Fix vision image freeze: reduce maxPixels, parallelize ViT attention

Remove stale build assets

Fix vision preprocessing: correct normalization and min_pixels

Remove debug tokenizer exposure, clean build

Fix thinking mode: always show tokens, auto-switch sampling params, expose toggle

Clean chat template: EOS before callback, skip_special_tokens, full-sequence decode for UTF-8

Fix smart_resize to match HF reference (floor/ceil rounding, correct min/max pixels)

Fix vision preprocessing: CLIP normalization and ViT RoPE frequencies

Add vision/multimodal support across all Qwen3.5 models

Lock to 2B model, instruct-general only, rep penalty 1.15

Default to 0.8B, cap at 2B for MacBooks, repetition penalty 1.15

Fix boot regression: render synchronously, probe GPU in background

Add WebGPU support check and hardware-adaptive model selection

Sync latest local build with updated model & GPU ops

Fix asset paths: use relative base for HF Spaces compatibility

Deploy TensorBend — browser-based LLM inference via WebGPU

initial commit