dreamlite-sd15-dummy-prequant (PLACEHOLDER / NOT A REAL MODEL)

⚠️ This is a dummy placeholder checkpoint. It is NOT a trained DreamLite model. It exists only to prove the w8a16 NPU quantize + export flow end-to-end.

What this is

A tiny, randomly-initialized SD1.5-shaped DreamLite checkpoint built locally to exercise the adapter → AIMET w8a16 → ONNX pipeline without real weights:

• unet/ — UNet2DConditionModel, SD1.5-shaped (block_out_channels=(320,320,320,320)), 8 input channels (latent + in-context reference latent concat), random weights.
• vae/ — AutoencoderKL SD1.5-shaped, random weights.
• text_encoder/ — a minimal Qwen3VLTextModel with the full Qwen tokenizer vocab (151936) but tiny hidden size (64), random weights. This is what makes the text-encoder calibration path runnable.
• tokenizer/ + tokenizer.json + tokenizer_config.json + chat_template.jinja — the real Qwen/Qwen3-0.6B tokenizer (needed so token ids fall inside the dummy embedding table).
• prompts.txt — 4 calibration prompts.

Why "pre-NPU-optimization"

This repo holds the unquantized raw checkpoint — the input to the quantize step. The NPU-optimized (w8a16, int8 weights / int16 activations) QNN context binaries are produced separately by the quantize flow and live in the local build/dummy_dreamlite/ output, not here.

Status

• Local w8a16 quantize + ONNX export: proven for unet, vae, text_encoder.
• Real DreamLite weights (Phase 1): not yet integrated — these are dummies.
• Cloud QNN context-binary compile (Phase 7): performed on the exported ONNX, not on this raw checkpoint.

Do not use this repo for inference; the weights are random.