Skip to content

Qualcomm AI Engine Direct - Enable Lookahead Decoding #11437

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged

Conversation

shewu-quic
Copy link
Collaborator

@shewu-quic shewu-quic commented Jun 6, 2025

Summary:

  • Add new eval_mode: lookahead
  • Add three arguments: ngram, window, gcap
  • Add lhd_token_generator

Command

python3 examples/qualcomm/oss_scripts/llama/llama.py -b build-android --checkpoint stories110M.pt --params params.json --tokenizer_model tokenizer.model --prompt "Once" --temperature 0 --tokenizer_bin tokenizer.bin --llama_model stories110m --model_mode lookahead --ptq 16a4w -m SM8650 -H ${host} -s ${device}  -a ${artifacts} --max_seq_len 4096 --kv_updater smart_mask  --prefill_ar_len 64 --ngram 3 --window 2 --gcap 2

Test Results

QNN SDK: 2.28
Device: SM8650
max_seq_len: 4096

Performance Improvement under different AR-N and different W/G/N

Llama 3.2 3B
image

Llama 3.2 1B
image

Story Llama 110M
image

Performance Improvement under different prompt

image

Reference

cc: @haowhsu-quic

Copy link

pytorch-bot bot commented Jun 6, 2025

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/11437

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure

As of commit 6bd4645 with merge base 56392aa (image):

NEW FAILURE - The following job has failed:

  • Lint / lintrunner / linux-job (gh)
    RuntimeError: Command docker exec -t e368a39b51fa7fd6f04d143c70d9df48f9a096e6c9c989f1abce042e72715bc5 /exec failed with exit code 127

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Jun 6, 2025
Copy link

github-actions bot commented Jun 6, 2025

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

@cccclai
Copy link
Contributor

cccclai commented Jun 12, 2025

Hey can you add a read me to explain how this specific lookahead decoding work? Also, does 0/0/0 mean no lookahead decoding?

Copy link
Contributor

@cccclai cccclai left a comment

Choose a reason for hiding this comment

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

It looks great!

summary:
- Add new eval_mode: lookahead
- Add three arguments: ngram, window, gcap
- Add lhd_token_generator
@shewu-quic shewu-quic force-pushed the dev1/hutton/enable_lookahead_decoding branch from 1bc9dcb to 12e1ae5 Compare June 16, 2025 02:56
@shewu-quic
Copy link
Collaborator Author

Hey can you add a read me to explain how this specific lookahead decoding work? Also, does 0/0/0 mean no lookahead decoding?

Sure. Let me work on it.
Yes, 0/0/0 means AR-N decoding without lookahead.
I manually change ar_len of decoding model in hybrid mode.

@facebook-github-bot
Copy link
Contributor

@cccclai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

@cccclai cccclai merged commit 16ffc96 into pytorch:main Jun 16, 2025
101 of 104 checks passed
abhinaykukkadapu pushed a commit to abhinaykukkadapu/executorch that referenced this pull request Jun 17, 2025
## Summary:
- Add new eval_mode: lookahead
- Add three arguments: ngram, window, gcap
- Add lhd_token_generator

## Command
```
python3 examples/qualcomm/oss_scripts/llama/llama.py -b build-android --checkpoint stories110M.pt --params params.json --tokenizer_model tokenizer.model --prompt "Once" --temperature 0 --tokenizer_bin tokenizer.bin --llama_model stories110m --model_mode lookahead --ptq 16a4w -m SM8650 -H ${host} -s ${device}  -a ${artifacts} --max_seq_len 4096 --kv_updater smart_mask  --prefill_ar_len 64 --ngram 3 --window 2 --gcap 2
```

## Test Results
QNN SDK: 2.28
Device: SM8650
max_seq_len: 4096

### Performance Improvement under different AR-N and different W/G/N
Llama 3.2 3B

![image](https://github.com/user-attachments/assets/98365f9f-ccb9-49b0-a4ab-e51b9880efc3)

Llama 3.2 1B

![image](https://github.com/user-attachments/assets/c21aba0a-ab2d-4f30-9fbe-cce439bd5f7e)

Story Llama 110M

![image](https://github.com/user-attachments/assets/debdf888-5ece-400f-b3ae-892c06ef352a)

### Performance Improvement under different prompt

![image](https://github.com/user-attachments/assets/cd072fa5-1eda-4390-9748-882baab442e0)

## Reference
- https://lmsys.org/blog/2023-11-21-lookahead-decoding/
- https://github.com/hao-ai-lab/LookaheadDecoding/tree/main/lade
-
https://github.com/ggml-org/llama.cpp/blob/master/examples/lookahead/lookahead.cpp

cc: @haowhsu-quic
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants