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 2026-05-15 06:07:21 +00:00

311b28bd9f fixey wixey

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 06:01:54 +00:00

685bce48b4 actually handle expert param mapping

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 05:48:42 +00:00

f17efa340d are the weights ever not zero?

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 05:41:15 +00:00

c5d800f133 can we see the wt in?

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 05:31:17 +00:00

6a4f52cedc god dam i just want the gemm in

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 05:21:53 +00:00

3b3c506af5 whoops

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 05:08:58 +00:00

76e9b078a2 more debug2

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 04:53:29 +00:00

912e4622d7 more debug

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 04:35:49 +00:00

c7f6a1dc4d fix: transpose B and SFB on the Python side at weight-load time, and adjust the SFB remap kernel to read from column-major source layout

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 04:30:32 +00:00

c56cc34ae1 fix: LayoutBTag is now RowMajor

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 03:59:08 +00:00

9975558c23 Add always-on alpha/x_sf debug prints for L1 and L2 GEMM calls

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 03:52:03 +00:00

9c318c3353 force no cache

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 03:32:21 +00:00

ff6bb32684 Plumb global scale as GEMM alpha instead of folding into UE4M3

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 03:27:56 +00:00

d547da2948 stage_activation: add per-tensor global scale matching NVFP4 spec

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 03:01:08 +00:00

108ff07569 debug: remove one-shot gate from logit dump, log every forward

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 02:37:24 +00:00

3600a4b06a debug: add logit quality dump in compute_logits (ungated, once)

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 02:17:17 +00:00

29f8b8c174 fix: load lm_head.weight in outer model before forwarding to inner

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 01:54:55 +00:00

46536e5ccf fix: hc param renames missing leading dot

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 01:29:01 +00:00

086f3fa5c5 fix: hc params dot→underscore + compressor position_bias→ape combined rule

biondizzle pushed to master at biondizzle/nvfp4-megamoe-kernel 2026-05-15 01:16:20 +00:00

44d4b6c225 fix: add missing renames for Hadamard coding + compressor.ape