diffusiongemma / app.py
multimodalart's picture
multimodalart HF Staff
Update app.py
6834b26 verified
Raw
History Blame Contribute Delete
10.7 kB
import spaces
import torch
import gradio as gr
from threading import Thread
from queue import Queue
from transformers import DiffusionGemmaForBlockDiffusion, AutoProcessor, TextDiffusionStreamer
MODEL_ID = "google/diffusiongemma-26B-A4B-it"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = DiffusionGemmaForBlockDiffusion.from_pretrained(MODEL_ID, dtype=torch.bfloat16)
model.to("cuda")
model.eval()
CANVAS_LENGTH = getattr(model.config, "canvas_length", 256)
_SENTINEL = object()
# Per-token denoising colors. DiffusionGemma uses random-token *renoising* (not [MASK]
# diffusion): the entropy sampler locks low-entropy positions while the rest are random
# noise each step. So a position's stability across steps is our confidence proxy.
COLOR_MAP = {
"done": "#66CC66", # committed block (final)
"stable": "#8FD18F", # settled for several steps
"mid": "#FFCC66", # settling
"noise": "#E8896B", # just changed / still noisy
}
_STABLE_STEPS = 3 # unchanged for >= this many steps -> "stable"
class CanvasStreamer(TextDiffusionStreamer):
"""Pushes (committed_text, draft_segments, draft_plain) snapshots to a queue.
`put_draft` fires every denoising step with the full argmax canvas of the block
being denoised. We track, per position, how many consecutive steps its token has
been unchanged ("settle" count) and color it accordingly, so the canvas visibly
condenses from noise into settled text. `put` fires when a block is committed.
"""
def __init__(self, tokenizer, **kwargs):
super().__init__(tokenizer, skip_special_tokens=True, **kwargs)
self.queue = Queue()
self.committed = ""
self.last_draft = ""
self.started = False
self._takes_logits = False
self.prev_ids = None
self.settle = None
self.special_ids = set(tokenizer.all_special_ids)
def _render(self, ids):
"""Color positions up to the furthest settled real token (the 'frontier')."""
frontier = -1
for i, tid in enumerate(ids):
if tid not in self.special_ids and self.settle[i] >= 2:
frontier = i
segments = []
plain = []
cur_text = ""
cur_cls = None
for i in range(frontier + 1):
tid = ids[i]
if tid in self.special_ids:
continue
piece = self.tokenizer.decode([tid], skip_special_tokens=True)
if not piece:
continue
s = self.settle[i]
cls = "stable" if s >= _STABLE_STEPS else ("mid" if s >= 1 else "noise")
plain.append(piece)
if cls == cur_cls:
cur_text += piece
else:
if cur_text:
segments.append((cur_text, cur_cls))
cur_text, cur_cls = piece, cls
if cur_text:
segments.append((cur_text, cur_cls))
return segments, "".join(plain)
def put_draft(self, value, **kwargs):
if len(value.shape) > 1:
value = value[0]
ids = value.tolist()
self.started = True
if self.prev_ids is None or len(self.prev_ids) != len(ids):
self.settle = [0] * len(ids)
else:
for i, tid in enumerate(ids):
self.settle[i] = self.settle[i] + 1 if tid == self.prev_ids[i] else 0
self.prev_ids = ids
segments, plain = self._render(ids)
self.last_draft = plain
self.queue.put((self.committed, segments, plain))
def put(self, value):
if len(value.shape) > 1 and value.shape[0] > 1:
raise ValueError("batch size 1 only")
elif len(value.shape) > 1:
value = value[0]
if not self.started: # prompt context, before any denoising step
return
self.committed += self.tokenizer.decode(value, skip_special_tokens=True)
self.last_draft = ""
self.prev_ids = None
self.settle = None
self.queue.put((self.committed, [], ""))
def end(self):
self.queue.put(_SENTINEL)
def build_display(committed, segments):
out = []
if committed:
out.append((committed, "done"))
out.extend(segments)
return out
@spaces.GPU(duration=150, size="xlarge")
@torch.no_grad()
def generate_streaming(messages, max_new_tokens, max_denoising_steps, enable_thinking):
inputs = processor.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt",
enable_thinking=enable_thinking,
).to("cuda")
streamer = CanvasStreamer(processor.tokenizer)
result = {}
def run():
try:
out = model.generate(
**inputs,
max_new_tokens=int(max_new_tokens),
max_denoising_steps=int(max_denoising_steps),
streamer=streamer,
)
result["ids"] = out
except Exception as e:
result["error"] = e
streamer.queue.put(_SENTINEL)
thread = Thread(target=run)
thread.start()
while True:
item = streamer.queue.get()
if item is _SENTINEL:
break
committed, segments, plain = item
yield build_display(committed, segments), (committed + plain), None
thread.join()
if "error" in result:
raise result["error"]
final_text = (streamer.committed + streamer.last_draft).strip()
yield [(final_text, "done")] if final_text else [], final_text, final_text
def _file_path(item):
"""Extract a local file path from a Gradio content part / file dict."""
for key in ("file", "path", "url"):
val = item.get(key)
if isinstance(val, dict):
val = val.get("path") or val.get("url")
if val:
return val
return None
def to_model_messages(history):
"""Convert Gradio (messages format) history into processor chat format with images."""
messages = []
for msg in history:
role = msg["role"]
content = msg["content"]
parts = []
if isinstance(content, str):
parts.append({"type": "text", "text": content})
elif isinstance(content, tuple): # legacy (path, alt) file tuple
parts.append({"type": "image", "url": content[0]})
elif isinstance(content, dict): # single file part, e.g. {"path": ...}
p = _file_path(content)
if p:
parts.append({"type": "image", "url": p})
elif isinstance(content, list):
for item in content:
if not isinstance(item, dict):
parts.append({"type": "text", "text": str(item)})
elif item.get("type") == "text" or "text" in item:
parts.append({"type": "text", "text": item.get("text", "")})
else:
p = _file_path(item)
if p:
parts.append({"type": "image", "url": p})
if parts:
messages.append({"role": role, "content": parts})
return messages
css = """
.category-legend{display:none}
.legend{margin-bottom: 5px}
.legend-item{height: 25px}
"""
def create_demo():
with gr.Blocks(title="DiffusionGemma", css=css) as demo:
gr.Markdown("# DiffusionGemma 26B-A4B — Block Diffusion Chat")
gr.Markdown(
"[model](https://huggingface.co/google/diffusiongemma-26B-A4B-it) · "
"Watch the canvas denoise in real time on the right — text condenses from "
"noise (orange) into settled output (green). Attach an image to ask about it."
)
with gr.Row():
with gr.Column(scale=3):
chatbot = gr.Chatbot(label="Conversation")
chat_input = gr.MultimodalTextbox(
interactive=True,
file_types=["image"],
placeholder="Type a message and/or attach an image…",
show_label=False,
)
with gr.Column(scale=2):
canvas = gr.HighlightedText(
label="Denoising canvas",
combine_adjacent=False,
show_legend=True,
color_map=COLOR_MAP,
)
with gr.Accordion("Generation settings", open=False):
with gr.Row():
enable_thinking = gr.Checkbox(value=False, label="Thinking mode")
with gr.Row():
max_new_tokens = gr.Slider(64, 1024, value=256, step=64, label="Max new tokens")
max_denoising_steps = gr.Slider(8, 64, value=48, step=4, label="Max denoising steps")
clear_btn = gr.Button("Clear conversation")
def add_message(message, history):
history = history or []
for f in message.get("files", []):
history.append({"role": "user", "content": {"path": f}})
if message.get("text"):
history.append({"role": "user", "content": message["text"]})
return history, gr.MultimodalTextbox(value=None, interactive=False)
def bot(history, max_new_tokens, max_denoising_steps, enable_thinking):
if not history:
yield history, []
return
messages = to_model_messages(history)
history = history + [{"role": "assistant", "content": ""}]
final = None
try:
for canvas_state, plain, text in generate_streaming(
messages, max_new_tokens, max_denoising_steps, enable_thinking
):
if text is not None:
final = text
history[-1]["content"] = final if final is not None else plain
yield history, canvas_state
except Exception as e:
history[-1]["content"] = f"Error: {e}"
yield history, [(str(e), "noise")]
def reenable():
return gr.MultimodalTextbox(interactive=True)
chat_msg = chat_input.submit(
add_message, [chat_input, chatbot], [chatbot, chat_input]
)
bot_msg = chat_msg.then(
bot,
[chatbot, max_new_tokens, max_denoising_steps, enable_thinking],
[chatbot, canvas],
)
bot_msg.then(reenable, None, [chat_input])
clear_btn.click(lambda: ([], []), None, [chatbot, canvas])
return demo
demo = create_demo()
demo.queue()
if __name__ == "__main__":
demo.launch()