transformers
🚨 NVFP4: MoE experts and ModelOpt exports on finegrained-kernels
#49851
Open
Open
🚨 NVFP4: MoE experts and ModelOpt exports on finegrained-kernels #49851
IlyasMoutawwakil
wants to merge
21 commits into
main
from
nvfp4-experts
NVFP4: MoE experts and ModelOpt exports on finegrained-kernels
35b8576e
NVFP4: save_pretrained support, parameterized tests
e39ee4b9
NVFP4: build on the quantizer API, port the stack's layout ops, fix TP
6d1360fc
NVFP4: review cleanups
f016cd4e
NVFP4: plan scales instead of special-casing the shard op, review cle…
9f1f68e7
NVFP4: one swizzle op on row-major scale params, no core change
12833c57
NVFP4: read MXFP8 checkpoints with NVFP4 experts, readers translate f…
005b75d6
NVFP4: shard experts under tensor parallelism instead of replicating …
681ba464
NVFP4: swizzle each rank's scales when they fit, warn when they don't
3ca9d26a
NVFP4: pass the linear's kernel a contiguous input
664bc049
NVFP4: let core build the converters and plan entries from serialized…
df42b9ad
NVFP4: say why the linear's input is made contiguous
5082fec5
NVFP4: pass the linear's input as it is
89ec7c19
NVFP4: keep the experts prefix on the expert stacks' serialized names
5d16de96
small nit
00617dcf
SunMarc
commented
on 2026-10-28
no need for utils for this
04671339
cleanup
7309135e
cleanup
1351a845
cleanup
98fb0683
NVFP4: quantize on the fly through the weight conversions
ed2295ea
NVFP4: address the review
02c47b16
Login to write a write a comment.
Login via GitHub