Spaces:
Running
on
Zero
Running
on
Zero
Commit
·
12437ad
1
Parent(s):
c1e0817
init
Browse files
app.py
CHANGED
@@ -10,7 +10,7 @@ import time
|
|
10 |
import os
|
11 |
from typing import List, Dict, Optional, Tuple, Iterator, Set
|
12 |
import gradio as gr
|
13 |
-
import spaces #
|
14 |
|
15 |
# Suppress some Hugging Face warnings
|
16 |
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
@@ -24,6 +24,7 @@ from model_cache.llada.configuration_llada import LLaDAConfig
|
|
24 |
def set_seed(seed):
|
25 |
torch.manual_seed(seed); random.seed(seed); np.random.seed(seed);
|
26 |
if torch.cuda.is_available(): torch.cuda.manual_seed_all(seed); torch.backends.cudnn.deterministic = True; torch.backends.cudnn.benchmark = False
|
|
|
27 |
def create_full_block_attention_mask(prompt_length, max_length, block_size, device=None, dtype=None):
|
28 |
if dtype is None: dtype = torch.bfloat16
|
29 |
attention_mask = torch.full((1, 1, max_length, max_length), -torch.inf, device=device, dtype=dtype)
|
@@ -38,12 +39,14 @@ def create_full_block_attention_mask(prompt_length, max_length, block_size, devi
|
|
38 |
attention_mask[:, :, block_start:block_end, prev_start:prev_end] = 0
|
39 |
attention_mask[:, :, block_start:block_end, block_start:block_end] = 0
|
40 |
return attention_mask
|
|
|
41 |
def extract_attention_mask(full_mask, start_pos, input_length, cache_length):
|
42 |
end_pos = start_pos + input_length; total_length = cache_length + input_length
|
43 |
extracted_mask = torch.full((1, 1, input_length, total_length), -torch.inf, device=full_mask.device, dtype=full_mask.dtype)
|
44 |
extracted_mask[:, :, :, :cache_length] = full_mask[:, :, start_pos:end_pos, :cache_length]
|
45 |
extracted_mask[:, :, :, cache_length:] = full_mask[:, :, start_pos:end_pos, start_pos:end_pos]
|
46 |
return extracted_mask
|
|
|
47 |
def top_p_logits(logits, top_p=None):
|
48 |
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
|
49 |
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
|
@@ -54,11 +57,13 @@ def top_p_logits(logits, top_p=None):
|
|
54 |
mask = mask.scatter_(-1, sorted_indices, sorted_indices_to_remove)
|
55 |
logits = logits.masked_fill(mask, torch.finfo(logits.dtype).min)
|
56 |
return logits
|
|
|
57 |
def top_k_logits(logits, top_k=None):
|
58 |
top_k = min(top_k, logits.size(-1))
|
59 |
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
|
60 |
logits = logits.masked_fill(indices_to_remove, torch.finfo(logits.dtype).min)
|
61 |
return logits
|
|
|
62 |
def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None, margin_confidence=False, neg_entropy=False):
|
63 |
if temperature > 0: logits = logits / temperature
|
64 |
if top_p is not None and top_p < 1: logits = top_p_logits(logits, top_p)
|
@@ -178,39 +183,46 @@ class DreamLoRAInference:
|
|
178 |
|
179 |
def __init__(self, **kwargs):
|
180 |
print("Initializing DreamLoRAInference...")
|
|
|
|
|
|
|
|
|
181 |
self.device = torch.device(kwargs.get("device", "cuda") if torch.cuda.is_available() else "cpu")
|
182 |
-
|
183 |
-
if
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
|
189 |
def _setup_model(self, pretrained_path, lora_path):
|
190 |
-
# --- MODIFICATION START ---
|
191 |
-
# The arguments `trust_remote_code=True` have been removed as they are not needed here
|
192 |
-
# and were causing warnings in the log.
|
193 |
config = LLaDAConfig.from_pretrained(pretrained_path)
|
194 |
self.model = LLaDAModelLM.from_pretrained(
|
195 |
pretrained_path,
|
196 |
config=config,
|
197 |
torch_dtype=self.target_dtype,
|
198 |
-
# device_map="auto" is handled by accelerate for better memory management on Spaces
|
199 |
device_map="auto"
|
200 |
).eval()
|
201 |
-
|
202 |
-
# THIS IS THE CRITICAL FIX: Tie the weights before loading the adapter.
|
203 |
-
# This resolves the error message from the log and allows `device_map="auto"` to work correctly.
|
204 |
-
# print("Tying model weights...")
|
205 |
-
# self.model.tie_weights()
|
206 |
-
# print("Weights tied.")
|
207 |
|
208 |
-
# Now, load the PEFT adapter on top of the correctly configured base model
|
209 |
self.model = PeftModel.from_pretrained(self.model, lora_path)
|
210 |
-
# --- MODIFICATION END ---
|
211 |
-
|
212 |
self.tokenizer = AutoTokenizer.from_pretrained(pretrained_path)
|
213 |
-
if self.tokenizer.pad_token is None:
|
|
|
214 |
|
215 |
def _apply_chat_template(self, prompt):
|
216 |
chat_history = [{"role": "user", "content": prompt}]
|
@@ -225,13 +237,11 @@ class DreamLoRAInference:
|
|
225 |
if (next_block_id := block_id + 1) in block_states:
|
226 |
block_states[next_block_id]['is_complete'] = True
|
227 |
|
228 |
-
# The rest of your class methods (_render_visualization_html, _render_status_html, stream_and_capture_for_gradio)
|
229 |
-
# remain completely unchanged.
|
230 |
def _render_visualization_html(self, step: int, x_t: torch.Tensor, block_states: Dict, cache_length: int, updated_block_ids: Set[int]) -> str:
|
231 |
timestamp = int(time.time() * 1000)
|
232 |
|
233 |
html_parts = []
|
234 |
-
for block_id in sorted(k for k in block_states.keys() if k > 0):
|
235 |
state = block_states[block_id]
|
236 |
container_classes = ["block-container"]
|
237 |
if block_id in updated_block_ids: container_classes.append("block-updating")
|
@@ -370,7 +380,7 @@ class DreamLoRAInference:
|
|
370 |
|
371 |
return complete_html
|
372 |
|
373 |
-
@spaces.GPU #
|
374 |
@torch.inference_mode()
|
375 |
def stream_and_capture_for_gradio(
|
376 |
self,
|
@@ -382,6 +392,9 @@ class DreamLoRAInference:
|
|
382 |
skip_threshold: float
|
383 |
) -> Iterator[Tuple[str, List[Tuple[str, str]], str, str, str]]:
|
384 |
|
|
|
|
|
|
|
385 |
start_time = time.time()
|
386 |
captured_frames: List[Tuple[str, str]] = []
|
387 |
|
@@ -396,7 +409,7 @@ class DreamLoRAInference:
|
|
396 |
|
397 |
# Capture initial state
|
398 |
initial_viz_html = self._render_visualization_html(0, x_t, block_states, 0, set())
|
399 |
-
initial_status_html = self._render_status_html(0,
|
400 |
captured_frames.append((initial_viz_html, initial_status_html))
|
401 |
|
402 |
yield "", captured_frames, "Initializing generation process...", "Initializing visualization...", "Initializing block status..."
|
@@ -507,6 +520,8 @@ if __name__ == "__main__":
|
|
507 |
"sampling_strategy": "default",
|
508 |
}
|
509 |
set_seed(42)
|
|
|
|
|
510 |
inference_engine = DreamLoRAInference(**config)
|
511 |
|
512 |
def animate_visualization(html_frames_list: List[Tuple[str, str]], delay: float) -> Iterator[Tuple[str, str]]:
|
|
|
10 |
import os
|
11 |
from typing import List, Dict, Optional, Tuple, Iterator, Set
|
12 |
import gradio as gr
|
13 |
+
import spaces # 导入 spaces 模块
|
14 |
|
15 |
# Suppress some Hugging Face warnings
|
16 |
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
|
|
24 |
def set_seed(seed):
|
25 |
torch.manual_seed(seed); random.seed(seed); np.random.seed(seed);
|
26 |
if torch.cuda.is_available(): torch.cuda.manual_seed_all(seed); torch.backends.cudnn.deterministic = True; torch.backends.cudnn.benchmark = False
|
27 |
+
|
28 |
def create_full_block_attention_mask(prompt_length, max_length, block_size, device=None, dtype=None):
|
29 |
if dtype is None: dtype = torch.bfloat16
|
30 |
attention_mask = torch.full((1, 1, max_length, max_length), -torch.inf, device=device, dtype=dtype)
|
|
|
39 |
attention_mask[:, :, block_start:block_end, prev_start:prev_end] = 0
|
40 |
attention_mask[:, :, block_start:block_end, block_start:block_end] = 0
|
41 |
return attention_mask
|
42 |
+
|
43 |
def extract_attention_mask(full_mask, start_pos, input_length, cache_length):
|
44 |
end_pos = start_pos + input_length; total_length = cache_length + input_length
|
45 |
extracted_mask = torch.full((1, 1, input_length, total_length), -torch.inf, device=full_mask.device, dtype=full_mask.dtype)
|
46 |
extracted_mask[:, :, :, :cache_length] = full_mask[:, :, start_pos:end_pos, :cache_length]
|
47 |
extracted_mask[:, :, :, cache_length:] = full_mask[:, :, start_pos:end_pos, start_pos:end_pos]
|
48 |
return extracted_mask
|
49 |
+
|
50 |
def top_p_logits(logits, top_p=None):
|
51 |
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
|
52 |
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
|
|
|
57 |
mask = mask.scatter_(-1, sorted_indices, sorted_indices_to_remove)
|
58 |
logits = logits.masked_fill(mask, torch.finfo(logits.dtype).min)
|
59 |
return logits
|
60 |
+
|
61 |
def top_k_logits(logits, top_k=None):
|
62 |
top_k = min(top_k, logits.size(-1))
|
63 |
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
|
64 |
logits = logits.masked_fill(indices_to_remove, torch.finfo(logits.dtype).min)
|
65 |
return logits
|
66 |
+
|
67 |
def sample_tokens(logits, temperature=0.0, top_p=None, top_k=None, margin_confidence=False, neg_entropy=False):
|
68 |
if temperature > 0: logits = logits / temperature
|
69 |
if top_p is not None and top_p < 1: logits = top_p_logits(logits, top_p)
|
|
|
183 |
|
184 |
def __init__(self, **kwargs):
|
185 |
print("Initializing DreamLoRAInference...")
|
186 |
+
# 只保存配置,不立即加载模型
|
187 |
+
self.config = kwargs
|
188 |
+
self.model = None
|
189 |
+
self.tokenizer = None
|
190 |
self.device = torch.device(kwargs.get("device", "cuda") if torch.cuda.is_available() else "cpu")
|
191 |
+
|
192 |
+
if kwargs.get("dtype") == "bfloat16" and torch.cuda.is_bf16_supported():
|
193 |
+
self.target_dtype = torch.bfloat16
|
194 |
+
elif kwargs.get("dtype") == "float16":
|
195 |
+
self.target_dtype = torch.float16
|
196 |
+
else:
|
197 |
+
self.target_dtype = torch.float32
|
198 |
+
|
199 |
+
# 从配置中获取其他属性
|
200 |
+
for key, value in kwargs.items():
|
201 |
+
if not hasattr(self, key):
|
202 |
+
setattr(self, key, value)
|
203 |
+
|
204 |
+
print("DreamLoRAInference initialized (model will be loaded on first use).")
|
205 |
+
|
206 |
+
def _ensure_model_loaded(self):
|
207 |
+
"""延迟加载模型,只在真正需要时加载"""
|
208 |
+
if self.model is None:
|
209 |
+
print("Loading model for the first time...")
|
210 |
+
self._setup_model(self.config["pretrained_path"], self.config["lora_path"])
|
211 |
+
print("Model and tokenizer setup complete.")
|
212 |
|
213 |
def _setup_model(self, pretrained_path, lora_path):
|
|
|
|
|
|
|
214 |
config = LLaDAConfig.from_pretrained(pretrained_path)
|
215 |
self.model = LLaDAModelLM.from_pretrained(
|
216 |
pretrained_path,
|
217 |
config=config,
|
218 |
torch_dtype=self.target_dtype,
|
|
|
219 |
device_map="auto"
|
220 |
).eval()
|
|
|
|
|
|
|
|
|
|
|
|
|
221 |
|
|
|
222 |
self.model = PeftModel.from_pretrained(self.model, lora_path)
|
|
|
|
|
223 |
self.tokenizer = AutoTokenizer.from_pretrained(pretrained_path)
|
224 |
+
if self.tokenizer.pad_token is None:
|
225 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
226 |
|
227 |
def _apply_chat_template(self, prompt):
|
228 |
chat_history = [{"role": "user", "content": prompt}]
|
|
|
237 |
if (next_block_id := block_id + 1) in block_states:
|
238 |
block_states[next_block_id]['is_complete'] = True
|
239 |
|
|
|
|
|
240 |
def _render_visualization_html(self, step: int, x_t: torch.Tensor, block_states: Dict, cache_length: int, updated_block_ids: Set[int]) -> str:
|
241 |
timestamp = int(time.time() * 1000)
|
242 |
|
243 |
html_parts = []
|
244 |
+
for block_id in sorted(k for k in block_states.keys() if k > 0):
|
245 |
state = block_states[block_id]
|
246 |
container_classes = ["block-container"]
|
247 |
if block_id in updated_block_ids: container_classes.append("block-updating")
|
|
|
380 |
|
381 |
return complete_html
|
382 |
|
383 |
+
@spaces.GPU # 关键修复:GPU 装饰器
|
384 |
@torch.inference_mode()
|
385 |
def stream_and_capture_for_gradio(
|
386 |
self,
|
|
|
392 |
skip_threshold: float
|
393 |
) -> Iterator[Tuple[str, List[Tuple[str, str]], str, str, str]]:
|
394 |
|
395 |
+
# 确保模型已加载
|
396 |
+
self._ensure_model_loaded()
|
397 |
+
|
398 |
start_time = time.time()
|
399 |
captured_frames: List[Tuple[str, str]] = []
|
400 |
|
|
|
409 |
|
410 |
# Capture initial state
|
411 |
initial_viz_html = self._render_visualization_html(0, x_t, block_states, 0, set())
|
412 |
+
initial_status_html = self._render_status_html(0, block_states, 0)
|
413 |
captured_frames.append((initial_viz_html, initial_status_html))
|
414 |
|
415 |
yield "", captured_frames, "Initializing generation process...", "Initializing visualization...", "Initializing block status..."
|
|
|
520 |
"sampling_strategy": "default",
|
521 |
}
|
522 |
set_seed(42)
|
523 |
+
|
524 |
+
# 创建推理引擎但不立即加载模型
|
525 |
inference_engine = DreamLoRAInference(**config)
|
526 |
|
527 |
def animate_visualization(html_frames_list: List[Tuple[str, str]], delay: float) -> Iterator[Tuple[str, str]]:
|