Spaces:
Running
on
Zero
Running
on
Zero
Update interruption method
Browse files- utils/models.py +37 -9
utils/models.py
CHANGED
@@ -62,6 +62,7 @@ def generate_summaries(example, model_a_name, model_b_name):
|
|
62 |
Generates summaries for the given example using the assigned models sequentially.
|
63 |
"""
|
64 |
if generation_interrupt.is_set():
|
|
|
65 |
return "", ""
|
66 |
|
67 |
context_text = ""
|
@@ -69,6 +70,11 @@ def generate_summaries(example, model_a_name, model_b_name):
|
|
69 |
|
70 |
if "full_contexts" in example and example["full_contexts"]:
|
71 |
for i, ctx in enumerate(example["full_contexts"]):
|
|
|
|
|
|
|
|
|
|
|
72 |
content = ""
|
73 |
|
74 |
# Extract content from either dict or string
|
@@ -92,17 +98,22 @@ def generate_summaries(example, model_a_name, model_b_name):
|
|
92 |
question = example.get("question", "")
|
93 |
|
94 |
if generation_interrupt.is_set():
|
|
|
95 |
return "", ""
|
96 |
|
|
|
97 |
# Run model A
|
98 |
summary_a = run_inference(models[model_a_name], context_text, question)
|
99 |
|
100 |
if generation_interrupt.is_set():
|
|
|
101 |
return summary_a, ""
|
102 |
|
|
|
103 |
# Run model B
|
104 |
summary_b = run_inference(models[model_b_name], context_text, question)
|
105 |
|
|
|
106 |
return summary_a, summary_b
|
107 |
|
108 |
|
@@ -114,6 +125,7 @@ def run_inference(model_name, context, question):
|
|
114 |
"""
|
115 |
# Check interrupt at the beginning
|
116 |
if generation_interrupt.is_set():
|
|
|
117 |
return ""
|
118 |
|
119 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
@@ -134,6 +146,11 @@ def run_inference(model_name, context, question):
|
|
134 |
if model_name in tokenizer_cache:
|
135 |
tokenizer = tokenizer_cache[model_name]
|
136 |
else:
|
|
|
|
|
|
|
|
|
|
|
137 |
# Common arguments for tokenizer loading
|
138 |
tokenizer_load_args = {"padding_side": "left", "token": True}
|
139 |
|
@@ -155,6 +172,7 @@ def run_inference(model_name, context, question):
|
|
155 |
|
156 |
# Check interrupt before loading the model
|
157 |
if generation_interrupt.is_set():
|
|
|
158 |
return ""
|
159 |
|
160 |
# Create interrupt criteria for this generation
|
@@ -162,19 +180,21 @@ def run_inference(model_name, context, question):
|
|
162 |
|
163 |
print("REACHED HERE BEFORE pipe")
|
164 |
print(f"Loading model {model_name}...")
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
if "bitnet" in model_name.lower():
|
166 |
bitnet_model = BitNetForCausalLM.from_pretrained(
|
167 |
model_name,
|
168 |
-
#device_map="auto",
|
169 |
torch_dtype=torch.bfloat16,
|
170 |
-
#trust_remote_code=True,
|
171 |
)
|
172 |
pipe = pipeline(
|
173 |
"text-generation",
|
174 |
model=bitnet_model,
|
175 |
tokenizer=tokenizer,
|
176 |
-
#device_map="auto",
|
177 |
-
#trust_remote_code=True,
|
178 |
torch_dtype=torch.bfloat16,
|
179 |
model_kwargs={
|
180 |
"attn_implementation": "eager",
|
@@ -206,13 +226,20 @@ def run_inference(model_name, context, question):
|
|
206 |
torch_dtype=torch.bfloat16,
|
207 |
)
|
208 |
|
|
|
|
|
|
|
|
|
|
|
209 |
text_input = format_rag_prompt(question, context, accepts_sys)
|
|
|
|
|
210 |
if "Gemma-3".lower() in model_name.lower():
|
211 |
print("REACHED HERE BEFORE GEN")
|
212 |
result = pipe(
|
213 |
text_input,
|
214 |
max_new_tokens=512,
|
215 |
-
stopping_criteria=[interrupt_criteria],
|
216 |
generation_kwargs={"skip_special_tokens": True}
|
217 |
)[0]["generated_text"]
|
218 |
|
@@ -238,6 +265,7 @@ def run_inference(model_name, context, question):
|
|
238 |
with torch.inference_mode():
|
239 |
# Check interrupt before generation
|
240 |
if generation_interrupt.is_set():
|
|
|
241 |
return ""
|
242 |
|
243 |
output_sequences = model.generate(
|
@@ -246,7 +274,7 @@ def run_inference(model_name, context, question):
|
|
246 |
max_new_tokens=512,
|
247 |
eos_token_id=tokenizer.eos_token_id,
|
248 |
pad_token_id=tokenizer.pad_token_id,
|
249 |
-
stopping_criteria=[interrupt_criteria]
|
250 |
)
|
251 |
|
252 |
generated_token_ids = output_sequences[0][prompt_tokens_length:]
|
@@ -278,17 +306,17 @@ def run_inference(model_name, context, question):
|
|
278 |
)
|
279 |
|
280 |
input_length = len(formatted)
|
281 |
-
# Check interrupt before generation
|
282 |
|
283 |
outputs = pipe(
|
284 |
formatted,
|
285 |
max_new_tokens=512,
|
286 |
-
stopping_criteria=[interrupt_criteria],
|
287 |
generation_kwargs={"skip_special_tokens": True}
|
288 |
)
|
289 |
-
# print(outputs[0]['generated_text'])
|
290 |
result = outputs[0]["generated_text"][input_length:]
|
291 |
|
|
|
|
|
292 |
except Exception as e:
|
293 |
print(f"Error in inference for {model_name}: {e}")
|
294 |
print(traceback.format_exc())
|
|
|
62 |
Generates summaries for the given example using the assigned models sequentially.
|
63 |
"""
|
64 |
if generation_interrupt.is_set():
|
65 |
+
print("Generation interrupted before starting")
|
66 |
return "", ""
|
67 |
|
68 |
context_text = ""
|
|
|
70 |
|
71 |
if "full_contexts" in example and example["full_contexts"]:
|
72 |
for i, ctx in enumerate(example["full_contexts"]):
|
73 |
+
# Check interrupt during context processing
|
74 |
+
if generation_interrupt.is_set():
|
75 |
+
print("Generation interrupted during context processing")
|
76 |
+
return "", ""
|
77 |
+
|
78 |
content = ""
|
79 |
|
80 |
# Extract content from either dict or string
|
|
|
98 |
question = example.get("question", "")
|
99 |
|
100 |
if generation_interrupt.is_set():
|
101 |
+
print("Generation interrupted before model A")
|
102 |
return "", ""
|
103 |
|
104 |
+
print(f"Starting inference for Model A: {model_a_name}")
|
105 |
# Run model A
|
106 |
summary_a = run_inference(models[model_a_name], context_text, question)
|
107 |
|
108 |
if generation_interrupt.is_set():
|
109 |
+
print("Generation interrupted after model A, before model B")
|
110 |
return summary_a, ""
|
111 |
|
112 |
+
print(f"Starting inference for Model B: {model_b_name}")
|
113 |
# Run model B
|
114 |
summary_b = run_inference(models[model_b_name], context_text, question)
|
115 |
|
116 |
+
print("Both models completed successfully")
|
117 |
return summary_a, summary_b
|
118 |
|
119 |
|
|
|
125 |
"""
|
126 |
# Check interrupt at the beginning
|
127 |
if generation_interrupt.is_set():
|
128 |
+
print(f"Inference interrupted before starting for {model_name}")
|
129 |
return ""
|
130 |
|
131 |
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
146 |
if model_name in tokenizer_cache:
|
147 |
tokenizer = tokenizer_cache[model_name]
|
148 |
else:
|
149 |
+
# Check interrupt before loading tokenizer
|
150 |
+
if generation_interrupt.is_set():
|
151 |
+
print(f"Inference interrupted before loading tokenizer for {model_name}")
|
152 |
+
return ""
|
153 |
+
|
154 |
# Common arguments for tokenizer loading
|
155 |
tokenizer_load_args = {"padding_side": "left", "token": True}
|
156 |
|
|
|
172 |
|
173 |
# Check interrupt before loading the model
|
174 |
if generation_interrupt.is_set():
|
175 |
+
print(f"Inference interrupted before loading model {model_name}")
|
176 |
return ""
|
177 |
|
178 |
# Create interrupt criteria for this generation
|
|
|
180 |
|
181 |
print("REACHED HERE BEFORE pipe")
|
182 |
print(f"Loading model {model_name}...")
|
183 |
+
|
184 |
+
# Check interrupt before model loading
|
185 |
+
if generation_interrupt.is_set():
|
186 |
+
print(f"Inference interrupted during model loading for {model_name}")
|
187 |
+
return ""
|
188 |
+
|
189 |
if "bitnet" in model_name.lower():
|
190 |
bitnet_model = BitNetForCausalLM.from_pretrained(
|
191 |
model_name,
|
|
|
192 |
torch_dtype=torch.bfloat16,
|
|
|
193 |
)
|
194 |
pipe = pipeline(
|
195 |
"text-generation",
|
196 |
model=bitnet_model,
|
197 |
tokenizer=tokenizer,
|
|
|
|
|
198 |
torch_dtype=torch.bfloat16,
|
199 |
model_kwargs={
|
200 |
"attn_implementation": "eager",
|
|
|
226 |
torch_dtype=torch.bfloat16,
|
227 |
)
|
228 |
|
229 |
+
# Final interrupt check before generation
|
230 |
+
if generation_interrupt.is_set():
|
231 |
+
print(f"Inference interrupted before generation for {model_name}")
|
232 |
+
return ""
|
233 |
+
|
234 |
text_input = format_rag_prompt(question, context, accepts_sys)
|
235 |
+
|
236 |
+
print(f"Starting generation for {model_name}")
|
237 |
if "Gemma-3".lower() in model_name.lower():
|
238 |
print("REACHED HERE BEFORE GEN")
|
239 |
result = pipe(
|
240 |
text_input,
|
241 |
max_new_tokens=512,
|
242 |
+
stopping_criteria=[interrupt_criteria],
|
243 |
generation_kwargs={"skip_special_tokens": True}
|
244 |
)[0]["generated_text"]
|
245 |
|
|
|
265 |
with torch.inference_mode():
|
266 |
# Check interrupt before generation
|
267 |
if generation_interrupt.is_set():
|
268 |
+
print(f"Inference interrupted before torch generation for {model_name}")
|
269 |
return ""
|
270 |
|
271 |
output_sequences = model.generate(
|
|
|
274 |
max_new_tokens=512,
|
275 |
eos_token_id=tokenizer.eos_token_id,
|
276 |
pad_token_id=tokenizer.pad_token_id,
|
277 |
+
stopping_criteria=[interrupt_criteria]
|
278 |
)
|
279 |
|
280 |
generated_token_ids = output_sequences[0][prompt_tokens_length:]
|
|
|
306 |
)
|
307 |
|
308 |
input_length = len(formatted)
|
|
|
309 |
|
310 |
outputs = pipe(
|
311 |
formatted,
|
312 |
max_new_tokens=512,
|
313 |
+
stopping_criteria=[interrupt_criteria],
|
314 |
generation_kwargs={"skip_special_tokens": True}
|
315 |
)
|
|
|
316 |
result = outputs[0]["generated_text"][input_length:]
|
317 |
|
318 |
+
print(f"Generation completed for {model_name}")
|
319 |
+
|
320 |
except Exception as e:
|
321 |
print(f"Error in inference for {model_name}: {e}")
|
322 |
print(traceback.format_exc())
|