HF RL Explorer

DPO training fails with tensor parallelism enabled

DPO training fails with tensor parallelism enabled: a task in LegoFlow-SWE (Harbor dataset). Symptom When tensor parallel degree is set to a value greater than 1, DPO training immediately aborts with a NotImplementedError . The error message states that tensor parallelism is not supported because…

The task

**Symptom** When `tensor_parallel_degree` is set to a value greater than 1, DPO training immediately aborts with a `NotImplementedError`. The error message states that tensor parallelism is not supported because “D Tensor view ops are incompatible with torch.compile”. This makes it impossible to use multi‑GPU TP with…

Part of Lego-X/LegoFlow-SWE.