Upload gpt_oss_vllm_harmony.py
Browse files- gpt_oss_vllm_harmony.py +608 -0
gpt_oss_vllm_harmony.py
ADDED
@@ -0,0 +1,608 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# /// script
|
2 |
+
# requires-python = ">=3.12,<3.13" # Required for vllm==0.10.1+gptoss
|
3 |
+
# dependencies = [
|
4 |
+
# "datasets",
|
5 |
+
# "huggingface-hub[hf_transfer]",
|
6 |
+
# "torch",
|
7 |
+
# "openai-harmony", # Official OpenAI harmony library
|
8 |
+
# "vllm==0.10.1+gptoss", # Specific version for GPT OSS models
|
9 |
+
# "tqdm",
|
10 |
+
# ]
|
11 |
+
#
|
12 |
+
# [[tool.uv.index]]
|
13 |
+
# url = "https://wheels.vllm.ai/gpt-oss/"
|
14 |
+
#
|
15 |
+
# [[tool.uv.index]]
|
16 |
+
# url = "https://download.pytorch.org/whl/nightly/cu128"
|
17 |
+
#
|
18 |
+
# [tool.uv]
|
19 |
+
# index-strategy = "unsafe-best-match"
|
20 |
+
# ///
|
21 |
+
"""
|
22 |
+
Generate responses with transparent reasoning using OpenAI GPT OSS models with harmony format.
|
23 |
+
|
24 |
+
This script uses the official openai_harmony library for proper message formatting
|
25 |
+
and channel parsing, as recommended in the OpenAI cookbook.
|
26 |
+
|
27 |
+
Example usage:
|
28 |
+
# Generate haiku with reasoning
|
29 |
+
uv run gpt_oss_vllm_harmony.py \\
|
30 |
+
--input-dataset davanstrien/haiku_dpo \\
|
31 |
+
--output-dataset username/haiku-reasoning \\
|
32 |
+
--prompt-column question
|
33 |
+
|
34 |
+
# Any prompt dataset with custom settings
|
35 |
+
uv run gpt_oss_vllm_harmony.py \\
|
36 |
+
--input-dataset username/prompts \\
|
37 |
+
--output-dataset username/responses-with-reasoning \\
|
38 |
+
--prompt-column prompt \\
|
39 |
+
--reasoning-level high \\
|
40 |
+
--max-samples 100
|
41 |
+
|
42 |
+
# HF Jobs execution
|
43 |
+
hf jobs uv run --flavor a10g-small \\
|
44 |
+
https://huggingface.co/datasets/uv-scripts/openai-reasoning/raw/main/gpt_oss_vllm_harmony.py \\
|
45 |
+
--input-dataset username/prompts \\
|
46 |
+
--output-dataset username/responses-with-reasoning
|
47 |
+
"""
|
48 |
+
|
49 |
+
import argparse
|
50 |
+
import json
|
51 |
+
import logging
|
52 |
+
import os
|
53 |
+
import sys
|
54 |
+
import time
|
55 |
+
from datetime import datetime
|
56 |
+
from typing import Dict, List, Optional
|
57 |
+
|
58 |
+
from datasets import Dataset, load_dataset
|
59 |
+
from huggingface_hub import DatasetCard, get_token, login
|
60 |
+
from openai_harmony import (
|
61 |
+
HarmonyEncodingName,
|
62 |
+
load_harmony_encoding,
|
63 |
+
Conversation,
|
64 |
+
Message,
|
65 |
+
Role,
|
66 |
+
SystemContent,
|
67 |
+
DeveloperContent,
|
68 |
+
)
|
69 |
+
from torch import cuda
|
70 |
+
from tqdm.auto import tqdm
|
71 |
+
from vllm import LLM, SamplingParams
|
72 |
+
|
73 |
+
# Enable HF Transfer for faster downloads
|
74 |
+
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
|
75 |
+
|
76 |
+
# TODO: Change logging level back to INFO after initial testing
|
77 |
+
logging.basicConfig(
|
78 |
+
level=logging.DEBUG, format="%(asctime)s - %(levelname)s - %(message)s"
|
79 |
+
)
|
80 |
+
logger = logging.getLogger(__name__)
|
81 |
+
|
82 |
+
|
83 |
+
def check_gpu_availability() -> int:
|
84 |
+
"""Check if CUDA is available and return the number of GPUs."""
|
85 |
+
if not cuda.is_available():
|
86 |
+
logger.error("CUDA is not available. This script requires a GPU.")
|
87 |
+
logger.error(
|
88 |
+
"Please run on a machine with NVIDIA GPU or use HF Jobs with GPU flavor."
|
89 |
+
)
|
90 |
+
sys.exit(1)
|
91 |
+
|
92 |
+
num_gpus = cuda.device_count()
|
93 |
+
for i in range(num_gpus):
|
94 |
+
gpu_name = cuda.get_device_name(i)
|
95 |
+
gpu_memory = cuda.get_device_properties(i).total_memory / 1024**3
|
96 |
+
logger.info(f"GPU {i}: {gpu_name} with {gpu_memory:.1f} GB memory")
|
97 |
+
|
98 |
+
return num_gpus
|
99 |
+
|
100 |
+
|
101 |
+
def parse_harmony_messages(entries: List, prompt: str) -> Dict[str, str]:
|
102 |
+
"""
|
103 |
+
Parse harmony message entries into think/content structure.
|
104 |
+
|
105 |
+
The harmony format produces structured messages with different channels:
|
106 |
+
- analysis: Chain of thought reasoning
|
107 |
+
- final: User-facing response
|
108 |
+
- commentary: Tool calls (if any)
|
109 |
+
"""
|
110 |
+
think = ""
|
111 |
+
content = ""
|
112 |
+
|
113 |
+
# Log what we received for debugging
|
114 |
+
logger.debug(f"[VERBOSE] Parsing {len(entries)} harmony entries")
|
115 |
+
|
116 |
+
for i, entry in enumerate(entries):
|
117 |
+
entry_dict = entry.to_dict()
|
118 |
+
logger.debug(f"[VERBOSE] Entry {i}: {json.dumps(entry_dict, indent=2)}")
|
119 |
+
|
120 |
+
# Extract content based on the message structure
|
121 |
+
if "content" in entry_dict:
|
122 |
+
if isinstance(entry_dict["content"], list):
|
123 |
+
for content_item in entry_dict["content"]:
|
124 |
+
if content_item.get("type") == "text":
|
125 |
+
text = content_item.get("text", "")
|
126 |
+
# Determine channel based on content or metadata
|
127 |
+
# This is a simplified approach - adjust based on actual harmony output
|
128 |
+
if "analysis" in str(entry_dict).lower() or i == 0:
|
129 |
+
think += text + "\n"
|
130 |
+
else:
|
131 |
+
content += text + "\n"
|
132 |
+
elif isinstance(entry_dict["content"], str):
|
133 |
+
# Simple string content
|
134 |
+
if i == 0: # First message is often reasoning
|
135 |
+
think = entry_dict["content"]
|
136 |
+
else:
|
137 |
+
content = entry_dict["content"]
|
138 |
+
|
139 |
+
# Clean up whitespace
|
140 |
+
think = think.strip()
|
141 |
+
content = content.strip()
|
142 |
+
|
143 |
+
# If we didn't parse anything, use the first entry as content
|
144 |
+
if not think and not content and entries:
|
145 |
+
content = str(entries[0].to_dict())
|
146 |
+
|
147 |
+
return {
|
148 |
+
"prompt": prompt,
|
149 |
+
"think": think,
|
150 |
+
"content": content,
|
151 |
+
"raw_output": json.dumps([e.to_dict() for e in entries], indent=2)
|
152 |
+
}
|
153 |
+
|
154 |
+
|
155 |
+
def create_dataset_card(
|
156 |
+
input_dataset: str,
|
157 |
+
model_id: str,
|
158 |
+
prompt_column: str,
|
159 |
+
reasoning_level: str,
|
160 |
+
num_examples: int,
|
161 |
+
generation_time: str,
|
162 |
+
tensor_parallel_size: int,
|
163 |
+
temperature: float,
|
164 |
+
max_tokens: int,
|
165 |
+
) -> str:
|
166 |
+
"""Create a dataset card documenting the generation process."""
|
167 |
+
return f"""---
|
168 |
+
tags:
|
169 |
+
- generated
|
170 |
+
- synthetic
|
171 |
+
- reasoning
|
172 |
+
- openai-gpt-oss
|
173 |
+
- harmony-format
|
174 |
+
---
|
175 |
+
|
176 |
+
# Generated Responses with Reasoning (Harmony Format)
|
177 |
+
|
178 |
+
This dataset contains AI-generated responses with transparent chain-of-thought reasoning using OpenAI GPT OSS models and the official harmony format.
|
179 |
+
|
180 |
+
## Generation Details
|
181 |
+
|
182 |
+
- **Source Dataset**: [{input_dataset}](https://huggingface.co/datasets/{input_dataset})
|
183 |
+
- **Model**: [{model_id}](https://huggingface.co/{model_id})
|
184 |
+
- **Reasoning Level**: {reasoning_level}
|
185 |
+
- **Number of Examples**: {num_examples:,}
|
186 |
+
- **Generation Date**: {generation_time}
|
187 |
+
- **Format**: Official OpenAI Harmony format
|
188 |
+
|
189 |
+
## Dataset Structure
|
190 |
+
|
191 |
+
Each example contains:
|
192 |
+
- `prompt`: The input prompt from the source dataset
|
193 |
+
- `think`: The model's internal reasoning process (analysis channel)
|
194 |
+
- `content`: The final response (final channel)
|
195 |
+
- `raw_output`: Complete harmony format output
|
196 |
+
- `reasoning_level`: The reasoning effort level used
|
197 |
+
- `model`: Model identifier
|
198 |
+
|
199 |
+
## Generation Script
|
200 |
+
|
201 |
+
Generated using [uv-scripts/openai-reasoning](https://huggingface.co/datasets/uv-scripts/openai-reasoning) with official harmony format.
|
202 |
+
|
203 |
+
To reproduce:
|
204 |
+
```bash
|
205 |
+
uv run gpt_oss_vllm_harmony.py \\
|
206 |
+
--input-dataset {input_dataset} \\
|
207 |
+
--output-dataset <your-dataset> \\
|
208 |
+
--prompt-column {prompt_column} \\
|
209 |
+
--model-id {model_id} \\
|
210 |
+
--reasoning-level {reasoning_level}
|
211 |
+
```
|
212 |
+
"""
|
213 |
+
|
214 |
+
|
215 |
+
def main(
|
216 |
+
input_dataset: str,
|
217 |
+
output_dataset_hub_id: str,
|
218 |
+
prompt_column: str = "prompt",
|
219 |
+
model_id: str = "openai/gpt-oss-20b",
|
220 |
+
reasoning_level: str = "high",
|
221 |
+
max_samples: Optional[int] = None,
|
222 |
+
temperature: float = 0.7,
|
223 |
+
max_tokens: int = 512,
|
224 |
+
gpu_memory_utilization: float = 0.90,
|
225 |
+
tensor_parallel_size: Optional[int] = None,
|
226 |
+
hf_token: Optional[str] = None,
|
227 |
+
):
|
228 |
+
"""
|
229 |
+
Main generation pipeline using official harmony format.
|
230 |
+
|
231 |
+
Args:
|
232 |
+
input_dataset: Source dataset on Hugging Face Hub
|
233 |
+
output_dataset_hub_id: Where to save results on Hugging Face Hub
|
234 |
+
prompt_column: Column containing the prompts
|
235 |
+
model_id: OpenAI GPT OSS model to use
|
236 |
+
reasoning_level: Reasoning effort level (high/medium/low)
|
237 |
+
max_samples: Maximum number of samples to process
|
238 |
+
temperature: Sampling temperature
|
239 |
+
max_tokens: Maximum tokens to generate
|
240 |
+
gpu_memory_utilization: GPU memory utilization factor
|
241 |
+
tensor_parallel_size: Number of GPUs to use (auto-detect if None)
|
242 |
+
hf_token: Hugging Face authentication token
|
243 |
+
"""
|
244 |
+
generation_start_time = datetime.now().isoformat()
|
245 |
+
|
246 |
+
# GPU check and configuration
|
247 |
+
num_gpus = check_gpu_availability()
|
248 |
+
if tensor_parallel_size is None:
|
249 |
+
tensor_parallel_size = num_gpus
|
250 |
+
logger.info(
|
251 |
+
f"Auto-detected {num_gpus} GPU(s), using tensor_parallel_size={tensor_parallel_size}"
|
252 |
+
)
|
253 |
+
|
254 |
+
# Authentication
|
255 |
+
HF_TOKEN = hf_token or os.environ.get("HF_TOKEN") or get_token()
|
256 |
+
|
257 |
+
if not HF_TOKEN:
|
258 |
+
logger.error("No HuggingFace token found. Please provide token via:")
|
259 |
+
logger.error(" 1. --hf-token argument")
|
260 |
+
logger.error(" 2. HF_TOKEN environment variable")
|
261 |
+
logger.error(" 3. Run 'huggingface-cli login'")
|
262 |
+
sys.exit(1)
|
263 |
+
|
264 |
+
logger.info("HuggingFace token found, authenticating...")
|
265 |
+
login(token=HF_TOKEN)
|
266 |
+
|
267 |
+
# Initialize harmony encoding
|
268 |
+
logger.info("Loading harmony encoding...")
|
269 |
+
encoding = load_harmony_encoding(HarmonyEncodingName.HARMONY_GPT_OSS)
|
270 |
+
|
271 |
+
# Get stop tokens from harmony
|
272 |
+
stop_token_ids = encoding.stop_tokens_for_assistant_action()
|
273 |
+
logger.info(f"[VERBOSE] Harmony stop token IDs: {stop_token_ids}")
|
274 |
+
|
275 |
+
# Initialize vLLM
|
276 |
+
logger.info(f"Loading model: {model_id}")
|
277 |
+
logger.info("Note: vLLM will handle batching automatically for optimal throughput")
|
278 |
+
try:
|
279 |
+
llm = LLM(
|
280 |
+
model=model_id,
|
281 |
+
tensor_parallel_size=tensor_parallel_size,
|
282 |
+
gpu_memory_utilization=gpu_memory_utilization,
|
283 |
+
trust_remote_code=True,
|
284 |
+
dtype="bfloat16",
|
285 |
+
)
|
286 |
+
logger.info("[VERBOSE] Model loaded successfully")
|
287 |
+
except Exception as e:
|
288 |
+
logger.error(f"Failed to load model with vLLM: {e}")
|
289 |
+
if "mxfp4" in str(e).lower():
|
290 |
+
logger.error("This appears to be a quantization format issue.")
|
291 |
+
logger.error("The model uses mxfp4 quantization which requires specific support.")
|
292 |
+
sys.exit(1)
|
293 |
+
|
294 |
+
# Create sampling parameters
|
295 |
+
sampling_params = SamplingParams(
|
296 |
+
temperature=temperature,
|
297 |
+
max_tokens=max_tokens,
|
298 |
+
stop_token_ids=stop_token_ids,
|
299 |
+
)
|
300 |
+
logger.info(f"[VERBOSE] Sampling params: temp={temperature}, max_tokens={max_tokens}")
|
301 |
+
|
302 |
+
# Load dataset
|
303 |
+
logger.info(f"Loading dataset: {input_dataset}")
|
304 |
+
dataset = load_dataset(input_dataset, split="train")
|
305 |
+
|
306 |
+
# Validate prompt column
|
307 |
+
if prompt_column not in dataset.column_names:
|
308 |
+
logger.error(
|
309 |
+
f"Column '{prompt_column}' not found. Available columns: {dataset.column_names}"
|
310 |
+
)
|
311 |
+
sys.exit(1)
|
312 |
+
|
313 |
+
# Limit samples if requested
|
314 |
+
if max_samples:
|
315 |
+
dataset = dataset.select(range(min(max_samples, len(dataset))))
|
316 |
+
total_examples = len(dataset)
|
317 |
+
logger.info(f"Processing {total_examples:,} examples")
|
318 |
+
|
319 |
+
# Prepare prompts using harmony format
|
320 |
+
logger.info(f"Preparing prompts with harmony format and reasoning_level={reasoning_level}...")
|
321 |
+
prefill_ids_list = []
|
322 |
+
prompts = []
|
323 |
+
|
324 |
+
for i, example in enumerate(tqdm(dataset, desc="Preparing prompts")):
|
325 |
+
prompt_text = example[prompt_column]
|
326 |
+
prompts.append(prompt_text)
|
327 |
+
|
328 |
+
# Create harmony conversation
|
329 |
+
# Inject reasoning level into developer message
|
330 |
+
developer_content = DeveloperContent.new()
|
331 |
+
if reasoning_level:
|
332 |
+
developer_content = developer_content.with_instructions(
|
333 |
+
f"Reasoning: {reasoning_level}"
|
334 |
+
)
|
335 |
+
|
336 |
+
convo = Conversation.from_messages([
|
337 |
+
Message.from_role_and_content(Role.SYSTEM, SystemContent.new()),
|
338 |
+
Message.from_role_and_content(Role.DEVELOPER, developer_content),
|
339 |
+
Message.from_role_and_content(Role.USER, prompt_text),
|
340 |
+
])
|
341 |
+
|
342 |
+
# Render to token IDs
|
343 |
+
prefill_ids = encoding.render_conversation_for_completion(convo, Role.ASSISTANT)
|
344 |
+
prefill_ids_list.append(prefill_ids)
|
345 |
+
|
346 |
+
# Log first few examples
|
347 |
+
if i < 10:
|
348 |
+
logger.info(f"[VERBOSE] Example {i} original text: {prompt_text[:200]}...")
|
349 |
+
logger.info(f"[VERBOSE] Example {i} prefill length: {len(prefill_ids)} tokens")
|
350 |
+
|
351 |
+
# Generate responses with vLLM
|
352 |
+
logger.info(f"Starting generation for {len(prefill_ids_list):,} prompts...")
|
353 |
+
logger.info("[VERBOSE] Using prompt_token_ids for generation")
|
354 |
+
|
355 |
+
start_time = time.time()
|
356 |
+
outputs = llm.generate(
|
357 |
+
prompt_token_ids=prefill_ids_list,
|
358 |
+
sampling_params=sampling_params,
|
359 |
+
)
|
360 |
+
end_time = time.time()
|
361 |
+
|
362 |
+
generation_time = end_time - start_time
|
363 |
+
logger.info(f"\n[VERBOSE] Generation Performance Metrics:")
|
364 |
+
logger.info(f"[VERBOSE] - Total time: {generation_time:.2f} seconds")
|
365 |
+
logger.info(f"[VERBOSE] - Throughput: {len(outputs) / generation_time:.2f} prompts/second")
|
366 |
+
logger.info(f"[VERBOSE] - Average time per prompt: {generation_time / len(outputs):.2f} seconds")
|
367 |
+
|
368 |
+
# Parse outputs using harmony format
|
369 |
+
logger.info("Parsing generated outputs with harmony format...")
|
370 |
+
results = []
|
371 |
+
|
372 |
+
# Track statistics
|
373 |
+
parse_stats = {"success": 0, "empty": 0, "error": 0}
|
374 |
+
|
375 |
+
for i, output in enumerate(tqdm(outputs, desc="Parsing outputs")):
|
376 |
+
gen = output.outputs[0]
|
377 |
+
text = gen.text
|
378 |
+
output_tokens = gen.token_ids
|
379 |
+
|
380 |
+
logger.debug(f"[VERBOSE] Output {i}: {len(output_tokens)} tokens, {len(text)} chars")
|
381 |
+
|
382 |
+
try:
|
383 |
+
# Parse with harmony
|
384 |
+
entries = encoding.parse_messages_from_completion_tokens(output_tokens, Role.ASSISTANT)
|
385 |
+
|
386 |
+
# Convert to our format
|
387 |
+
parsed = parse_harmony_messages(entries, prompts[i])
|
388 |
+
|
389 |
+
if parsed["think"] or parsed["content"]:
|
390 |
+
parse_stats["success"] += 1
|
391 |
+
else:
|
392 |
+
parse_stats["empty"] += 1
|
393 |
+
|
394 |
+
# Verbose logging for first 10 examples
|
395 |
+
if i < 10:
|
396 |
+
logger.info(f"\n[VERBOSE] ========== Example {i} Output ==========")
|
397 |
+
logger.info(f"[VERBOSE] Original prompt: {prompts[i][:200]}...")
|
398 |
+
logger.info(f"[VERBOSE] Raw text output: {text}")
|
399 |
+
logger.info(f"[VERBOSE] Harmony entries: {len(entries)}")
|
400 |
+
for j, entry in enumerate(entries):
|
401 |
+
logger.info(f"[VERBOSE] Entry {j}: {json.dumps(entry.to_dict(), indent=2)}")
|
402 |
+
logger.info(f"[VERBOSE] Parsed think ({len(parsed['think'])} chars): {parsed['think'][:500]}...")
|
403 |
+
logger.info(f"[VERBOSE] Parsed content ({len(parsed['content'])} chars): {parsed['content'][:500]}...")
|
404 |
+
logger.info(f"[VERBOSE] ====================================\n")
|
405 |
+
|
406 |
+
except Exception as e:
|
407 |
+
logger.error(f"[VERBOSE] Error parsing output {i}: {e}")
|
408 |
+
parse_stats["error"] += 1
|
409 |
+
# Fallback: use raw text
|
410 |
+
parsed = {
|
411 |
+
"prompt": prompts[i],
|
412 |
+
"think": "",
|
413 |
+
"content": text,
|
414 |
+
"raw_output": text
|
415 |
+
}
|
416 |
+
|
417 |
+
result = {
|
418 |
+
"prompt": parsed["prompt"],
|
419 |
+
"think": parsed["think"],
|
420 |
+
"content": parsed["content"],
|
421 |
+
"raw_output": parsed["raw_output"],
|
422 |
+
"reasoning_level": reasoning_level,
|
423 |
+
"model": model_id,
|
424 |
+
}
|
425 |
+
results.append(result)
|
426 |
+
|
427 |
+
# Log parsing statistics
|
428 |
+
logger.info(f"\n[VERBOSE] Parsing Statistics:")
|
429 |
+
logger.info(f"[VERBOSE] - Successfully parsed: {parse_stats['success']} ({parse_stats['success']/len(outputs)*100:.1f}%)")
|
430 |
+
logger.info(f"[VERBOSE] - Empty results: {parse_stats['empty']} ({parse_stats['empty']/len(outputs)*100:.1f}%)")
|
431 |
+
logger.info(f"[VERBOSE] - Parse errors: {parse_stats['error']} ({parse_stats['error']/len(outputs)*100:.1f}%)")
|
432 |
+
|
433 |
+
# Create dataset
|
434 |
+
logger.info("Creating output dataset...")
|
435 |
+
output_dataset = Dataset.from_list(results)
|
436 |
+
|
437 |
+
# Create dataset card
|
438 |
+
logger.info("Creating dataset card...")
|
439 |
+
card_content = create_dataset_card(
|
440 |
+
input_dataset=input_dataset,
|
441 |
+
model_id=model_id,
|
442 |
+
prompt_column=prompt_column,
|
443 |
+
reasoning_level=reasoning_level,
|
444 |
+
num_examples=total_examples,
|
445 |
+
generation_time=generation_start_time,
|
446 |
+
tensor_parallel_size=tensor_parallel_size,
|
447 |
+
temperature=temperature,
|
448 |
+
max_tokens=max_tokens,
|
449 |
+
)
|
450 |
+
|
451 |
+
# Push to hub
|
452 |
+
logger.info(f"Pushing dataset to: {output_dataset_hub_id}")
|
453 |
+
output_dataset.push_to_hub(output_dataset_hub_id, token=HF_TOKEN)
|
454 |
+
|
455 |
+
# Push dataset card
|
456 |
+
card = DatasetCard(card_content)
|
457 |
+
card.push_to_hub(output_dataset_hub_id, token=HF_TOKEN)
|
458 |
+
|
459 |
+
logger.info("✅ Generation complete!")
|
460 |
+
logger.info(
|
461 |
+
f"Dataset available at: https://huggingface.co/datasets/{output_dataset_hub_id}"
|
462 |
+
)
|
463 |
+
|
464 |
+
# Final summary
|
465 |
+
logger.info(f"\n[VERBOSE] ========== FINAL SUMMARY ==========")
|
466 |
+
logger.info(f"[VERBOSE] Model: {model_id}")
|
467 |
+
logger.info(f"[VERBOSE] Reasoning level: {reasoning_level}")
|
468 |
+
logger.info(f"[VERBOSE] Examples processed: {total_examples}")
|
469 |
+
logger.info(f"[VERBOSE] Temperature: {temperature}")
|
470 |
+
logger.info(f"[VERBOSE] Max tokens: {max_tokens}")
|
471 |
+
logger.info(f"[VERBOSE] GPU config: {tensor_parallel_size} GPU(s)")
|
472 |
+
logger.info(f"[VERBOSE] ====================================")
|
473 |
+
|
474 |
+
|
475 |
+
if __name__ == "__main__":
|
476 |
+
if len(sys.argv) > 1:
|
477 |
+
parser = argparse.ArgumentParser(
|
478 |
+
description="Generate responses with reasoning using OpenAI GPT OSS models (Harmony format)",
|
479 |
+
formatter_class=argparse.RawDescriptionHelpFormatter,
|
480 |
+
epilog="""
|
481 |
+
Examples:
|
482 |
+
# Generate haiku with reasoning
|
483 |
+
uv run gpt_oss_vllm_harmony.py \\
|
484 |
+
--input-dataset davanstrien/haiku_dpo \\
|
485 |
+
--output-dataset username/haiku-reasoning \\
|
486 |
+
--prompt-column question
|
487 |
+
|
488 |
+
# Any prompt dataset
|
489 |
+
uv run gpt_oss_vllm_harmony.py \\
|
490 |
+
--input-dataset username/prompts \\
|
491 |
+
--output-dataset username/responses-reasoning \\
|
492 |
+
--reasoning-level high \\
|
493 |
+
--max-samples 100
|
494 |
+
|
495 |
+
# Use larger 120B model (requires 4x L40S GPUs)
|
496 |
+
uv run gpt_oss_vllm_harmony.py \\
|
497 |
+
--input-dataset username/prompts \\
|
498 |
+
--output-dataset username/responses-reasoning \\
|
499 |
+
--model-id openai/gpt-oss-120b \\
|
500 |
+
--tensor-parallel-size 4
|
501 |
+
""",
|
502 |
+
)
|
503 |
+
|
504 |
+
parser.add_argument(
|
505 |
+
"--input-dataset",
|
506 |
+
type=str,
|
507 |
+
required=True,
|
508 |
+
help="Input dataset on Hugging Face Hub",
|
509 |
+
)
|
510 |
+
parser.add_argument(
|
511 |
+
"--output-dataset",
|
512 |
+
type=str,
|
513 |
+
required=True,
|
514 |
+
help="Output dataset name on Hugging Face Hub",
|
515 |
+
)
|
516 |
+
parser.add_argument(
|
517 |
+
"--prompt-column",
|
518 |
+
type=str,
|
519 |
+
default="prompt",
|
520 |
+
help="Column containing prompts (default: prompt)",
|
521 |
+
)
|
522 |
+
parser.add_argument(
|
523 |
+
"--model-id",
|
524 |
+
type=str,
|
525 |
+
default="openai/gpt-oss-20b",
|
526 |
+
help="Model to use (default: openai/gpt-oss-20b)",
|
527 |
+
)
|
528 |
+
parser.add_argument(
|
529 |
+
"--reasoning-level",
|
530 |
+
type=str,
|
531 |
+
choices=["high", "medium", "low"],
|
532 |
+
default="high",
|
533 |
+
help="Reasoning effort level (default: high)",
|
534 |
+
)
|
535 |
+
parser.add_argument(
|
536 |
+
"--max-samples", type=int, help="Maximum number of samples to process"
|
537 |
+
)
|
538 |
+
parser.add_argument(
|
539 |
+
"--temperature",
|
540 |
+
type=float,
|
541 |
+
default=0.7,
|
542 |
+
help="Sampling temperature (default: 0.7)",
|
543 |
+
)
|
544 |
+
parser.add_argument(
|
545 |
+
"--max-tokens",
|
546 |
+
type=int,
|
547 |
+
default=512,
|
548 |
+
help="Maximum tokens to generate (default: 512)",
|
549 |
+
)
|
550 |
+
parser.add_argument(
|
551 |
+
"--gpu-memory-utilization",
|
552 |
+
type=float,
|
553 |
+
default=0.90,
|
554 |
+
help="GPU memory utilization (default: 0.90)",
|
555 |
+
)
|
556 |
+
parser.add_argument(
|
557 |
+
"--tensor-parallel-size",
|
558 |
+
type=int,
|
559 |
+
help="Number of GPUs to use (default: auto-detect)",
|
560 |
+
)
|
561 |
+
parser.add_argument(
|
562 |
+
"--hf-token",
|
563 |
+
type=str,
|
564 |
+
help="Hugging Face token (can also use HF_TOKEN env var)",
|
565 |
+
)
|
566 |
+
|
567 |
+
args = parser.parse_args()
|
568 |
+
|
569 |
+
main(
|
570 |
+
input_dataset=args.input_dataset,
|
571 |
+
output_dataset_hub_id=args.output_dataset,
|
572 |
+
prompt_column=args.prompt_column,
|
573 |
+
model_id=args.model_id,
|
574 |
+
reasoning_level=args.reasoning_level,
|
575 |
+
max_samples=args.max_samples,
|
576 |
+
temperature=args.temperature,
|
577 |
+
max_tokens=args.max_tokens,
|
578 |
+
gpu_memory_utilization=args.gpu_memory_utilization,
|
579 |
+
tensor_parallel_size=args.tensor_parallel_size,
|
580 |
+
hf_token=args.hf_token,
|
581 |
+
)
|
582 |
+
else:
|
583 |
+
# Show HF Jobs example when run without arguments
|
584 |
+
print("""
|
585 |
+
OpenAI GPT OSS Reasoning Generation Script (Harmony Format)
|
586 |
+
==========================================================
|
587 |
+
|
588 |
+
This script requires arguments. For usage information:
|
589 |
+
uv run gpt_oss_vllm_harmony.py --help
|
590 |
+
|
591 |
+
Example HF Jobs command for 20B model:
|
592 |
+
hf jobs uv run \\
|
593 |
+
--flavor a10g-large \\ # 20B model requires ~40GB memory
|
594 |
+
https://huggingface.co/datasets/uv-scripts/openai-reasoning/raw/main/gpt_oss_vllm_harmony.py \\
|
595 |
+
--input-dataset davanstrien/haiku_dpo \\
|
596 |
+
--output-dataset username/haiku-reasoning \\
|
597 |
+
--prompt-column question \\
|
598 |
+
--reasoning-level high
|
599 |
+
|
600 |
+
Example HF Jobs command for 120B model:
|
601 |
+
hf jobs uv run \\
|
602 |
+
--flavor l40s-4x \\ # 120B model requires ~240GB memory
|
603 |
+
https://huggingface.co/datasets/uv-scripts/openai-reasoning/raw/main/gpt_oss_vllm_harmony.py \\
|
604 |
+
--input-dataset username/prompts \\
|
605 |
+
--output-dataset username/responses-reasoning \\
|
606 |
+
--model-id openai/gpt-oss-120b \\
|
607 |
+
--reasoning-level high
|
608 |
+
""")
|