Spaces:
Sleeping
Sleeping
Commit History
-
Fix inference: switch to OpenAI-compatible /v1/chat/completions endpoint
The previous code used the legacy text-generation API with a manually formatted prompt string. That format does not work reliably with PEFT adapter models. The /v1/chat/completions endpoint applies the model's own chat template automatically and streams tokens in standard OpenAI SSE format (choices[0].delta.content). Co-authored-by: Cursor <cursoragent@cursor.com>b02f0c1 -
Initial setup: React + Express streaming chat for Qwen 2.5 7B
- Vite/React frontend with SSE streaming and direct DOM rendering at 50 tok/s - Express backend proxy with Qwen ChatML template and HF Inference API - Multi-stage Dockerfile optimised for Hugging Face Spaces (port 7860) Co-authored-by: Cursor <cursoragent@cursor.com>9b54db2