Spaces:
Running on Zero
Running on Zero
Commit History
-
Raw tab gets a working scrollbar (transparent by default); all scrollbars reveal on rail-strip hover instead of thumb-only hover. Running
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>f06b3ef -
One block for chat + input: the input line moves inside the message block behind a single divider; send/clear become flat line-form icons like the attach icon.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>f914b10 -
Dark mode: example chips follow the theme text color (white on black / black on white).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>e07dda8 -
Input bar: one flat box (outer block + inset stripped); pending media float in a strip at the message block's bottom instead of inside the bar.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>c0e5343 -
True multi-turn: the model sees the whole session via offline_batch_generate.
chat_generate wraps the model's session-state API: prior messages (with their media, referenced by path) round-trip through gr.State and are re-encoded on every stateless ZeroGPU call, so follow-up turns can reference earlier turns and earlier media. A 🗑 button resets the conversation. The engine is also exposed as the /chat API/MCP tool (session in -> reply + session out); analyze_image / analyze_video / analyze_text stay as single-turn tools. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>50e5fdf -
Multi-turn media chat: full-width stage, multimodal input bar, per-turn media.
- Left media block removed; the chat/raw stage spans full width with the input console glued directly beneath it (no gap). - gr.MultimodalTextbox input bar: attach button inside at the right, pending media thumbnail in a row just above the text (click opens the full-res file; corner ✕ removes), round send button beside the bar. - Each turn sends what is staged: media + text, media alone (default prompt), or text alone -> new analyze_text path (empty images/videos query via the model's one-shot helper), also exposed as an API/MCP tool. - Examples are multimodal (text + file) and fill the input bar. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>d2c0ac9 -
Console: round send button beside the input bar (explicitly wired); input scrolls with no visible scrollbar.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>280d066 -
EN/中文 language switcher next to the header links, same pattern as the live demo.
Value-stable radio choices; toggles the title, file label, clear-media button, chat/raw toggle, input placeholder, and Advanced Settings labels. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>4e88db1 -
Input bar: center the placeholder; drop the Analyze button — the embedded round send is the only trigger (offline one-shot chat needs no session controls).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>ccd28ff -
Console like the live demo: prompt bar under the chat, examples fill it; prompt block removed.
- Right column now mirrors the live demo: chat/raw stage + a bottom console row (input bar with embedded send + Analyze). Enter or Analyze sends the staged media + prompt; each ask APPENDS to the ongoing chat (media bubble re-shown only when the staged media changes) and the input clears. - Left column: media stage + Clear media button, strictly aligned with the right stage and console row (same heights as the live demo pattern). - Examples' prompts populate the console input bar. - Raw event view auto-scrolls to the newest entry. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>c573ff7 -
gradio 6: api_name=False -> api_visibility=private; annotate analyze_* returns as str so the API/MCP endpoints return the response text.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>d6f508d -
Fix TypeError on image requests: bridge resample->interpolation for the fast processor.
transformers 4.57 fast image processors pass interpolation= into _preprocess, but the model's remote-code MossVLImageProcessorFast still names the parameter resample (stale name, same torchvision type). Every image analysis died with "TypeError: _preprocess() missing 1 required positional argument: 'resample'". Shim maps the kwarg at call time; no-ops once the model repo renames it (the MOSS-VL-Realtime repo already carries the corrected signature). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>2630de4
-
Single-stage layout: one image/video block, chat/raw response view, fixed equal heights.
- One media block (gr.File + scaled-to-fit previews) replaces the two tabs; examples merged into one table (3 image rows + 1 video row). - Right response block mirrors the live demo: segmented Chat/{ } Raw toggle, chatbot + raw request/response JSON, pinned to a fixed 680px height that matches the left column exactly; long text scrolls inside with transparent hover-only scrollbars. - analyze_image / analyze_video stay exposed as API/MCP tools; a MOCK mode (MOSS_DEMO_MOCK=1) allows local UI work without torch/spaces. - sdk bumped to gradio 6.15.1 (same as the realtime Space). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>65be679
-
Delete tmp/hugging-demos-build-model_OpenMOSS-Team_MOSS-VL-Instruct-0708-t4yujp0c/build/requirements.txt with huggingface_hub
Verified signature7b97603 -
Upload requirements.txt with huggingface_hub
Verified signature04034e3 -
Upload README.md with huggingface_hub
Verified signature690e20e -
Upload /tmp/hugging-demos-build-model_OpenMOSS-Team_MOSS-VL-Instruct-0708-t4yujp0c/build/requirements.txt with huggingface_hub
Verified signatureb27b4e7 -
Upload requirements.txt with huggingface_hub
Verified signaturee5e173a -
Upload requirements.txt with huggingface_hub
Verified signaturecf201cf -
Upload folder using huggingface_hub
Verified signaturec346860 -
initial commit
Verified signature670fb9c
