Upload model card

README.md

@@ -1,198 +1,153 @@
 ---
+license: openrail++
 library_name: diffusers
+pipeline_tag: text-to-image
+base_model: stabilityai/sd-turbo
+tags:
+  - diffusion
+  - text-to-image
+  - sd-turbo
+  - quantization
+  - pruning
+  - distillation
+  - edge-ai
+  - mixed-precision
 ---
 
-# Model Card for Model ID
+# EdgeDiffuse
 
-<!-- Provide a quick summary of what the model is/does. -->
+Edge-deployable SD-Turbo via multi-stage compression: structural pruning → distillation → sensitivity-aware mixed-precision quantization (GPTQ) → QLoRA recovery.
 
+**Code & paper-style writeup**: [github.com/SeanHe727/EdgeDiffusion](https://github.com/SeanHe727/EdgeDiffusion)
 
+---
 
-## Model Details
-
-### Model Description
-
-<!-- Provide a longer summary of what this model is. -->
-
-This is the model card of a 🧨 diffusers model that has been pushed on the Hub. This model card has been automatically generated.
-
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
-
-### Model Sources [optional]
-
-<!-- Provide the basic links for the model. -->
-
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
-
-## Uses
-
-<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
-
-### Direct Use
-
-<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
-
-[More Information Needed]
-
-### Downstream Use [optional]
-
-<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
-
-[More Information Needed]
-
-### Out-of-Scope Use
-
-<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
-
-[More Information Needed]
-
-## Bias, Risks, and Limitations
-
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-
-[More Information Needed]
-
-### Recommendations
-
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
-
-## How to Get Started with the Model
-
-Use the code below to get started with the model.
-
-[More Information Needed]
-
-## Training Details
-
-### Training Data
-
-<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
-
-[More Information Needed]
-
-### Training Procedure
-
-<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
-
-#### Preprocessing [optional]
-
-[More Information Needed]
-
-
-#### Training Hyperparameters
-
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
-
-#### Speeds, Sizes, Times [optional]
-
-<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
-
-[More Information Needed]
-
-## Evaluation
-
-<!-- This section describes the evaluation protocols and provides the results. -->
-
-### Testing Data, Factors & Metrics
-
-#### Testing Data
-
-<!-- This should link to a Dataset Card if possible. -->
-
-[More Information Needed]
-
-#### Factors
-
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
-
-[More Information Needed]
-
-#### Metrics
-
-<!-- These are the evaluation metrics being used, ideally with a description of why. -->
-
-[More Information Needed]
-
-### Results
-
-[More Information Needed]
-
-#### Summary
-
-
-
-## Model Examination [optional]
-
-<!-- Relevant interpretability work for the model goes here -->
-
-[More Information Needed]
-
-## Environmental Impact
+## What's in this repo
 
-<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+| File / dir | What it is |
+|---|---|
+| `unet/` | Mixed-precision quantized UNet (GPTQ-applied). 152 of 192 Linear layers quantized to INT4 (45) / INT8 (107); the rest stay fp16. Fake-quantized: values rounded to int grid, stored as bf16. |
+| `text_encoder/`, `vae/`, `tokenizer/`, `scheduler/`, `model_index.json` | Standard `stabilityai/sd-turbo` components, unmodified |
+| `lora_adapter.pt` | (Optional) QLoRA recovery adapter trained on top of the quantized UNet. Improves LPIPS by ~8 % when applied. See "Advanced: QLoRA recovery" below. |
+| `mp_quant_metadata.json` | Per-layer bit-width assignment + GPTQ hyper-parameters for full reproducibility |
 
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+---
 
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
+## Quick start
 
-## Technical Specifications [optional]
+```python
+from diffusers import StableDiffusionPipeline
+import torch
 
-### Model Architecture and Objective
+pipe = StableDiffusionPipeline.from_pretrained(
+    "ChenHe727/EdgeDiffusion",
+    torch_dtype=torch.bfloat16,        # required: INT4 layers use bf16 dtype
+)
+pipe = pipe.to("cuda")
 
-[More Information Needed]
+image = pipe(
+    "a photo of a tabby cat sitting on a wooden chair, sharp focus",
+    num_inference_steps=2,             # 2-step is the sweet spot for SD-Turbo derivatives
+    guidance_scale=0.0,                # SD-Turbo doesn't use CFG
+).images[0]
 
-### Compute Infrastructure
+image.save("output.png")
+```
 
-[More Information Needed]
+### Why 2 inference steps?
 
-#### Hardware
+SD-Turbo is fundamentally trained with **adversarial diffusion distillation** for 1-step generation. Empirically, 2 steps gives the best quality/speed trade-off for our compressed model: 28 % faster than 4 steps with marginally better LPIPS.
 
-[More Information Needed]
+---
 
-#### Software
+## Results
 
-[More Information Needed]
+Benchmark on RTX 5070 (Blackwell), 512 × 512, 2-step inference:
 
-## Citation [optional]
+| Variant | Params | Latency | VRAM | LPIPS vs original SD-Turbo | LPIPS vs fp16 baseline |
+|---|---:|---:|---:|---:|---:|
+| stabilityai/sd-turbo (original) | 860 M | 0.146 s | 3.05 GB | 0 | 0.278 |
+| fp16 baseline (pruned + distilled) | 642 M | 0.142 s | 2.64 GB | 0.278 | 0 |
+| **this repo (mp_quant PTQ)** | 642 M | 0.145 s | 2.64 GB | 0.277 | 0.062 |
+| with LoRA adapter loaded | 642 M + 9 MB | 0.171 s | 2.65 GB | 0.278 | **0.057** |
 
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+**Key takeaway**: mixed-precision quantization adds essentially **zero perceptual cost** on top of the pruned + distilled baseline (LPIPS 0.062 vs fp16). The dominant quality cost in the pipeline is the pruning stage; quantization is "free".
 
-**BibTeX:**
+---
 
-[More Information Needed]
+## Advanced: QLoRA recovery adapter
+
+The included `lora_adapter.pt` was trained for 500 steps with step-wise teacher-student distillation to recover residual PTQ quality loss. It reduces the LPIPS gap from 0.062 to 0.057 (~8 % improvement).
+
+```python
+import torch
+from peft import LoraConfig, get_peft_model
+from diffusers import StableDiffusionPipeline
+from huggingface_hub import hf_hub_download
+import json
+
+# Load base pipeline
+pipe = StableDiffusionPipeline.from_pretrained(
+    "ChenHe727/EdgeDiffusion", torch_dtype=torch.bfloat16,
+).to("cuda")
+
+# Discover which layers were quantized (LoRA targets these)
+meta_path = hf_hub_download("ChenHe727/EdgeDiffusion", "mp_quant_metadata.json")
+with open(meta_path) as f:
+    meta = json.load(f)
+target_fqns = [fqn for fqn, bit in meta["quantization"]["assignment"].items() if bit != "fp16"]
+
+# Re-attach LoRA structure and load adapter weights
+lora_state = torch.load(hf_hub_download("ChenHe727/EdgeDiffusion", "lora_adapter.pt"),
+                        weights_only=False, map_location="cuda")
+sample_key = next(k for k in lora_state if "lora_A" in k)
+rank = lora_state[sample_key].shape[0]
+
+pipe.unet = get_peft_model(pipe.unet, LoraConfig(
+    r=rank, lora_alpha=rank * 2, target_modules=target_fqns,
+    lora_dropout=0.0, bias="none",
+))
+own = pipe.unet.state_dict()
+for k, v in lora_state.items():
+    if k in own:
+        own[k].copy_(v.to(own[k].device, dtype=own[k].dtype))
+pipe.unet.eval()
+
+# Generate as usual
+image = pipe("a cat", num_inference_steps=2, guidance_scale=0.0).images[0]
+```
 
-**APA:**
+---
 
-[More Information Needed]
+## Pipeline overview
 
-## Glossary [optional]
+The model in this repo is the output of a three-stage compression pipeline applied to `stabilityai/sd-turbo`:
 
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+```
+stabilityai/sd-turbo (860 M)
+    ↓ structural pruning + step-wise distillation
+ChenHe727/EdgeDiffusion_distilled_feat_attn (642 M, fp16)
+    ↓ sensitivity-aware mixed-precision GPTQ (this repo's UNet)
+    ↓ QLoRA recovery training (this repo's lora_adapter.pt)
+ChenHe727/EdgeDiffusion (this repo)
+```
 
-[More Information Needed]
+Full design rationale, ablations, and reproducibility instructions: see the [GitHub repo](https://github.com/SeanHe727/EdgeDiffusion).
 
-## More Information [optional]
+---
 
-[More Information Needed]
+## Limitations
 
-## Model Card Authors [optional]
+- **Conv2d layers are not quantized in v1** — only `nn.Linear` (attention projections, FFN). Conv2d holds ~70 % of UNet parameters; full quantization is planned for v2.
+- **Fake-quant storage**: weights are rounded to INT4/INT8 grids but stored as bf16 (2 bytes/value). Real packed INT4/INT8 storage would shrink the file from 1.22 GB to ~900 MB but requires a separate packing step.
+- **LPIPS vs original SD-Turbo ≈ 0.28** mostly comes from the upstream pruning + distillation stage. The quantization stage itself adds only 0.005-0.062.
+- **2-step inference is the recommended default.** 1-step works (faster) but quality drops noticeably; 4-step is slower and not better.
 
-[More Information Needed]
+---
 
-## Model Card Contact
+## Acknowledgments
 
-[More Information Needed]
+- **LD-Pruner** ([Castells et al. 2024](https://arxiv.org/abs/2404.11936)) — sensitivity metric
+- **GPTQ** ([Frantar et al. 2023](https://arxiv.org/abs/2210.17323)) — Hessian-based PTQ (re-implemented from the paper in this repo)
+- **QLoRA** ([Dettmers et al. 2023](https://arxiv.org/abs/2305.14314)) — parameter-efficient recovery
+- **SD-Turbo** ([Sauer et al. 2023](https://stability.ai/research/adversarial-diffusion-distillation)) — base model