huggingface/transformers

Pull Requests Commits

Format EP implementation changes

IlyasMoutawwakil committed 3 days ago

6ca9df45

Use EP implementation naming in Mega MoE

IlyasMoutawwakil committed 3 days ago

b6bbb4b7

Note that Mega MoE needs data-parallel inference to run as designed

IlyasMoutawwakil committed 3 days ago

ff15b657

Hand each rank's Mega MoE kernel a slice of the shared batch

IlyasMoutawwakil committed 3 days ago

7cb77529

Keep the Mega MoE forward in its original order

IlyasMoutawwakil committed 3 days ago

cd057eb7

Allocate the Mega MoE buffer into the cache, then read it back

IlyasMoutawwakil committed 3 days ago

9d8c4d2c

`deepgemm_megamoe` strategy owns the Mega MoE symmetric buffers, shared by every layer

IlyasMoutawwakil committed 3 days ago

a0038bd5

Remove the Mega MoE router and experts parallel styles

IlyasMoutawwakil committed 3 days ago

359335d2

Deprecate the Mega MoE router and experts parallel styles

IlyasMoutawwakil committed 3 days ago

6a1f6668

`passthrough` expert-parallel strategy, selected for DeepGEMM Mega MoE

IlyasMoutawwakil committed 3 days ago

a466c57c

Format EP implementation changes

IlyasMoutawwakil committed 3 days ago

9863215b

Rename EP strategies to implementations

IlyasMoutawwakil committed 3 days ago

8495c60b

test checkpointing

IlyasMoutawwakil committed 3 days ago

0b965e48

Merge remote-tracking branch 'origin/main' into ep-masked-when-shared-batch

IlyasMoutawwakil committed 4 days ago

ccfc1335

Expert-parallel layers in `expert_parallel.py`, `ep_implementation`, and the EP tests on their own

IlyasMoutawwakil committed 4 days ago

6bfa58b3

Add Unlimited OCR (#46836)

guarin committed 4 days ago

Verified 5789b6dd

FSDP x TP Saving / Loading (#48802)

michaelbenayoun committed 4 days ago

Verified 89d1c880

Use aiter-rope version 3 on ROCm (#49595)

Abdennacer-Badaoui committed 4 days ago

Verified 2b8b231f

Don't set a device per rank for gloo tensor-parallel tests (#49636)

IlyasMoutawwakil committed 4 days ago

Verified 276d3fac

fix(distributed): keep enable_expert_parallel as EP alone (#49633)

dacorvo committed 4 days ago

Verified d21c0d9e

Remove missing methods from docs (#49637)

guarin committed 4 days ago

Verified 16cad282

Translate `torch.mm(..., out_dtype=...)` to ONNX (#49635)

IlyasMoutawwakil committed 4 days ago

Verified 0e85988b

[docs] Hardware section (#49541)

michaelbenayoun committed 4 days ago

Verified c9e5f15a

Raise when a torch fallback would drop packed-sequence boundaries (#49474)

qgallouedec committed 4 days ago

Verified 01f16c0e

Fa packed sequence decode (#49610)

SunMarc committed 4 days ago

Verified da23990a

fix nested set_attn_implementation (#49138)

eustlb committed 4 days ago

Verified 4e3fc5e8

feat: Add MTP support to Mellum2/2.1 (#49531)

shadeMe committed 4 days ago

Verified 14300401

Fall back to the AITER kernel for flash_attention_2 on ROCm (#49608)

Abdennacer-Badaoui committed 4 days ago

Verified 2126f65b

extend onnx exporters tests to non-cuda devices (#49627)

kaixuanliu committed 4 days ago

Verified cd1d4736

Document video processor backends (#49470)

DimensionSTP committed 4 days ago

Verified b4c24288

Older