Commit
·
89351df
1
Parent(s):
bb33774
pin xet
Browse files- gpt_oss_transformers.py +143 -97
gpt_oss_transformers.py
CHANGED
@@ -3,40 +3,42 @@
|
|
3 |
# dependencies = [
|
4 |
# "datasets",
|
5 |
# "huggingface-hub[hf_transfer]",
|
|
|
6 |
# "torch",
|
7 |
-
# "transformers>=4.
|
8 |
# "tqdm",
|
9 |
# "accelerate",
|
10 |
-
# "kernels>=0.9.0", # For Flash Attention 3 support (optional but recommended)
|
11 |
# ]
|
12 |
# ///
|
13 |
"""
|
14 |
-
Generate responses with transparent reasoning using OpenAI's
|
15 |
|
16 |
-
This implementation
|
17 |
-
The models
|
18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
Example usage:
|
|
|
|
|
|
|
21 |
# Generate haiku with reasoning
|
22 |
uv run gpt_oss_transformers.py \\
|
23 |
--input-dataset davanstrien/haiku_dpo \\
|
24 |
--output-dataset username/haiku-reasoning \\
|
25 |
--prompt-column question
|
26 |
|
27 |
-
#
|
28 |
-
uv run gpt_oss_transformers.py \\
|
29 |
-
--input-dataset username/prompts \\
|
30 |
-
--output-dataset username/responses-with-reasoning \\
|
31 |
-
--prompt-column prompt \\
|
32 |
-
--reasoning-level high \\
|
33 |
-
--max-samples 100
|
34 |
-
|
35 |
-
# HF Jobs execution
|
36 |
hf jobs uv run --flavor a10g-small \\
|
37 |
https://huggingface.co/datasets/uv-scripts/openai-oss/raw/main/gpt_oss_transformers.py \\
|
38 |
-
--input-dataset
|
39 |
-
--output-dataset username/
|
|
|
40 |
"""
|
41 |
|
42 |
import argparse
|
@@ -89,34 +91,79 @@ def parse_channels(raw_output: str) -> Dict[str, str]:
|
|
89 |
"""
|
90 |
Extract think/content from GPT OSS channel-based output.
|
91 |
|
92 |
-
|
93 |
-
|
94 |
-
|
|
|
|
|
95 |
"""
|
96 |
-
|
97 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
|
99 |
-
# Extract
|
100 |
-
|
101 |
-
r"<\|start\|>assistant<\|channel\|>analysis<\|message\|>(.*?)<\|end\|>"
|
102 |
-
)
|
103 |
-
analysis_match = re.search(analysis_pattern, raw_output, re.DOTALL)
|
104 |
-
if analysis_match:
|
105 |
-
think = analysis_match.group(1).strip()
|
106 |
|
107 |
-
#
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
# If no channels found, treat entire output as content
|
116 |
-
if not think and not content:
|
117 |
-
content = raw_output.strip()
|
118 |
|
119 |
-
return
|
120 |
|
121 |
|
122 |
def create_dataset_card(
|
@@ -227,10 +274,11 @@ def main(
|
|
227 |
logger.info("HuggingFace token found, authenticating...")
|
228 |
login(token=HF_TOKEN)
|
229 |
|
230 |
-
# Load tokenizer
|
231 |
logger.info(f"Loading tokenizer: {model_id}")
|
232 |
tokenizer = AutoTokenizer.from_pretrained(
|
233 |
-
model_id,
|
|
|
234 |
)
|
235 |
|
236 |
# Add padding token if needed
|
@@ -252,58 +300,49 @@ def main(
|
|
252 |
|
253 |
# Load model
|
254 |
logger.info(f"Loading model: {model_id}")
|
255 |
-
logger.info("
|
256 |
-
|
257 |
-
#
|
258 |
-
logger.info("Note: MXFP4
|
259 |
-
|
260 |
# Check available GPU memory
|
261 |
if num_gpus > 0:
|
262 |
gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1024**3
|
263 |
if gpu_memory < 40 and "20b" in model_id.lower():
|
264 |
-
logger.
|
265 |
-
|
|
|
|
|
266 |
|
267 |
try:
|
268 |
-
#
|
269 |
-
|
270 |
model = AutoModelForCausalLM.from_pretrained(
|
271 |
model_id,
|
272 |
-
torch_dtype=torch.bfloat16,
|
273 |
-
attn_implementation="kernels-community/vllm-flash-attn3"
|
274 |
**model_kwargs,
|
275 |
)
|
276 |
model.eval()
|
277 |
-
logger.info("Successfully loaded
|
|
|
|
|
|
|
|
|
|
|
|
|
278 |
except torch.cuda.OutOfMemoryError as e:
|
279 |
logger.error(f"Out of memory error: {e}")
|
280 |
-
logger.error("\
|
281 |
-
logger.error("- 20B model: ~40GB VRAM (use
|
282 |
-
logger.error("- 120B model: ~240GB VRAM (use
|
|
|
|
|
|
|
283 |
sys.exit(1)
|
284 |
except Exception as e:
|
285 |
-
|
286 |
-
|
287 |
-
logger.info("Using eager attention instead (standard implementation)")
|
288 |
-
try:
|
289 |
-
model = AutoModelForCausalLM.from_pretrained(
|
290 |
-
model_id,
|
291 |
-
torch_dtype=torch.bfloat16,
|
292 |
-
attn_implementation="eager", # Fallback to eager attention
|
293 |
-
**model_kwargs,
|
294 |
-
)
|
295 |
-
model.eval()
|
296 |
-
logger.info("Successfully loaded with eager attention")
|
297 |
-
except torch.cuda.OutOfMemoryError as oom_error:
|
298 |
-
logger.error(f"Out of memory with eager attention: {oom_error}")
|
299 |
-
logger.error("The model requires more GPU memory than available")
|
300 |
-
sys.exit(1)
|
301 |
-
except Exception as eager_error:
|
302 |
-
logger.error(f"Failed with eager attention: {eager_error}")
|
303 |
-
sys.exit(1)
|
304 |
-
else:
|
305 |
-
logger.error(f"Unexpected error loading model: {e}")
|
306 |
-
sys.exit(1)
|
307 |
|
308 |
# Generation configuration
|
309 |
generation_config = GenerationConfig(
|
@@ -336,29 +375,36 @@ def main(
|
|
336 |
prompts = []
|
337 |
original_prompts = []
|
338 |
|
|
|
|
|
|
|
|
|
|
|
339 |
for example in tqdm(dataset, desc="Preparing prompts"):
|
340 |
prompt_text = example[prompt_column]
|
341 |
original_prompts.append(prompt_text)
|
342 |
|
343 |
-
# Create
|
344 |
-
messages = [
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
|
359 |
-
|
360 |
-
|
361 |
-
|
|
|
|
|
362 |
prompts.append(prompt)
|
363 |
|
364 |
# Generate responses in batches
|
|
|
3 |
# dependencies = [
|
4 |
# "datasets",
|
5 |
# "huggingface-hub[hf_transfer]",
|
6 |
+
# "hf-xet >= 1.1.7",
|
7 |
# "torch",
|
8 |
+
# "transformers>=4.55.0",
|
9 |
# "tqdm",
|
10 |
# "accelerate",
|
|
|
11 |
# ]
|
12 |
# ///
|
13 |
"""
|
14 |
+
Generate responses with transparent reasoning using OpenAI's GPT OSS models.
|
15 |
|
16 |
+
This implementation works on regular GPUs (L4, A100, A10G, T4) without requiring H100s.
|
17 |
+
The models automatically dequantize MXFP4 to bf16 when needed, making them accessible
|
18 |
+
on standard datacenter hardware.
|
19 |
+
|
20 |
+
Key features:
|
21 |
+
- Works on regular GPUs without special hardware
|
22 |
+
- Extracts reasoning from analysis/commentary channels
|
23 |
+
- Handles the simplified channel output format
|
24 |
+
- No Flash Attention 3 or special kernels needed
|
25 |
|
26 |
Example usage:
|
27 |
+
# Quick test with a single prompt
|
28 |
+
uv run gpt_oss_transformers.py --prompt "Write a haiku about mountains"
|
29 |
+
|
30 |
# Generate haiku with reasoning
|
31 |
uv run gpt_oss_transformers.py \\
|
32 |
--input-dataset davanstrien/haiku_dpo \\
|
33 |
--output-dataset username/haiku-reasoning \\
|
34 |
--prompt-column question
|
35 |
|
36 |
+
# HF Jobs execution (A10G for $1.50/hr)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
hf jobs uv run --flavor a10g-small \\
|
38 |
https://huggingface.co/datasets/uv-scripts/openai-oss/raw/main/gpt_oss_transformers.py \\
|
39 |
+
--input-dataset davanstrien/haiku_dpo \\
|
40 |
+
--output-dataset username/haiku-reasoning \\
|
41 |
+
--prompt-column question
|
42 |
"""
|
43 |
|
44 |
import argparse
|
|
|
91 |
"""
|
92 |
Extract think/content from GPT OSS channel-based output.
|
93 |
|
94 |
+
The actual output format is simpler than expected:
|
95 |
+
analysisREASONING_TEXTassistantfinalRESPONSE_TEXT
|
96 |
+
|
97 |
+
Sometimes includes commentary channel:
|
98 |
+
commentaryMETA_TEXTanalysisREASONING_TEXTassistantfinalRESPONSE_TEXT
|
99 |
"""
|
100 |
+
result = {"think": "", "content": "", "raw_output": raw_output}
|
101 |
+
|
102 |
+
# Clean up the text - remove system prompt if present
|
103 |
+
if "user" in raw_output:
|
104 |
+
# Take everything after the last user prompt
|
105 |
+
parts = raw_output.split("user")
|
106 |
+
if len(parts) > 1:
|
107 |
+
text = parts[-1]
|
108 |
+
# Find where the assistant response starts
|
109 |
+
for marker in ["analysis", "commentary", "assistant"]:
|
110 |
+
if marker in text:
|
111 |
+
idx = text.find(marker)
|
112 |
+
if idx > 0:
|
113 |
+
text = text[idx:]
|
114 |
+
raw_output = text
|
115 |
+
break
|
116 |
+
else:
|
117 |
+
text = raw_output
|
118 |
|
119 |
+
# Extract reasoning (analysis and/or commentary)
|
120 |
+
reasoning_parts = []
|
|
|
|
|
|
|
|
|
|
|
121 |
|
122 |
+
# Try to extract analysis
|
123 |
+
if "analysis" in text:
|
124 |
+
match = re.search(
|
125 |
+
r"analysis(.*?)(?:commentary|assistantfinal|final|$)", text, re.DOTALL
|
126 |
+
)
|
127 |
+
if match:
|
128 |
+
reasoning_parts.append(("Analysis", match.group(1).strip()))
|
129 |
+
|
130 |
+
# Try to extract commentary
|
131 |
+
if "commentary" in text:
|
132 |
+
match = re.search(
|
133 |
+
r"commentary(.*?)(?:analysis|assistantfinal|final|$)", text, re.DOTALL
|
134 |
+
)
|
135 |
+
if match:
|
136 |
+
reasoning_parts.append(("Commentary", match.group(1).strip()))
|
137 |
+
|
138 |
+
# Combine reasoning
|
139 |
+
if reasoning_parts:
|
140 |
+
result["think"] = "\n\n".join(
|
141 |
+
f"[{label}] {content}" for label, content in reasoning_parts
|
142 |
+
)
|
143 |
+
|
144 |
+
# Extract final response
|
145 |
+
if "assistantfinal" in text:
|
146 |
+
parts = text.split("assistantfinal")
|
147 |
+
if len(parts) > 1:
|
148 |
+
result["content"] = parts[-1].strip()
|
149 |
+
elif "final" in text:
|
150 |
+
# Fallback - look for "final" keyword
|
151 |
+
parts = text.split("final")
|
152 |
+
if len(parts) > 1:
|
153 |
+
result["content"] = parts[-1].strip()
|
154 |
+
|
155 |
+
# Clean up any remaining tokens
|
156 |
+
for key in ["think", "content"]:
|
157 |
+
result[key] = result[key].replace("<|end|>", "").replace("<|return|>", "")
|
158 |
+
result[key] = (
|
159 |
+
result[key].replace("<|message|>", "").replace("assistant", "").strip()
|
160 |
+
)
|
161 |
|
162 |
# If no channels found, treat entire output as content
|
163 |
+
if not result["think"] and not result["content"]:
|
164 |
+
result["content"] = raw_output.strip()
|
165 |
|
166 |
+
return result
|
167 |
|
168 |
|
169 |
def create_dataset_card(
|
|
|
274 |
logger.info("HuggingFace token found, authenticating...")
|
275 |
login(token=HF_TOKEN)
|
276 |
|
277 |
+
# Load tokenizer (always use padding_side="left" for generation)
|
278 |
logger.info(f"Loading tokenizer: {model_id}")
|
279 |
tokenizer = AutoTokenizer.from_pretrained(
|
280 |
+
model_id,
|
281 |
+
padding_side="left", # Always use left padding for generation
|
282 |
)
|
283 |
|
284 |
# Add padding token if needed
|
|
|
300 |
|
301 |
# Load model
|
302 |
logger.info(f"Loading model: {model_id}")
|
303 |
+
logger.info("Using standard configuration (no Flash Attention 3 needed)")
|
304 |
+
|
305 |
+
# Note about MXFP4
|
306 |
+
logger.info("Note: MXFP4 will auto-dequantize to bf16 on non-Hopper GPUs")
|
307 |
+
|
308 |
# Check available GPU memory
|
309 |
if num_gpus > 0:
|
310 |
gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1024**3
|
311 |
if gpu_memory < 40 and "20b" in model_id.lower():
|
312 |
+
logger.info(
|
313 |
+
f"GPU has {gpu_memory:.1f}GB. 20B model needs ~40GB when dequantized"
|
314 |
+
)
|
315 |
+
logger.info("Model will still load but may use CPU offloading if needed")
|
316 |
|
317 |
try:
|
318 |
+
# Load with standard configuration (no Flash Attention 3)
|
319 |
+
# This works on L4, A100, A10G, T4 GPUs
|
320 |
model = AutoModelForCausalLM.from_pretrained(
|
321 |
model_id,
|
322 |
+
torch_dtype=torch.bfloat16, # Can also use "auto"
|
323 |
+
# DO NOT USE: attn_implementation="kernels-community/vllm-flash-attn3"
|
324 |
**model_kwargs,
|
325 |
)
|
326 |
model.eval()
|
327 |
+
logger.info("Successfully loaded model")
|
328 |
+
|
329 |
+
# Report memory usage
|
330 |
+
if torch.cuda.is_available():
|
331 |
+
memory_gb = torch.cuda.memory_allocated() / 1024**3
|
332 |
+
logger.info(f"GPU memory used: {memory_gb:.1f}GB")
|
333 |
+
|
334 |
except torch.cuda.OutOfMemoryError as e:
|
335 |
logger.error(f"Out of memory error: {e}")
|
336 |
+
logger.error("\nMemory requirements:")
|
337 |
+
logger.error("- 20B model: ~40GB VRAM (use A100-40GB or 2xL4)")
|
338 |
+
logger.error("- 120B model: ~240GB VRAM (use 4xA100-80GB)")
|
339 |
+
logger.error("\nFor HF Jobs, try:")
|
340 |
+
logger.error("- 20B: --flavor a10g-large or a100-large")
|
341 |
+
logger.error("- 120B: --flavor 4xa100")
|
342 |
sys.exit(1)
|
343 |
except Exception as e:
|
344 |
+
logger.error(f"Error loading model: {e}")
|
345 |
+
sys.exit(1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
346 |
|
347 |
# Generation configuration
|
348 |
generation_config = GenerationConfig(
|
|
|
375 |
prompts = []
|
376 |
original_prompts = []
|
377 |
|
378 |
+
# Get current date for system prompt
|
379 |
+
from datetime import datetime
|
380 |
+
|
381 |
+
current_date = datetime.now().strftime("%Y-%m-%d")
|
382 |
+
|
383 |
for example in tqdm(dataset, desc="Preparing prompts"):
|
384 |
prompt_text = example[prompt_column]
|
385 |
original_prompts.append(prompt_text)
|
386 |
|
387 |
+
# Create messages with reasoning level in system prompt
|
388 |
+
messages = [
|
389 |
+
{
|
390 |
+
"role": "system",
|
391 |
+
"content": f"""You are ChatGPT, a large language model trained by OpenAI.
|
392 |
+
Knowledge cutoff: 2024-06
|
393 |
+
Current date: {current_date}
|
394 |
+
|
395 |
+
Reasoning: {reasoning_level}
|
396 |
+
|
397 |
+
# Valid channels: analysis, commentary, final. Channel must be included for every message.""",
|
398 |
+
},
|
399 |
+
{"role": "user", "content": prompt_text},
|
400 |
+
]
|
401 |
+
|
402 |
+
# Apply chat template
|
403 |
+
prompt = tokenizer.apply_chat_template(
|
404 |
+
messages,
|
405 |
+
add_generation_prompt=True,
|
406 |
+
tokenize=False,
|
407 |
+
)
|
408 |
prompts.append(prompt)
|
409 |
|
410 |
# Generate responses in batches
|