Spaces:
Runtime error
Runtime error
Commit History
-
Remove authentication for ungated Qwen model
Prashanth Ramakrishna- Qwen2.5-7B-Instruct is not gated - Token in headers was causing 401 errors - Commenting out auth for public model access5846074 -
Force rebuild with Qwen2.5-7B
Prashanth Ramakrishna- Added comment to trigger HF Space rebuild - Qwen2.5-7B is ungated and reliable6d1a742 -
Switch to Qwen2.5-7B-Instruct
Prashanth Ramakrishna- Ungated model, no authentication issues - Proven XML compliance (72B version worked perfectly) - Fast 7B model with same architecture benefits - Should resolve both access and truncation issues296d6c6 -
Switch to Mistral-7B-Instruct-v0.3
Prashanth Ramakrishna- Now using HF token for authentication (already configured in Space secrets) - Mistral models have proven reliability with XML formatting - Should resolve synthesis truncation issues seen with Zephyr973b3fe -
Switch to Zephyr-7B-beta (truly ungated model)
Prashanth Ramakrishna- Zephyr-7B-beta is completely open and ungated - Maintained by HuggingFaceH4, no authentication required - High quality instruction-following model - Already supported in our model_formats.py libraryf00bbe6 -
Switch to Mistral-7B-Instruct-v0.1 (ungated model)
Prashanth Ramakrishna- v0.3 and v0.2 are gated models requiring authentication - v0.1 is fully open and ungated - Same performance characteristics for our use casef9c7412 -
Switch to Mistral-7B-Instruct-v0.3 for faster inference
Prashanth Ramakrishna- Using single GPU (NUM_SHARD=1) for optimal 7B model performance - No quantization needed with 24GB L4 VRAM - Optimized batch settings for high throughput - 10-20x faster than Qwen-72B configuration40b044c
-
Force rebuild - Deploy Qwen2.5-72B with 4xL4 tensor parallelism
Prashanth Ramakrishna0a09fc5 -
Switch to Qwen2.5-72B-Instruct with 4xL4 GPU tensor parallelism and NF4 quantization
Prashanth Ramakrishna82e4ccc -
Fix TGI Numba caching issues - comprehensive environment setup
Prashanth Ramakrishnaab82ad3 -
Fix TGI Docker image - use v1.4 with Python dependencies
Prashanth Ramakrishna457a3f0 -
Replace Ollama with TGI for high-performance inference
Prashanth Ramakrishna206ee47 -
Revert to ENTRYPOINT and fix Ollama permissions with /tmp directory
Prashanth Ramakrishna7d1b0bb -
Fix startup by using CMD instead of ENTRYPOINT for HF Spaces
Prashanth Ramakrishnaad3212e -
Fix Ollama permissions by using /tmp directory
Prashanth Ramakrishnad997b8f -
Fix entrypoint to ensure startup script executes
Prashanth Ramakrishnaa411574 -
Initial deployment of Annotation Army Model Zoo with Ollama and multiple LLMs
Prashanth Ramakrishna25eda95