[Feature Request] mlp linear to conv,not use up_proj and gate_proj

[cuiyu1227]

qai_hub_models/models/_shared/qwen3/model_adaptations.py
class QCQwen3MLP(Qwen3MLP):
def prepare_conv(self) -> None:
# TODO (https://github.com/qcom-ai-hub/tetracode/issues/17113)
# Temporarily commented out due to AISW-148745.
# self.up_proj = ConvInplaceLinear(self.up_proj) # type: ignore[has-type, arg-type, unused-ignore]
self.down_proj = ConvInplaceLinear(self.down_proj) # type: ignore[has-type, arg-type, unused-ignore]
# self.gate_proj = ConvInplaceLinear(self.gate_proj) # type: ignore[has-type, arg-type, unused-ignore]

hi,why not use self.up_proj = ConvInplaceLinear and self.gate_proj = ConvInplaceLinear.https://github.com/qcom-ai-hub/tetracode/issues/17113 could not get in.could you tell me the reason?

[gustavla]

Great question! This actually doesn't matter for all the devices that we currently support. It does matter however for QCS6490. We added these ConvInplaceLinear in an effort to support QCS6490, but have not developed a model we feel good about yet (due to precision constraints in that particular chipset). However, due to a compiler bug in QAIRT SDK that affects only up_proj/gate_proj, we had to disable this. This means we can't support QCS6490, but we didn't have a model ready for that anyway.

So, long story short, it looks like the QAIRT bug has been resolved, so we could uncomment these (although we would have to re-quantize the model, since encodings names will changes). This won't actually make any difference on the devices we do support, but it would be a necessary condition for trying QCS6490 again.

[qaihm-bot]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.