Spaces:
Running
Running
Upload README_space.md with huggingface_hub
Browse files- README_space.md +243 -0
README_space.md
ADDED
|
@@ -0,0 +1,243 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- bn
|
| 5 |
+
- en
|
| 6 |
+
tags:
|
| 7 |
+
- transformers
|
| 8 |
+
- safetensors
|
| 9 |
+
- stable-diffusion
|
| 10 |
+
- bangla
|
| 11 |
+
- text-to-video
|
| 12 |
+
- lora
|
| 13 |
+
- scene-planning
|
| 14 |
+
- computer-vision
|
| 15 |
+
- natural-language-processing
|
| 16 |
+
- mlops
|
| 17 |
+
- production-grade
|
| 18 |
+
pipeline_tag: text-to-video
|
| 19 |
+
model-index:
|
| 20 |
+
- name: memo
|
| 21 |
+
results: []
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
# Memo: Production-Grade Transformers + Safetensors Implementation
|
| 25 |
+
|
| 26 |
+

|
| 27 |
+

|
| 28 |
+

|
| 29 |
+

|
| 30 |
+
|
| 31 |
+
## Overview
|
| 32 |
+
|
| 33 |
+
This is the complete transformation of Memo to use **Transformers + Safetensors** properly, replacing unsafe pickle files and toy logic with enterprise-grade machine learning infrastructure.
|
| 34 |
+
|
| 35 |
+
## What We've Built
|
| 36 |
+
|
| 37 |
+
### β
Core Requirements Met
|
| 38 |
+
|
| 39 |
+
1. **Transformers Integration**
|
| 40 |
+
- Bangla text parsing using `google/mt5-small`
|
| 41 |
+
- Proper tokenization and model loading
|
| 42 |
+
- Deterministic scene extraction with controlled parameters
|
| 43 |
+
- Memory optimization with device mapping
|
| 44 |
+
|
| 45 |
+
2. **Safetensors Security**
|
| 46 |
+
- **MANDATORY** `use_safetensors=True` for all model loading
|
| 47 |
+
- No .bin, .ckpt, or pickle files anywhere
|
| 48 |
+
- Model weight validation and security checks
|
| 49 |
+
- Signature verification for LoRA files
|
| 50 |
+
|
| 51 |
+
3. **Production Architecture**
|
| 52 |
+
- Tier-based model management (Free/Pro/Enterprise)
|
| 53 |
+
- Memory optimization and performance tuning
|
| 54 |
+
- Background processing for long-running tasks
|
| 55 |
+
- Proper error handling and logging
|
| 56 |
+
|
| 57 |
+
## File Structure
|
| 58 |
+
|
| 59 |
+
```
|
| 60 |
+
π Memo/
|
| 61 |
+
βββ π requirements.txt # Production dependencies
|
| 62 |
+
βββ π models/
|
| 63 |
+
β βββ π text/
|
| 64 |
+
β βββ π bangla_parser.py # Transformer-based Bangla parser
|
| 65 |
+
βββ π core/
|
| 66 |
+
β βββ π scene_planner.py # ML-based scene planning
|
| 67 |
+
βββ π models/
|
| 68 |
+
β βββ π image/
|
| 69 |
+
β βββ π sd_generator.py # Stable Diffusion + Safetensors
|
| 70 |
+
βββ π data/
|
| 71 |
+
β βββ π lora/
|
| 72 |
+
β βββ π README.md # LoRA configuration (safetensors only)
|
| 73 |
+
βββ π scripts/
|
| 74 |
+
β βββ π train_scene_lora.py # Training with safetensors output
|
| 75 |
+
βββ π config/
|
| 76 |
+
β βββ π model_tiers.py # Tier management system
|
| 77 |
+
βββ π api/
|
| 78 |
+
βββ π main.py # Production API endpoint
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
## Key Features
|
| 82 |
+
|
| 83 |
+
### π Security (Non-Negotiable)
|
| 84 |
+
- **Safetensors-only model loading** - No unsafe formats
|
| 85 |
+
- **Model signature validation** - Verify weight integrity
|
| 86 |
+
- **LoRA security checks** - Ensure only .safetensors files
|
| 87 |
+
- **Memory-safe loading** - Prevent buffer overflows
|
| 88 |
+
|
| 89 |
+
### π Performance
|
| 90 |
+
- **Memory optimization** - xFormers, attention slicing, CPU offload
|
| 91 |
+
- **FP16 precision** - 50% memory reduction with maintained quality
|
| 92 |
+
- **LCM acceleration** - Faster inference when available
|
| 93 |
+
- **Device mapping** - Optimal GPU/CPU utilization
|
| 94 |
+
|
| 95 |
+
### π’ Enterprise Features
|
| 96 |
+
- **Tier-based pricing** - Free/Pro/Enterprise configurations
|
| 97 |
+
- **Resource management** - Memory limits and concurrent request handling
|
| 98 |
+
- **Security compliance** - Audit trails and validation
|
| 99 |
+
- **Scalability** - Background processing and proper async handling
|
| 100 |
+
|
| 101 |
+
## Model Tiers
|
| 102 |
+
|
| 103 |
+
### Free Tier
|
| 104 |
+
- Base SDXL model (512x512)
|
| 105 |
+
- 15 inference steps
|
| 106 |
+
- No LoRA
|
| 107 |
+
- 1 concurrent request
|
| 108 |
+
|
| 109 |
+
### Pro Tier
|
| 110 |
+
- Base SDXL model (768x768)
|
| 111 |
+
- 25 inference steps
|
| 112 |
+
- Scene LoRA enabled
|
| 113 |
+
- LCM acceleration
|
| 114 |
+
- 3 concurrent requests
|
| 115 |
+
|
| 116 |
+
### Enterprise Tier
|
| 117 |
+
- Base SDXL model (1024x1024)
|
| 118 |
+
- 30 inference steps
|
| 119 |
+
- Custom LoRA support
|
| 120 |
+
- LCM acceleration
|
| 121 |
+
- 10 concurrent requests
|
| 122 |
+
|
| 123 |
+
## Usage Examples
|
| 124 |
+
|
| 125 |
+
### Basic Scene Planning
|
| 126 |
+
```python
|
| 127 |
+
from core.scene_planner import plan_scenes
|
| 128 |
+
|
| 129 |
+
scenes = plan_scenes(
|
| 130 |
+
text_bn="ΰ¦ΰ¦ΰ¦ΰ§ΰ¦° দিনΰ¦ΰ¦Ώ ΰ¦ΰ§ΰ¦¬ ΰ¦Έΰ§ΰ¦¨ΰ§ΰ¦¦ΰ¦° ΰ¦ΰ¦Ώΰ¦²ΰ₯€",
|
| 131 |
+
duration=15
|
| 132 |
+
)
|
| 133 |
+
```
|
| 134 |
+
|
| 135 |
+
### Tier-Based Generation
|
| 136 |
+
```python
|
| 137 |
+
from config.model_tiers import get_tier_config
|
| 138 |
+
from models.image.sd_generator import get_generator
|
| 139 |
+
|
| 140 |
+
config = get_tier_config("pro")
|
| 141 |
+
generator = get_generator(
|
| 142 |
+
model_id=config.image_model_id,
|
| 143 |
+
lora_path=config.lora_path,
|
| 144 |
+
use_lcm=config.lcm_enabled
|
| 145 |
+
)
|
| 146 |
+
|
| 147 |
+
frames = generator.generate_frames(
|
| 148 |
+
prompt="Beautiful landscape scene",
|
| 149 |
+
frames=5
|
| 150 |
+
)
|
| 151 |
+
```
|
| 152 |
+
|
| 153 |
+
### API Usage
|
| 154 |
+
```bash
|
| 155 |
+
curl -X POST "http://localhost:8000/generate" \\
|
| 156 |
+
-H "Content-Type: application/json" \\
|
| 157 |
+
-d '{
|
| 158 |
+
"text": "ΰ¦ΰ¦ΰ¦ΰ§ΰ¦° দিনΰ¦ΰ¦Ώ ΰ¦ΰ§ΰ¦¬ ΰ¦Έΰ§ΰ¦¨ΰ§ΰ¦¦ΰ¦° ΰ¦ΰ¦Ώΰ¦²ΰ₯€",
|
| 159 |
+
"duration": 15,
|
| 160 |
+
"tier": "pro"
|
| 161 |
+
}'
|
| 162 |
+
```
|
| 163 |
+
|
| 164 |
+
## Training Custom LoRA
|
| 165 |
+
|
| 166 |
+
```python
|
| 167 |
+
from scripts.train_scene_lora import SceneLoRATrainer, TrainingConfig
|
| 168 |
+
|
| 169 |
+
config = TrainingConfig(
|
| 170 |
+
base_model="google/mt5-small",
|
| 171 |
+
rank=32,
|
| 172 |
+
alpha=64,
|
| 173 |
+
save_safetensors=True # MANDATORY
|
| 174 |
+
)
|
| 175 |
+
|
| 176 |
+
trainer = SceneLoRATrainer(config)
|
| 177 |
+
trainer.load_model()
|
| 178 |
+
trainer.setup_lora()
|
| 179 |
+
trainer.train(training_data)
|
| 180 |
+
```
|
| 181 |
+
|
| 182 |
+
## Security Validation
|
| 183 |
+
|
| 184 |
+
```python
|
| 185 |
+
from config.model_tiers import validate_model_weights_security
|
| 186 |
+
|
| 187 |
+
result = validate_model_weights_security("data/lora/memo-scene-lora.safetensors")
|
| 188 |
+
print(f"Secure: {result['is_secure']}")
|
| 189 |
+
print(f"Issues: {result['issues']}")
|
| 190 |
+
```
|
| 191 |
+
|
| 192 |
+
## What This Guarantees
|
| 193 |
+
|
| 194 |
+
β
**Transformers-based** - Real ML, not toy logic
|
| 195 |
+
β
**Safetensors-only** - No security vulnerabilities
|
| 196 |
+
β
**Production-ready** - Enterprise architecture
|
| 197 |
+
β
**Memory optimized** - Proper resource management
|
| 198 |
+
β
**Tier-based** - Scalable pricing model
|
| 199 |
+
β
**Audit compliant** - Security validation built-in
|
| 200 |
+
|
| 201 |
+
## What This Doesn't Do
|
| 202 |
+
|
| 203 |
+
β Make GPUs cheap
|
| 204 |
+
β Fix bad prompts
|
| 205 |
+
β Read your mind
|
| 206 |
+
β Guarantee perfect results
|
| 207 |
+
|
| 208 |
+
## Next Steps
|
| 209 |
+
|
| 210 |
+
If you're serious about production deployment:
|
| 211 |
+
|
| 212 |
+
1. **Cold-start optimization** - Preload frequently used models
|
| 213 |
+
2. **Model versioning** - Track changes per tier
|
| 214 |
+
3. **A/B testing** - Compare model performance
|
| 215 |
+
4. **Monitoring** - Track usage and performance metrics
|
| 216 |
+
5. **Load balancing** - Distribute across multiple GPUs
|
| 217 |
+
|
| 218 |
+
## Running the System
|
| 219 |
+
|
| 220 |
+
```bash
|
| 221 |
+
# Install dependencies
|
| 222 |
+
pip install -r requirements.txt
|
| 223 |
+
|
| 224 |
+
# Train custom LoRA
|
| 225 |
+
python scripts/train_scene_lora.py
|
| 226 |
+
|
| 227 |
+
# Start API server
|
| 228 |
+
python api/main.py
|
| 229 |
+
|
| 230 |
+
# Check health
|
| 231 |
+
curl http://localhost:8000/health
|
| 232 |
+
```
|
| 233 |
+
|
| 234 |
+
## Reality Check
|
| 235 |
+
|
| 236 |
+
This implementation is now:
|
| 237 |
+
- β
**Correct** - Uses proper ML frameworks
|
| 238 |
+
- β
**Modern** - Transformers + Safetensors
|
| 239 |
+
- β
**Secure** - No unsafe model formats
|
| 240 |
+
- β
**Scalable** - Tier-based architecture
|
| 241 |
+
- β
**Defensible** - Production-grade security
|
| 242 |
+
|
| 243 |
+
If your API claims "state-of-the-art" without these features, you're lying. Memo now actually delivers on that promise.
|