Oct 15 20:13 fix(vulkan): normalize fused dtypes and resize casts Oct 14 17:45 docs: update model card, context lengths, and config specs Oct 13 19:24 fix(vulkan): use F.embedding for sliding window masks 18:40 fix(vulkan): reduce context length to 2048 to avoid device OOM 16:07 fix(vulkan): bind graph input start_pos directly in custom attention transforms 15:31 fix(vulkan): restore correct prefill attention masks and positions 14:44 fix(vulkan): pass scalar graph input input_pos to layers 14:12 fix(vulkan): re-export gemma_4_e2b_vulkan_8da4w with initialized mutable buffers Oct 12 13:28 update models Sep 3 14:48 Stop restating quantized, and drop the unused default quantized duplicated precisions.json, which is authoritative for the
quantized/non-quantized partition. default expressed a per-repository
recommendation nothing consumed: the library resolves its own per-platform
default. Both are gone from the schema. Jun 22 09:43 Use min/max dynamic shape bounds for forward inputs/outputs 08:39 Backfill constant metadata method values in config.json from .pte Jun 10 11:53 Delete e2b/vulkan/tokenizer_config.json 11:53 Delete e2b/vulkan/tokenizer.json May 25 08:45 Upload folder using huggingface_hub May 22 13:13 Upload folder using huggingface_hub