Skip to content

Conversation

@sayakpaul
Copy link
Member

Running into this issue when trying to apply quants:
https://pastebin.com/VeX2v7Wt

Command:

CKPT="black-forest-labs/FLUX.1-Kontext-dev"
IMAGE="yarn-art-pikachu.png"
PROMPT="Make Pikachu hold a sign that says 'Black Forest Labs is awesome', yarn art style, detailed, vibrant colors"
CACHE_DIR="CHANGEME"

python run_benchmark.py \
    --ckpt $CKPT --image $IMAGE --prompt "$PROMPT" \
    --compile_export_mode compile \
    --disable_quant \
    --disable_inductor_tuning_flags \
    --output-file bf16_compile_qkv_chan_fa3.png \
    --num_inference_steps 28 \
    --cache-dir $CACHE_DIR \
    > bf16_compile_qkv_chan_fa3.txt 2>&1

I am on this branch of diffusers: huggingface/diffusers#11818

@sayakpaul sayakpaul requested a review from jbschlosser June 27, 2025 08:10
@jbschlosser
Copy link
Collaborator

@sayakpaul which version of torchao are you using? I think you'll need this fix: pytorch/ao#2293

@sayakpaul
Copy link
Member Author

I am using the

pip install --pre torchao==0.12.0.dev20250609+cu126 --index-url https://download.pytorch.org/whl/nightly/cu126

@jbschlosser
Copy link
Collaborator

@sayakpaul I think that version is too old by a day.. IIRC the fix landed in the 06/10 nightly:

torchao==0.12.0.dev20250610+cu126 - note that we rely on a fix in the 06/10 nightly

@sayakpaul
Copy link
Member Author

Rerunning.

@sayakpaul sayakpaul marked this pull request as ready for review June 28, 2025 03:34
@sayakpaul
Copy link
Member Author

@jbschlosser ready to go! Please review.

Copy link
Collaborator

@jbschlosser jbschlosser left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

awesome! thanks for adding this :)

@jbschlosser jbschlosser merged commit 40abe4f into main Jun 30, 2025
@sayakpaul sayakpaul deleted the kontext-update branch July 4, 2025 01:30
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants