biondizzle

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

2026-06-01 01:27:19 +00:00

7fbbdc5204 diag: validate router output before MoE

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

2026-06-01 01:26:53 +00:00

f5fa84016e diag: sync+error check after each layer on first token

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

2026-06-01 01:14:41 +00:00

91b3929605 fix: call moe_runner.run() and se_runner.run() (not __call__)

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

2026-06-01 01:08:05 +00:00

03c45d4bfb fix: pass int32 token_ids to hash router (was int64)

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

2026-06-01 01:01:16 +00:00

62efde5c9f fix: router — use cuBLAS BF16 GEMM + activation_topk CUDA kernel (production path, not CuTeDSL fused)

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

2026-06-01 00:59:21 +00:00

5591a725e1 fix: router kernel — infer OperandMajorMode from tensor layout (same pattern as MoE GEMM)

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

2026-06-01 00:56:02 +00:00

0ab5d8c317 fix: disable broken CuTeDSL fused router — use BF16 linear + activation_topk (both are production paths)

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

2026-06-01 00:54:20 +00:00

c339fe7ad9 fix: router A operand major mode MN (not K) — fixes CuTeDSL local_tile coord error

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

2026-06-01 00:42:13 +00:00

b7a8c44d26 single_shot: eager MoE/SE weight processing, stale GPU cleanup, --prefill-tokens flag

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

2026-06-01 00:32:48 +00:00

15f45b57c3 fix: correct Nvfp4Linear dimension inference from checkpoint weights

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

2026-06-01 00:30:40 +00:00

e671780008 fix: transpose checkpoint weights before make_b_k_major in Nvfp4Linear/SharedExpert

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

2026-06-01 00:18:37 +00:00

e8a7a9256f fix: convert uint8 checkpoint weights to float4_e2m1fn_x2 for CuTeDSL GEMM

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

2026-06-01 00:10:54 +00:00

172448514c fix: fold weight_scale_2 into global_scale_b for NVFP4 GEMM

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

2026-06-01 00:04:50 +00:00

563df02aef fix: import SF_VEC_SIZE from quantize in gemm_runner (was NameError)

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

2026-06-01 00:00:09 +00:00

be476b2ce2 router: catch CuTeDSL warmup failures fast, don't let MLIR errors slow down init

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

2026-05-31 23:55:18 +00:00

56dff8d185 fix: W_gate is (H, E) but F.linear expects (E, H), transpose before linear

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

2026-05-31 23:54:17 +00:00

5396a04c28 router: broaden except to catch all CuTeDSL errors, fall through to cuBLAS+activation_topk path

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

2026-05-31 23:53:15 +00:00

3b5b9f487c fix: compute num_tma_load_bytes inside cute.compile context

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

2026-05-31 23:51:44 +00:00

1bc0da0f35 fix: properly scope swap code inside else/guard blocks, replace continue with if guard

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

2026-05-31 23:50:41 +00:00

d0d765e1f2 fix: replace break statements with flag-based loops in router kernel (CuTeDSL restriction)