Repository navigation
Are there plans to add multi-LoRA support to the Qualcomm backend #16999
Description
🚀 The feature, motivation and pitch
Are there plans to add multi-LoRA support to the Qualcomm backend and support runtime LoRA switching to meet the needs of multi-LoRA scenarios?
The Feature
Add support for multiple LoRA adapters on the Qualcomm backend, including the ability to dynamically switch LoRA adapters at runtime without reloading or recompiling the base model.
Specifically, this feature would enable:
Loading and managing multiple LoRA adapters simultaneously
Selecting or switching the active LoRA adapter during inference
Keeping the base model static while applying different LoRA weights on demand
This is particularly useful for scenarios where a single base model serves multiple tasks, domains, or user profiles.
Alternatives
No response
Additional context
No response
RFC (Optional)
No response
cc @cccclai @winskuo-quic @shewu-quic @haowhsu-quic @DannyYuyang-quic @cbilgin
Activity
cccclai commentedon Jan 29, 2026
Yeah we've been discussing it for a while, there are a few options currently
32kilometre commentedon Mar 27, 2026
Hi, just following up on this issue.
I recall there were previous discussions and several candidate solutions for this feature. I'm wondering if there's been any further progress or a preferred direction at this point?
If the approach hasn't been finalized yet, I'd also appreciate learning more about the candidate solutions that were mentioned earlier. I'm happy to help with testing or provide feedback based on use cases if that would be useful.
Thanks!
lucylq commentedon Apr 2, 2026
Hi @cxn-selfie thanks for following up. We are in discussions with Qualcomm about this and I hope we will have a solution and can get back to you soon. Qualcomm currently has an internal way of enabling multiple loras, and there's some work to enable it with the executorch APIs.
cc @abhinaykukkadapu
abhinaykukkadapu commentedon Apr 2, 2026
@cxn-selfie thanks for being proactive for Qualcomm LoRA, we are currently discussing the LoRA deliverables from QCOM and i think we will have something in the next release, will keep you updated. Are you trying to enable any specific use case or device?
qti-horodnic commentedon Jul 22, 2026
@32kilometre
Sorry for the late reply. To answer your question, yes, multi-LoRA support for the Qualcomm backend is actively in development. We're implementing two tiers: one would allow runtime adapter switching by supplying LoRA weights as graph inputs and another which would use QNN's patch binary mechanism to support multiple adapters with different ranks and fast switching with lower runtime overhead. In both cases the base model is compiled once and remains static; adapters are packed separately and selected at inference time.
We expect these features to likely be available by the end of the year.
@qti-mmadhava
32kilometre commentedon Nov 3, 2026
@qti-horodnic Really glad to hear this! @abhinaykukkadapu Our use case is an on-device assistant that uses a single base model with task-specific LoRA adapters, switching between them for tasks like summarization and translation.
Thanks to the team for all the work on this!