[Bugfix] Fix InternVL/InternVL3 LoRA loading TypeError in adapter fallback

[waynehacking8]

Summary

Fixes #4105. Serving OpenGVLab/InternVL3-14B (or any InternVL) with a LoRA adapter on the PyTorch backend crashes with:

TypeError: load_lora_weights() missing 1 required positional argument: 'adapter_id'

Root cause

Both InternVLChatModel.load_lora_weights (internvl.py) and InternVLForConditionalGeneration.load_lora_weights (internvl3_hf.py) fall back to the standalone lmdeploy.pytorch.adapter.adapter.load_lora_weights when the inner language model does not define load_lora_weights. Its signature is:

def load_lora_weights(model: nn.Module, weights, adapter_id): ...

but both call it as load_lora_weights(weights, adapter_id) — weights binds to model, adapter_id binds to weights, and adapter_id is unbound → TypeError.

internvl3_hf.py has a second bug that makes the broken branch fire unconditionally: it gates on hasattr(self.model.language_model, 'load_lora_weights'), but the constructor stores self.language_model (there is no self.model). hasattr swallows the AttributeError and always takes the else branch.

Fix

Pass the top-level model (self) to the standalone loader, and use self.language_model for the hasattr/delegate path in internvl3_hf.py.

self is correct because the loader does dict(model.named_parameters()) and maps de-prefixed adapter keys like language_model.model.layers.0.mlp.down_proj.lora_adapters.down_proj.lora_A — i.e. names that start with language_model.. Only the top-level VLM (which has language_model as a child) produces that namespace. This mirrors the canonical caller internlm2.py:

load_lora_weights(self, weights_iter, adapter_id)

Relation to #4106

The earlier (closed) #4106 passed self.language_model (internvl.py) / self.model.language_model (internvl3_hf.py). self.language_model.named_parameters() yields model.layers... (no language_model. prefix), so the loader's params_dict[param_name] lookup misses — which is the KeyError/param-name mismatch the reporter hit after trying #4106. Passing self gives the loader the namespace it actually maps against, and also fixes the non-existent self.model reference.

Test

Added tests/pytorch/test_internvl_lora.py (GPU-free; mocks the standalone loader):

pytest tests/pytorch/test_internvl_lora.py

• WITH fix: 3 passed — both classes pass (model, weights, adapter_id) to the loader, and internvl3_hf delegates correctly when the language model implements load_lora_weights.
• WITHOUT fix: 3 failed (loader called with the wrong (weights, adapter_id) arity; delegation never happens).

ruff check clean on all three files.

Honesty / scope

The TypeError and self.model bugs are verified by code analysis + the GPU-free unit tests above. I do not have the InternVL3-14B + LoRA serving setup to run the full end-to-end here; the parameter-namespace analysis indicates passing self also resolves the downstream KeyError from #4106, but an e2e confirmation by someone with that model/adapter would be welcome. cc @windreamer (you triaged the original issue).

---

AI-assisted (Claude); every line human-reviewed.