pytorch / pytorch/executorch

[QNN ISSUE] QNN 16a8w Quantization Method for the Llama-3.2 Model.

Open
#6,684 5 comments 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

module: qnn triaged
Dominant language
Python
Stars
5k
Forks
1.2k
Avg merge
2d 10h
Merged PRs (30d)
581

Description

Dear @shewu-quic

We are continuously performing QNN quantization on the Llama-3.2-1B & 3B-Instruct models and testing them on the device.
We have confirmed that the model works correctly with qnn_16a16w, but we are not getting accurate results with qnn_8a8w.
Therefore, we believe that 16a8w might be a reasonable compromise.

Are you testing 16a8w internally? Or could we create a 16a8w config directly at the Python module stage?

Thank you

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reviewing the Python-module quantization configuration for the Llama-3.2-1B and 3B-Instruct models and reproduce the qnn_8a8w results on the device. Determine whether qnn_16a8w is supported or can be configured at that stage; done means answering that capability question with a tested configuration or a clear limitation.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
embedded-iot, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.