Upload folder using huggingface_hub

README.md

@@ -1,13 +1,23 @@
 ---
-title: Diffusiongemma Demo
-emoji: 🔥
-colorFrom: green
-colorTo: blue
+title: DiffusionGemma 26B Chat
+emoji: 🌀
+colorFrom: blue
+colorTo: purple
 sdk: gradio
 sdk_version: 6.17.3
-python_version: '3.13'
 app_file: app.py
 pinned: false
+license: apache-2.0
+suggested_hardware: zero-a10g
+models:
+  - google/diffusiongemma-26B-A4B-it
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# DiffusionGemma 26B Chat
+
+Chat demo for [google/diffusiongemma-26B-A4B-it](https://huggingface.co/google/diffusiongemma-26B-A4B-it),
+a 26B Mixture-of-Experts (3.8B active) multimodal model that generates text via
+discrete block diffusion — denoising 256-token blocks in parallel instead of
+generating token by token.
+
+Supports text and image inputs. Runs on ZeroGPU.

app.py

@@ -0,0 +1,72 @@
+import gradio as gr
+import spaces
+import torch
+from transformers import AutoProcessor, DiffusionGemmaForBlockDiffusion
+
+MODEL_ID = "google/diffusiongemma-26B-A4B-it"
+
+processor = AutoProcessor.from_pretrained(MODEL_ID)
+model = DiffusionGemmaForBlockDiffusion.from_pretrained(
+    MODEL_ID,
+    dtype="auto",
+    device_map="auto",
+)
+
+
+def build_messages(message, history):
+    messages = []
+    for turn in history:
+        content = turn["content"]
+        if isinstance(content, str):
+            messages.append(
+                {"role": turn["role"], "content": [{"type": "text", "text": content}]}
+            )
+        else:
+            # Gradio stores uploaded files in history as (filepath,) tuples
+            for path in content:
+                messages.append(
+                    {"role": turn["role"], "content": [{"type": "image", "image": path}]}
+                )
+
+    user_content = [{"type": "image", "image": f} for f in message.get("files", [])]
+    user_content.append({"type": "text", "text": message["text"]})
+    messages.append({"role": "user", "content": user_content})
+    return messages
+
+
+@spaces.GPU(duration=120)
+def respond(message, history, max_new_tokens):
+    messages = build_messages(message, history)
+    inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=True,
+        return_tensors="pt",
+    ).to(model.device)
+    input_len = inputs["input_ids"].shape[-1]
+
+    with torch.inference_mode():
+        output = model.generate(**inputs, max_new_tokens=max_new_tokens)
+
+    return processor.decode(output[0][input_len:], skip_special_tokens=True)
+
+
+demo = gr.ChatInterface(
+    respond,
+    type="messages",
+    multimodal=True,
+    title="DiffusionGemma 26B-A4B-it",
+    description=(
+        "Chat with [google/diffusiongemma-26B-A4B-it]"
+        "(https://huggingface.co/google/diffusiongemma-26B-A4B-it) — a 26B MoE "
+        "(3.8B active) model that generates text via discrete block diffusion. "
+        "Supports image input."
+    ),
+    additional_inputs=[
+        gr.Slider(minimum=256, maximum=2048, value=512, step=256, label="Max new tokens"),
+    ],
+)
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,4 @@
+git+https://github.com/huggingface/transformers.git
+torch
+accelerate
+spaces