Migrate DiT to dedicated Inference Endpoint (L40S, scale-to-zero)

app.py: /api/vraminfo probes actual ZeroGPU VRAM slice

Pipeline: DiT resident on GPU + async job endpoints

app.py: add /api/envinfo for flash-attn wheel selection

app.py: bump @spaces.GPU duration 300s→900s

DiT: bypass diffusers entirely, use internal PSM pipeline byte-identically

Rebuild QwenEmbedRope pos/neg freqs on CPU after meta init

Attach cache_context no-op to internal transformer

PE: SSE streaming (/api/generate_sp/stream) + SPA live token render

Attach .dtype/.device properties to internal transformer

Swap diffusers' transformer for internal modeling.qwenimage impl

DiT: compact-single-quote SP normalization on the Space side too

DiT: only overlay dit_model.*, skip text_encoder (matches EvalKit)

PE input_template: '<prompt>' (raw), drop [width/height] suffix

PE JSON extractor: split duplicate-id dicts + json_repair fallback

Space /api/generate_sp: retry-with-backoff on endpoint 503 (cold start)

Align PE inference with EvalKit serve_and_eval defaults

Overlay SP-fine-tuned DiT + text-encoder onto base Qwen-Image

Wire /api/generate_sp to HF Inference Endpoint (vLLM+Qwen3.5-35B-A3B)

Prepend a Route('/') instead of middleware (fastapi frozen post-launch)

Drop show_api (gradio 6 removed it)

Let Gradio be primary launcher; attach fastapi routes to demo.app

Revert to sdk: gradio, mount fastapi via ssr_mode=False

Switch to sdk: docker (Gradio SDK monkey-patches uvicorn.run)

Add hidden gradio.Blocks so HF ZeroGPU detects @spaces.GPU

Switch to FastAPI + SPA + diffusers DiT

PE model smoke test: vLLM FP8 + @spaces.GPU

initial commit