pytorch / pytorch/executorch

[qnn] about the new arch chipset and larger models

Open
#18,964 2 comments 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

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

Description

Hi Teams,

forgive me about "too noisy" with so many issues. For I am so excited with the activate community with edge deployment with qualcomm chips.

I found some code about 8397/8797 with multiple NPUs with 320 tops (320 tops int8 dense). So we can image that larger models used in the chipset.

So I have some questions about it:

  1. To speed up the prefill, can we use two/or more npu in the prefill stage, like tensor parallelism on GPU? If it can be achieved, how to?

  2. about the MOE models (low computation, lower bandwith, powerful), can we use moe model with executorch/qnn sdk?

  3. lora/multi lora support

  4. SSD for larger models

Looking forward to you insights.
cc @cccclai @cbilgin @abhinaykukkadapu @winskuo-quic @shewu-quic @haowhsu-quic @DannyYuyang-quic

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

No files or tests are named. Start by reviewing ExecuTorch’s QNN SDK integration and its documented support for multi-NPU prefill, MoE models, LoRA, and SSD-backed model storage; the issue needs to be narrowed into a scoped capability or implementation task before work can be considered done.

Written by the indexing model from the issue text.

Assessment

Tech stack
machine-learning
Domain
embedded-iot, machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.