Upload folder using huggingface_hub

README.md

@@ -1,5 +1,5 @@
 ---
-title: Qwen Image 2.1 Studio
+title: Qwen Image 2.1 GGUF Studio
 emoji: 🎨
 colorFrom: indigo
 colorTo: purple
@@ -7,15 +7,16 @@ sdk: gradio
 sdk_version: 6.28.0
 python_version: "3.12"
 app_file: app.py
-short_description: Create, edit, and generate transparent images with Qwen
+short_description: Qwen Image 2.1 Uncensored GGUF demo with Q4_K_M weights
 startup_duration_timeout: 1h
 models:
+  - abenzerps/Qwen-Image-2.1-Uncensored-GGUF
   - Qwen/Qwen-Image-2.1
 ---
 
-# Qwen Image 2.1 Studio
+# Qwen Image 2.1 GGUF Studio
 
-A Gradio demo of [Qwen/Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1), running the model directly on Hugging Face ZeroGPU.
+A Gradio demo of [abenzerps/Qwen-Image-2.1-Uncensored-GGUF](https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-GGUF), using its recommended **Q4_K_M** checkpoint on Hugging Face ZeroGPU. This replaces the original diffusion transformer in this Space.
 
 - **Create an image:** describe a scene or choose an example.
 - **Edit an image:** upload one reference image and describe the change.
@@ -26,7 +27,13 @@ The default is the model card's recommended 40 inference steps; lower this for q
 
 ## Hosting
 
-This Space requires `zero-a10g` hardware. Configure hardware in the Space settings; README metadata does not select hardware. No external inference API key is needed. The model weights and Diffusers revision are pinned in the source. The initial download is approximately 33 GB and can take several minutes.
+This Space requires `zero-a10g` hardware. Configure hardware in the Space settings; README metadata does not select hardware. No external inference API key is needed. Model and library revisions are pinned in the source. The initial download is approximately 24 GB and can take several minutes.
+
+The 4.60 GB `qwen-image-2.1-Q4_K_M.gguf` file is checked against the publisher's SHA256 checksum. All 297 tensors are loaded with strict key and shape validation. The GGUF values are expanded to BF16 once at startup so ZeroGPU can pack and stream ordinary PyTorch tensors; the model does not remain compressed in GPU memory. This trades the runtime memory saving for predictable inference latency within the 60-second GPU reservation.
+
+The compatible text encoder, VAE, processor, scheduler, and transformer configuration come from the pinned upstream `Qwen/Qwen-Image-2.1` repository. Its diffusion transformer weights are neither downloaded nor used as a fallback. The publisher describes the GGUF as a quantization of the original upstream weights, not a separately trained model. The Q8_0 variant is not used because the model card reports a shape-mismatch issue.
+
+Example caching uses a separate directory for this GGUF revision, preventing results from the previous model from being served. PNG metadata records the exact repository, revision, filename, checksum, and generation settings.
 
 The model loads at startup and is registered with ZeroGPU. The Gradio queue runs one generation at a time. Visitors use Hugging Face's daily ZeroGPU quota. Uploaded and generated files are temporary; download results you want to keep. The app does not send images to an external API or publish a community gallery. Gradio cached files expire after 24 hours; cached examples can be reused across visitors.
 

app.py

@@ -1,21 +1,31 @@
 import os
 
 os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+os.environ["GRADIO_EXAMPLES_CACHE"] = ".gradio/gguf-q4-k-m-e92386d3-examples"
 
 import spaces
 import gradio as gr
 import torch
+import hashlib
 import json
 import random
 import tempfile
 import time
 from pathlib import Path
 
-from diffusers import QwenImage21Pipeline
+from accelerate import init_empty_weights
+from diffusers import QwenImage21Pipeline, QwenImage21Transformer2DModel
+from diffusers.models.model_loading_utils import load_gguf_checkpoint
+from diffusers.quantizers.gguf.utils import dequantize_gguf_tensor
+from huggingface_hub import hf_hub_download
 from PIL import Image, ImageOps, PngImagePlugin
 
-MODEL_ID = "Qwen/Qwen-Image-2.1"
-MODEL_REVISION = "790c92633540aa0cb11d9abf19eb46d861714758"
+MODEL_ID = "abenzerps/Qwen-Image-2.1-Uncensored-GGUF"
+MODEL_REVISION = "e92386d3b86fd77ae1650534815ab3f6b85d46ba"
+CHECKPOINT = "qwen-image-2.1-Q4_K_M.gguf"
+CHECKPOINT_SHA256 = "833439e91bc1152d28f37aa198c7f6f4218b7de95754c2f7a318a2422ab4b2f8"
+COMPANION_ID = "Qwen/Qwen-Image-2.1"
+COMPANION_REVISION = "790c92633540aa0cb11d9abf19eb46d861714758"
 MODES = ["Create an image", "Edit an image", "Transparent PNG"]
 SIZES = {
     "Square · 1:1": (1024, 1024),
@@ -27,10 +37,34 @@ SIZES = {
 MAX_SEED = 2**31 - 1
 
 print(f"Loading {MODEL_ID} at {MODEL_REVISION}", flush=True)
+checkpoint_path = hf_hub_download(MODEL_ID, CHECKPOINT, revision=MODEL_REVISION)
+with open(checkpoint_path, "rb") as checkpoint_file:
+    checksum = hashlib.file_digest(checkpoint_file, "sha256").hexdigest()
+if checksum != CHECKPOINT_SHA256:
+    raise RuntimeError("The GGUF checkpoint checksum does not match the published SHA256SUMS.")
+
+# Expand the requested GGUF once for ZeroGPU's eager tensor packing. This
+# preserves the quantized checkpoint's values without per-step dequantization.
+weights = load_gguf_checkpoint(checkpoint_path)
+for name in weights:
+    weights[name] = dequantize_gguf_tensor(weights[name]).to(torch.bfloat16)
+config = QwenImage21Transformer2DModel.load_config(
+    COMPANION_ID, subfolder="transformer", revision=COMPANION_REVISION
+)
+with init_empty_weights():
+    transformer = QwenImage21Transformer2DModel.from_config(config)
+transformer.load_state_dict(weights, strict=True, assign=True)
+transformer.eval().requires_grad_(False)
+print(f"Verified {CHECKPOINT}: loaded all {len(weights)} tensors; sha256={checksum}", flush=True)
+del weights
+
+# Supplying the transformer prevents the upstream diffusion weights from
+# being downloaded. Only its compatible text encoder, VAE and config are used.
 pipe = QwenImage21Pipeline.from_pretrained(
-    MODEL_ID, revision=MODEL_REVISION, torch_dtype=torch.bfloat16
+    COMPANION_ID, revision=COMPANION_REVISION, transformer=transformer,
+    torch_dtype=torch.bfloat16,
 ).to("cuda")
-print("Qwen-Image-2.1 pipeline loaded on CUDA via ZeroGPU.", flush=True)
+print(f"{MODEL_ID} / {CHECKPOINT} loaded on CUDA via ZeroGPU.", flush=True)
 
 
 @spaces.GPU(duration=60)
@@ -44,7 +78,7 @@ def generate(
     randomize_seed: bool = True,
     progress: gr.Progress = gr.Progress(track_tqdm=True),
 ) -> tuple[str, str, int, str]:
-    """Generate or edit an image with Qwen-Image-2.1 and return a PNG.
+    """Generate or edit an image with abenzerps' Qwen-Image-2.1 Q4_K_M GGUF.
 
     Choose Create an image, Edit an image (requires a reference), or
     Transparent PNG. Disable randomize_seed to reuse a seed. The result
@@ -94,6 +128,9 @@ def generate(
     metadata = {
         "model": MODEL_ID,
         "revision": MODEL_REVISION,
+        "checkpoint": CHECKPOINT,
+        "checkpoint_sha256": CHECKPOINT_SHA256,
+        "runtime_dtype": "bfloat16",
         "prompt": prompt,
         "effective_prompt": effective_prompt,
         "mode": mode,
@@ -106,12 +143,12 @@ def generate(
     png_info = PngImagePlugin.PngInfo()
     png_info.add_text("generation", json.dumps(metadata, ensure_ascii=False))
     # Gradio owns its output cache and expires files after 24 hours.
-    output_dir = Path(tempfile.mkdtemp(prefix="qwen21-", dir=gr.utils.get_upload_folder()))
-    path = output_dir / f"qwen-image-2.1-{actual_seed}.png"
+    output_dir = Path(tempfile.mkdtemp(prefix="qwen21-gguf-", dir=gr.utils.get_upload_folder()))
+    path = output_dir / f"qwen-image-2.1-gguf-{actual_seed}.png"
     result.save(path, pnginfo=png_info)
     details = (
         f"{result.width} × {result.height} · {int(steps)} steps · "
-        f"{elapsed:.1f}s · seed {actual_seed} · {result.mode} PNG"
+        f"{elapsed:.1f}s · seed {actual_seed} · {result.mode} PNG · Q4_K_M"
     )
     if mode == "Transparent PNG" and (
         "A" not in result.getbands() or result.getchannel("A").getextrema()[0] == 255
@@ -135,11 +172,11 @@ CSS = """
 #generate { min-height: 48px; }
 """
 
-with gr.Blocks(title="Qwen Image 2.1 Studio", delete_cache=(3600, 86400)) as demo:
+with gr.Blocks(title="Qwen Image 2.1 GGUF Studio", delete_cache=(3600, 86400)) as demo:
     gr.Markdown(
-        "# Qwen Image 2.1 Studio\n"
+        "# Qwen Image 2.1 GGUF Studio\n"
         "Turn an idea into an image. Reimagine a photo. Create a transparent asset.\n\n"
-        "[Model](https://huggingface.co/Qwen/Qwen-Image-2.1) · "
+        "[abenzerps / Qwen-Image-2.1-Uncensored-GGUF](https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-GGUF) · **Q4_K_M** · "
         "[Qwen Research License](https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE)",
         elem_id="intro",
     )

requirements.txt

@@ -6,3 +6,4 @@ safetensors
 Pillow
 sentencepiece
 mcp
+gguf==0.19.0