huggingface/transformers

Pull Requests Commits

Bring two comment blocks under the noisy-comment limits

IlyasMoutawwakil committed 25 days ago

21c0a070

Point everything at the fine-grained integration; deprecate the frozen files

IlyasMoutawwakil committed 25 days ago

f6878604

Cover the groups feature, and export what it asks callers to build

IlyasMoutawwakil committed 25 days ago

4cbbe7ff

IlyasMoutawwakil committed 25 days ago

f7a8e543

IlyasMoutawwakil committed 25 days ago

2392e6a6

IlyasMoutawwakil committed 25 days ago

95b3e44f

Address the per-scheme review

IlyasMoutawwakil committed 25 days ago

9fc683de

Carry the quantized-load fixes onto the per-scheme arms, and the same simplification

IlyasMoutawwakil committed 26 days ago

a67e7acf

Quantize per module group; split the arms by producer, not by format

IlyasMoutawwakil committed 26 days ago

4cb5d9cb

Key on capability not name; drop the last op boilerplate

IlyasMoutawwakil committed 26 days ago

60f904e6

Collapse the conversion-op boilerplate; shorten the longest units

IlyasMoutawwakil committed 27 days ago

e83eadf5

Fix a destructive Qwen3 TP-plan override; split out the conversions

IlyasMoutawwakil committed 27 days ago

9ddb2140

Finish the finegrained integration: reload, TP sharding, dsv4 norms

IlyasMoutawwakil committed 27 days ago

f7081ecf

IlyasMoutawwakil committed 30 days ago

b4dcb7b0

Load a calibrated checkpoint's input_scale, and fix what per-tensor weights reach

IlyasMoutawwakil committed 30 days ago

daa24205

IlyasMoutawwakil committed 30 days ago

f9302cc2

fix multimodal and add some testing

IlyasMoutawwakil committed 31 days ago

22b1001c

fixes for the new model

IlyasMoutawwakil committed 31 days ago

fe0b7182

Shard quantized expert companions under EP and TP

IlyasMoutawwakil committed 31 days ago

43b9a748

deepgemm: only Mega MoE cannot apply a post-expert norm

IlyasMoutawwakil committed 32 days ago

e8e6b879

finegrained: drop the DTensor unwrap, and pass the kernel operands by name

IlyasMoutawwakil committed 32 days ago

53a8970a

IlyasMoutawwakil committed 32 days ago

788d02dd

is_sm100: answer False without CUDA, and fake both halves in the tests

IlyasMoutawwakil committed 32 days ago

97dfa106

Pin that every modelopt converter target is a slot the module holds

IlyasMoutawwakil committed 32 days ago

1265f0cc

Tolerate an experts module without the norm, and a machine without CUDA

IlyasMoutawwakil committed 32 days ago

f63ad97d

FineGrainedFP8Config: drop the frozen-notice docstring line

IlyasMoutawwakil committed 33 days ago

29cca042

FineGrainedFP8Config is its own class again, not an alias of the new one

IlyasMoutawwakil committed 33 days ago

c6656441

IlyasMoutawwakil committed 33 days ago

e139d2a0

Merge branch 'main' into finegrained-unified

IlyasMoutawwakil committed 33 days ago

d59f9037

finegrained: freeze the mxfp4 and nvfp4 integrations, and serve what they served

IlyasMoutawwakil committed 33 days ago

09ea21c0

Older