Commit History

  1. fix(vulkan): normalize fused dtypes and resize casts

    4ce8047

  1. docs: update model card, context lengths, and config specs

    89b9cd0

  1. fix(vulkan): use F.embedding for sliding window masks

    6b255d5
  2. fix(vulkan): reduce context length to 2048 to avoid device OOM

    abee758
  3. fix(vulkan): bind graph input start_pos directly in custom attention transforms

    fd538bd
  4. fix(vulkan): restore correct prefill attention masks and positions

    6f00a20
  5. fix(vulkan): pass scalar graph input input_pos to layers

    0e81595
  6. fix(vulkan): re-export gemma_4_e2b_vulkan_8da4w with initialized mutable buffers

    5e754b7

  1. update models

    3302ece

  1. Stop restating quantized, and drop the unused default⁠

    d6f7580

  1. Use min/max dynamic shape bounds for forward inputs/outputs

    3cdead0
  2. Backfill constant metadata method values in config.json from .pte

    193632d

  1. Delete e2b/vulkan/tokenizer_config.json

    Verified signature
    f717f7a
  2. Delete e2b/vulkan/tokenizer.json

    Verified signature
    a4ee19f

  1. Upload folder using huggingface_hub

    Verified signature
    8b1973a

  1. Upload folder using huggingface_hub

    Verified signature
    d5e81f2