organization · inferred from evidence
tgaddair
A maintainer or contributor identified in the text as spotting the problem.
- Lorax server hangs on warming up the server when total_tokens is larger than 4096 with LLama2-7b-hf on a l4 GPU
A practical guide for Predibase. [tgaddair (contributor)]: The server hanging during warmup was a real hang, not just slow warmup. Fixed in PR #352; the reporter confirmed the server starts up without issues after the fix.
- LoRAX: No local weights error when loading an adapter from a local directory
The `No local weights` error almost always means the adapter path you passed is not visible inside the LoRAX container. The path in `adapter_id` is resolved by the server, so mount your adapter directory with `-v` and use the container-side
- LoRAX fails to start with microsoft/Phi-3-mini-128k-instruct
If LoRAX refuses to start with Phi-3-mini-128k-instruct and the logs mention RoPE scaling or model type detection, you need the fix from PR #499. Pull the latest main Docker image, which the reporter confirmed starts the model successfully.
- LoRAX: TensorParallelHead object has no attribute base_layer when serving Mixtral with an adapter
This crash is a LoRAX bug in adapter loading for Mixtral's tensor-parallel head, fixed in PR #131. Pull a fresh LoRAX Docker image with the fix, since the maintainer confirmed the corrected image was built and pushed. If you pinned an image