What this does

Fixes #3056

The LoRA adapter class claimed to support nn.Conv layers but assumed nn.Linear semantics, crashing with RuntimeError: mat1 and mat2 shapes cannot be multiplied on any convolutional layer (e.g. the wav2vec2-style repro in the issue).

This PR implements real convolutional support instead of removing the claim, following the approach used by HF peft:

• Down projection: a conv of the same type mirroring the pretrained layer's geometry (kernel_size, stride, padding, dilation, padding_mode), mapping onto rank channels — so both branches produce outputs of identical shape
• Up projection: a pointwise (1×1) conv from rank to out_channels, zero-initialized (standard LoRA init)
• Grouped convolutions (groups != 1): clear ValueError instead of a cryptic crash
• No behavior change for nn.Linear (and other weight-matrix modules); checkpoint attribute names (adapter_down_proj/adapter_up_proj) are preserved so existing checkpoints keep loading

Testing

New tests/unittests/test_adapters.py with 9 tests verifying the core LoRA properties:

• Identity at init (zero up-projection ⇒ adapted output == pretrained output), for Linear, Conv1d, Conv2d and Conv3d
• Exact match with the manual formula base(x) + up(down(x)) * alpha/rank
• Pretrained weights frozen (no grads) while adapter weights receive gradients and actually learn (one SGD step changes the output, pretrained output unchanged)
• Shape equality across geometries (stride=5, padding='same', dilation=2)
• groups != 1 raises ValueError
• The exact AdaptedModel(all_conv=True) repro from the issue, forward + backward, with only adapter params trainable

Also added a Conv1d doctest to the LoRA docstring. All 9 tests + 3 doctests pass; ruff check, ruff format and codespell pass on the changed files.

Co-written by a human (@oliver0006) and Claude AI (Fable 5) working together.

🤖 Generated with Claude Code