davanstrien HF Staff commited on
Commit
eb5a4ea
·
1 Parent(s): a875142

ruff format

Browse files
Files changed (2) hide show
  1. classify-dataset.py +46 -40
  2. generate-responses.py +15 -12
classify-dataset.py CHANGED
@@ -54,7 +54,9 @@ from toolz import concat, keymap, partition_all
54
  from tqdm.auto import tqdm
55
  from vllm import LLM
56
 
57
- logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
 
 
58
  logger = logging.getLogger(__name__)
59
 
60
 
@@ -62,9 +64,11 @@ def check_gpu_availability():
62
  """Check if CUDA is available and log GPU information."""
63
  if not torch.cuda.is_available():
64
  logger.error("CUDA is not available. This script requires a GPU.")
65
- logger.error("Please run on a machine with NVIDIA GPU or use HF Jobs with GPU flavor.")
 
 
66
  sys.exit(1)
67
-
68
  gpu_name = torch.cuda.get_device_name(0)
69
  gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1024**3
70
  logger.info(f"GPU detected: {gpu_name} with {gpu_memory:.1f} GB memory")
@@ -75,25 +79,24 @@ def get_model_id2label(hub_model_id: str) -> Optional[dict[int, str]]:
75
  """Extract label mapping from model's config.json on Hugging Face Hub."""
76
  try:
77
  response = httpx.get(
78
- hf_hub_url(hub_model_id, filename="config.json"),
79
- follow_redirects=True
80
  )
81
  if response.status_code != 200:
82
  logger.warning(f"Could not fetch config.json for {hub_model_id}")
83
  return None
84
-
85
  data = response.json()
86
  id2label = data.get("id2label")
87
-
88
  if id2label is None:
89
  logger.info("No id2label mapping found in config.json")
90
  return None
91
-
92
  # Convert string keys to integers
93
  label_map = keymap(int, id2label)
94
  logger.info(f"Found label mapping: {label_map}")
95
  return label_map
96
-
97
  except Exception as e:
98
  logger.warning(f"Failed to parse config.json: {e}")
99
  return None
@@ -102,11 +105,11 @@ def get_model_id2label(hub_model_id: str) -> Optional[dict[int, str]]:
102
  def get_top_label(output, label_map: Optional[dict[int, str]] = None):
103
  """
104
  Extract the top predicted label and confidence score from vLLM output.
105
-
106
  Args:
107
  output: vLLM ClassificationRequestOutput
108
  label_map: Optional mapping from label indices to label names
109
-
110
  Returns:
111
  Tuple of (label, confidence_score)
112
  """
@@ -114,7 +117,7 @@ def get_top_label(output, label_map: Optional[dict[int, str]] = None):
114
  probs = F.softmax(logits, dim=0)
115
  top_idx = torch.argmax(probs).item()
116
  top_prob = probs[top_idx].item()
117
-
118
  # Use label name if mapping available, otherwise use index
119
  label = label_map.get(top_idx, str(top_idx)) if label_map else str(top_idx)
120
  return label, top_prob
@@ -130,7 +133,7 @@ def main(
130
  ):
131
  """
132
  Main classification pipeline.
133
-
134
  Args:
135
  hub_model_id: Hugging Face model ID for classification
136
  src_dataset_hub_id: Input dataset on Hugging Face Hub
@@ -141,58 +144,62 @@ def main(
141
  """
142
  # GPU check
143
  check_gpu_availability()
144
-
145
  # Authentication
146
  HF_TOKEN = hf_token or os.environ.get("HF_TOKEN")
147
  if HF_TOKEN:
148
  login(token=HF_TOKEN)
149
  else:
150
- logger.error("HF_TOKEN is required. Set via --hf-token or HF_TOKEN environment variable.")
 
 
151
  sys.exit(1)
152
-
153
  # Initialize vLLM with classification task
154
  logger.info(f"Loading model: {hub_model_id}")
155
  llm = LLM(model=hub_model_id, task="classify")
156
-
157
  # Get label mapping if available
158
  id2label = get_model_id2label(hub_model_id)
159
-
160
  # Load dataset
161
  logger.info(f"Loading dataset: {src_dataset_hub_id}")
162
  dataset = load_dataset(src_dataset_hub_id, split="train")
163
  total_examples = len(dataset)
164
  logger.info(f"Dataset loaded with {total_examples:,} examples")
165
-
166
  # Extract text column
167
  if inference_column not in dataset.column_names:
168
- logger.error(f"Column '{inference_column}' not found. Available columns: {dataset.column_names}")
 
 
169
  sys.exit(1)
170
-
171
  prompts = dataset[inference_column]
172
-
173
  # Process in batches
174
  logger.info(f"Starting classification with batch size {batch_size:,}")
175
  all_results = []
176
-
177
  for batch in tqdm(
178
  list(partition_all(batch_size, prompts)),
179
  desc="Processing batches",
180
- unit="batch"
181
  ):
182
  batch_results = llm.classify(batch)
183
  all_results.append(batch_results)
184
-
185
  # Flatten results
186
  outputs = list(concat(all_results))
187
-
188
  # Extract labels and probabilities
189
  logger.info("Extracting predictions...")
190
  labels_and_probs = [get_top_label(output, id2label) for output in outputs]
191
-
192
  # Add results to dataset
193
  dataset = dataset.add_column("label", [label for label, _ in labels_and_probs])
194
  dataset = dataset.add_column("prob", [prob for _, prob in labels_and_probs])
195
-
196
  # Push to hub
197
  logger.info(f"Pushing results to: {output_dataset_hub_id}")
198
  dataset.push_to_hub(output_dataset_hub_id, token=HF_TOKEN)
@@ -216,41 +223,40 @@ Examples:
216
 
217
  # Using environment variable for token
218
  HF_TOKEN=hf_xxx uv run classify-dataset.py model/name input-dataset output-dataset
219
- """
220
  )
221
-
222
  parser.add_argument(
223
  "hub_model_id",
224
- help="Hugging Face model ID for classification (e.g., bert-base-uncased)"
225
  )
226
  parser.add_argument(
227
  "src_dataset_hub_id",
228
- help="Input dataset on Hugging Face Hub (e.g., username/dataset-name)"
229
  )
230
  parser.add_argument(
231
- "output_dataset_hub_id",
232
- help="Output dataset name on Hugging Face Hub"
233
  )
234
  parser.add_argument(
235
  "--inference-column",
236
  type=str,
237
  default="text",
238
- help="Column containing text to classify (default: text)"
239
  )
240
  parser.add_argument(
241
  "--batch-size",
242
  type=int,
243
  default=10_000,
244
- help="Batch size for inference (default: 10,000)"
245
  )
246
  parser.add_argument(
247
  "--hf-token",
248
  type=str,
249
- help="Hugging Face token (can also use HF_TOKEN env var)"
250
  )
251
-
252
  args = parser.parse_args()
253
-
254
  main(
255
  hub_model_id=args.hub_model_id,
256
  src_dataset_hub_id=args.src_dataset_hub_id,
@@ -283,4 +289,4 @@ Example HF Jobs command:
283
  ' \\
284
  --project vllm-classify \\
285
  --name my-classification-job
286
- """)
 
54
  from tqdm.auto import tqdm
55
  from vllm import LLM
56
 
57
+ logging.basicConfig(
58
+ level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
59
+ )
60
  logger = logging.getLogger(__name__)
61
 
62
 
 
64
  """Check if CUDA is available and log GPU information."""
65
  if not torch.cuda.is_available():
66
  logger.error("CUDA is not available. This script requires a GPU.")
67
+ logger.error(
68
+ "Please run on a machine with NVIDIA GPU or use HF Jobs with GPU flavor."
69
+ )
70
  sys.exit(1)
71
+
72
  gpu_name = torch.cuda.get_device_name(0)
73
  gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1024**3
74
  logger.info(f"GPU detected: {gpu_name} with {gpu_memory:.1f} GB memory")
 
79
  """Extract label mapping from model's config.json on Hugging Face Hub."""
80
  try:
81
  response = httpx.get(
82
+ hf_hub_url(hub_model_id, filename="config.json"), follow_redirects=True
 
83
  )
84
  if response.status_code != 200:
85
  logger.warning(f"Could not fetch config.json for {hub_model_id}")
86
  return None
87
+
88
  data = response.json()
89
  id2label = data.get("id2label")
90
+
91
  if id2label is None:
92
  logger.info("No id2label mapping found in config.json")
93
  return None
94
+
95
  # Convert string keys to integers
96
  label_map = keymap(int, id2label)
97
  logger.info(f"Found label mapping: {label_map}")
98
  return label_map
99
+
100
  except Exception as e:
101
  logger.warning(f"Failed to parse config.json: {e}")
102
  return None
 
105
  def get_top_label(output, label_map: Optional[dict[int, str]] = None):
106
  """
107
  Extract the top predicted label and confidence score from vLLM output.
108
+
109
  Args:
110
  output: vLLM ClassificationRequestOutput
111
  label_map: Optional mapping from label indices to label names
112
+
113
  Returns:
114
  Tuple of (label, confidence_score)
115
  """
 
117
  probs = F.softmax(logits, dim=0)
118
  top_idx = torch.argmax(probs).item()
119
  top_prob = probs[top_idx].item()
120
+
121
  # Use label name if mapping available, otherwise use index
122
  label = label_map.get(top_idx, str(top_idx)) if label_map else str(top_idx)
123
  return label, top_prob
 
133
  ):
134
  """
135
  Main classification pipeline.
136
+
137
  Args:
138
  hub_model_id: Hugging Face model ID for classification
139
  src_dataset_hub_id: Input dataset on Hugging Face Hub
 
144
  """
145
  # GPU check
146
  check_gpu_availability()
147
+
148
  # Authentication
149
  HF_TOKEN = hf_token or os.environ.get("HF_TOKEN")
150
  if HF_TOKEN:
151
  login(token=HF_TOKEN)
152
  else:
153
+ logger.error(
154
+ "HF_TOKEN is required. Set via --hf-token or HF_TOKEN environment variable."
155
+ )
156
  sys.exit(1)
157
+
158
  # Initialize vLLM with classification task
159
  logger.info(f"Loading model: {hub_model_id}")
160
  llm = LLM(model=hub_model_id, task="classify")
161
+
162
  # Get label mapping if available
163
  id2label = get_model_id2label(hub_model_id)
164
+
165
  # Load dataset
166
  logger.info(f"Loading dataset: {src_dataset_hub_id}")
167
  dataset = load_dataset(src_dataset_hub_id, split="train")
168
  total_examples = len(dataset)
169
  logger.info(f"Dataset loaded with {total_examples:,} examples")
170
+
171
  # Extract text column
172
  if inference_column not in dataset.column_names:
173
+ logger.error(
174
+ f"Column '{inference_column}' not found. Available columns: {dataset.column_names}"
175
+ )
176
  sys.exit(1)
177
+
178
  prompts = dataset[inference_column]
179
+
180
  # Process in batches
181
  logger.info(f"Starting classification with batch size {batch_size:,}")
182
  all_results = []
183
+
184
  for batch in tqdm(
185
  list(partition_all(batch_size, prompts)),
186
  desc="Processing batches",
187
+ unit="batch",
188
  ):
189
  batch_results = llm.classify(batch)
190
  all_results.append(batch_results)
191
+
192
  # Flatten results
193
  outputs = list(concat(all_results))
194
+
195
  # Extract labels and probabilities
196
  logger.info("Extracting predictions...")
197
  labels_and_probs = [get_top_label(output, id2label) for output in outputs]
198
+
199
  # Add results to dataset
200
  dataset = dataset.add_column("label", [label for label, _ in labels_and_probs])
201
  dataset = dataset.add_column("prob", [prob for _, prob in labels_and_probs])
202
+
203
  # Push to hub
204
  logger.info(f"Pushing results to: {output_dataset_hub_id}")
205
  dataset.push_to_hub(output_dataset_hub_id, token=HF_TOKEN)
 
223
 
224
  # Using environment variable for token
225
  HF_TOKEN=hf_xxx uv run classify-dataset.py model/name input-dataset output-dataset
226
+ """,
227
  )
228
+
229
  parser.add_argument(
230
  "hub_model_id",
231
+ help="Hugging Face model ID for classification (e.g., bert-base-uncased)",
232
  )
233
  parser.add_argument(
234
  "src_dataset_hub_id",
235
+ help="Input dataset on Hugging Face Hub (e.g., username/dataset-name)",
236
  )
237
  parser.add_argument(
238
+ "output_dataset_hub_id", help="Output dataset name on Hugging Face Hub"
 
239
  )
240
  parser.add_argument(
241
  "--inference-column",
242
  type=str,
243
  default="text",
244
+ help="Column containing text to classify (default: text)",
245
  )
246
  parser.add_argument(
247
  "--batch-size",
248
  type=int,
249
  default=10_000,
250
+ help="Batch size for inference (default: 10,000)",
251
  )
252
  parser.add_argument(
253
  "--hf-token",
254
  type=str,
255
+ help="Hugging Face token (can also use HF_TOKEN env var)",
256
  )
257
+
258
  args = parser.parse_args()
259
+
260
  main(
261
  hub_model_id=args.hub_model_id,
262
  src_dataset_hub_id=args.src_dataset_hub_id,
 
289
  ' \\
290
  --project vllm-classify \\
291
  --name my-classification-job
292
+ """)
generate-responses.py CHANGED
@@ -42,9 +42,8 @@ import logging
42
  import os
43
  import sys
44
  from datetime import datetime
45
- from typing import List, Optional
46
 
47
- import torch
48
  from datasets import load_dataset
49
  from huggingface_hub import DatasetCard, get_token, login
50
  from torch import cuda
@@ -105,7 +104,7 @@ def create_dataset_card(
105
  - **Max Model Length Used**: {max_model_len_used:,} tokens
106
 
107
  Note: Prompts exceeding the maximum model length were skipped and have empty responses."""
108
-
109
  return f"""---
110
  viewer: false
111
  tags:
@@ -160,7 +159,7 @@ uv run https://huggingface.co/datasets/uv-scripts/vllm/raw/main/generate-respons
160
  --temperature {sampling_params.temperature} \\
161
  --top-p {sampling_params.top_p} \\
162
  --top-k {sampling_params.top_k} \\
163
- --max-tokens {sampling_params.max_tokens}{f' \\\\\\n --max-model-len {max_model_len_used}' if max_model_len_used else ''}
164
  ```
165
  """
166
 
@@ -287,7 +286,7 @@ def main(
287
  valid_prompts = []
288
  valid_indices = []
289
  skipped_info = []
290
-
291
  for i, example in enumerate(tqdm(dataset, desc="Processing messages")):
292
  messages = example[messages_column]
293
  # Apply chat template
@@ -295,7 +294,7 @@ def main(
295
  messages, tokenize=False, add_generation_prompt=True
296
  )
297
  all_prompts.append(prompt)
298
-
299
  # Count tokens if filtering is enabled
300
  if skip_long_prompts:
301
  tokens = tokenizer.encode(prompt)
@@ -307,20 +306,24 @@ def main(
307
  else:
308
  valid_prompts.append(prompt)
309
  valid_indices.append(i)
310
-
311
  # Log filtering results
312
  if skip_long_prompts and skipped_info:
313
- logger.warning(f"Skipped {len(skipped_info)} prompts that exceed max_model_len ({effective_max_len} tokens)")
 
 
314
  logger.info("Skipped prompt details (first 10):")
315
  for idx, (prompt_idx, token_count) in enumerate(skipped_info[:10]):
316
- logger.info(f" - Example {prompt_idx}: {token_count} tokens (exceeds by {token_count - effective_max_len})")
 
 
317
  if len(skipped_info) > 10:
318
  logger.info(f" ... and {len(skipped_info) - 10} more")
319
-
320
  skip_percentage = (len(skipped_info) / total_examples) * 100
321
  if skip_percentage > 10:
322
  logger.warning(f"WARNING: {skip_percentage:.1f}% of prompts were skipped!")
323
-
324
  if not valid_prompts:
325
  logger.error("No valid prompts to process after filtering!")
326
  sys.exit(1)
@@ -334,7 +337,7 @@ def main(
334
  # Extract generated text and create full response list
335
  logger.info("Extracting generated responses...")
336
  responses = [""] * total_examples # Initialize with empty strings
337
-
338
  for idx, output in enumerate(outputs):
339
  original_idx = valid_indices[idx]
340
  response = output.outputs[0].text.strip()
 
42
  import os
43
  import sys
44
  from datetime import datetime
45
+ from typing import Optional
46
 
 
47
  from datasets import load_dataset
48
  from huggingface_hub import DatasetCard, get_token, login
49
  from torch import cuda
 
104
  - **Max Model Length Used**: {max_model_len_used:,} tokens
105
 
106
  Note: Prompts exceeding the maximum model length were skipped and have empty responses."""
107
+
108
  return f"""---
109
  viewer: false
110
  tags:
 
159
  --temperature {sampling_params.temperature} \\
160
  --top-p {sampling_params.top_p} \\
161
  --top-k {sampling_params.top_k} \\
162
+ --max-tokens {sampling_params.max_tokens}{f" \\\\\\n --max-model-len {max_model_len_used}" if max_model_len_used else ""}
163
  ```
164
  """
165
 
 
286
  valid_prompts = []
287
  valid_indices = []
288
  skipped_info = []
289
+
290
  for i, example in enumerate(tqdm(dataset, desc="Processing messages")):
291
  messages = example[messages_column]
292
  # Apply chat template
 
294
  messages, tokenize=False, add_generation_prompt=True
295
  )
296
  all_prompts.append(prompt)
297
+
298
  # Count tokens if filtering is enabled
299
  if skip_long_prompts:
300
  tokens = tokenizer.encode(prompt)
 
306
  else:
307
  valid_prompts.append(prompt)
308
  valid_indices.append(i)
309
+
310
  # Log filtering results
311
  if skip_long_prompts and skipped_info:
312
+ logger.warning(
313
+ f"Skipped {len(skipped_info)} prompts that exceed max_model_len ({effective_max_len} tokens)"
314
+ )
315
  logger.info("Skipped prompt details (first 10):")
316
  for idx, (prompt_idx, token_count) in enumerate(skipped_info[:10]):
317
+ logger.info(
318
+ f" - Example {prompt_idx}: {token_count} tokens (exceeds by {token_count - effective_max_len})"
319
+ )
320
  if len(skipped_info) > 10:
321
  logger.info(f" ... and {len(skipped_info) - 10} more")
322
+
323
  skip_percentage = (len(skipped_info) / total_examples) * 100
324
  if skip_percentage > 10:
325
  logger.warning(f"WARNING: {skip_percentage:.1f}% of prompts were skipped!")
326
+
327
  if not valid_prompts:
328
  logger.error("No valid prompts to process after filtering!")
329
  sys.exit(1)
 
337
  # Extract generated text and create full response list
338
  logger.info("Extracting generated responses...")
339
  responses = [""] * total_examples # Initialize with empty strings
340
+
341
  for idx, output in enumerate(outputs):
342
  original_idx = valid_indices[idx]
343
  response = output.outputs[0].text.strip()