biondizzle

0 Followers · 0 Following

• Joined on 2025-12-10

Repositories

25

Projects Packages Public Activity Starred Repositories

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:25:22 +00:00

1eb9c43217 Rewrite CUTLASS kernel based on NVIDIA example 72b (nv_float4_t, CollectiveBuilder, OpClassBlockScaledTensorOp)

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:23:03 +00:00

8a9af441dc Fix includes: use cutlass/float_subbyte.h (has float_e2m1_t and float_ue4m3_t), point to latest CUTLASS

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:18:28 +00:00

d789f5e3e0 Add CCCL include path for CUTLASS 3.x

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:17:31 +00:00

12588047fd Fix setup.py: use include_dirs and extra_compile_args (correct PyTorch extension API)

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:14:07 +00:00

1b1c3a42fe Fix setup.py source paths

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:12:48 +00:00

f375c80bfe feat: CUTLASS NVFP4 block-scaled GEMM kernel (native SM100 Blackwell)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:11:17 +00:00

f375c80bfe feat: CUTLASS NVFP4 block-scaled GEMM kernel (native SM100 Blackwell)

56c7880296 Native NVFP4 TileLang kernel: tcgen05 block-scaled MMA

bf13665dbe Implement TileLang NVFP4 mega_moe L1/L2 kernels

ebc0ab0cac Fix: keep scales as float8_e4m3fn, don't pack to uint32 (min_all_cuda unsupported)

94233c4dd3 Fix __init__.py: remove private imports

Compare 6 commits »

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 23:02:08 +00:00

56c7880296 Native NVFP4 TileLang kernel: tcgen05 block-scaled MMA

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 22:37:00 +00:00

bf13665dbe Implement TileLang NVFP4 mega_moe L1/L2 kernels

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 21:54:42 +00:00

ebc0ab0cac Fix: keep scales as float8_e4m3fn, don't pack to uint32 (min_all_cuda unsupported)

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 21:43:48 +00:00

94233c4dd3 Fix __init__.py: remove private imports

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 21:41:45 +00:00

1a452ffabd Fix weight_transform signature to match nightly vLLM finalize_weights call

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 16:11:11 +00:00

47ca5631d8 Fix __init__.py: only import from package modules

c2b752c2fe Initial: TileLang NVFP4 mega_moe kernel package

Compare 2 commits »

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-13 16:08:39 +00:00

47ca5631d8 Fix __init__.py: only import from package modules

biondizzle created branch master in biondizzle/nvfp4-megamoe-kernel 2026-05-13 15:44:58 +00:00

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-13 15:44:58 +00:00

c2b752c2fe Initial: TileLang NVFP4 mega_moe kernel package

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 15:22:36 +00:00

673f67681f Add vLLM integration layer and packaging

biondizzle created branch main in biondizzle/nvfp4-megamoe-kernel 2026-05-13 14:51:14 +00:00

biondizzle pushed to main at biondizzle/nvfp4-megamoe-kernel 2026-05-13 14:51:14 +00:00

a4b90b5780 Initial commit: NVFP4 mega_moe kernel in TileLang

biondizzle created repository biondizzle/nvfp4-megamoe-kernel 2026-05-13 14:49:50 +00:00