Spaces:
Paused
Paused
Commit History
-
generated more worlds
630dbf1 -
added submission tags to README
24491a4 -
added social links
c618683 -
replaced MiniCPM5-1B with MiniCPM4.1-8B
a354cf0 -
Fix q8 quant
1e97283 -
disabled reasoning; changed default quant from q4 to q8
9a82af1 -
updated model from MiniCPM3 to MiniCPM5
b0dd749 -
colors fixes; updated README
170c331 -
Fix ZeroGPU: preload CUDA runtime libs for the llama.cpp wheel
The cu124 llama-cpp-python wheel is dynamically linked against libcudart/libcublas, which are not on the loader path in the HF Spaces image (OSError: libcudart.so.12). Ship them as nvidia-* pip wheels and dlopen them with RTLD_GLOBAL (cudart first) before importing llama_cpp so libllama.so resolves its symbols. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>76e2a61 -
Fix Python 3.10 startup: tomllib fallback + pin python_version 3.12
HF Spaces' default image is Python 3.10, where the stdlib `tomllib` (3.11+) is missing. Pin python_version to 3.12 in the Space metadata and fall back to `tomli` in config.py so it loads on 3.10 too. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>83b2601 -
Run a local small model (MiniCPM3-4B via llama.cpp) instead of a cloud API
The Space had no path to the VPN-only API used by the CLI, and the hackathon requires a <=32B model with no user credentials. Run OpenBMB MiniCPM3-4B (~4B) in-process through llama.cpp on ZeroGPU: - src/id/llm/local.py: in-process llama.cpp runtime that downloads the GGUF from the Hub and generates inside an @spaces.GPU function, behind an OpenAI-shaped adapter so the engine/client logic is unchanged. - client.py: route the "local" provider to that adapter. - app.py/config.toml: default every tier to the local provider + MiniCPM3-4B; prefetch the GGUF on CPU at startup. - requirements.txt: llama-cpp-python (CUDA 12.4 wheels), spaces, huggingface_hub. - README: document the no-API/ZeroGPU setup and override knobs. Stacks the Off-the-Grid + Llama Champion + Tiny Titan badges and the OpenBMB sponsor track. Worlds are pre-generated, so play time only runs the light tiers (character/extractor/guard/environment/judge). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>d000651 -
Fix dark-mode color clash by overriding Gradio design tokens
The force-light JS redirect did not take effect inside the HF Spaces iframe, leaving dark-mode browsers with dark dropdowns/transcript and unreadable text. Override Gradio's CSS design tokens (backgrounds, text, borders, buttons) for both the default and .dark scopes — and on the outer html/body/gradio-app elements — so the light paper palette renders identically regardless of the visitor's system theme. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>b99262c -
Move to Gradio 6, fix Space YAML, improve readability
- Upgrade to gradio 6.18.0 (theme/css/js moved to launch(); drop removed Chatbot type= arg; use GoogleFont objects in the theme) - Fix Space metadata rejected by HF: colorTo stone -> blue, emoji -> 🔍 - Force light theme and tighten CSS so text/panels stay readable in dark-mode browsers Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>5af5776 -
Merge branch 'main' of https://huggingface.co/spaces/build-small-hackathon/f-id
0c21845 -
Initial Gradio Space
0423b99 -
initial commit
Verified signatureb1b8837
