Fix ZeroGPU compat: torch 2.6.0 -> 2.8.0, update causal_conv1d wheel

Fix ZeroGPU compat: torch 2.6.0 -> 2.8.0, update causal_conv1d wheel

Fix ZeroGPU compat: torch 2.6.0 -> 2.8.0, update causal_conv1d wheel

Drop redundant Pred tag from viewer header

Stream into stable iframe via postMessage; add live/done badges

Stream OCR output: render blocks as </div> closes

Add HF model link at top

Rename Accordion to Generated HTML / Markdown

Remove zoom buttons + bbox label overlays

Add Markdown tab via chandra parse_markdown; drop redundant title/description

Use full progress bar

Use minimal progress indicator (spinner only)

Hide progress bar during OCR execution

Narrow OCR result panel from 460px to 360px

Drop remaining open-chandra references

Update model reference to Qwen3.5-ocr-jp-2b

Fix PIL fp closed error on image preview (eagerly load pixel data)

Match victor/chandra-ocr-2 inference pattern exactly (batch wrap + padding + simpler eos handling)

Use Dao-AILab official causal_conv1d wheel (victor/misc was private/401)

Add flash-linear-attention + causal_conv1d (Qwen3.5 has 18 linear-attn layers); pin versions; area-based image cap

Disable pre-resize (let processor token bounds match training); bump duration to 300s

Use SDPA instead of flash-attn (sdist build fails in HF Space due to torch install order)

Revert vLLM (ZeroGPU MIG incompatible); use HF transformers + flash-attn 2

Switch inference backend from HF transformers to vLLM

Fix max_new_tokens to 12288 (training max_length); drop slider

Translate UI to English; remove JDocQA sample; preview on upload

Eager-load model at module level for ZeroGPU; bump duration to 180s

Add torchvision (required by Qwen3VLVideoProcessor auto-load)

Initial Space upload (Open Chandra OCR + ZeroGPU)

initial commit