Update lora def

[lucylq]

Summary

Update lora def to use nn.linear instead of torch.nn.functional.linear

1. Remove debug string from FQN. Otherwise we get different quant indices for lora and non-lora, after this change, and the foundation weights are no longer shareable...
2. Having nn.Linear as a submodule allows torchao quant to capture it, and we don't need custom logic filtering LoraLinear. lora_a and lora_b are also captured, which changes our results slightly. Note that only lora_a is quantized (shape [16, dim], where dim %32 ==0), as lora_b has shape [dim, 16], where 16 %32 != 0 and fails the group size check.
3. Having nn.Linear as submodule also means we need to remap lora weights, as weight names have an extra 'linear' in them.
4. Add @Property for weight, bias so that it remains BC and can be treated as a regular linear module. This is used in [load_weights_from_attention.py](

   executorch/examples/models/llama/static_attention.py

   Line 1154 in eb92cec

   def load_weights_from_attention_mha(

Test plan

CI

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18247

• 📄 Preview Python docs built from this PR

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

⏳ No Failures, 251 Pending

As of commit 2e04f71 with merge base eb92cec ():
💚 Looks good so far! There are no failures yet. 💚

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

[github-actions]

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.

[Copilot]

Pull request overview

Updates the LLaMA example LoRA linear implementation to call an nn.Linear module in forward() (instead of torch.nn.functional.linear), presumably to align with module-based patterns.

Changes:

• Replace torch.nn.functional.linear(x, self.weight, self.bias) with self.linear(x).
• Introduce self.linear = nn.Linear(...) during LoRALinear initialization.

Comments suppressed due to low confidence (1)

examples/models/llama/lora.py:36

• linear and weight are now undefined after switching to self.linear; bias = linear.bias ... and register_parameter("weight", ...) will raise at init time. Either derive bias/weight from self.linear (or remove the extra register_parameter calls entirely) so construction works.

        self.linear = nn.Linear(in_dim, out_dim, bias=use_bias)
        bias = linear.bias if self.use_bias else None
        self.register_parameter("weight", nn.Parameter(weight))
        self.register_parameter(
            "bias", nn.Parameter(bias) if bias is not None else None
        )

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

Pull request overview

Updates the Llama LoRA implementation to call an nn.Linear submodule directly (instead of torch.nn.functional.linear) and simplifies quantization filtering accordingly.

Changes:

• Refactors LoRALinear to wrap a real nn.Linear module and use it in forward.
• Adds partial state-dict backward compatibility by remapping the legacy weight key.
• Simplifies the 8da*w quantization filter_fn to only consider nn.Linear modules.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
examples/models/llama/source_transformation/quantize.py	Simplifies the 8da4w/8da8w quantization filter to target nn.Linear modules only.
examples/models/llama/lora.py	Refactors LoRALinear to delegate to an internal nn.Linear and updates forward/state-dict behavior.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[Copilot]

Pull request overview

Updates the LLaMA LoRA implementation to wrap a real nn.Linear submodule (instead of calling torch.nn.functional.linear directly), enabling TorchAO quantization tooling to recognize and quantize the base linear layer consistently across LoRA and non-LoRA models.

Changes:

• Refactors LoRALinear to contain self.linear: nn.Linear, adds weight/bias properties for backward-compatible access, and remaps old checkpoint keys on load.
• Simplifies TorchAO 8da*xw quantization filtering to target nn.Linear modules only (no special-casing LoRA modules).
• Makes XNNPACK constant serialization keys content-hash-based (removes tensor-name prefix) and updates the LoRA CI expectation string accordingly.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
examples/models/llama/source_transformation/quantize.py	Simplifies TorchAO quantization filter to match nn.Linear modules and apply group-size compatibility logic.
examples/models/llama/lora.py	Refactors LoRALinear to wrap an nn.Linear submodule; adds BC weight/bias accessors and state-dict key remapping.
backends/xnnpack/operators/node_visitor.py	Changes named constant key generation to be solely SHA256-based to stabilize dedup/indexing behavior.
.ci/scripts/test_lora.sh	Updates expected output prefix text for the quantized LoRA test case.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.