Spaces:
Build error
Build error
Upload folder using huggingface_hub
Browse files- .gitignore +3 -0
- DEPLOY_CHECKLIST.md +52 -0
- README.md +283 -12
- app.py +180 -0
- install_requirements.bat +35 -0
- install_requirements.py +83 -0
- requirements-base.txt +8 -0
- requirements-flash.txt +2 -0
- requirements.txt +17 -0
- run_transformers_training.py +964 -0
- transformers_config.json +162 -0
- update_space.py +278 -0
.gitignore
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
.env
|
2 |
+
*.pyc
|
3 |
+
__pycache__
|
DEPLOY_CHECKLIST.md
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Phi-4 Training Critical Deployment Checklist
|
2 |
+
|
3 |
+
## Essential Configuration Requirements
|
4 |
+
|
5 |
+
### 1. Model Configuration
|
6 |
+
- [ ] Model name: `unsloth/phi-4-unsloth-bnb-4bit`
|
7 |
+
- [ ] BF16 precision enabled, FP16 disabled
|
8 |
+
- [ ] Appropriate sequence length (2048)
|
9 |
+
- [ ] LoRA parameters correctly configured (r: 32, alpha: 16)
|
10 |
+
|
11 |
+
### 2. Hardware & Resource Management
|
12 |
+
- [ ] Per-device batch size ≤ 16
|
13 |
+
- [ ] Gradient accumulation steps ≥ 3
|
14 |
+
- [ ] Gradient checkpointing enabled
|
15 |
+
- [ ] Memory usage limits properly set (85% of GPU capacity)
|
16 |
+
|
17 |
+
### 3. Critical Dataset Handling Rules
|
18 |
+
- [ ] **NO REORDERING of dataset entries** - original order must be preserved
|
19 |
+
- [ ] **NO COMBINING of separate entries** - each entry must remain distinct
|
20 |
+
- [ ] **SEQUENTIAL PROCESSING required** - entries must be processed one after another
|
21 |
+
- [ ] `sort_by_id` and `maintain_paper_order` flags properly set to preserve data sequence
|
22 |
+
- [ ] Sequential sampler used with no shuffling (`"shuffle": false`)
|
23 |
+
- [ ] Dataset sequential integrity verified with validation samples
|
24 |
+
- [ ] Conversation structure preserved (original format maintained)
|
25 |
+
|
26 |
+
### 4. Essential Error Handling
|
27 |
+
- [ ] Clear error catching for dataset loading issues
|
28 |
+
- [ ] Memory tracking at key training points
|
29 |
+
- [ ] Low-verbosity logging for HF Space compatibility
|
30 |
+
|
31 |
+
### 5. Training Core Requirements
|
32 |
+
- [ ] Appropriate learning rate (2e-5)
|
33 |
+
- [ ] Proper checkpointing frequency
|
34 |
+
- [ ] Hub settings correctly configured for model saving
|
35 |
+
|
36 |
+
---
|
37 |
+
|
38 |
+
## Pre-Deployment Verification
|
39 |
+
|
40 |
+
| Requirement | Status | Notes |
|
41 |
+
|-------------|--------|-------|
|
42 |
+
| Data sequential integrity | | Confirm entries processed in order |
|
43 |
+
| GPU memory within limits | | Check peak memory doesn't exceed 20GB per GPU |
|
44 |
+
| Training batch verification | | Verify first few batches maintain proper order |
|
45 |
+
|
46 |
+
---
|
47 |
+
|
48 |
+
**Current Hardware**: 4× NVIDIA L4 GPUs (24GB VRAM each)
|
49 |
+
|
50 |
+
**CRITICAL REMINDER**: Data sequence preservation is the highest priority - any shuffling, reordering, or combining of entries will compromise model quality.
|
51 |
+
|
52 |
+
*Last Updated: 2025-03-09*
|
README.md
CHANGED
@@ -1,12 +1,283 @@
|
|
1 |
-
---
|
2 |
-
title:
|
3 |
-
emoji:
|
4 |
-
colorFrom: blue
|
5 |
-
colorTo: purple
|
6 |
-
sdk: gradio
|
7 |
-
sdk_version: 5.
|
8 |
-
app_file: app.py
|
9 |
-
pinned: false
|
10 |
-
|
11 |
-
|
12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
title: Phi-4 Unsloth Training
|
3 |
+
emoji: 🧠
|
4 |
+
colorFrom: blue
|
5 |
+
colorTo: purple
|
6 |
+
sdk: gradio
|
7 |
+
sdk_version: 5.17.0
|
8 |
+
app_file: app.py
|
9 |
+
pinned: false
|
10 |
+
license: mit
|
11 |
+
---
|
12 |
+
|
13 |
+
# Phi-4 Unsloth Optimized Training
|
14 |
+
|
15 |
+
This space is dedicated to training Microsoft's Phi-4 model using Unsloth optimizations for enhanced performance and efficiency. The training process utilizes 4-bit quantization and advanced memory optimizations.
|
16 |
+
|
17 |
+
## Installation
|
18 |
+
|
19 |
+
This Hugging Face Space automatically installs dependencies from requirements.txt. The following packages are included:
|
20 |
+
|
21 |
+
### Installation Process
|
22 |
+
|
23 |
+
For clearer dependency management, the installation is split into multiple files:
|
24 |
+
|
25 |
+
1. **Base Dependencies (requirements-base.txt)**:
|
26 |
+
- Core packages like torch, transformers, accelerate, etc.
|
27 |
+
- Install with: `pip install -r requirements-base.txt`
|
28 |
+
|
29 |
+
2. **Standard Dependencies (requirements.txt)**:
|
30 |
+
- References base requirements and adds additional packages
|
31 |
+
- Install with: `pip install -r requirements.txt`
|
32 |
+
|
33 |
+
3. **Flash Attention (requirements-flash.txt)** (Optional):
|
34 |
+
- For faster attention computation
|
35 |
+
- Install with: `pip install -r requirements-flash.txt --no-build-isolation`
|
36 |
+
|
37 |
+
Using this staged approach helps prevent dependency conflicts and installation issues.
|
38 |
+
|
39 |
+
### Essential Dependencies
|
40 |
+
|
41 |
+
- **unsloth** (>=2024.3): Required for optimized 4-bit training
|
42 |
+
- **peft** (>=0.9.0): Required for parameter-efficient fine-tuning
|
43 |
+
- **transformers** (>=4.36.0): For model architecture and tokenization
|
44 |
+
- **einops**: Required by Unsloth for tensor manipulation
|
45 |
+
- **sentencepiece**: Required for tokenization
|
46 |
+
|
47 |
+
### Optional Dependencies
|
48 |
+
|
49 |
+
- **flash-attn**: Optional for faster attention computation (not included by default as it can cause build issues)
|
50 |
+
|
51 |
+
## Features
|
52 |
+
|
53 |
+
- 4-bit quantization using Unsloth
|
54 |
+
- Optimized training pipeline
|
55 |
+
- Cognitive dataset integration
|
56 |
+
- Advanced memory management
|
57 |
+
- Gradient checkpointing
|
58 |
+
- Sequential data processing
|
59 |
+
|
60 |
+
## Configuration Files
|
61 |
+
|
62 |
+
- `transformers_config.json`: Model and training parameters
|
63 |
+
- `hardware_config.json`: Hardware-specific optimizations
|
64 |
+
- `dataset_config.json`: Dataset processing settings
|
65 |
+
- `requirements.txt`: Required dependencies
|
66 |
+
|
67 |
+
## Training Process
|
68 |
+
|
69 |
+
The training utilizes the following optimizations:
|
70 |
+
- Unsloth's 4-bit quantization
|
71 |
+
- Custom chat templates for Phi-4
|
72 |
+
- Paper-order preservation
|
73 |
+
- Efficient memory usage
|
74 |
+
- Gradient accumulation
|
75 |
+
|
76 |
+
## Dataset
|
77 |
+
|
78 |
+
Training uses the cognitive dataset with:
|
79 |
+
- Maintained paper order
|
80 |
+
- Proper metadata handling
|
81 |
+
- Optimized sequence length
|
82 |
+
- Efficient batching
|
83 |
+
|
84 |
+
## Hardware Requirements
|
85 |
+
|
86 |
+
- GPU: A10G or better
|
87 |
+
- VRAM: 24GB minimum
|
88 |
+
- RAM: 32GB recommended
|
89 |
+
|
90 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
91 |
+
|
92 |
+
# Phase 1: Domain Adaptation (Unsupervised)
|
93 |
+
|
94 |
+
This directory contains the code and configuration for domain adaptation of the phi-4-unsloth-bnb-4bit model to the cognitive science domain. This phase produces our domain-adapted model: [George-API/phi-4-research-assistant](https://huggingface.co/George-API/phi-4-research-assistant).
|
95 |
+
|
96 |
+
## Overview
|
97 |
+
|
98 |
+
Domain adaptation is the first phase of our training process, where we expose the model to a large corpus of cognitive science texts to help it learn domain-specific vocabulary, concepts, and patterns. This phase prepares the model for the more focused supervised fine-tuning in Phase 2.
|
99 |
+
|
100 |
+
## Files
|
101 |
+
|
102 |
+
### Core Training Files
|
103 |
+
- `run_transformers_training.py`: Main script for domain adaptation
|
104 |
+
- `transformers_config.json`: Model and training parameters
|
105 |
+
- `hardware_config.json`: Hardware-specific optimizations
|
106 |
+
- `dataset_config.json`: Dataset loading and processing settings
|
107 |
+
- `requirements.txt`: Required Python packages
|
108 |
+
|
109 |
+
### Analysis & Utilities
|
110 |
+
- `check_tokenization.py`: Script to analyze token distributions
|
111 |
+
- `update_space.py`: Hugging Face Space update utility
|
112 |
+
- `.env`: Environment variables (API tokens, etc.)
|
113 |
+
|
114 |
+
## Setup
|
115 |
+
|
116 |
+
1. **Environment Setup**:
|
117 |
+
```bash
|
118 |
+
python -m venv venv
|
119 |
+
source venv/bin/activate # or `venv\Scripts\activate` on Windows
|
120 |
+
pip install -r requirements.txt
|
121 |
+
```
|
122 |
+
|
123 |
+
2. **Environment Variables**:
|
124 |
+
Create `.env` file with:
|
125 |
+
```
|
126 |
+
HUGGINGFACE_TOKEN=your_token_here
|
127 |
+
```
|
128 |
+
|
129 |
+
3. **Verify Setup**:
|
130 |
+
```bash
|
131 |
+
python check_tokenization.py # Ensures tokenizer works
|
132 |
+
```
|
133 |
+
|
134 |
+
## How It Works
|
135 |
+
|
136 |
+
1. **Data Loading**: Loads pre-tokenized data from the Hugging Face dataset
|
137 |
+
2. **Sequential Processing**: Processes data in order, maintaining the integrity of research papers
|
138 |
+
3. **Efficient Training**: Uses pre-quantized Unsloth 4-bit model for memory-efficient and faster training
|
139 |
+
4. **Checkpointing**: Saves regular checkpoints and pushes to Hub
|
140 |
+
5. **Monitoring**: Logs detailed metrics and statistics during training
|
141 |
+
6. **Model Publishing**: Pushes the trained model to Hugging Face Hub
|
142 |
+
|
143 |
+
## Key Features
|
144 |
+
|
145 |
+
### Memory-Efficient Training
|
146 |
+
|
147 |
+
The training setup is optimized for A10G GPUs:
|
148 |
+
- Uses pre-quantized 4-bit model (no additional quantization needed)
|
149 |
+
- Gradient checkpointing for memory efficiency
|
150 |
+
- Flash attention for faster training
|
151 |
+
- bfloat16 mixed precision training
|
152 |
+
- Optimized batch sizes for maximum throughput
|
153 |
+
|
154 |
+
### Sequential Processing
|
155 |
+
|
156 |
+
The training script ensures that chunks from the same research paper are processed together by:
|
157 |
+
- Sorting the dataset by ID
|
158 |
+
- Using a SequentialSampler to maintain order
|
159 |
+
- Processing chunks sequentially (average 1,673 tokens per chunk)
|
160 |
+
|
161 |
+
### Data Collator
|
162 |
+
|
163 |
+
The `SimpleDataCollator` class:
|
164 |
+
- Preserves pre-tokenized data format
|
165 |
+
- Processes each entry independently
|
166 |
+
- Provides detailed logging of processing statistics
|
167 |
+
- Handles errors gracefully
|
168 |
+
|
169 |
+
### Checkpointing
|
170 |
+
|
171 |
+
The training process saves checkpoints:
|
172 |
+
- Every 200 steps
|
173 |
+
- Pushes to Hub on every save
|
174 |
+
- Maintains up to 5 recent checkpoints
|
175 |
+
- Automatically resumes from the latest checkpoint if interrupted
|
176 |
+
|
177 |
+
## Hardware Requirements
|
178 |
+
|
179 |
+
This training setup is optimized for:
|
180 |
+
- 2x NVIDIA A10G GPUs (24GB VRAM each)
|
181 |
+
- 92GB System RAM
|
182 |
+
- CUDA 11.8 or higher
|
183 |
+
|
184 |
+
Memory breakdown per GPU:
|
185 |
+
- Model (4-bit): ~3.5GB
|
186 |
+
- Optimizer states: ~1GB
|
187 |
+
- Batch memory: ~2GB
|
188 |
+
- Peak usage: 18-20GB
|
189 |
+
- Safe headroom: 4-6GB
|
190 |
+
|
191 |
+
## Configuration
|
192 |
+
|
193 |
+
Key parameters in `transformers_config.json`:
|
194 |
+
|
195 |
+
- `model_name`: unsloth/phi-4-unsloth-bnb-4bit
|
196 |
+
- `learning_rate`: 2e-5
|
197 |
+
- `num_train_epochs`: 3
|
198 |
+
- `per_device_train_batch_size`: 16
|
199 |
+
- `gradient_accumulation_steps`: 4
|
200 |
+
- `effective_batch_size`: 128 (16 * 4 * 2 GPUs)
|
201 |
+
- `max_seq_length`: 2048
|
202 |
+
- `lr_scheduler_type`: "cosine"
|
203 |
+
- `warmup_ratio`: 0.03
|
204 |
+
- `neftune_noise_alpha`: 5
|
205 |
+
|
206 |
+
The configuration is optimized for:
|
207 |
+
- Maximum memory efficiency with pre-quantized model
|
208 |
+
- Stable training with cosine learning rate schedule
|
209 |
+
- Effective gradient updates with accumulation
|
210 |
+
- Regular checkpointing and Hub updates
|
211 |
+
|
212 |
+
## Running Domain Adaptation
|
213 |
+
|
214 |
+
To start domain adaptation:
|
215 |
+
|
216 |
+
```bash
|
217 |
+
python run_transformers_training.py
|
218 |
+
```
|
219 |
+
|
220 |
+
The script will:
|
221 |
+
1. Load the pre-quantized model and dataset
|
222 |
+
2. Apply optimized training parameters
|
223 |
+
3. Process the data sequentially
|
224 |
+
4. Train the model for 3 epochs
|
225 |
+
5. Save and push checkpoints to Hub regularly
|
226 |
+
|
227 |
+
## Using the Model
|
228 |
+
|
229 |
+
After training, you can use the domain-adapted model:
|
230 |
+
|
231 |
+
```python
|
232 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
233 |
+
|
234 |
+
# Load the domain-adapted model
|
235 |
+
model_name = "George-API/phi-4-research-assistant"
|
236 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
237 |
+
model = AutoModelForCausalLM.from_pretrained(model_name,
|
238 |
+
device_map="auto",
|
239 |
+
torch_dtype="bfloat16")
|
240 |
+
|
241 |
+
# Generate text
|
242 |
+
input_text = "The hippocampus is involved in"
|
243 |
+
inputs = tokenizer(input_text, return_tensors="pt")
|
244 |
+
outputs = model.generate(**inputs, max_length=100)
|
245 |
+
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
246 |
+
```
|
247 |
+
|
248 |
+
## Chat Format Example
|
249 |
+
|
250 |
+
Phi-4 works best with its native chat template:
|
251 |
+
|
252 |
+
```python
|
253 |
+
from transformers import pipeline
|
254 |
+
|
255 |
+
pipeline = pipeline(
|
256 |
+
"text-generation",
|
257 |
+
model="George-API/phi-4-research-assistant",
|
258 |
+
model_kwargs={"torch_dtype": "bfloat16"},
|
259 |
+
device_map="auto",
|
260 |
+
)
|
261 |
+
|
262 |
+
messages = [
|
263 |
+
{"role": "system", "content": "You are an expert in cognitive science."},
|
264 |
+
{"role": "user", "content": "Explain the role of the hippocampus in memory formation."},
|
265 |
+
]
|
266 |
+
|
267 |
+
outputs = pipeline(messages, max_new_tokens=256)
|
268 |
+
print(outputs[0]["generated_text"])
|
269 |
+
```
|
270 |
+
|
271 |
+
## Expected Outcomes
|
272 |
+
|
273 |
+
After domain adaptation, the model should:
|
274 |
+
- Have a better understanding of cognitive science terminology
|
275 |
+
- Show improved performance on domain-specific tasks
|
276 |
+
- Be ready for supervised fine-tuning in Phase 2
|
277 |
+
|
278 |
+
## Next Steps
|
279 |
+
|
280 |
+
After completing domain adaptation:
|
281 |
+
1. Evaluate the model's performance on cognitive science texts
|
282 |
+
2. Proceed to Phase 2 (Supervised Fine-Tuning)
|
283 |
+
3. Use TensorBoard to analyze training metrics
|
app.py
ADDED
@@ -0,0 +1,180 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding=utf-8
|
3 |
+
|
4 |
+
import os
|
5 |
+
import sys
|
6 |
+
import json
|
7 |
+
import logging
|
8 |
+
import subprocess
|
9 |
+
import time
|
10 |
+
from datetime import datetime
|
11 |
+
|
12 |
+
# Configure logging to match HF Space logs
|
13 |
+
logging.basicConfig(
|
14 |
+
level=logging.INFO,
|
15 |
+
format="%(asctime)s - %(levelname)s - %(message)s",
|
16 |
+
handlers=[logging.StreamHandler(sys.stdout)]
|
17 |
+
)
|
18 |
+
logger = logging.getLogger(__name__)
|
19 |
+
|
20 |
+
# Set other loggers to WARNING to reduce noise and ensure our logs are visible
|
21 |
+
logging.getLogger("transformers").setLevel(logging.WARNING)
|
22 |
+
logging.getLogger("datasets").setLevel(logging.WARNING)
|
23 |
+
logging.getLogger("accelerate").setLevel(logging.WARNING)
|
24 |
+
logging.getLogger("torch").setLevel(logging.WARNING)
|
25 |
+
logging.getLogger("bitsandbytes").setLevel(logging.WARNING)
|
26 |
+
|
27 |
+
# Define a clean logging function for HF Space compatibility
|
28 |
+
def log_info(message):
|
29 |
+
"""Log information in a format compatible with Hugging Face Spaces"""
|
30 |
+
logger.info(message)
|
31 |
+
# Ensure output is flushed immediately for streaming
|
32 |
+
sys.stdout.flush()
|
33 |
+
|
34 |
+
# Configuration paths
|
35 |
+
CONFIG_DIR = "."
|
36 |
+
TRANSFORMERS_CONFIG = os.path.join(CONFIG_DIR, "transformers_config.json")
|
37 |
+
|
38 |
+
def load_config(config_path):
|
39 |
+
"""Load configuration from JSON file."""
|
40 |
+
try:
|
41 |
+
if os.path.exists(config_path):
|
42 |
+
with open(config_path, 'r') as f:
|
43 |
+
return json.load(f)
|
44 |
+
else:
|
45 |
+
log_info(f"Config file not found: {config_path}")
|
46 |
+
return None
|
47 |
+
except Exception as e:
|
48 |
+
log_info(f"Error loading config: {str(e)}")
|
49 |
+
return None
|
50 |
+
|
51 |
+
def display_config():
|
52 |
+
"""Display current training configuration."""
|
53 |
+
config = load_config(TRANSFORMERS_CONFIG)
|
54 |
+
|
55 |
+
if not config:
|
56 |
+
return "Error loading configuration file."
|
57 |
+
|
58 |
+
# Extract sub-configurations
|
59 |
+
transformers_config = config
|
60 |
+
hardware_config = config.get("hardware", {})
|
61 |
+
dataset_config = config.get("dataset", {})
|
62 |
+
|
63 |
+
model_name = transformers_config.get("model", {}).get("name") or transformers_config.get("model_name_or_path", "")
|
64 |
+
|
65 |
+
# Training parameters
|
66 |
+
training_config = transformers_config.get("training", {})
|
67 |
+
batch_size = training_config.get("per_device_train_batch_size", 16)
|
68 |
+
grad_accum = training_config.get("gradient_accumulation_steps", 3)
|
69 |
+
epochs = training_config.get("num_train_epochs", 3)
|
70 |
+
learning_rate = training_config.get("learning_rate", 2e-5)
|
71 |
+
|
72 |
+
# Hardware settings
|
73 |
+
gpu_count = hardware_config.get("specs", {}).get("gpu_count", 4)
|
74 |
+
gpu_type = hardware_config.get("specs", {}).get("gpu_type", "L4")
|
75 |
+
vram = hardware_config.get("specs", {}).get("vram_per_gpu", 24)
|
76 |
+
|
77 |
+
# Dataset info
|
78 |
+
dataset_name = dataset_config.get("dataset", {}).get("name", "")
|
79 |
+
|
80 |
+
# Format response as HTML for better display
|
81 |
+
html = f"""
|
82 |
+
<h2>Training Configuration</h2>
|
83 |
+
<h3>Model</h3>
|
84 |
+
<ul>
|
85 |
+
<li><b>Model:</b> {model_name}</li>
|
86 |
+
<li><b>Learning Rate:</b> {training_config.get('learning_rate', '2e-5')}</li>
|
87 |
+
<li><b>Per-Device Batch Size:</b> {batch_size}</li>
|
88 |
+
<li><b>Gradient Accumulation:</b> {grad_accum}</li>
|
89 |
+
<li><b>Total Effective Batch Size:</b> {batch_size} × {gpu_count} × {grad_accum} = {batch_size * gpu_count * grad_accum}</li>
|
90 |
+
<li><b>Epochs:</b> {epochs}</li>
|
91 |
+
<li><b>Precision:</b> {'BF16' if transformers_config.get('bf16', True) else 'FP16' if transformers_config.get('fp16', False) else 'FP32'}</li>
|
92 |
+
<li><b>Max Sequence Length:</b> {transformers_config.get('tokenizer', {}).get('max_seq_length', 2048)}</li>
|
93 |
+
</ul>
|
94 |
+
|
95 |
+
<h3>Hardware</h3>
|
96 |
+
<ul>
|
97 |
+
<li><b>GPU:</b> {gpu_count}× {gpu_type} ({vram} GB VRAM per GPU, total: {vram * gpu_count} GB)</li>
|
98 |
+
<li><b>Multi-GPU Strategy:</b> {hardware_config.get('training_optimizations', {}).get('multi_gpu_strategy', 'data_parallel')}</li>
|
99 |
+
<li><b>Memory Optimizations:</b> {'Gradient Checkpointing' if hardware_config.get('training_optimizations', {}).get('memory_optimizations', {}).get('use_gradient_checkpointing', True) else 'None'}</li>
|
100 |
+
</ul>
|
101 |
+
|
102 |
+
<h3>Dataset</h3>
|
103 |
+
<ul>
|
104 |
+
<li><b>Dataset:</b> {dataset_name}</li>
|
105 |
+
<li><b>Dataset Split:</b> {dataset_config.get('dataset', {}).get('split', 'train')}</li>
|
106 |
+
</ul>
|
107 |
+
"""
|
108 |
+
|
109 |
+
return html
|
110 |
+
|
111 |
+
def start_training():
|
112 |
+
"""Start the training process."""
|
113 |
+
try:
|
114 |
+
# Log configuration check
|
115 |
+
log_info("Preparing to start training process...")
|
116 |
+
log_info("Using consolidated configuration from transformers_config.json")
|
117 |
+
|
118 |
+
# Start training
|
119 |
+
log_info("Starting training process...")
|
120 |
+
|
121 |
+
# Run in a background process for HF Space
|
122 |
+
cmd = "python run_transformers_training.py"
|
123 |
+
|
124 |
+
# In HF Spaces, we don't need to handle process management ourselves
|
125 |
+
subprocess.Popen(cmd, shell=True, stdout=sys.stdout, stderr=sys.stderr)
|
126 |
+
|
127 |
+
log_info("Training process has been started. You can monitor progress in the logs.")
|
128 |
+
|
129 |
+
return "Training started successfully. Monitor progress in the Hugging Face Space logs."
|
130 |
+
|
131 |
+
except Exception as e:
|
132 |
+
error_msg = f"Error starting training: {str(e)}"
|
133 |
+
log_info(error_msg)
|
134 |
+
return error_msg
|
135 |
+
|
136 |
+
# Interface setup for gradio
|
137 |
+
def create_interface():
|
138 |
+
import gradio as gr
|
139 |
+
|
140 |
+
with gr.Blocks(title="Phi-4 Training Center") as demo:
|
141 |
+
gr.Markdown("# Phi-4 Research Assistant Training")
|
142 |
+
|
143 |
+
with gr.Row():
|
144 |
+
with gr.Column():
|
145 |
+
gr.Markdown("## Control Panel")
|
146 |
+
|
147 |
+
# Display current config
|
148 |
+
config_html = gr.HTML(display_config())
|
149 |
+
refresh_btn = gr.Button("Refresh Configuration")
|
150 |
+
|
151 |
+
# Training controls
|
152 |
+
train_btn = gr.Button("Start Training", variant="primary")
|
153 |
+
train_output = gr.Textbox(label="Status", interactive=False)
|
154 |
+
|
155 |
+
with gr.Column():
|
156 |
+
gr.Markdown("## Training Information")
|
157 |
+
gr.Markdown("""
|
158 |
+
### Hardware:
|
159 |
+
- 4× NVIDIA L4 GPUs (24GB VRAM per GPU, 96GB total)
|
160 |
+
- Training with BF16 precision
|
161 |
+
- Using Data Parallel for multi-GPU
|
162 |
+
- Effective batch size: 16 (per device) × 4 (GPUs) × 3 (gradient accumulation) = 192
|
163 |
+
|
164 |
+
### Notes:
|
165 |
+
- Training may take several hours depending on dataset size
|
166 |
+
- Check the Space logs for real-time progress
|
167 |
+
- Model checkpoints will be saved to ./results directory
|
168 |
+
""")
|
169 |
+
|
170 |
+
# Connect buttons to functions
|
171 |
+
refresh_btn.click(lambda: gr.update(value=display_config()), outputs=config_html)
|
172 |
+
train_btn.click(start_training, outputs=train_output)
|
173 |
+
|
174 |
+
return demo
|
175 |
+
|
176 |
+
if __name__ == "__main__":
|
177 |
+
# If run directly, create and launch the Gradio interface
|
178 |
+
demo = create_interface()
|
179 |
+
demo.queue()
|
180 |
+
demo.launch()
|
install_requirements.bat
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
@echo off
|
2 |
+
echo Installing Phi-4 Training Requirements
|
3 |
+
echo =====================================
|
4 |
+
echo.
|
5 |
+
|
6 |
+
REM Check if Python is available
|
7 |
+
where python >nul 2>&1
|
8 |
+
if %ERRORLEVEL% neq 0 (
|
9 |
+
echo Python not found! Please make sure Python is installed and in your PATH.
|
10 |
+
exit /b 1
|
11 |
+
)
|
12 |
+
|
13 |
+
echo Step 1: Installing base requirements...
|
14 |
+
python -m pip install -r requirements-base.txt
|
15 |
+
if %ERRORLEVEL% neq 0 (
|
16 |
+
echo Failed to install base requirements.
|
17 |
+
exit /b 1
|
18 |
+
)
|
19 |
+
echo Base requirements installed successfully.
|
20 |
+
echo.
|
21 |
+
|
22 |
+
echo Step 2: Installing additional requirements...
|
23 |
+
python -m pip install -r requirements.txt
|
24 |
+
if %ERRORLEVEL% neq 0 (
|
25 |
+
echo Failed to install additional requirements.
|
26 |
+
exit /b 1
|
27 |
+
)
|
28 |
+
echo Additional requirements installed successfully.
|
29 |
+
echo.
|
30 |
+
|
31 |
+
echo All required packages installed successfully!
|
32 |
+
echo To install optional flash-attention, run: python -m pip install -r requirements-flash.txt --no-build-isolation
|
33 |
+
echo.
|
34 |
+
|
35 |
+
pause
|
install_requirements.py
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding=utf-8
|
3 |
+
|
4 |
+
"""
|
5 |
+
Script to install requirements in the correct order for the Phi-4 training project.
|
6 |
+
This ensures base requirements are installed first, followed by additional requirements.
|
7 |
+
"""
|
8 |
+
|
9 |
+
import os
|
10 |
+
import sys
|
11 |
+
import subprocess
|
12 |
+
import argparse
|
13 |
+
import logging
|
14 |
+
from pathlib import Path
|
15 |
+
|
16 |
+
# Configure logging
|
17 |
+
logging.basicConfig(
|
18 |
+
level=logging.INFO,
|
19 |
+
format="%(asctime)s - %(levelname)s - %(message)s",
|
20 |
+
handlers=[logging.StreamHandler(sys.stdout)]
|
21 |
+
)
|
22 |
+
logger = logging.getLogger(__name__)
|
23 |
+
|
24 |
+
def install_requirements(include_flash=False):
|
25 |
+
"""Install requirements in the correct order."""
|
26 |
+
current_dir = Path(__file__).parent
|
27 |
+
base_req_path = current_dir / "requirements-base.txt"
|
28 |
+
main_req_path = current_dir / "requirements.txt"
|
29 |
+
flash_req_path = current_dir / "requirements-flash.txt"
|
30 |
+
|
31 |
+
if not base_req_path.exists():
|
32 |
+
logger.error(f"Base requirements file not found: {base_req_path}")
|
33 |
+
return False
|
34 |
+
|
35 |
+
if not main_req_path.exists():
|
36 |
+
logger.error(f"Main requirements file not found: {main_req_path}")
|
37 |
+
return False
|
38 |
+
|
39 |
+
logger.info("Installing dependencies in sequential order...")
|
40 |
+
|
41 |
+
try:
|
42 |
+
# Step 1: Install base requirements
|
43 |
+
logger.info(f"Step 1: Installing base requirements from {base_req_path}")
|
44 |
+
subprocess.run([sys.executable, "-m", "pip", "install", "-r", str(base_req_path)],
|
45 |
+
check=True)
|
46 |
+
logger.info("Base requirements installed successfully")
|
47 |
+
|
48 |
+
# Step 2: Install main requirements
|
49 |
+
logger.info(f"Step 2: Installing additional requirements from {main_req_path}")
|
50 |
+
subprocess.run([sys.executable, "-m", "pip", "install", "-r", str(main_req_path)],
|
51 |
+
check=True)
|
52 |
+
logger.info("Additional requirements installed successfully")
|
53 |
+
|
54 |
+
# Step 3: Optionally install flash-attention
|
55 |
+
if include_flash and flash_req_path.exists():
|
56 |
+
logger.info(f"Step 3: Installing flash-attention from {flash_req_path}")
|
57 |
+
subprocess.run([sys.executable, "-m", "pip", "install", "-r", str(flash_req_path), "--no-build-isolation"],
|
58 |
+
check=True)
|
59 |
+
logger.info("Flash-attention installed successfully")
|
60 |
+
elif include_flash:
|
61 |
+
logger.warning(f"Flash requirements file not found: {flash_req_path}")
|
62 |
+
|
63 |
+
logger.info("All required packages installed successfully!")
|
64 |
+
return True
|
65 |
+
|
66 |
+
except subprocess.CalledProcessError as e:
|
67 |
+
logger.error(f"Error installing dependencies: {str(e)}")
|
68 |
+
return False
|
69 |
+
|
70 |
+
def main():
|
71 |
+
parser = argparse.ArgumentParser(description="Install requirements for Phi-4 training")
|
72 |
+
parser.add_argument("--flash", action="store_true", help="Also install flash-attention (optional)")
|
73 |
+
args = parser.parse_args()
|
74 |
+
|
75 |
+
success = install_requirements(include_flash=args.flash)
|
76 |
+
if success:
|
77 |
+
logger.info("Installation completed successfully!")
|
78 |
+
else:
|
79 |
+
logger.error("Installation failed. Please check the logs for details.")
|
80 |
+
sys.exit(1)
|
81 |
+
|
82 |
+
if __name__ == "__main__":
|
83 |
+
main()
|
requirements-base.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch>=2.0.0
|
2 |
+
accelerate>=0.27.0
|
3 |
+
bitsandbytes>=0.41.0
|
4 |
+
datasets>=2.15.0
|
5 |
+
gradio>=5.17.0
|
6 |
+
huggingface-hub>=0.19.0
|
7 |
+
tensorboard>=2.15.0
|
8 |
+
transformers>=4.36.0
|
requirements-flash.txt
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
-r requirements-base.txt
|
2 |
+
flash-attn==2.5.2
|
requirements.txt
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
-r requirements-base.txt
|
2 |
+
einops>=0.7.0
|
3 |
+
filelock>=3.13.1
|
4 |
+
matplotlib>=3.7.0
|
5 |
+
numpy>=1.24.0
|
6 |
+
packaging>=23.0
|
7 |
+
peft>=0.9.0
|
8 |
+
psutil>=5.9.0
|
9 |
+
python-dotenv>=1.0.0
|
10 |
+
pyyaml>=6.0.1
|
11 |
+
regex>=2023.0.0
|
12 |
+
requests>=2.31.0
|
13 |
+
safetensors>=0.4.1
|
14 |
+
sentencepiece>=0.1.99
|
15 |
+
tqdm>=4.65.0
|
16 |
+
typing-extensions>=4.8.0
|
17 |
+
unsloth>=2024.3
|
run_transformers_training.py
ADDED
@@ -0,0 +1,964 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding=utf-8
|
3 |
+
|
4 |
+
# Basic Python imports
|
5 |
+
import os
|
6 |
+
import sys
|
7 |
+
import json
|
8 |
+
import argparse
|
9 |
+
import logging
|
10 |
+
from datetime import datetime
|
11 |
+
import time
|
12 |
+
import warnings
|
13 |
+
from importlib.util import find_spec
|
14 |
+
|
15 |
+
# Check hardware capabilities first
|
16 |
+
import torch
|
17 |
+
CUDA_AVAILABLE = torch.cuda.is_available()
|
18 |
+
NUM_GPUS = torch.cuda.device_count() if CUDA_AVAILABLE else 0
|
19 |
+
DEVICE_TYPE = "cuda" if CUDA_AVAILABLE else "cpu"
|
20 |
+
|
21 |
+
# Configure logging early
|
22 |
+
logging.basicConfig(
|
23 |
+
level=logging.INFO,
|
24 |
+
format="%(asctime)s - %(levelname)s - %(message)s",
|
25 |
+
handlers=[logging.StreamHandler(sys.stdout)]
|
26 |
+
)
|
27 |
+
logger = logging.getLogger(__name__)
|
28 |
+
|
29 |
+
# Set other loggers to WARNING to reduce noise and ensure our logs are visible
|
30 |
+
logging.getLogger("transformers").setLevel(logging.WARNING)
|
31 |
+
logging.getLogger("datasets").setLevel(logging.WARNING)
|
32 |
+
logging.getLogger("accelerate").setLevel(logging.WARNING)
|
33 |
+
logging.getLogger("torch").setLevel(logging.WARNING)
|
34 |
+
logging.getLogger("bitsandbytes").setLevel(logging.WARNING)
|
35 |
+
|
36 |
+
# Import Unsloth first, before other ML imports
|
37 |
+
try:
|
38 |
+
from unsloth import FastLanguageModel
|
39 |
+
from unsloth.chat_templates import get_chat_template
|
40 |
+
unsloth_available = True
|
41 |
+
logger.info("Unsloth successfully imported")
|
42 |
+
except ImportError:
|
43 |
+
unsloth_available = False
|
44 |
+
logger.warning("Unsloth not available. Please install with: pip install unsloth")
|
45 |
+
|
46 |
+
# Now import other ML libraries
|
47 |
+
try:
|
48 |
+
import transformers
|
49 |
+
from transformers import (
|
50 |
+
AutoModelForCausalLM,
|
51 |
+
AutoTokenizer,
|
52 |
+
TrainingArguments,
|
53 |
+
Trainer,
|
54 |
+
TrainerCallback,
|
55 |
+
set_seed,
|
56 |
+
BitsAndBytesConfig
|
57 |
+
)
|
58 |
+
logger.info(f"Transformers version: {transformers.__version__}")
|
59 |
+
except ImportError:
|
60 |
+
logger.error("Transformers not available. This is a critical dependency.")
|
61 |
+
|
62 |
+
# Check availability of libraries
|
63 |
+
peft_available = find_spec("peft") is not None
|
64 |
+
if peft_available:
|
65 |
+
import peft
|
66 |
+
logger.info(f"PEFT version: {peft.__version__}")
|
67 |
+
else:
|
68 |
+
logger.warning("PEFT not available. Parameter-efficient fine-tuning will not be used.")
|
69 |
+
|
70 |
+
# Import datasets library after the main ML libraries
|
71 |
+
try:
|
72 |
+
from datasets import load_dataset
|
73 |
+
logger.info("Datasets library successfully imported")
|
74 |
+
except ImportError:
|
75 |
+
logger.error("Datasets library not available. This is required for loading training data.")
|
76 |
+
|
77 |
+
# Define a clean logging function for HF Space compatibility
|
78 |
+
def log_info(message):
|
79 |
+
"""Log information in a format compatible with Hugging Face Spaces"""
|
80 |
+
# Just use the logger, but ensure consistent formatting
|
81 |
+
logger.info(message)
|
82 |
+
# Also ensure output is flushed immediately for streaming
|
83 |
+
sys.stdout.flush()
|
84 |
+
|
85 |
+
# Check for BitsAndBytes
|
86 |
+
try:
|
87 |
+
from transformers import BitsAndBytesConfig
|
88 |
+
bitsandbytes_available = True
|
89 |
+
except ImportError:
|
90 |
+
bitsandbytes_available = False
|
91 |
+
logger.warning("BitsAndBytes not available. 4-bit quantization will not be used.")
|
92 |
+
|
93 |
+
# Check for PEFT
|
94 |
+
try:
|
95 |
+
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
|
96 |
+
peft_available = True
|
97 |
+
except ImportError:
|
98 |
+
peft_available = False
|
99 |
+
logger.warning("PEFT not available. Parameter-efficient fine-tuning will not be used.")
|
100 |
+
|
101 |
+
def load_env_variables():
|
102 |
+
"""Load environment variables from system, .env file, or Hugging Face Space variables."""
|
103 |
+
# Check if we're running in a Hugging Face Space
|
104 |
+
if os.environ.get("SPACE_ID"):
|
105 |
+
logging.info("Running in Hugging Face Space")
|
106 |
+
|
107 |
+
# Log the presence of variables (without revealing values)
|
108 |
+
logging.info(f"HF_TOKEN available: {bool(os.environ.get('HF_TOKEN'))}")
|
109 |
+
logging.info(f"HF_USERNAME available: {bool(os.environ.get('HF_USERNAME'))}")
|
110 |
+
|
111 |
+
# If username is not set, try to extract from SPACE_ID
|
112 |
+
if not os.environ.get("HF_USERNAME") and "/" in os.environ.get("SPACE_ID", ""):
|
113 |
+
username = os.environ.get("SPACE_ID").split("/")[0]
|
114 |
+
os.environ["HF_USERNAME"] = username
|
115 |
+
logging.info(f"Set HF_USERNAME from SPACE_ID: {username}")
|
116 |
+
else:
|
117 |
+
# Try to load from .env file if not in a Space
|
118 |
+
try:
|
119 |
+
from dotenv import load_dotenv
|
120 |
+
# Updated path to .env file in the new directory structure
|
121 |
+
env_path = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "shared", ".env")
|
122 |
+
if os.path.exists(env_path):
|
123 |
+
load_dotenv(env_path)
|
124 |
+
logging.info(f"Loaded environment variables from {env_path}")
|
125 |
+
logging.info(f"HF_TOKEN loaded from .env file: {bool(os.environ.get('HF_TOKEN'))}")
|
126 |
+
logging.info(f"HF_USERNAME loaded from .env file: {bool(os.environ.get('HF_USERNAME'))}")
|
127 |
+
logging.info(f"HF_SPACE_NAME loaded from .env file: {bool(os.environ.get('HF_SPACE_NAME'))}")
|
128 |
+
else:
|
129 |
+
logging.warning(f"No .env file found at {env_path}")
|
130 |
+
except ImportError:
|
131 |
+
logging.warning("python-dotenv not installed, not loading from .env file")
|
132 |
+
|
133 |
+
if not os.environ.get("HF_USERNAME"):
|
134 |
+
logger.warning("HF_USERNAME is not set. Using default username.")
|
135 |
+
|
136 |
+
if not os.environ.get("HF_SPACE_NAME"):
|
137 |
+
logger.warning("HF_SPACE_NAME is not set. Using default space name.")
|
138 |
+
|
139 |
+
# Set HF_TOKEN for huggingface_hub
|
140 |
+
if os.environ.get("HF_TOKEN"):
|
141 |
+
os.environ["HUGGING_FACE_HUB_TOKEN"] = os.environ.get("HF_TOKEN")
|
142 |
+
|
143 |
+
def load_configs(base_path):
|
144 |
+
"""Load configuration from transformers_config.json file."""
|
145 |
+
# Using a single consolidated config file
|
146 |
+
config_file = base_path
|
147 |
+
|
148 |
+
try:
|
149 |
+
with open(config_file, "r") as f:
|
150 |
+
config = json.load(f)
|
151 |
+
logger.info(f"Loaded configuration from {config_file}")
|
152 |
+
return config
|
153 |
+
except Exception as e:
|
154 |
+
logger.error(f"Error loading {config_file}: {e}")
|
155 |
+
raise
|
156 |
+
|
157 |
+
def parse_args():
|
158 |
+
parser = argparse.ArgumentParser(description="Fine-tune a language model on a text dataset")
|
159 |
+
parser.add_argument("--config", type=str, default="transformers_config.json", help="Path to configuration file")
|
160 |
+
return parser.parse_args()
|
161 |
+
|
162 |
+
def load_model_and_tokenizer(config):
|
163 |
+
"""Load model and tokenizer with proper error handling and optimizations."""
|
164 |
+
try:
|
165 |
+
if not unsloth_available:
|
166 |
+
logger.error("Unsloth is required for training with pre-quantized model")
|
167 |
+
logger.error("Please ensure unsloth is in requirements.txt")
|
168 |
+
raise ImportError("Unsloth is required for this training setup")
|
169 |
+
|
170 |
+
# Get model name correctly from config
|
171 |
+
model_name = config.get("model_name") or config.get("model", {}).get("name")
|
172 |
+
logger.info(f"Loading model: {model_name}")
|
173 |
+
|
174 |
+
if not model_name:
|
175 |
+
raise ValueError("Model name not found in configuration. Please check your transformers_config.json file.")
|
176 |
+
|
177 |
+
logger.info("Using Unsloth optimizations with pre-quantized model")
|
178 |
+
|
179 |
+
# First detect if we have a GPU
|
180 |
+
if torch.cuda.is_available():
|
181 |
+
gpu_count = torch.cuda.device_count()
|
182 |
+
logger.info(f"Found {gpu_count} CUDA devices")
|
183 |
+
else:
|
184 |
+
logger.warning("No CUDA devices detected. Training will be slow on CPU!")
|
185 |
+
gpu_count = 0
|
186 |
+
|
187 |
+
# Set default dtype for better numerics
|
188 |
+
if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8:
|
189 |
+
# Use bfloat16 for Ampere or newer
|
190 |
+
dtype = torch.bfloat16
|
191 |
+
logger.info("Using bfloat16 precision (Ampere+ GPU)")
|
192 |
+
elif torch.cuda.is_available():
|
193 |
+
# Use float16 for older GPUs
|
194 |
+
dtype = torch.float16
|
195 |
+
logger.info("Using float16 precision (pre-Ampere GPU)")
|
196 |
+
else:
|
197 |
+
# CPU, use default dtype
|
198 |
+
dtype = None
|
199 |
+
logger.info("Using default precision (CPU)")
|
200 |
+
|
201 |
+
# Check for flash attention as the last dependency check
|
202 |
+
use_flash_attention = config.get("use_flash_attention", True)
|
203 |
+
if use_flash_attention and not find_spec("flash_attn"):
|
204 |
+
logger.warning("flash-attn not found. Will continue without flash attention.")
|
205 |
+
logger.warning("To use flash attention, install with: pip install flash-attn --no-build-isolation")
|
206 |
+
use_flash_attention = False
|
207 |
+
|
208 |
+
# Set device map based on config or default to "auto"
|
209 |
+
device_map = config.get("hardware", {}).get("hardware_setup", {}).get("device_map", "auto")
|
210 |
+
|
211 |
+
# Calculate max memory settings if multiple GPUs are available
|
212 |
+
max_memory = None
|
213 |
+
if gpu_count > 1:
|
214 |
+
memory_per_gpu = config.get("hardware", {}).get("specs", {}).get("vram_per_gpu", 24)
|
215 |
+
max_memory = {i: f"{int(memory_per_gpu * 0.85)}GiB" for i in range(gpu_count)}
|
216 |
+
max_memory["cpu"] = "64GiB" # Allow CPU offloading if needed
|
217 |
+
|
218 |
+
# Load model with proper error handling for out-of-memory
|
219 |
+
try:
|
220 |
+
# Improved memory settings for multi-GPU setup
|
221 |
+
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
|
222 |
+
|
223 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
224 |
+
model_name=model_name,
|
225 |
+
max_seq_length=config.get("max_seq_length", 2048) or config.get("tokenizer", {}).get("max_seq_length", 2048),
|
226 |
+
dtype=dtype,
|
227 |
+
device_map=device_map,
|
228 |
+
max_memory=max_memory,
|
229 |
+
# Don't explicitly use flash attention config here, let Unsloth handle it
|
230 |
+
)
|
231 |
+
except RuntimeError as e:
|
232 |
+
if "CUDA out of memory" in str(e):
|
233 |
+
logger.error("Out of GPU memory. Consider using a smaller batch size or gradient accumulation steps.")
|
234 |
+
raise
|
235 |
+
else:
|
236 |
+
# Try again with CPU placement to see if it's a memory issue
|
237 |
+
logger.warning(f"Error loading model on default device: {str(e)}")
|
238 |
+
logger.warning("Attempting to load with device_map='cpu' and no specific dtype")
|
239 |
+
model, tokenizer = FastLanguageModel.from_pretrained(
|
240 |
+
model_name=model_name,
|
241 |
+
max_seq_length=config.get("max_seq_length", 2048) or config.get("tokenizer", {}).get("max_seq_length", 2048),
|
242 |
+
dtype=None,
|
243 |
+
device_map={"": "cpu"},
|
244 |
+
)
|
245 |
+
logger.warning("Model loaded on CPU. Training will be very slow.")
|
246 |
+
|
247 |
+
# Ensure model and optimizer init is on the same device
|
248 |
+
logger.info(f"Model device map: {model.hf_device_map if hasattr(model, 'hf_device_map') else 'Not available'}")
|
249 |
+
|
250 |
+
# Apply Unsloth's training optimizations with config parameters
|
251 |
+
unsloth_config = config.get("unsloth", {})
|
252 |
+
model = FastLanguageModel.get_peft_model(
|
253 |
+
model,
|
254 |
+
r=unsloth_config.get("r", 32),
|
255 |
+
target_modules=unsloth_config.get("target_modules",
|
256 |
+
["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]),
|
257 |
+
lora_alpha=unsloth_config.get("alpha", 16),
|
258 |
+
lora_dropout=unsloth_config.get("dropout", 0.05),
|
259 |
+
bias="none",
|
260 |
+
use_gradient_checkpointing=config.get("gradient_checkpointing", True) or config.get("training", {}).get("gradient_checkpointing", True),
|
261 |
+
random_state=config.get("seed", 42),
|
262 |
+
)
|
263 |
+
logger.info("Unsloth optimizations applied successfully")
|
264 |
+
|
265 |
+
# Set up tokenizer settings
|
266 |
+
chat_template = config.get("chat_template") or config.get("tokenizer", {}).get("chat_template")
|
267 |
+
if chat_template:
|
268 |
+
try:
|
269 |
+
template = get_chat_template("phi")
|
270 |
+
tokenizer.chat_template = template
|
271 |
+
logger.info("Set phi chat template")
|
272 |
+
except Exception as e:
|
273 |
+
logger.warning(f"Failed to set chat template: {str(e)}")
|
274 |
+
|
275 |
+
# Ensure proper token settings
|
276 |
+
if tokenizer.pad_token_id is None:
|
277 |
+
tokenizer.pad_token_id = tokenizer.eos_token_id
|
278 |
+
logger.info(f"Set pad_token_id to eos_token_id: {tokenizer.pad_token_id}")
|
279 |
+
|
280 |
+
return model, tokenizer
|
281 |
+
|
282 |
+
except Exception as e:
|
283 |
+
logger.error(f"Error in model/tokenizer loading: {str(e)}")
|
284 |
+
logger.error("If missing dependencies, check the requirements.txt file")
|
285 |
+
raise
|
286 |
+
|
287 |
+
def load_dataset_with_mapping(dataset_config):
|
288 |
+
"""Load dataset and apply appropriate column mappings."""
|
289 |
+
try:
|
290 |
+
# Load dataset
|
291 |
+
dataset_name = dataset_config.get("dataset", {}).get("name", "")
|
292 |
+
dataset_split = dataset_config.get("dataset", {}).get("split", "train")
|
293 |
+
|
294 |
+
if not dataset_name:
|
295 |
+
raise ValueError("Dataset name not provided in configuration")
|
296 |
+
|
297 |
+
logger.info(f"Loading pre-processed dataset {dataset_name}, split {dataset_split}")
|
298 |
+
dataset = load_dataset(dataset_name, split=dataset_split)
|
299 |
+
|
300 |
+
# Apply minimal processing since the dataset has already been properly structured
|
301 |
+
# Just perform validation to ensure required fields exist
|
302 |
+
|
303 |
+
# Check for required fields
|
304 |
+
required_fields = ["prompt_number", "article_id", "conversations"]
|
305 |
+
missing_fields = [field for field in required_fields if field not in dataset.column_names]
|
306 |
+
|
307 |
+
if missing_fields:
|
308 |
+
logger.warning(f"Dataset is missing required fields: {missing_fields}")
|
309 |
+
logger.warning("This may cause issues with sequence integrity and metadata management")
|
310 |
+
else:
|
311 |
+
logger.info(f"Dataset has all required fields: {required_fields}")
|
312 |
+
|
313 |
+
# Verify that column order matches our expectation
|
314 |
+
expected_order = ["prompt_number", "article_id", "conversations"]
|
315 |
+
actual_order = dataset.column_names
|
316 |
+
|
317 |
+
if actual_order == expected_order:
|
318 |
+
logger.info("Dataset column order matches expected order (prompt_number, article_id, conversations)")
|
319 |
+
else:
|
320 |
+
logger.warning(f"Dataset column order ({', '.join(actual_order)}) differs from expected order ({', '.join(expected_order)})")
|
321 |
+
logger.warning("This should not affect processing but is noted for debugging purposes")
|
322 |
+
|
323 |
+
# Log a few samples for verification
|
324 |
+
if len(dataset) > 0:
|
325 |
+
sample_indices = range(min(5, len(dataset)))
|
326 |
+
sample_records = []
|
327 |
+
|
328 |
+
for i in sample_indices:
|
329 |
+
record = {}
|
330 |
+
record["prompt_number"] = dataset[i].get("prompt_number", "N/A")
|
331 |
+
record["article_id"] = dataset[i].get("article_id", "N/A")
|
332 |
+
if "conversations" in dataset[i]:
|
333 |
+
record["conversations_length"] = len(dataset[i]["conversations"])
|
334 |
+
sample_records.append(record)
|
335 |
+
|
336 |
+
logger.info(f"Sample records: {sample_records}")
|
337 |
+
|
338 |
+
# Verify sequential integrity
|
339 |
+
if "prompt_number" in dataset.column_names and len(dataset) > 1:
|
340 |
+
first_prompt_numbers = [dataset[i]["prompt_number"] for i in range(min(10, len(dataset)))]
|
341 |
+
is_sequential = all(first_prompt_numbers[i] == i + 1 for i in range(len(first_prompt_numbers)))
|
342 |
+
|
343 |
+
if is_sequential:
|
344 |
+
logger.info("Dataset prompt numbers are sequential (1-indexed) - sequence integrity preserved")
|
345 |
+
else:
|
346 |
+
logger.warning("Dataset prompt numbers are not sequential - sequence integrity may be compromised")
|
347 |
+
logger.info(f"First few prompt numbers: {first_prompt_numbers}")
|
348 |
+
|
349 |
+
logger.info(f"Dataset loaded successfully with {len(dataset)} examples")
|
350 |
+
logger.info(f"Dataset columns: {dataset.column_names}")
|
351 |
+
|
352 |
+
# Data loading configuration - ensure shuffle is disabled
|
353 |
+
data_loading_config = dataset_config.get("data_loading", {})
|
354 |
+
if data_loading_config.get("shuffle", False):
|
355 |
+
logger.error("CRITICAL: shuffle is enabled in the dataset config!")
|
356 |
+
logger.error("This will RANDOMIZE your dataset and break sequential order.")
|
357 |
+
logger.error("Setting shuffle to False to preserve order")
|
358 |
+
data_loading_config["shuffle"] = False
|
359 |
+
|
360 |
+
return dataset
|
361 |
+
|
362 |
+
except Exception as e:
|
363 |
+
logger.error(f"Error loading dataset: {str(e)}")
|
364 |
+
raise
|
365 |
+
|
366 |
+
def format_phi_chat(messages, dataset_config):
|
367 |
+
"""Format messages according to phi-4's chat template and dataset config."""
|
368 |
+
formatted_chat = ""
|
369 |
+
|
370 |
+
# Get role templates from config
|
371 |
+
roles = dataset_config.get("data_formatting", {}).get("roles", {
|
372 |
+
"system": "System: {content}\n\n",
|
373 |
+
"human": "Human: {content}\n\n",
|
374 |
+
"user": "Human: {content}\n\n",
|
375 |
+
"assistant": "Assistant: {content}\n\n"
|
376 |
+
})
|
377 |
+
|
378 |
+
# Handle research introduction metadata first
|
379 |
+
metadata = next((msg for msg in messages if isinstance(msg, dict) and
|
380 |
+
"[RESEARCH INTRODUCTION]" in msg.get("content", "")), None)
|
381 |
+
if metadata:
|
382 |
+
system_template = roles.get("system", "System: {content}\n\n")
|
383 |
+
formatted_chat = system_template.format(content=metadata['content'])
|
384 |
+
messages = [msg for msg in messages if msg != metadata]
|
385 |
+
|
386 |
+
# Process remaining messages
|
387 |
+
for message in messages:
|
388 |
+
if not isinstance(message, dict) or "content" not in message:
|
389 |
+
logger.warning(f"Skipping invalid message format: {message}")
|
390 |
+
continue
|
391 |
+
|
392 |
+
role = message.get("role", "").lower()
|
393 |
+
content = message.get("content", "")
|
394 |
+
|
395 |
+
# Format based on role
|
396 |
+
if role == "human" or role == "user":
|
397 |
+
template = roles.get("user", roles.get("human", "Human: {content}\n\n"))
|
398 |
+
formatted_chat += template.format(content=content)
|
399 |
+
elif role == "assistant" or role == "bot":
|
400 |
+
template = roles.get("assistant", "Assistant: {content}\n\n")
|
401 |
+
formatted_chat += template.format(content=content)
|
402 |
+
elif role == "system":
|
403 |
+
# For system messages, prepend them
|
404 |
+
template = roles.get("system", "System: {content}\n\n")
|
405 |
+
formatted_chat = template.format(content=content) + formatted_chat
|
406 |
+
else:
|
407 |
+
# Default to system for unknown roles
|
408 |
+
logger.warning(f"Unknown role '{role}' - treating as system message")
|
409 |
+
template = roles.get("system", "System: {content}\n\n")
|
410 |
+
formatted_chat += template.format(content=content)
|
411 |
+
|
412 |
+
return formatted_chat.strip()
|
413 |
+
|
414 |
+
class SimpleDataCollator:
|
415 |
+
def __init__(self, tokenizer, dataset_config):
|
416 |
+
self.tokenizer = tokenizer
|
417 |
+
self.dataset_config = dataset_config
|
418 |
+
self.stats = {"processed": 0, "skipped": 0, "total_tokens": 0}
|
419 |
+
self.pad_token_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else 0
|
420 |
+
self.max_seq_length = dataset_config.get("dataset", {}).get("processing", {}).get("max_seq_length", 2048)
|
421 |
+
logger.info(f"SimpleDataCollator initialized - using pre-audited dataset with max_seq_length={self.max_seq_length}")
|
422 |
+
logger.info("Using exact dataset structure without reformatting")
|
423 |
+
|
424 |
+
# Check if we're on GPU
|
425 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
426 |
+
logger.info(f"SimpleDataCollator using device: {self.device}")
|
427 |
+
|
428 |
+
def __call__(self, features):
|
429 |
+
"""Process examples preserving exact JSONL structure"""
|
430 |
+
batch = {"input_ids": [], "attention_mask": [], "labels": []}
|
431 |
+
|
432 |
+
for example in features:
|
433 |
+
try:
|
434 |
+
# Get ID
|
435 |
+
paper_id = example.get("id", "")
|
436 |
+
|
437 |
+
# Get conversations - these should already contain role and content
|
438 |
+
conversations = example.get("conversations", [])
|
439 |
+
if not conversations:
|
440 |
+
self.stats["skipped"] += 1
|
441 |
+
continue
|
442 |
+
|
443 |
+
# Directly use the conversations array as input to the model's chat template
|
444 |
+
# This preserves the exact structure with roles and content as they are
|
445 |
+
try:
|
446 |
+
# Let tokenizer handle the content with the model's chat template
|
447 |
+
inputs = self.tokenizer.apply_chat_template(
|
448 |
+
conversations,
|
449 |
+
return_tensors=None,
|
450 |
+
add_generation_prompt=False
|
451 |
+
)
|
452 |
+
except Exception as chat_error:
|
453 |
+
# Fallback if apply_chat_template fails
|
454 |
+
logger.warning(f"Chat template application failed for example {paper_id}: {str(chat_error)[:100]}")
|
455 |
+
|
456 |
+
# Create a basic representation of the conversation
|
457 |
+
conversation_text = ""
|
458 |
+
for msg in conversations:
|
459 |
+
if isinstance(msg, dict) and 'content' in msg:
|
460 |
+
conversation_text += msg.get('content', '') + "\n\n"
|
461 |
+
|
462 |
+
# Basic tokenization
|
463 |
+
inputs = self.tokenizer(
|
464 |
+
conversation_text,
|
465 |
+
add_special_tokens=True,
|
466 |
+
return_tensors=None
|
467 |
+
)
|
468 |
+
|
469 |
+
# Apply length cap if needed (shouldn't be necessary for pre-audited data)
|
470 |
+
if self.max_seq_length > 0 and len(inputs) > self.max_seq_length:
|
471 |
+
logger.warning(f"Example {paper_id} exceeds max_seq_length ({len(inputs)} > {self.max_seq_length})")
|
472 |
+
inputs = inputs[:self.max_seq_length]
|
473 |
+
|
474 |
+
# Create attention mask (1 for all tokens)
|
475 |
+
attention_mask = [1] * len(inputs)
|
476 |
+
|
477 |
+
if len(inputs) > 0:
|
478 |
+
# For causal language modeling, labels are the same as inputs
|
479 |
+
labels = inputs.copy()
|
480 |
+
|
481 |
+
batch["input_ids"].append(inputs)
|
482 |
+
batch["attention_mask"].append(attention_mask)
|
483 |
+
batch["labels"].append(labels)
|
484 |
+
|
485 |
+
self.stats["processed"] += 1
|
486 |
+
self.stats["total_tokens"] += len(inputs)
|
487 |
+
|
488 |
+
# Debug logging for first few examples
|
489 |
+
log_samples = self.dataset_config.get("validation", {}).get("log_samples", 3)
|
490 |
+
if self.stats["processed"] <= log_samples:
|
491 |
+
logger.info(f"Example {self.stats['processed']}:")
|
492 |
+
logger.info(f"Paper ID: {paper_id}")
|
493 |
+
logger.info(f"Token count: {len(inputs)}")
|
494 |
+
logger.info(f"Conversation entries: {len(conversations)}")
|
495 |
+
else:
|
496 |
+
self.stats["skipped"] += 1
|
497 |
+
except Exception as e:
|
498 |
+
logger.warning(f"Error processing example: {str(e)[:100]}...")
|
499 |
+
logger.warning(f"Problematic example ID: {example.get('id', 'unknown')}")
|
500 |
+
self.stats["skipped"] += 1
|
501 |
+
continue
|
502 |
+
|
503 |
+
if not batch["input_ids"]:
|
504 |
+
logger.warning("Empty batch, returning dummy tensors")
|
505 |
+
return {
|
506 |
+
"input_ids": torch.zeros((1, 1), dtype=torch.long),
|
507 |
+
"attention_mask": torch.zeros((1, 1), dtype=torch.long),
|
508 |
+
"labels": torch.zeros((1, 1), dtype=torch.long)
|
509 |
+
}
|
510 |
+
|
511 |
+
# Pad the batch
|
512 |
+
max_length = max(len(ids) for ids in batch["input_ids"])
|
513 |
+
|
514 |
+
for i in range(len(batch["input_ids"])):
|
515 |
+
padding_length = max_length - len(batch["input_ids"][i])
|
516 |
+
if padding_length > 0:
|
517 |
+
batch["input_ids"][i].extend([self.pad_token_id] * padding_length)
|
518 |
+
batch["attention_mask"][i].extend([0] * padding_length)
|
519 |
+
batch["labels"][i].extend([-100] * padding_length)
|
520 |
+
|
521 |
+
# Convert to tensors
|
522 |
+
batch = {k: torch.tensor(v, dtype=torch.long) for k, v in batch.items()}
|
523 |
+
|
524 |
+
# Log stats periodically
|
525 |
+
log_interval = self.dataset_config.get("validation", {}).get("log_interval", 100)
|
526 |
+
if self.stats["processed"] % log_interval == 0 and self.stats["processed"] > 0:
|
527 |
+
logger.info(f"Data collator stats: processed={self.stats['processed']}, "
|
528 |
+
f"skipped={self.stats['skipped']}, "
|
529 |
+
f"avg_tokens={self.stats['total_tokens']/self.stats['processed']:.1f}")
|
530 |
+
|
531 |
+
return batch
|
532 |
+
|
533 |
+
class LoggingCallback(TrainerCallback):
|
534 |
+
def __init__(self):
|
535 |
+
super().__init__()
|
536 |
+
self.training_started = time.time()
|
537 |
+
self.last_log_time = time.time()
|
538 |
+
self.last_step = 0
|
539 |
+
self.verify_sequence = None
|
540 |
+
self.sequence_samples = None
|
541 |
+
self.sample_indices = None
|
542 |
+
|
543 |
+
def on_train_begin(self, args, state, control, **kwargs):
|
544 |
+
log_info(f"=== Training started at {time.strftime('%Y-%m-%d %H:%M:%S')} ===")
|
545 |
+
log_info(f"Model parameters: {sum(p.numel() for p in model.parameters())/1e6:.2f}M")
|
546 |
+
|
547 |
+
# Disable sequence verification
|
548 |
+
self.verify_sequence = False
|
549 |
+
|
550 |
+
log_info("=== Training is starting ===")
|
551 |
+
|
552 |
+
# Log important training parameters for visibility
|
553 |
+
total_batch_size = args.per_device_train_batch_size * args.gradient_accumulation_steps * NUM_GPUS
|
554 |
+
total_steps = int(len(dataset) / (args.per_device_train_batch_size * NUM_GPUS * args.gradient_accumulation_steps) * args.num_train_epochs)
|
555 |
+
log_info(f"Training plan: {len(dataset)} examples over {args.num_train_epochs} epochs ≈ {total_steps} steps")
|
556 |
+
log_info(f"Batch size: {args.per_device_train_batch_size} × {args.gradient_accumulation_steps} steps × {NUM_GPUS} GPUs = {total_batch_size} total")
|
557 |
+
log_info(f"Learning rate: {args.learning_rate}")
|
558 |
+
log_info(f"Epochs: {args.num_train_epochs}")
|
559 |
+
|
560 |
+
# Log memory information in compact format
|
561 |
+
if CUDA_AVAILABLE:
|
562 |
+
memory_info = []
|
563 |
+
for i in range(NUM_GPUS):
|
564 |
+
allocated = torch.cuda.memory_allocated(i) / 1024**2
|
565 |
+
max_mem = torch.cuda.max_memory_allocated(i) / 1024**2
|
566 |
+
memory_info.append(f"GPU {i}: {allocated:.1f}MB (max: {max_mem:.1f}MB)")
|
567 |
+
|
568 |
+
log_info(f"Initial memory usage - {', '.join(memory_info)}")
|
569 |
+
|
570 |
+
def on_step_end(self, args, state, control, **kwargs):
|
571 |
+
# Log every 50 steps or every 5 minutes, whichever comes first
|
572 |
+
current_time = time.time()
|
573 |
+
|
574 |
+
# Sequence verification removed
|
575 |
+
|
576 |
+
# Log progress at regular intervals
|
577 |
+
if (state.global_step % 50 == 0) or (current_time - self.last_log_time > 300):
|
578 |
+
if state.log_history:
|
579 |
+
loss = state.log_history[-1].get('loss', 'N/A')
|
580 |
+
# Use simple formatting for better Space log compatibility
|
581 |
+
log_info(f"Step {state.global_step}: Loss {loss}")
|
582 |
+
else:
|
583 |
+
log_info(f"Step {state.global_step}: No loss data available")
|
584 |
+
self.last_log_time = current_time
|
585 |
+
|
586 |
+
def on_train_end(self, args, state, control, **kwargs):
|
587 |
+
training_time = time.strftime("%H:%M:%S", time.gmtime(time.time() - self.training_started))
|
588 |
+
log_info(f"=== Training completed in {training_time} ===")
|
589 |
+
|
590 |
+
# Log final memory usage
|
591 |
+
if CUDA_AVAILABLE:
|
592 |
+
for i in range(NUM_GPUS):
|
593 |
+
max_mem = torch.cuda.max_memory_allocated(i) / 1024**3 # GB
|
594 |
+
log_info(f"GPU {i} max memory: {max_mem:.2f} GB")
|
595 |
+
|
596 |
+
# Clear GPU memory
|
597 |
+
torch.cuda.empty_cache()
|
598 |
+
log_info("GPU memory cleared")
|
599 |
+
|
600 |
+
log_info(f"Total steps: {state.global_step}")
|
601 |
+
log_info(f"Final loss: {state.log_history[-1].get('loss', 'N/A') if state.log_history else 'N/A'}")
|
602 |
+
|
603 |
+
def check_dependencies():
|
604 |
+
"""Check if all required dependencies are installed and in the correct order."""
|
605 |
+
missing_packages = []
|
606 |
+
order_issues = []
|
607 |
+
|
608 |
+
# Check critical packages in the required order
|
609 |
+
|
610 |
+
# 1. First check for unsloth as it should be imported before transformers
|
611 |
+
if not unsloth_available:
|
612 |
+
missing_packages.append("unsloth>=2024.3")
|
613 |
+
|
614 |
+
# 2. Check transformers (imported at module level)
|
615 |
+
try:
|
616 |
+
import transformers
|
617 |
+
logger.info(f"Using transformers version {transformers.__version__}")
|
618 |
+
except ImportError:
|
619 |
+
missing_packages.append("transformers>=4.38.0")
|
620 |
+
|
621 |
+
# 3. Check for peft
|
622 |
+
if not peft_available:
|
623 |
+
missing_packages.append("peft>=0.9.0")
|
624 |
+
|
625 |
+
# 4. Check for accelerate
|
626 |
+
try:
|
627 |
+
import accelerate
|
628 |
+
logger.info(f"Using accelerate version {accelerate.__version__}")
|
629 |
+
except ImportError:
|
630 |
+
missing_packages.append("accelerate>=0.27.0")
|
631 |
+
|
632 |
+
# Check for order-specific issues
|
633 |
+
try:
|
634 |
+
import sys
|
635 |
+
modules = sys.modules.keys()
|
636 |
+
|
637 |
+
# Unsloth should be imported before transformers for optimal performance
|
638 |
+
if 'transformers' in modules and 'unsloth' in modules:
|
639 |
+
if modules.index('transformers') < modules.index('unsloth'):
|
640 |
+
order_issues.append("For optimal performance, unsloth should be imported before transformers")
|
641 |
+
except Exception:
|
642 |
+
# If we can't check order, just skip this check
|
643 |
+
pass
|
644 |
+
|
645 |
+
# If critical packages are missing, exit with instructions
|
646 |
+
if missing_packages:
|
647 |
+
logger.error("Critical dependencies missing:")
|
648 |
+
for pkg in missing_packages:
|
649 |
+
logger.error(f" - {pkg}")
|
650 |
+
logger.error("Please install the missing dependencies with:")
|
651 |
+
logger.error(f" pip install {' '.join(missing_packages)}")
|
652 |
+
return False
|
653 |
+
|
654 |
+
# Report order issues as warnings
|
655 |
+
for issue in order_issues:
|
656 |
+
logger.warning(issue)
|
657 |
+
|
658 |
+
# Optional packages - moved to the end
|
659 |
+
if find_spec("flash_attn"):
|
660 |
+
logger.info("flash-attn found. Flash attention will be used for faster training.")
|
661 |
+
else:
|
662 |
+
logger.warning("flash-attn not found. Training will work but may be slower.")
|
663 |
+
logger.warning("To use flash attention, install with: pip install flash-attn --no-build-isolation")
|
664 |
+
|
665 |
+
# Additional optional packages that improve performance
|
666 |
+
if find_spec("bitsandbytes"):
|
667 |
+
logger.info("bitsandbytes found. Quantization will be available.")
|
668 |
+
else:
|
669 |
+
logger.warning("bitsandbytes not found. Quantization may not be available.")
|
670 |
+
logger.warning("To use quantization, install with: pip install bitsandbytes")
|
671 |
+
|
672 |
+
return True
|
673 |
+
|
674 |
+
def main():
|
675 |
+
# Set up logging
|
676 |
+
logger.info("Starting training process")
|
677 |
+
|
678 |
+
# Check dependencies first, before any other operations
|
679 |
+
if not check_dependencies():
|
680 |
+
logger.error("Aborting due to missing critical dependencies")
|
681 |
+
return 1
|
682 |
+
|
683 |
+
# Parse arguments
|
684 |
+
args = parse_args()
|
685 |
+
|
686 |
+
# Load environment variables
|
687 |
+
load_env_variables()
|
688 |
+
|
689 |
+
# Load configuration
|
690 |
+
try:
|
691 |
+
transformers_config = load_configs(args.config)
|
692 |
+
hardware_config = transformers_config.get("hardware", {})
|
693 |
+
dataset_config = transformers_config.get("dataset", {})
|
694 |
+
logger.info("Configuration loaded successfully")
|
695 |
+
except Exception as e:
|
696 |
+
logger.error(f"Error loading configuration: {e}")
|
697 |
+
return 1
|
698 |
+
|
699 |
+
# Check if we're in distributed mode
|
700 |
+
is_distributed = "WORLD_SIZE" in os.environ and int(os.environ.get("WORLD_SIZE", "1")) > 1
|
701 |
+
if is_distributed:
|
702 |
+
local_rank = int(os.environ.get("LOCAL_RANK", "0"))
|
703 |
+
log_info(f"Running in distributed mode with {os.environ.get('WORLD_SIZE')} processes, local_rank: {local_rank}")
|
704 |
+
else:
|
705 |
+
log_info("Running in non-distributed mode (single process)")
|
706 |
+
|
707 |
+
# Set random seed for reproducibility
|
708 |
+
seed = transformers_config.get("seed", 42)
|
709 |
+
set_seed(seed)
|
710 |
+
logger.info(f"Set random seed to {seed}")
|
711 |
+
|
712 |
+
# Load model and tokenizer using the consolidated config
|
713 |
+
model, tokenizer = load_model_and_tokenizer(transformers_config)
|
714 |
+
|
715 |
+
# Empty CUDA cache to ensure clean state
|
716 |
+
if CUDA_AVAILABLE:
|
717 |
+
torch.cuda.empty_cache()
|
718 |
+
log_info("Cleared CUDA cache")
|
719 |
+
|
720 |
+
# Setup environment variable for CUDA memory allocation
|
721 |
+
if CUDA_AVAILABLE:
|
722 |
+
system_settings = hardware_config.get("system_settings", {})
|
723 |
+
cuda_memory_fraction = system_settings.get("cuda_memory_fraction", 0.85)
|
724 |
+
|
725 |
+
if cuda_memory_fraction < 1.0:
|
726 |
+
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = f"max_split_size_mb:128,expandable_segments:True"
|
727 |
+
log_info(f"Set CUDA memory allocation limit to expandable with max_split_size_mb:128")
|
728 |
+
|
729 |
+
try:
|
730 |
+
log_info("Loading dataset...")
|
731 |
+
dataset = load_dataset_with_mapping(dataset_config)
|
732 |
+
log_info(f"Dataset loaded with {len(dataset)} examples")
|
733 |
+
|
734 |
+
# Minimal validation before proceeding
|
735 |
+
if dataset is None or len(dataset) == 0:
|
736 |
+
logger.error("Dataset is empty or None! Cannot proceed with training.")
|
737 |
+
return 1
|
738 |
+
|
739 |
+
# Create data collator
|
740 |
+
data_collator = SimpleDataCollator(tokenizer, dataset_config)
|
741 |
+
|
742 |
+
# Verify precision settings - ensure only one of bf16/fp16 is set, with bf16 taking precedence
|
743 |
+
# First check hardware config, then transformers config
|
744 |
+
use_bf16 = False
|
745 |
+
use_fp16 = False
|
746 |
+
|
747 |
+
# Check hardware config first
|
748 |
+
hardware_precision = hardware_config.get("training_optimizations", {}).get("mixed_precision", "")
|
749 |
+
if hardware_precision.lower() == "bf16":
|
750 |
+
use_bf16 = True
|
751 |
+
log_info("Using BF16 precision from hardware config")
|
752 |
+
elif hardware_precision.lower() == "fp16":
|
753 |
+
use_fp16 = True
|
754 |
+
log_info("Using FP16 precision from hardware config")
|
755 |
+
else:
|
756 |
+
# Fall back to transformers config
|
757 |
+
use_bf16 = transformers_config.get("bf16", False) or transformers_config.get("torch_dtype", "") == "bfloat16"
|
758 |
+
use_fp16 = transformers_config.get("fp16", False) and not use_bf16 # Only use fp16 if bf16 is not set
|
759 |
+
log_info(f"Using precision: {'bf16' if use_bf16 else 'fp16' if use_fp16 else 'full precision'}")
|
760 |
+
|
761 |
+
# Get per device batch size - from transformers config, but possibly overridden by hardware config
|
762 |
+
per_device_batch_size = transformers_config.get("training", {}).get("per_device_train_batch_size", 16)
|
763 |
+
gradient_accumulation_steps = transformers_config.get("training", {}).get("gradient_accumulation_steps", 3)
|
764 |
+
|
765 |
+
# Get multi-GPU strategy from hardware config (default to data_parallel)
|
766 |
+
multi_gpu_strategy = hardware_config.get("training_optimizations", {}).get("multi_gpu_strategy", "data_parallel")
|
767 |
+
logger.info(f"Multi-GPU strategy: {multi_gpu_strategy}")
|
768 |
+
|
769 |
+
# For multi-GPU setup, adjust for better balance
|
770 |
+
if CUDA_AVAILABLE and NUM_GPUS > 1:
|
771 |
+
log_info(f"Multi-GPU setup: Adjusting for {NUM_GPUS} GPUs")
|
772 |
+
|
773 |
+
# Set up FSDP for multi-GPU training if specified and in distributed mode
|
774 |
+
fsdp_config = None
|
775 |
+
if multi_gpu_strategy == "fsdp" and is_distributed and NUM_GPUS > 1:
|
776 |
+
try:
|
777 |
+
from torch.distributed.fsdp import (
|
778 |
+
FullyShardedDataParallel as FSDP,
|
779 |
+
MixedPrecision,
|
780 |
+
BackwardPrefetch,
|
781 |
+
ShardingStrategy,
|
782 |
+
CPUOffload,
|
783 |
+
)
|
784 |
+
from torch.distributed.fsdp.wrap import (
|
785 |
+
transformer_auto_wrap_policy,
|
786 |
+
enable_wrap,
|
787 |
+
wrap,
|
788 |
+
)
|
789 |
+
|
790 |
+
log_info("Using FSDP for distributed training")
|
791 |
+
|
792 |
+
# Configure FSDP
|
793 |
+
fsdp_config = {
|
794 |
+
"fsdp_transformer_layer_cls_to_wrap": ["LlamaDecoderLayer"],
|
795 |
+
"fsdp_offload_params": False,
|
796 |
+
"fsdp_backward_prefetch": "BACKWARD_PRE",
|
797 |
+
"fsdp_min_num_params": 1e6,
|
798 |
+
"fsdp_sharding_strategy": 1, # FULL_SHARD
|
799 |
+
}
|
800 |
+
|
801 |
+
if use_bf16 or use_fp16:
|
802 |
+
precision_type = "bf16" if use_bf16 else "fp16"
|
803 |
+
fsdp_config["fsdp_state_dict_type"] = "FULL_STATE_DICT"
|
804 |
+
log_info(f"FSDP using mixed precision: {precision_type}")
|
805 |
+
except ImportError:
|
806 |
+
log_info("FSDP imports failed, falling back to standard DDP")
|
807 |
+
fsdp_config = None
|
808 |
+
elif multi_gpu_strategy == "fsdp" and not is_distributed:
|
809 |
+
log_info("FSDP disabled: requires distributed environment (use torchrun or accelerate)")
|
810 |
+
log_info("Using DataParallel for multi-GPU training instead")
|
811 |
+
else:
|
812 |
+
log_info(f"Using {multi_gpu_strategy} for multi-GPU training")
|
813 |
+
|
814 |
+
# Get system settings from hardware config
|
815 |
+
dataloader_workers = hardware_config.get("system_settings", {}).get("dataloader_num_workers", 2)
|
816 |
+
pin_memory = hardware_config.get("system_settings", {}).get("dataloader_pin_memory", True)
|
817 |
+
|
818 |
+
# Set up training arguments
|
819 |
+
log_info("Setting up training arguments")
|
820 |
+
training_args = TrainingArguments(
|
821 |
+
output_dir=transformers_config.get("output_dir", "./results") or transformers_config.get("checkpointing", {}).get("output_dir", "./results"),
|
822 |
+
num_train_epochs=transformers_config.get("training", {}).get("num_train_epochs", 3),
|
823 |
+
per_device_train_batch_size=per_device_batch_size,
|
824 |
+
gradient_accumulation_steps=gradient_accumulation_steps,
|
825 |
+
learning_rate=transformers_config.get("training", {}).get("learning_rate", 2e-5),
|
826 |
+
weight_decay=transformers_config.get("training", {}).get("weight_decay", 0.01),
|
827 |
+
warmup_ratio=transformers_config.get("training", {}).get("warmup_ratio", 0.05),
|
828 |
+
lr_scheduler_type=transformers_config.get("training", {}).get("lr_scheduler_type", "cosine"),
|
829 |
+
logging_steps=transformers_config.get("training", {}).get("logging_steps", 10),
|
830 |
+
save_strategy=transformers_config.get("checkpointing", {}).get("save_strategy", "steps"),
|
831 |
+
save_steps=transformers_config.get("checkpointing", {}).get("save_steps", 100),
|
832 |
+
save_total_limit=transformers_config.get("checkpointing", {}).get("save_total_limit", 3),
|
833 |
+
fp16=use_fp16,
|
834 |
+
bf16=use_bf16,
|
835 |
+
max_grad_norm=transformers_config.get("training", {}).get("max_grad_norm", 1.0),
|
836 |
+
push_to_hub=transformers_config.get("huggingface_hub", {}).get("push_to_hub", False),
|
837 |
+
hub_model_id=transformers_config.get("huggingface_hub", {}).get("hub_model_id", None),
|
838 |
+
hub_token=os.environ.get("HF_TOKEN", None),
|
839 |
+
report_to="tensorboard",
|
840 |
+
remove_unused_columns=False, # Keep all columns
|
841 |
+
gradient_checkpointing=transformers_config.get("training", {}).get("gradient_checkpointing", True),
|
842 |
+
dataloader_pin_memory=pin_memory,
|
843 |
+
optim=transformers_config.get("training", {}).get("optim", "adamw_torch"),
|
844 |
+
ddp_find_unused_parameters=False, # Improve distributed training efficiency
|
845 |
+
dataloader_drop_last=False, # Process all examples
|
846 |
+
dataloader_num_workers=dataloader_workers,
|
847 |
+
no_cuda=False if CUDA_AVAILABLE else True, # Use CUDA if available
|
848 |
+
# Only add FSDP if we're in distributed mode with FSDP strategy
|
849 |
+
fsdp=fsdp_config if is_distributed and multi_gpu_strategy == "fsdp" else None,
|
850 |
+
)
|
851 |
+
|
852 |
+
# Create sequential sampler to maintain original dataset order
|
853 |
+
sequential_sampler = torch.utils.data.SequentialSampler(dataset)
|
854 |
+
|
855 |
+
# Initialize trainer first
|
856 |
+
log_info("Initializing Trainer")
|
857 |
+
trainer = Trainer(
|
858 |
+
model=model,
|
859 |
+
args=training_args,
|
860 |
+
train_dataset=dataset, # We'll override this with our custom dataloader
|
861 |
+
data_collator=data_collator,
|
862 |
+
callbacks=[LoggingCallback()],
|
863 |
+
)
|
864 |
+
|
865 |
+
# Then override the get_train_dataloader method
|
866 |
+
def custom_get_train_dataloader():
|
867 |
+
"""Custom dataloader that preserves original dataset order"""
|
868 |
+
log_info("Creating sequential dataloader to maintain original dataset order")
|
869 |
+
|
870 |
+
# Create a simple sequential sampler
|
871 |
+
sequential_sampler = torch.utils.data.SequentialSampler(dataset)
|
872 |
+
|
873 |
+
# Verification of sequence preservation flags - simplified
|
874 |
+
data_loading_config = dataset_config.get("data_loading", {})
|
875 |
+
shuffle_enabled = data_loading_config.get("shuffle", False)
|
876 |
+
|
877 |
+
if shuffle_enabled:
|
878 |
+
log_info("WARNING: Shuffle is enabled in configuration! This will be overridden to preserve order.")
|
879 |
+
# We enforce sequential processing regardless of config
|
880 |
+
|
881 |
+
# Log our approach clearly
|
882 |
+
log_info("Using SequentialSampler to guarantee dataset order is preserved based on prompt_number")
|
883 |
+
|
884 |
+
# Verify column order
|
885 |
+
expected_order = ["prompt_number", "article_id", "conversations"]
|
886 |
+
if hasattr(dataset, 'column_names'):
|
887 |
+
actual_order = dataset.column_names
|
888 |
+
if actual_order == expected_order:
|
889 |
+
log_info(f"Confirmed dataset columns are in expected order: {', '.join(expected_order)}")
|
890 |
+
else:
|
891 |
+
log_info(f"Note: Dataset columns ({', '.join(actual_order)}) are not in expected order ({', '.join(expected_order)})")
|
892 |
+
log_info("This is handled correctly by field-based access, but noting for clarity")
|
893 |
+
|
894 |
+
log_info("Dataset is pre-processed with prompt_number field indicating the correct sequence")
|
895 |
+
|
896 |
+
# Calculate batch size based on device availability
|
897 |
+
if getattr(training_args, "no_cuda", False):
|
898 |
+
batch_size = training_args.per_device_train_batch_size
|
899 |
+
else:
|
900 |
+
batch_size = max(training_args.per_device_train_batch_size * max(1, NUM_GPUS), 1)
|
901 |
+
|
902 |
+
log_info(f"Using sequential sampler with batch size {batch_size}")
|
903 |
+
|
904 |
+
# Return DataLoader with sequential sampler
|
905 |
+
return torch.utils.data.DataLoader(
|
906 |
+
dataset,
|
907 |
+
batch_size=batch_size,
|
908 |
+
sampler=sequential_sampler, # Always use sequential sampler
|
909 |
+
collate_fn=data_collator,
|
910 |
+
drop_last=training_args.dataloader_drop_last,
|
911 |
+
num_workers=training_args.dataloader_num_workers,
|
912 |
+
pin_memory=training_args.dataloader_pin_memory,
|
913 |
+
)
|
914 |
+
|
915 |
+
# Override the get_train_dataloader method
|
916 |
+
trainer.get_train_dataloader = custom_get_train_dataloader
|
917 |
+
|
918 |
+
# Start training
|
919 |
+
log_info("=== Starting Training ===")
|
920 |
+
try:
|
921 |
+
# Empty cache again right before training
|
922 |
+
if CUDA_AVAILABLE:
|
923 |
+
torch.cuda.empty_cache()
|
924 |
+
log_info("Cleared CUDA cache before training")
|
925 |
+
|
926 |
+
# Display compact training info
|
927 |
+
total_steps = int(len(dataset) / (per_device_batch_size * NUM_GPUS * gradient_accumulation_steps) * training_args.num_train_epochs)
|
928 |
+
log_info(f"Training plan: {len(dataset)} examples over {training_args.num_train_epochs} epochs ≈ {total_steps} steps")
|
929 |
+
|
930 |
+
trainer.train()
|
931 |
+
log_info("Training completed successfully!")
|
932 |
+
|
933 |
+
# Save the final model
|
934 |
+
log_info("Saving final model...")
|
935 |
+
trainer.save_model()
|
936 |
+
log_info(f"Model saved to {training_args.output_dir}")
|
937 |
+
|
938 |
+
# Push to hub if enabled
|
939 |
+
if transformers_config.get("huggingface_hub", {}).get("push_to_hub", False):
|
940 |
+
hub_id = transformers_config.get("huggingface_hub", {}).get("hub_model_id", "model")
|
941 |
+
log_info(f"Pushing model to Hugging Face Hub as {hub_id}...")
|
942 |
+
trainer.push_to_hub()
|
943 |
+
log_info("Model successfully pushed to Hub")
|
944 |
+
|
945 |
+
return 0
|
946 |
+
except Exception as e:
|
947 |
+
logger.error(f"Training failed with error: {str(e)}")
|
948 |
+
# Log CUDA memory info if available in compact format
|
949 |
+
if CUDA_AVAILABLE:
|
950 |
+
memory_info = []
|
951 |
+
for i in range(NUM_GPUS):
|
952 |
+
allocated = torch.cuda.memory_allocated(i) / 1024**2
|
953 |
+
reserved = torch.cuda.memory_reserved(i) / 1024**2
|
954 |
+
max_mem = torch.cuda.max_memory_allocated(i) / 1024**2
|
955 |
+
memory_info.append(f"GPU {i}: {allocated:.1f}MB/{reserved:.1f}MB (max: {max_mem:.1f}MB)")
|
956 |
+
logger.error(f"GPU memory at failure: {', '.join(memory_info)}")
|
957 |
+
raise
|
958 |
+
|
959 |
+
except Exception as e:
|
960 |
+
logger.error(f"Error in main training loop: {str(e)}")
|
961 |
+
return 1
|
962 |
+
|
963 |
+
if __name__ == "__main__":
|
964 |
+
sys.exit(main())
|
transformers_config.json
ADDED
@@ -0,0 +1,162 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"model": {
|
3 |
+
"name": "unsloth/phi-4-unsloth-bnb-4bit",
|
4 |
+
"trust_remote_code": true,
|
5 |
+
"use_fast_tokenizer": true
|
6 |
+
},
|
7 |
+
|
8 |
+
"tokenizer": {
|
9 |
+
"chat_template": "phi",
|
10 |
+
"max_seq_length": 2048,
|
11 |
+
"padding_side": "right",
|
12 |
+
"add_eos_token": true
|
13 |
+
},
|
14 |
+
|
15 |
+
"training": {
|
16 |
+
"per_device_train_batch_size": 16,
|
17 |
+
"gradient_accumulation_steps": 3,
|
18 |
+
"learning_rate": 2e-5,
|
19 |
+
"num_train_epochs": 3,
|
20 |
+
"max_steps": -1,
|
21 |
+
"logging_steps": 10,
|
22 |
+
"save_steps": 200,
|
23 |
+
"save_total_limit": 5,
|
24 |
+
"push_to_hub": true,
|
25 |
+
"hub_strategy": "every_save",
|
26 |
+
"gradient_checkpointing": true,
|
27 |
+
"optim": "adamw_torch",
|
28 |
+
"lr_scheduler_type": "cosine",
|
29 |
+
"warmup_ratio": 0.05,
|
30 |
+
"weight_decay": 0.01,
|
31 |
+
"max_grad_norm": 1.0,
|
32 |
+
"neftune_noise_alpha": 5,
|
33 |
+
"fp16": false,
|
34 |
+
"bf16": true
|
35 |
+
},
|
36 |
+
|
37 |
+
"checkpointing": {
|
38 |
+
"output_dir": "./results",
|
39 |
+
"save_strategy": "steps",
|
40 |
+
"save_steps": 100,
|
41 |
+
"save_total_limit": 3,
|
42 |
+
"hub_strategy": "every_save"
|
43 |
+
},
|
44 |
+
|
45 |
+
"unsloth": {
|
46 |
+
"enabled": true,
|
47 |
+
"r": 32,
|
48 |
+
"alpha": 16,
|
49 |
+
"dropout": 0.05,
|
50 |
+
"target_modules": [
|
51 |
+
"q_proj",
|
52 |
+
"k_proj",
|
53 |
+
"v_proj",
|
54 |
+
"o_proj",
|
55 |
+
"gate_proj",
|
56 |
+
"up_proj",
|
57 |
+
"down_proj"
|
58 |
+
]
|
59 |
+
},
|
60 |
+
|
61 |
+
"distributed_training": {
|
62 |
+
"fsdp_config": {
|
63 |
+
"enabled": false,
|
64 |
+
"sharding_strategy": "FULL_SHARD",
|
65 |
+
"mixed_precision": "BF16",
|
66 |
+
"activation_checkpointing": true,
|
67 |
+
"offload_params": false
|
68 |
+
},
|
69 |
+
"ddp_find_unused_parameters": false,
|
70 |
+
"dataloader_num_workers": 2
|
71 |
+
},
|
72 |
+
|
73 |
+
"logging": {
|
74 |
+
"logging_steps": 50,
|
75 |
+
"log_level": "info"
|
76 |
+
},
|
77 |
+
|
78 |
+
"huggingface_hub": {
|
79 |
+
"push_to_hub": true,
|
80 |
+
"hub_model_id": "phi-4-cognitive-assistant",
|
81 |
+
"hub_private_repo": true
|
82 |
+
},
|
83 |
+
|
84 |
+
"model_name_or_path": "unsloth/phi-4-unsloth-bnb-4bit",
|
85 |
+
"model_revision": "main",
|
86 |
+
"use_flash_attention": true,
|
87 |
+
"torch_dtype": "bfloat16",
|
88 |
+
"bf16": true,
|
89 |
+
"fp16": false,
|
90 |
+
|
91 |
+
"hardware": {
|
92 |
+
"hardware_name": "4xL4",
|
93 |
+
"specs": {
|
94 |
+
"gpu_count": 4,
|
95 |
+
"gpu_type": "L4",
|
96 |
+
"vram_per_gpu": 24,
|
97 |
+
"total_vram": 96,
|
98 |
+
"vcpu_count": 48,
|
99 |
+
"ram": 186
|
100 |
+
},
|
101 |
+
"hardware_setup": {
|
102 |
+
"use_cpu": false,
|
103 |
+
"num_gpus": 4,
|
104 |
+
"device_map": "auto"
|
105 |
+
},
|
106 |
+
"training_optimizations": {
|
107 |
+
"per_device_batch_size": 16,
|
108 |
+
"gradient_accumulation_steps": 3,
|
109 |
+
"mixed_precision": "bf16",
|
110 |
+
"torch_compile": false,
|
111 |
+
"memory_optimizations": {
|
112 |
+
"use_gradient_checkpointing": true,
|
113 |
+
"use_flash_attention": true
|
114 |
+
},
|
115 |
+
"multi_gpu_strategy": "data_parallel"
|
116 |
+
},
|
117 |
+
"system_settings": {
|
118 |
+
"cuda_memory_fraction": 0.85,
|
119 |
+
"dataloader_num_workers": 2,
|
120 |
+
"dataloader_pin_memory": true
|
121 |
+
},
|
122 |
+
"memory_breakdown": {
|
123 |
+
"model_size": "~3.5GB (pre-quantized 4-bit)",
|
124 |
+
"optimizer_states": "~1GB",
|
125 |
+
"batch_memory_per_gpu": "~3GB",
|
126 |
+
"peak_memory_estimate": "~18GB",
|
127 |
+
"safe_headroom": "~6GB"
|
128 |
+
},
|
129 |
+
"compute_environment": "L4_CLOUD"
|
130 |
+
},
|
131 |
+
|
132 |
+
"dataset": {
|
133 |
+
"dataset": {
|
134 |
+
"name": "George-API/phi4-cognitive-dataset",
|
135 |
+
"split": "train"
|
136 |
+
},
|
137 |
+
"data_formatting": {
|
138 |
+
"chat_template": "phi",
|
139 |
+
"roles": {
|
140 |
+
"system": "System: {content}\n\n",
|
141 |
+
"human": "Human: {content}\n\n",
|
142 |
+
"assistant": "Assistant: {content}\n\n",
|
143 |
+
"user": "Human: {content}\n\n"
|
144 |
+
}
|
145 |
+
},
|
146 |
+
"data_loading": {
|
147 |
+
"batch_size": 24,
|
148 |
+
"shuffle": false,
|
149 |
+
"sequential_processing": true,
|
150 |
+
"drop_last": false,
|
151 |
+
"num_workers": 4,
|
152 |
+
"pin_memory": true,
|
153 |
+
"prefetch_factor": 4
|
154 |
+
},
|
155 |
+
"validation": {
|
156 |
+
"log_samples": 3,
|
157 |
+
"log_interval": 50,
|
158 |
+
"verify_sequence_integrity": true,
|
159 |
+
"metrics": ["processed", "skipped", "avg_tokens", "unique_articles"]
|
160 |
+
}
|
161 |
+
}
|
162 |
+
}
|
update_space.py
ADDED
@@ -0,0 +1,278 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
"""
|
4 |
+
Quick script to update your Hugging Face Space for phi-4-unsloth-bnb-4bit training.
|
5 |
+
This script handles the specific requirements for the 4-bit quantized Phi-4 model training,
|
6 |
+
including proper configuration and dependency management.
|
7 |
+
"""
|
8 |
+
|
9 |
+
import os
|
10 |
+
import sys
|
11 |
+
import json
|
12 |
+
import subprocess
|
13 |
+
import argparse
|
14 |
+
import logging
|
15 |
+
from pathlib import Path
|
16 |
+
from huggingface_hub import HfApi, login
|
17 |
+
import getpass
|
18 |
+
|
19 |
+
# Configure logging
|
20 |
+
logging.basicConfig(
|
21 |
+
level=logging.INFO,
|
22 |
+
format="%(asctime)s - %(levelname)s - %(message)s",
|
23 |
+
handlers=[logging.StreamHandler(sys.stdout)]
|
24 |
+
)
|
25 |
+
logger = logging.getLogger(__name__)
|
26 |
+
|
27 |
+
def load_env_variables():
|
28 |
+
"""Load environment variables from system or .env file."""
|
29 |
+
# First try to load from local .env file
|
30 |
+
try:
|
31 |
+
from dotenv import load_dotenv
|
32 |
+
env_path = Path(__file__).parent / ".env"
|
33 |
+
if env_path.exists():
|
34 |
+
# Load and explicitly set environment variables
|
35 |
+
with open(env_path) as f:
|
36 |
+
for line in f:
|
37 |
+
if line.strip() and not line.startswith('#'):
|
38 |
+
key, value = line.strip().split('=', 1)
|
39 |
+
os.environ[key] = value.strip()
|
40 |
+
logger.info(f"Loaded environment variables from {env_path}")
|
41 |
+
else:
|
42 |
+
logger.warning(f"No .env file found at {env_path}")
|
43 |
+
except ImportError:
|
44 |
+
logger.warning("python-dotenv not installed, skipping .env loading")
|
45 |
+
|
46 |
+
# Check if we're running in a Hugging Face Space
|
47 |
+
if os.environ.get("SPACE_ID"):
|
48 |
+
logger.info("Running in Hugging Face Space")
|
49 |
+
if "/" in os.environ.get("SPACE_ID", ""):
|
50 |
+
username = os.environ.get("SPACE_ID").split("/")[0]
|
51 |
+
os.environ["HF_USERNAME"] = username
|
52 |
+
logger.info(f"Set HF_USERNAME from SPACE_ID: {username}")
|
53 |
+
|
54 |
+
# Verify required variables
|
55 |
+
required_vars = {
|
56 |
+
"HF_TOKEN": os.environ.get("HF_TOKEN"),
|
57 |
+
"HF_USERNAME": os.environ.get("HF_USERNAME"),
|
58 |
+
"HF_SPACE_NAME": os.environ.get("HF_SPACE_NAME", "phi4training")
|
59 |
+
}
|
60 |
+
|
61 |
+
# Ensure the space name is set correctly
|
62 |
+
if "HF_SPACE_NAME" not in os.environ:
|
63 |
+
os.environ["HF_SPACE_NAME"] = "phi4training"
|
64 |
+
|
65 |
+
missing_vars = [k for k, v in required_vars.items() if not v]
|
66 |
+
if missing_vars:
|
67 |
+
raise ValueError(f"Missing required environment variables: {', '.join(missing_vars)}")
|
68 |
+
|
69 |
+
logger.info(f"Using environment variables: USERNAME={required_vars['HF_USERNAME']}, SPACE_NAME={required_vars['HF_SPACE_NAME']}")
|
70 |
+
return required_vars
|
71 |
+
|
72 |
+
def verify_configs():
|
73 |
+
"""Verify that all necessary configuration files exist and are valid."""
|
74 |
+
current_dir = Path(__file__).parent
|
75 |
+
required_files = [
|
76 |
+
"transformers_config.json",
|
77 |
+
"requirements.txt",
|
78 |
+
"run_transformers_training.py"
|
79 |
+
]
|
80 |
+
|
81 |
+
missing_files = []
|
82 |
+
for file in required_files:
|
83 |
+
if not (current_dir / file).exists():
|
84 |
+
missing_files.append(file)
|
85 |
+
|
86 |
+
if missing_files:
|
87 |
+
raise FileNotFoundError(f"Missing required files: {', '.join(missing_files)}")
|
88 |
+
|
89 |
+
# Verify JSON configs
|
90 |
+
json_files = [f for f in required_files if f.endswith('.json')]
|
91 |
+
for json_file in json_files:
|
92 |
+
try:
|
93 |
+
with open(current_dir / json_file) as f:
|
94 |
+
json.load(f)
|
95 |
+
logger.info(f"Verified {json_file} is valid JSON")
|
96 |
+
except json.JSONDecodeError as e:
|
97 |
+
raise ValueError(f"Invalid JSON in {json_file}: {e}")
|
98 |
+
|
99 |
+
def update_requirements():
|
100 |
+
"""Update requirements.txt with necessary packages using a two-stage installation process."""
|
101 |
+
logger.info("Setting up requirements files for sequential installation...")
|
102 |
+
current_dir = Path(__file__).parent
|
103 |
+
base_req_path = current_dir / "requirements-base.txt"
|
104 |
+
main_req_path = current_dir / "requirements.txt"
|
105 |
+
flash_req_path = current_dir / "requirements-flash.txt"
|
106 |
+
|
107 |
+
# First ensure base requirements exist
|
108 |
+
required_base_packages = {
|
109 |
+
"torch>=2.0.0",
|
110 |
+
"transformers>=4.36.0",
|
111 |
+
"accelerate>=0.27.0",
|
112 |
+
"bitsandbytes>=0.41.0",
|
113 |
+
"tensorboard>=2.15.0",
|
114 |
+
"gradio>=5.17.0",
|
115 |
+
"huggingface-hub>=0.19.0",
|
116 |
+
"datasets>=2.15.0"
|
117 |
+
}
|
118 |
+
|
119 |
+
# Additional packages for main requirements
|
120 |
+
required_additional_packages = {
|
121 |
+
"einops>=0.7.0",
|
122 |
+
"filelock>=3.13.1",
|
123 |
+
"matplotlib>=3.7.0",
|
124 |
+
"numpy>=1.24.0",
|
125 |
+
"packaging>=23.0",
|
126 |
+
"peft>=0.9.0",
|
127 |
+
"psutil>=5.9.0",
|
128 |
+
"python-dotenv>=1.0.0",
|
129 |
+
"pyyaml>=6.0.1",
|
130 |
+
"regex>=2023.0.0",
|
131 |
+
"requests>=2.31.0",
|
132 |
+
"safetensors>=0.4.1",
|
133 |
+
"sentencepiece>=0.1.99",
|
134 |
+
"tqdm>=4.65.0",
|
135 |
+
"typing-extensions>=4.8.0",
|
136 |
+
"unsloth>=2024.3"
|
137 |
+
}
|
138 |
+
|
139 |
+
# Read existing base requirements
|
140 |
+
existing_requirements = set()
|
141 |
+
if base_req_path.exists():
|
142 |
+
with open(base_req_path) as f:
|
143 |
+
existing_requirements = {line.strip() for line in f if line.strip() and not line.startswith('-r')}
|
144 |
+
|
145 |
+
# Add new requirements
|
146 |
+
updated_requirements = existing_requirements.union(required_base_packages)
|
147 |
+
|
148 |
+
# 1. Write updated base requirements
|
149 |
+
with open(base_req_path, 'w') as f:
|
150 |
+
# Ensure torch is first
|
151 |
+
torch_req = next((req for req in updated_requirements if req.startswith("torch")), "torch>=2.0.0")
|
152 |
+
f.write(f"{torch_req}\n")
|
153 |
+
|
154 |
+
# Write all other requirements (excluding torch)
|
155 |
+
for req in sorted(r for r in updated_requirements if not r.startswith("torch")):
|
156 |
+
f.write(f"{req}\n")
|
157 |
+
|
158 |
+
# 2. Create main requirements file (references base)
|
159 |
+
with open(main_req_path, 'w') as f:
|
160 |
+
f.write("-r requirements-base.txt\n")
|
161 |
+
for req in sorted(required_additional_packages):
|
162 |
+
f.write(f"{req}\n")
|
163 |
+
|
164 |
+
# 3. Create or update flash-attn requirements
|
165 |
+
with open(flash_req_path, 'w') as f:
|
166 |
+
f.write("-r requirements-base.txt\n")
|
167 |
+
f.write("flash-attn==2.5.2\n")
|
168 |
+
|
169 |
+
logger.info("Updated requirements files for sequential installation:")
|
170 |
+
logger.info(f"1. Base requirements in {base_req_path}")
|
171 |
+
logger.info(f"2. Main requirements in {main_req_path}")
|
172 |
+
logger.info(f"3. Flash-attention requirements in {flash_req_path}")
|
173 |
+
logger.info("This ensures packages are installed in the correct order")
|
174 |
+
|
175 |
+
def create_space(username, space_name):
|
176 |
+
"""Create or get a Hugging Face Space."""
|
177 |
+
try:
|
178 |
+
api = HfApi()
|
179 |
+
space_id = f"{username}/{space_name}"
|
180 |
+
logger.info(f"Checking Space {space_id}...")
|
181 |
+
|
182 |
+
# First try to get the space
|
183 |
+
try:
|
184 |
+
space_info = api.space_info(repo_id=space_id)
|
185 |
+
logger.info(f"Space {space_id} already exists")
|
186 |
+
return space_info
|
187 |
+
except Exception as e:
|
188 |
+
logger.info(f"Space {space_id} does not exist, creating new space...")
|
189 |
+
|
190 |
+
# Create new space
|
191 |
+
try:
|
192 |
+
api.create_repo(
|
193 |
+
repo_id=space_id,
|
194 |
+
private=False,
|
195 |
+
repo_type="space",
|
196 |
+
space_sdk="gradio"
|
197 |
+
)
|
198 |
+
logger.info(f"Created new space: {space_id}")
|
199 |
+
return api.space_info(repo_id=space_id)
|
200 |
+
except Exception as e:
|
201 |
+
logger.error(f"Failed to create space: {str(e)}")
|
202 |
+
raise
|
203 |
+
except Exception as e:
|
204 |
+
raise RuntimeError(f"Error with Space {space_id}: {str(e)}")
|
205 |
+
|
206 |
+
def main():
|
207 |
+
parser = argparse.ArgumentParser(description='Update Hugging Face Space for Phi-4 training')
|
208 |
+
parser.add_argument('--space_name', type=str, help='Space name (default: from env)')
|
209 |
+
parser.add_argument('--force', action='store_true', help='Skip confirmation')
|
210 |
+
args = parser.parse_args()
|
211 |
+
|
212 |
+
if not args.force:
|
213 |
+
print("\n" + "!"*80)
|
214 |
+
print("WARNING: Updating the Space will INTERRUPT any ongoing training!")
|
215 |
+
print("Make sure all checkpoints are saved before proceeding.")
|
216 |
+
print("!"*80 + "\n")
|
217 |
+
|
218 |
+
confirm = input("Type 'update' to confirm: ")
|
219 |
+
if confirm.lower() != 'update':
|
220 |
+
logger.info("Update cancelled")
|
221 |
+
return False
|
222 |
+
|
223 |
+
try:
|
224 |
+
# Load environment variables
|
225 |
+
env_vars = load_env_variables()
|
226 |
+
logger.info(f"Environment variables loaded: USERNAME={env_vars['HF_USERNAME']}, SPACE_NAME={env_vars['HF_SPACE_NAME']}")
|
227 |
+
|
228 |
+
# Verify configurations
|
229 |
+
verify_configs()
|
230 |
+
logger.info("All configuration files verified successfully")
|
231 |
+
|
232 |
+
# Update requirements
|
233 |
+
update_requirements()
|
234 |
+
logger.info("Requirements updated successfully")
|
235 |
+
|
236 |
+
# Get space name from args or env, prioritize args
|
237 |
+
space_name = args.space_name if args.space_name else env_vars["HF_SPACE_NAME"]
|
238 |
+
logger.info(f"Using space name: {space_name}")
|
239 |
+
|
240 |
+
# Login to Hugging Face
|
241 |
+
logger.info("Logging in to Hugging Face...")
|
242 |
+
login(token=env_vars["HF_TOKEN"])
|
243 |
+
logger.info("Successfully logged in to Hugging Face")
|
244 |
+
|
245 |
+
# Create/get space
|
246 |
+
space_info = create_space(env_vars["HF_USERNAME"], space_name)
|
247 |
+
logger.info(f"Space info: {space_info}")
|
248 |
+
|
249 |
+
# Upload files
|
250 |
+
current_dir = Path(__file__).parent
|
251 |
+
logger.info(f"Uploading files from {current_dir} to Space {env_vars['HF_USERNAME']}/{space_name}...")
|
252 |
+
|
253 |
+
# Create .gitignore
|
254 |
+
with open(current_dir / ".gitignore", "w") as f:
|
255 |
+
f.write(".env\n*.pyc\n__pycache__\n")
|
256 |
+
logger.info("Created .gitignore file")
|
257 |
+
|
258 |
+
api = HfApi()
|
259 |
+
api.upload_folder(
|
260 |
+
folder_path=str(current_dir),
|
261 |
+
repo_id=f"{env_vars['HF_USERNAME']}/{space_name}",
|
262 |
+
repo_type="space",
|
263 |
+
ignore_patterns=[".env", "*.pyc", "__pycache__", "TRAINING_IN_PROGRESS.lock"]
|
264 |
+
)
|
265 |
+
|
266 |
+
logger.info(f"Files uploaded successfully")
|
267 |
+
space_url = f"https://huggingface.co/spaces/{env_vars['HF_USERNAME']}/{space_name}"
|
268 |
+
logger.info(f"Space URL: {space_url}")
|
269 |
+
print(f"\nSpace created successfully! You can view it at:\n{space_url}")
|
270 |
+
return True
|
271 |
+
|
272 |
+
except Exception as e:
|
273 |
+
logger.error(f"Error updating Space: {str(e)}")
|
274 |
+
return False
|
275 |
+
|
276 |
+
if __name__ == "__main__":
|
277 |
+
success = main()
|
278 |
+
sys.exit(0 if success else 1)
|