Commit
·
bdf45c3
1
Parent(s):
34b0078
Feat (script): Added initial reproduction README.
Browse files- quant_sdxl/README.md +33 -0
quant_sdxl/README.md
ADDED
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Quantizing SDXL for MLPerf
|
2 |
+
|
3 |
+
The purpose of this document is to allow other users to quantize SDXL in various different ways, corresponding to AMD's 2024-07 MLPerf submission.
|
4 |
+
|
5 |
+
## Environmental Setup
|
6 |
+
|
7 |
+
The environment can be set up as follows:
|
8 |
+
|
9 |
+
```bash
|
10 |
+
conda env create -n quant_sdxl -f env/brv_sdxl_mlperf_minimal.yml # Use env/brv_sdxl_mlperf_rocm_minimal.yml is using a ROCm compatible GPU
|
11 |
+
conda activate quant_sdxl
|
12 |
+
```
|
13 |
+
|
14 |
+
## Fetching and Pre-Processing Datasets
|
15 |
+
|
16 |
+
Information on fetching and pre-processing the data can be found [here](https://github.com/mlcommons/inference/tree/master/text_to_image).
|
17 |
+
Note, the latents generated here will be used in the next step.
|
18 |
+
|
19 |
+
## Quantize Int8 / FP16 Model
|
20 |
+
|
21 |
+
```bash
|
22 |
+
python quant_sdxl.py --model stabilityai/stable-diffusion-xl-base-1.0 --device <device> --calibration-prompt-path ./captions.tsv --checkpoint-name unet.ckpt --path-to-latents <path/to/latents/latents.pt> --guidance-scale 7.5 --exclude-blacklist-act-eq [--path-to-coco <path/to/coco> --validation-prompts 5000]
|
23 |
+
```
|
24 |
+
|
25 |
+
Note, you can optionally validate on the MLPerf validation set, but be aware this will take ~48hrs.
|
26 |
+
|
27 |
+
## Quantize Int8 / FP8 Model
|
28 |
+
|
29 |
+
The Int8 / FP8 model can be quantized as follows:
|
30 |
+
|
31 |
+
```bash
|
32 |
+
python quant_sdxl.py --model stabilityai/stable-diffusion-xl-base-1.0 --device <device> --calibration-prompt-path ./captions.tsv --checkpoint-name unet.ckpt --path-to-latents <path/to/latents/latents.pt> --guidance-scale 7.5 --quantize-sdp --exclude-blacklist-act-eq [--path-to-coco <path/to/coco> --validation-prompts 5000]
|
33 |
+
```
|