Spaces:
Sleeping
Sleeping
Commit History
-
fix: Sage timeout — gunicorn 300s, max_tokens 150, ZeroGPU pattern, correct torch_dtype
Charan Suresh- Procfile + render.yaml: gunicorn --timeout 120 → 300 - knowledge/main.py: /api/sage passes max_tokens=150 - knowledge/ollama_client.py: chat() accepts max_tokens kwarg - hf_space/inference.py: full ZeroGPU rewrite (Gemma4Processor, bfloat16, no device_map, CUDA log) - hf_space/requirements.txt: transformers>=5.0.0, torch>=2.4.0, hub>=1.3.0 - hf_space/README.md: hardware: zero-gpuab04b85
-
fix: replace git transformers with stable release, update torch for Gemma 4 build
Charan Sureshc399b4d
-
Fix Sage/Lens silent failures and model loading on HF Space
Charan Suresh- health_check() no longer blocks on load_model(); returns current load state instantly so /api/health never times out - generate_text() moves tokenizer inputs to correct device (GPU/CPU) - generate_with_image() tries real Gemma 4 vision via AutoProcessor; falls back to text-only if unavailable - /api/generate and /api/generate_vision raise HTTP 503/500 instead of returning {error:...} with status 200 -- raise_for_status() now fires - Model preloaded in background thread at app startup (no cold-start) - HF Space POST timeout raised to 300s in ollama_client.py - 503 handled alongside 500 with detail extracted from response body6180f4c
-
Fix HF Space: switch from llama-cpp-python to transformers for better compatibility
dee67dc -
Switch to HF Space as default backend, make Ollama optional
e1b76ea
-
Replace cloud inference with local Ollama
4c42b27
-
feat: add anti-gaming comprehension pipeline and harden HF space integration
b3edcc3
-
Add HF Space model error diagnostics
372105d -
Switch HF Space to quantized Gemma 2B GGUF
1e693b7 -
Switch HF Space to transformers backend
0b7846f -
Harden HF Space llama-cpp fallback
0912d7b -
Add HF Space inference fallback
c34b1b1 -
Fix HF Space llama-cpp startup crash
205a0d7 -
fix: use pre-built CPU wheel for llama-cpp-python to speed up HF space build
97be553 -
fix: cast model_path and prompt to string to resolve ValueError('not a string')
b6392ce -
chore: migrate hf_space to CPU-friendly GGUF via llama-cpp-python
65fcecf
-
Fix HF Space inference runtime and error handling
fc09209 -
Fix HF Space backend contract and production health diagnostics
0dc30d2 -
Update app defaults and database to use US Curriculum Guide, fix HF requirements
1b03d62