biondizzle

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

2026-05-17 08:22:53 +00:00

7073daaffa fix: allocate token_indices on CPU, move to GPU AFTER JIT compilation

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

2026-05-17 08:22:14 +00:00

0e7b06b55c debug: clone + sync token indices before JIT

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

2026-05-17 08:20:42 +00:00

70c0618361 fix: allocate token_indices before CuTeDSL JIT compilation

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

2026-05-17 08:19:47 +00:00

2bbe04efd8 debug: remove assert, test token corruption

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

2026-05-17 08:18:38 +00:00

66627926c5 debug: int32 token indices with sync verify

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

2026-05-17 08:16:11 +00:00

da02a5dc11 debug: assert token indices are correct after allocation

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

2026-05-17 08:11:03 +00:00

c0d016a472 feat: compute_activation_global_scales warmup method

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

2026-05-17 08:07:12 +00:00

8c9a51e006 fix: call _ensure_stacked in warmup test

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

2026-05-17 08:06:34 +00:00

5ba77e355f test: warmup gs computation with safety margin sweep

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

2026-05-17 07:59:02 +00:00

ae6b879d38 fix: pass expert_offsets without leading 0 to GEMM (matches pipeline)

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

2026-05-17 07:57:02 +00:00

a1e6f5f891 fix: searchsorted right=True for correct expert assignment

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

2026-05-17 07:53:59 +00:00

ddffb7d8df docs: current bug analysis — scale_a layout vs expert_offsets mismatch

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

2026-05-17 07:47:16 +00:00

ed90341ea9 fix: scatter+per-expert-swizzle scale assembly (cudagraph-safe)

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

2026-05-17 07:43:11 +00:00

37fecb588f fix: separate L1/L2 scale buffers (different K_sf), fix assembly calls

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

2026-05-17 07:39:51 +00:00

b824b838a9 fix: 128-row-align each expert's scales in padded buffer

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

2026-05-17 07:37:48 +00:00

8dadd9a723 test: scale assembly debug

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

2026-05-17 07:37:05 +00:00

8642946274 fix: padded x_sf buffer for fixed-shape scale assembly

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

2026-05-17 07:35:50 +00:00

418e29f7f5 fix: per-expert scale assembly (match assemble_scales_2d_side)

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

2026-05-17 07:33:21 +00:00

7b95e76723 test: runner vs pipeline comparison + scale assembly comparison

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

2026-05-17 07:15:01 +00:00

366a0240a5 vllm tweaks