Adam Belfki
Adam Belfki
To improve the user experience, disabling remote logging shouldn't affect LOG responses received as a result of explicit log calls defined by the user in the intervention graph.
Handle value injection recursively for Collection data defined outside of tracing. e.g. ```py arr = list() with tiny_model.trace(tiny_input): arr.append(tiny_model.output.save()) assert isinstance(arr[0], torch.Tensor) ```
Possible fixes: - Leverage `setup.py` to only install required dependencies dynamically. - Create optional dependencies stages, e.g. `nnsight[vllm]`, `nnsight[remote]`, `nnsight[dev]`...etc
Check if the frame of the data type creation is not within an nnsight trace.