Spaces:
Paused
Paused
Commit History
-
Fix: Use single vLLM engine with continuous batching
- Detect available GPUs dynamically - Use vLLM's built-in continuous batching for concurrent requests - Remove multiprocessing approach that failed on HF Spaces - Use tensor_parallel_size for multi-GPU when available Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>4161fbd -
Reorder requirements to prioritize pydantic version
Put pydantic and vllm first to ensure pydantic>=2.12.0 is installed before gradio pulls in older version Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>0ab0126 -
Fix pydantic version for vllm compatibility
vllm 0.15.0 requires pydantic>=2.12.0 Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>099f1c5 -
Initial commit: Parallel request processing for JPharmatron
- 8 parallel input/output boxes with streaming - Multi-GPU architecture (one vLLM engine per GPU) - Multiprocessing workers for true parallelism - Individual and batch execution modes Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>86da8ad
