FP8 on the fine-grained machinery (`Fp8Config`) #49777
IlyasMoutawwakil
wants to merge
21 commits into
fg-split/4-machinery
from
fg-split/5-fp8
Add Fp8Config: block and per-tensor FP8 on the fine-grained machinery
51e3a37b
Document FP8 and the migration from FineGrainedFP8Config
73941c31
Keep the legacy FP8 tests while the legacy path ships
1151fab3
Quantize partial blocks in the legacy FP8 integration
315ed24e
Re-record the DeepSeek-V3 FP8 logits slice for the fixed block-FP8 ke…
3ec10303
Keep fine-grained member lookups and FP8 checkpoint reads to their ow…
1c576253
Explain why FP8 reads, pads and plans the way it does
012c2378
Test the FP8 experts layout on both sides of SM100
d945d16d
Document prefer_deepgemm_linear instead of the env var
824e9a81
Check main's local-tensor flag on the FP8 embedding
fcbb337f
Update the fine-grained plan tests for the generic host
ba103c86
Test that an embedding table loads through the member that owns it
da2d3535
Ask the member for a copied model's param ops
f610081b
Keep an FP8 table's scale in the model dtype
b589b897
Review cleanups: one embedding check, a private plan reader, short er…
008a9d6c
Name embedding tables with targeted plan entries only
1662f90b
Read DeepSeek-V4's quantization through an fp8 reader, not a model co…
8278aea4
Read FP8 embedding tables into plan entries and drop Fp8Config.module…
7114bcdd
Leave the ModelOpt and compressed-tensors readers to their own PR
1feb8aad
Leave a DeepSeek-V4 quantization config that transformers saved as it is
cdfe320e
Restack on the simpler plan resolution and merging
516802b7
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub