Performance report - 2023 M2 Max 96GB

[explorigin]

(Using my CLI options here )

time python main.py --prompt "Luxury food photograph" --steps 2 --seed 2 3.94s user 10.78s system 56% cpu 26.109 total

time python main.py --prompt "detailed cinematic dof render of an old dusty detailed CRT monitor on a wooden desk in a dim room with items around, messy dirty room. On the screen are the letters "FLUX" glowing softly. High detail hard surface render" --steps 2 --seed 2 3.94s user 11.75s system 61% cpu 25.565 total

So about 25-26 seconds.

[filipstrand]

@explorigin Thanks for the report (and nice picture with the food for --seed 2 :) ). I have looked at your PR and will include this performance result and some other numbers in the README once this PR is merged.

[7enChan]

(Using my CLI options here )

time python main.py --prompt "Luxury food photograph" --steps 2 --seed 2 3.94s user 10.78s system 56% cpu 26.109 total

time python main.py --prompt "detailed cinematic dof render of an old dusty detailed CRT monitor on a wooden desk in a dim room with items around, messy dirty room. On the screen are the letters "FLUX" glowing softly. High detail hard surface render" --steps 2 --seed 2 3.94s user 11.75s system 61% cpu 25.565 total

So about 25-26 seconds.

On the same machine (M2 Max 96GB MBP), the same prompt takes about 8 seconds longer to complete. I'm not sure what the reason is. I'm running this within a conda virtual environment with python 3.12.4

time python main.py --prompt "Luxury food photograph" --steps 2 --seed 2
Fetching 8 files: 100%|███████████████████████| 8/8 [00:00<00:00, 124737.67it/s]
Fetching 7 files: 100%|█████████████████████████| 7/7 [00:00<00:00, 9670.66it/s]
100%|█████████████████████████████████████████████| 2/2 [00:27<00:00, 13.83s/it]
python main.py --prompt "Luxury food photograph" --steps 2 --seed 2 4.19s user 12.84s system 49% cpu 34.180 total

time python main.py --prompt "detailed cinematic dof render of an old dusty detailed CRT monitor on a wooden desk in a dim room with items around, messy dirty room. On the screen are the letters "FLUX" glowing softly. High detail hard surface render" --steps 2 --seed 2
Fetching 8 files: 100%|███████████████████████| 8/8 [00:00<00:00, 127100.12it/s]
Fetching 7 files: 100%|████████████████████████| 7/7 [00:00<00:00, 74518.09it/s]
100%|█████████████████████████████████████████████| 2/2 [00:28<00:00, 14.00s/it]
python main.py --prompt --steps 2 --seed 2 4.26s user 13.21s system 49% cpu 35.015 total

[explorigin]

There could be a loading time difference. Let's compare a few more things:

• virtualenv environment, Python 3.12.1
• Sonoma 14.6.1
• 4TB harddrive (APFS Encrypted)

[7enChan]

There could be a loading time difference. Let's compare a few more things:

• virtualenv environment, Python 3.12.1
• Sonoma 14.6.1
• 4TB harddrive (APFS Encrypted)

Thanks for the reply. Here is mine:

• virtualenv environment, Python 3.12.4
• Sonoma 14.6.1
• 1TB harddrive (APFS Not encrypted)

[7enChan]

There could be a loading time difference. Let's compare a few more things:

• virtualenv environment, Python 3.12.1
• Sonoma 14.6.1
• 4TB harddrive (APFS Encrypted)

Thanks for the reply. Here is mine:

• virtualenv environment, Python 3.12.4
• Sonoma 14.6.1
• 1TB harddrive (APFS Not encrypted)

Forgot one thing, I'm using the latest version of mflux

[explorigin]

Encryption also includes compression. For loading large models into memory, that might be the difference if the drive is the bottleneck. I'm speculating.

[7enChan]

Encryption also includes compression. For loading large models into memory, that might be the difference if the drive is the bottleneck. I'm speculating.

Interesting. If the drive is the bottleneck, then the time difference for FLUX.1 dev, which takes 20 steps, should be less.

[7enChan]

Encryption also includes compression. For loading large models into memory, that might be the difference if the drive is the bottleneck. I'm speculating.

Interesting. If the drive is the bottleneck, then the time difference for FLUX.1 dev, which takes 20 steps, should be less.

I tried turning on encryption and it didn't take significantly less time, it was still around 34s 😂

[explorigin]

I'm speculating

That leaves drive size as the remaining differentiator.

[7enChan]

I'm speculating

That leaves drive size as the remaining differentiator.

Yeah, maybe the 4T harddrive doesn't have the same read/write speeds as 1T...