Implement get_dummy_lora_warmup_rank

[sstamenk]

vLLM v1's LoRAModelRunnerMixin.maybe_setup_dummy_loras calls self.lora_manager.get_dummy_lora_warmup_rank(...) during KV-cache profiling / cudagraph warmup. Unsloth replaces vLLM's LoRA manager with unsloth_zoo.vllm_lora_worker_manager.LRUCacheWorkerLoRAManager, which does not implement this method. This results in the following error:

AttributeError: 'LRUCacheWorkerLoRAManager' object has no attribute 'get_dummy_lora_warmup_rank'

get_dummy_lora_warmup_rank implementation picked up from #608

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[gemini-code-assist]

Code Review

This pull request adds the get_dummy_lora_warmup_rank method to vllm_lora_worker_manager.py to support vLLM v1's dummy LoRA warmup. The review feedback recommends defensively checking for the existence of _adapter_manager before accessing it to avoid potential AttributeError exceptions due to lazy initialization.

[gemini-code-assist]

Directly accessing self._adapter_manager can raise an AttributeError if get_dummy_lora_warmup_rank is called before create_lora_manager has initialized it (since it is lazily initialized and only annotated in __init__). It is safer to use getattr(self, '_adapter_manager', None) to defensively guard against this.

        manager = getattr(self, "_adapter_manager", None)
        manager_fn = getattr(
            manager, "get_dummy_lora_warmup_rank", None
        ) if manager is not None else None

[sstamenk]

@danielhanchen can you take a look at this? I've picked this up from #608 in hopes of getting it merged sooner since its a genuine bug and the original PR has been open for some time. Thanks

[sstamenk]

Created an issue to track this fix unslothai/unsloth#6114