huggingface/transformers

Pull Requests Commits

Leave a DeepSeek-V4 quantization config that transformers saved as it is

IlyasMoutawwakil committed 10 days ago

Verified cdfe320e

Leave the ModelOpt and compressed-tensors readers to their own PR

IlyasMoutawwakil committed 10 days ago

Verified 1feb8aad

Read FP8 embedding tables into plan entries and drop Fp8Config.modules_to_convert

IlyasMoutawwakil committed 10 days ago

Verified 7114bcdd

Read DeepSeek-V4's quantization through an fp8 reader, not a model config hook

IlyasMoutawwakil committed 10 days ago

Verified 8278aea4

Name embedding tables with targeted plan entries only

IlyasMoutawwakil committed 10 days ago

Verified 1662f90b

Review cleanups: one embedding check, a private plan reader, short error messages

IlyasMoutawwakil committed 10 days ago

Verified 008a9d6c

Keep an FP8 table's scale in the model dtype

IlyasMoutawwakil committed 10 days ago

Verified b589b897

Ask the member for a copied model's param ops

IlyasMoutawwakil committed 10 days ago

Verified f610081b

Test that an embedding table loads through the member that owns it

IlyasMoutawwakil committed 10 days ago

Verified da2d3535

Update the fine-grained plan tests for the generic host

IlyasMoutawwakil committed 10 days ago

Verified ba103c86

Check main's local-tensor flag on the FP8 embedding

IlyasMoutawwakil committed 10 days ago

Verified fcbb337f

Document prefer_deepgemm_linear instead of the env var

IlyasMoutawwakil committed 10 days ago

Verified 824e9a81

Test the FP8 experts layout on both sides of SM100

IlyasMoutawwakil committed 10 days ago

Verified d945d16d

Explain why FP8 reads, pads and plans the way it does

IlyasMoutawwakil committed 10 days ago

Verified 012c2378

Keep fine-grained member lookups and FP8 checkpoint reads to their own format

IlyasMoutawwakil committed 10 days ago

Verified 1c576253

Re-record the DeepSeek-V3 FP8 logits slice for the fixed block-FP8 kernel

IlyasMoutawwakil committed 10 days ago

Verified 3ec10303

Quantize partial blocks in the legacy FP8 integration

IlyasMoutawwakil committed 10 days ago

Verified 315ed24e

Keep the legacy FP8 tests while the legacy path ships

IlyasMoutawwakil committed 10 days ago

Verified 1151fab3

Document FP8 and the migration from FineGrainedFP8Config

IlyasMoutawwakil committed 10 days ago

Verified 73941c31

Add Fp8Config: block and per-tensor FP8 on the fine-grained machinery

IlyasMoutawwakil committed 10 days ago

Verified 51e3a37b

Document how a passed config merges into a plan, and fused modules

IlyasMoutawwakil committed 10 days ago

Verified 49cbba81

Test that only a format config lends to a plan

IlyasMoutawwakil committed 10 days ago

Verified c6b3d110

Let a block-scaled op read its own member's config

IlyasMoutawwakil committed 10 days ago

Verified 57fd5839

Read skip lists into None entries and drop config_for

IlyasMoutawwakil committed 10 days ago

Verified 2ffdc512

Let the block-scaled formats extend the standard quantizer hooks

IlyasMoutawwakil committed 10 days ago

Verified 019a5f02

Say in the docs that a plan entry can be any method's config

IlyasMoutawwakil committed 10 days ago

Verified 610c2f0b

Find a shared op's member through the host's member_for

IlyasMoutawwakil committed 10 days ago

Verified ceaa9769

Add the block-scaled quantizer base and the fine-grained modules

IlyasMoutawwakil committed 10 days ago

Verified 8817bc56

Pick a module's entry from the patterns that matched it

IlyasMoutawwakil committed 10 days ago

Verified 5441a389

Merge a passed config into the plan in FineGrainedConfig

IlyasMoutawwakil committed 10 days ago

Verified 2c6264c4

Newer Older