Use FQNs instead of hash for tensor/data names
Open
@lucylq is already working on this.
Since Nov 4, 2025.
- Dominant language
- Python
- Stars
- 5k
- Forks
- 1.2k
- Avg merge
- 2d 10h
- Merged PRs (30d)
- 581
Description
Currently XNNPACK and Vulkan hash tensors/data blobs and use that as the key for program-data separation.
This makes it difficult to perform weight-only updates, as the hash changes. Instead, use the FQN (fully qualified name) of the tensor as the key.
The FQN currently causes issues in a few areas:
- Quantized weights have the name overwritten by quant_param_{n}, or frozen_param_{n}, so we lose the FQN information
- XNNPACK weight cache is a singleton, so names may clash for different models.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.