skram/bioclip-2.5-vith14-endpoint

Roll back: remove OpenAI shim test

Roll back handler.py to pre-OpenAI version (restore commit 7875631 content: fp16 autocast handler, no OpenAI shim)

Add OpenAI-shim test script (same shim validated on bioclip-2; 1024-dim variant)

Add OpenAI embeddings-format compatibility (input/messages -> OpenAI envelope, float/base64 encoding) alongside existing inputs format

Fix fp16 CUDA dtype mismatch: run encoders under torch.autocast instead of bare .half() weights (found by GPU smoke test)

Pin handler dependency versions resolved and validated in sandbox test

Add reproducible handler test script (validated on CPU: classify/embed/embed_text pass)

Add requirements for the custom handler

Add custom Inference Endpoints handler for BioCLIP 2.5 (open_clip ViT-H/14)

Remove duplicate .bin weights (identical to safetensors) to halve endpoint cold-start download

Duplicate from imageomics/bioclip-2.5-vith14