[Bug] model.fast_generate() with lora_request fails with TypeError

[jqug]

I'm getting this error when calling model.fast_generate() and specifying a LoRA adapter path via the lora_request parameter:

File /venv/main/lib/python3.12/site-packages/unsloth_zoo/vllm_lora_worker_manager.py:147, in WorkerLoRAManager._load_adapter(self, lora_request)
    144         kwargs["embedding_modules"] = self.embedding_modules
    145         kwargs["embedding_padding_modules"] = self.embedding_padding_modules
--> 147     lora = load_method(**kwargs)
    149 except FileNotFoundError as e:
    150     # FileNotFoundError should be raised if both
    151     # - No adapter found to download from huggingface (or in
    152     #       offline mode)
    153     # - No local adapter files found at `lora_request.lora_path`
    154     # For NotFoundError
    155     raise ValueError(
    156         f"Loading lora {lora_request.lora_name} failed: No adapter "
    157         f"found for {lora_path}") from e

TypeError: LoRAModel.from_local_checkpoint() got an unexpected keyword argument 'lora_path'

This seems to break e.g. DeepSeek_R1_0528_Qwen3_(8B)_GRPO.ipynb when I run locally on H100 with latest unsloth and vllm. Here's a minimal repro:

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = 'Qwen/Qwen3-0.6B', # can be anything
    max_seq_length = 512,
    load_in_4bit = False,
    fast_inference = True, 
    max_lora_rank = 8,
    gpu_memory_utilization = 0.6, 
)

model = FastLanguageModel.get_peft_model(
    model,
    r = 8,
    target_modules = [
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    lora_alpha = 16
)

# Save the empty adapter as a dummy
model.save_lora('saved_lora_adapter')

# Fails with TypeError
outputs = model.fast_generate(
    ['dummy prompt'],
    lora_request=model.load_lora('saved_lora_adapter'),
)

I think the culprit is this change from 3 days ago. The code now has:

 kwargs["lora_path"] = lora_path
# [...]
lora = load_method(**kwargs)

But vllm.lora.LoraModel.from_local_checkpoint() (source) expects lora_dir, not lora_path.

Otherwise thanks unsloth team for the amazing work 🤩

[jqug]

For anyone else getting this, rolling back with pip install unsloth==2025.11.4 unsloth-zoo==2025.11.5 worked for me.

[Datta0]

Hey @jqug thanks for noticing the issue.
I will fix it up asap :)

[jqug]

Great, thanks @Datta0