biondizzle
0 Followers · 0 Following
  • Joined on 2025-12-10
Repositories
25
 Projects Packages Public Activity Starred Repositories
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 06:10:04 +00:00
26da6d33af Fix graph replay: remove extra token_id arg from forward_attention call
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 05:58:27 +00:00
ae26f6b83c Fix dense router BF16 dispatch: use torch.matmul instead of F.linear
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 05:50:21 +00:00
e46b615873 Fix dense router BF16 dispatch for CUDA graph capture
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 05:13:54 +00:00
b4a59d0940 Update CUDA graph docs with current status, A/B split, buffer fixes, remaining blockers
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 04:49:10 +00:00
ffa7842b58 Fix dense router: run GEMM in BF16, convert to FP32 only for activation
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 04:32:02 +00:00
119e6d471e Add safety check for swizzled buffers: fall through to Python path if None
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 04:13:43 +00:00
fae61d3ef7 Add c10/cuda/CUDAStream.h include for getCurrentCUDAStream
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 03:58:01 +00:00
ee86969f6c Fix CUDA stream: use c10::cuda::getCurrentCUDAStream() directly in kernel launch
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 03:43:08 +00:00
e26c28a1ce Fix CUDA stream API: getCurrentCUDAStream().stream()
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 03:29:13 +00:00
9b3917e248 Fix blackwell_swizzle.cu: add pybind11 bindings for torch extension loader
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 03:14:29 +00:00
5487a58df4 Fix NameError: add rows/cols variables to MoE swizzle
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 03:03:04 +00:00
a434545d12 Blackwell swizzle CUDA kernel for CUDA graph capture
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 02:42:02 +00:00
e7766254b7 Pre-allocate ALL GEMM output buffers for CUDA graph capture
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 02:06:20 +00:00
676a0448c0 CRITICAL FIX: _l1_out_buf was 2x too narrow — caused GPU memory corruption
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 01:49:14 +00:00
0890e578f4 DEBUG: print l1_out shape before gate/up split
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 01:38:26 +00:00
8546ed725f DEBUG: check SE input magnitude
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 01:30:31 +00:00
26ecf96328 DEBUG: check intermediate magnitude before SE L2
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 01:27:28 +00:00
5303d6a82f DEBUG: test copy_ with contiguous slice vs scalar assign for gsa
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 01:16:40 +00:00
ccbc713658 DEBUG: check gsa values and pinpoint exact failing operation
biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-06-04 01:05:49 +00:00
e77455c3ba DEBUG: add sync inside quantize_nvfp4_gpu_fused to catch async errors