**Description:**
- Add the `lora_request` parameter to the VLLM class to support LoRA
model configurations. This enhancement allows users to specify LoRA
requests directly when using VLLM, enabling more flexible and efficient
model customization.
**Issue:**
- No existing issue for `lora_adapter` in VLLM. This PR addresses the
need for configuring LoRA requests within the VLLM framework.
- Reference : [Using LoRA Adapters in
vLLM](https://docs.vllm.ai/en/stable/models/lora.html#using-lora-adapters)
**Example Code :**
Before this change, the `lora_request` parameter was not applied
correctly:
```python
ADAPTER_PATH = "/path/of/lora_adapter"
llm = VLLM(model="Bllossom/llama-3.2-Korean-Bllossom-3B",
max_new_tokens=512,
top_k=2,
top_p=0.90,
temperature=0.1,
vllm_kwargs={
"gpu_memory_utilization":0.5,
"enable_lora":True,
"max_model_len":1024,
}
)
print(llm.invoke(
["...prompt_content..."],
lora_request=LoRARequest("lora_adapter", 1, ADAPTER_PATH)
))
```
**Before Change Output:**
```bash
response was not applied lora_request
```
So, I attempted to apply the lora_adapter to
langchain_community.llms.vllm.VLLM.
**current output:**
```bash
response applied lora_request
```
**Dependencies:**
- None
**Lint and test:**
- All tests and lint checks have passed.
---------
Co-authored-by: Um Changyong <changyong.um@sfa.co.kr>
**Description:**
- This pull request addresses a bug in Langchain's VLLM integration,
where the use_beam_search parameter was erroneously passed to
SamplingParams. The SamplingParams class in vLLM does not support the
use_beam_search argument, which caused a TypeError.
- This PR introduces logic to filter out unsupported parameters,
ensuring that only valid parameters are passed to SamplingParams. As a
result, the integration now functions as expected without errors.
- The bug was reproduced by running the code sample from Langchain’s
documentation, which triggered the error due to the invalid parameter.
This fix resolves that error by implementing proper parameter filtering.
**VLLM Sampling Params Class:**
https://github.com/vllm-project/vllm/blob/main/vllm/sampling_params.py
**Issue:**
I could not found an Issue that belongs to this. Fixes "TypeError:
Unexpected keyword argument 'use_beam_search'" error when using VLLM
from Langchain.
**Dependencies:**
None.
**Tests and Documentation**:
Tests:
No new functionality was added, but I tested the changes by running
multiple prompts through the VLLM integration with various parameter
configurations. All tests passed successfully without breaking
compatibility.
Docs
No documentation changes were necessary as this is a bug fix.
**Reproducing the Error:**
https://python.langchain.com/docs/integrations/llms/vllm/
The code sample from the original documentation can be used to reproduce
the error I got.
from langchain_community.llms import VLLM
llm = VLLM(
model="mosaicml/mpt-7b",
trust_remote_code=True, # mandatory for hf models
max_new_tokens=128,
top_k=10,
top_p=0.95,
temperature=0.8,
)
print(llm.invoke("What is the capital of France ?"))
This PR resolves the issue by ensuring that only valid parameters are
passed to SamplingParams.
