huggingface/transformers

Pull Requests Commits

Let the swap hand every module the same storage

IlyasMoutawwakil committed 31 days ago

40e6b600

Simplify the quantize op's two new paths

IlyasMoutawwakil committed 31 days ago

b1f38a22

Quantize per module group; fix four defects in the quantized load

IlyasMoutawwakil committed 31 days ago

8a0dc74c

Key on capability not name; drop the last op boilerplate

IlyasMoutawwakil committed 32 days ago

60f904e6

Collapse the conversion-op boilerplate; shorten the longest units

IlyasMoutawwakil committed 32 days ago

e83eadf5

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

IlyasMoutawwakil committed 32 days ago

9ddb2140

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

IlyasMoutawwakil committed 33 days ago

f7081ecf

IlyasMoutawwakil committed 36 days ago

b4dcb7b0

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

IlyasMoutawwakil committed 36 days ago

daa24205

IlyasMoutawwakil committed 36 days ago

f9302cc2

fix multimodal and add some testing

IlyasMoutawwakil committed 36 days ago

22b1001c

fixes for the new model

IlyasMoutawwakil committed 36 days ago

fe0b7182

Shard quantized expert companions under EP and TP

IlyasMoutawwakil committed 37 days ago

43b9a748

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

IlyasMoutawwakil committed 37 days ago

e8e6b879

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

IlyasMoutawwakil committed 37 days ago

53a8970a

IlyasMoutawwakil committed 38 days ago

788d02dd

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

IlyasMoutawwakil committed 38 days ago

97dfa106

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

IlyasMoutawwakil committed 38 days ago

1265f0cc

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

IlyasMoutawwakil committed 38 days ago

f63ad97d

FineGrainedFP8Config: drop the frozen-notice docstring line

IlyasMoutawwakil committed 38 days ago

29cca042

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

IlyasMoutawwakil committed 38 days ago

c6656441

IlyasMoutawwakil committed 38 days ago

e139d2a0

Merge branch 'main' into finegrained-unified

IlyasMoutawwakil committed 38 days ago

d59f9037

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

IlyasMoutawwakil committed 38 days ago

09ea21c0

finegrained: address the integration review

IlyasMoutawwakil committed 38 days ago

89267bf3

Relax static-cache tolerance in test_generate_with_static_cache (1e-5 → 5e-5) (#48815)

ydshieh committed 38 days ago

Verified a4aa9890

Pin utf-8 in test_can_init_all_missing_weights source read (Windows non-UTF-8 locale fix) (#48819)

dltsum committed 39 days ago

Verified 3f162ca5

[generate] Drop attention mask early without padding (#48814)

Cyrilvallez committed 39 days ago

Verified d65fdc5f

[utils] Add MUSA support for Flash Attention 2 (#48612)

XiaomingFun233 committed 39 days ago

Verified a56a23f3

finegrained: consume modelopt's calibrated globals, and fuse a model's per-expert output norm

IlyasMoutawwakil committed 39 days ago

a562fb8c

Older