge for cosine comparison

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

2026-05-16 19:38:34 +00:00

0069769d12 debug: print global scales

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

2026-05-16 19:31:56 +00:00

84589fe984 debug: more prints

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

2026-05-16 19:29:43 +00:00

fa2d5708c5 debug: add L1 GEMM and SiLU output debug prints

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

2026-05-16 19:28:24 +00:00

4c06c51ec3 fix: moe_pipeline.py gate/up split — L1 output is 2*intermediate, not intermediate

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

2026-05-16 19:23:45 +00:00

da31ce7e1a allow for cuda graphs again

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

2026-05-16 19:07:37 +00:00

d15c43294b fix: test L2 weight N dim should be hidden_size, not hidden_size//2

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

2026-05-16 19:07:10 +00:00

28788c6f55 fix: L1 weight N dimension is 2*intermediate (gate+up), not intermediate

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

2026-05-16 18:55:49 +00:00

f7e29fdf1e docs: update README with cudagraph compatibility work and decisions