huggingface/transformers

Pull Requests Commits

Restack on the simpler plan resolution and merging

IlyasMoutawwakil committed 1 day ago

Verified 9ee98c88

Keep the saved-FP8 DeepSeek-V4 check in the plan test that stays

IlyasMoutawwakil committed 1 day ago

Verified 16910d45

Leave NVFP4's compressed-tensors and ModelOpt export support to the producer readers' PR

IlyasMoutawwakil committed 1 day ago

Verified bc73e938

Give the table test a model that holds the table

IlyasMoutawwakil committed 1 day ago

Verified 1a1bfd5a

Let a post-expert norm declare its fused kernel form

IlyasMoutawwakil committed 1 day ago

Verified 883435d2

Keep the old NVFP4 integration importable for a deprecation cycle

IlyasMoutawwakil committed 1 day ago

Verified 7bfee090

Name the embedding tables of the NVFP4 tests with targeted plan entries

IlyasMoutawwakil committed 1 day ago

Verified f62e0ec7

Read the NVFP4 producer layouts through get_weight_conversions

IlyasMoutawwakil committed 1 day ago

Verified cb8ecaf3

Fix the plan tests the host rebase merged twice

IlyasMoutawwakil committed 1 day ago

Verified b6defe6c

Update the NVFP4 plan tests for the generic host

IlyasMoutawwakil committed 1 day ago

Verified e7f30fc2

Test that DeepGEMM grouped experts refuse the Triton layout

IlyasMoutawwakil committed 1 day ago

Verified 1b7d4f44

Explain why NVFP4 inverts its globals and checks weight-only loads

IlyasMoutawwakil committed 1 day ago

Verified 4477b933

Skip the NVFP4 activation check when a plan load dequantizes

IlyasMoutawwakil committed 1 day ago

Verified 1f728460

Give the DeepGEMM refusal test's experts two K bands so MX scales swizzle

IlyasMoutawwakil committed 1 day ago

Verified d1b7081c

Lend NVFP4's activation format like its other loading attributes

IlyasMoutawwakil committed 1 day ago

Verified 164c37e5

Delete the frozen NVFP4 integration

IlyasMoutawwakil committed 1 day ago

Verified d83fbbf2

Document NVFP4 on the fine-grained machinery and ModelOpt mixed-precision plans

IlyasMoutawwakil committed 1 day ago

Verified 43ddf8d7

Move NVFP4Config onto the fine-grained machinery, with the producers' NVFP4 exports

IlyasMoutawwakil committed 1 day ago

Verified fb79484a

Restack on the simpler plan resolution and merging

IlyasMoutawwakil committed 1 day ago

Verified b0750fd3

Give the DeepSeek-V4 load-path fixture its plan explicitly

IlyasMoutawwakil committed 1 day ago

Verified c3e232f5

Leave MXFP4's ModelOpt layout to the producer readers' PR

IlyasMoutawwakil committed 1 day ago

Verified 3e3b270e

Read DeepSeek-V4's MXFP4 experts plan in the fp8 reader

IlyasMoutawwakil committed 1 day ago

Verified 6c840c54

Keep main's MXFP4 integration for GPUs below compute capability 8.9

IlyasMoutawwakil committed 1 day ago

Verified 0411198f

Read GPT-OSS's experts keys through get_weight_conversions

IlyasMoutawwakil committed 1 day ago

Verified ea7a1ece

Keep DeepSeek-V4's Triton sweeps on Triton linears through prefer_deepgemm_linear

IlyasMoutawwakil committed 1 day ago

Verified acfa6992

Keep DeepGEMM grouped experts out of DeepSeek-V4's SM100 sweeps

IlyasMoutawwakil committed 1 day ago

Verified afeff1c8

Explain why MXFP4 defaults to DeepSeek's layout and W4A4 experts

IlyasMoutawwakil committed 1 day ago

Verified 4d1cc69d

Dequantize MXFP4 experts for use_kernels on every device

IlyasMoutawwakil committed 1 day ago

Verified 3b1979e5

Require compute capability 8.9 for every MXFP4 activation format

IlyasMoutawwakil committed 1 day ago

Verified e4e0ee00

Test that Mxfp4Config takes main's positional arguments

IlyasMoutawwakil committed 1 day ago

Verified aa7faff3

Older