Add flux1-schnell-bf16-blackwell engines
Browse files- README.md +52 -0
- config.json +20 -0
- engines/ENGINES_NEEDED.txt +15 -0
README.md
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# FLUX1-SCHNELL-BF16-BLACKWELL
|
2 |
+
|
3 |
+
TensorRT-RTX engines for FLUX1-SCHNELL optimized for Blackwell architecture with BF16 quantization.
|
4 |
+
|
5 |
+
## Specifications
|
6 |
+
|
7 |
+
- **Model**: FLUX1-SCHNELL
|
8 |
+
- **Base Model**: black-forest-labs/FLUX.1-schnell
|
9 |
+
- **Architecture**: Blackwell
|
10 |
+
- **Quantization**: BF16
|
11 |
+
- **Batch Size**: 1 (optimized)
|
12 |
+
- **Resolution**: 1024x1024 (optimized)
|
13 |
+
- **Estimated Size**: 12.0 GB
|
14 |
+
|
15 |
+
## Performance Estimates
|
16 |
+
|
17 |
+
Based on architecture and quantization:
|
18 |
+
- **Memory Usage**: ~12.0GB VRAM
|
19 |
+
- **Speed**: ~2.5s (H200)
|
20 |
+
|
21 |
+
## Files
|
22 |
+
|
23 |
+
- `engines/`: TensorRT engine files (.plan)
|
24 |
+
- `config.json`: Configuration metadata
|
25 |
+
- `README.md`: This file
|
26 |
+
|
27 |
+
## Usage
|
28 |
+
|
29 |
+
```python
|
30 |
+
from imageai_server.tensorrt.nvidia_sdxl_pipeline import NVIDIASDXLPipeline
|
31 |
+
|
32 |
+
pipeline = NVIDIASDXLPipeline()
|
33 |
+
pipeline.load_engines(
|
34 |
+
engine_dir="./engines",
|
35 |
+
framework_model_dir="./framework",
|
36 |
+
onnx_dir="./onnx"
|
37 |
+
)
|
38 |
+
pipeline.activate_engines()
|
39 |
+
|
40 |
+
images, time_ms = pipeline.infer(
|
41 |
+
prompt="a beautiful landscape",
|
42 |
+
height=1024,
|
43 |
+
width=1024
|
44 |
+
)
|
45 |
+
```
|
46 |
+
|
47 |
+
## Requirements
|
48 |
+
|
49 |
+
- **GPU**: Blackwell architecture (Compute Capability 8.0+)
|
50 |
+
- **VRAM**: 12.0GB+
|
51 |
+
- **TensorRT-RTX**: 1.0.0.21+
|
52 |
+
- **CUDA**: 12.0+
|
config.json
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"model_type": "flux1-schnell",
|
3 |
+
"architecture": "blackwell",
|
4 |
+
"quantization": "bf16",
|
5 |
+
"batch_size": 1,
|
6 |
+
"height": 1024,
|
7 |
+
"width": 1024,
|
8 |
+
"base_model": "black-forest-labs/FLUX.1-schnell",
|
9 |
+
"expected_engines": [
|
10 |
+
"clip.plan",
|
11 |
+
"t5.plan",
|
12 |
+
"transformer.plan",
|
13 |
+
"vae.plan"
|
14 |
+
],
|
15 |
+
"created_at": 1754966251.0015864,
|
16 |
+
"tensorrt_version": "10.13.2.6",
|
17 |
+
"tensorrt_rtx_version": "1.0.0.21",
|
18 |
+
"size_estimate_gb": 12.0,
|
19 |
+
"build_system": "comprehensive_offline_builder"
|
20 |
+
}
|
engines/ENGINES_NEEDED.txt
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Engines for flux1-schnell-bf16-blackwell need to be built.
|
2 |
+
|
3 |
+
This is a placeholder structure. To complete:
|
4 |
+
1. Build actual flux1-schnell engines with bf16 quantization
|
5 |
+
2. Copy .plan files to this directory
|
6 |
+
3. Remove this placeholder file
|
7 |
+
|
8 |
+
Expected files:
|
9 |
+
- clip.plan
|
10 |
+
- t5.plan
|
11 |
+
- transformer.plan
|
12 |
+
- vae.plan
|
13 |
+
|
14 |
+
Build command:
|
15 |
+
python build_tensorrt_engines.py --single flux1-schnell blackwell bf16
|