Skip to content

fix: get_attn_backend_cls() missing **kwargs — breaks vLLM 0.16+ #23

Description

@dmvevents

Bug

`NeuronPlatform.get_attn_backend_cls()` in `vllm_neuron/platform.py` has a fixed parameter list that does not include `**kwargs`. vLLM 0.16 added new keyword arguments to this method, causing `TypeError` during attention backend initialization.

Environment

  • Instance: trn1.32xlarge
  • Neuron SDK: 2.28.0
  • vllm-neuron: 0.5.0 (release-0.5.0 branch)
  • vLLM: 0.16.0

Error

```
TypeError: NeuronPlatform.get_attn_backend_cls() got an unexpected keyword argument '...'
```

Steps to Reproduce

Install vLLM 0.16.0 with vllm-neuron 0.5.0 and start any model. The crash occurs during attention backend initialization.

Suggested Fix

Add `**kwargs` to the method signature for forward compatibility:

```python
@classmethod
def get_attn_backend_cls(
cls,
selected_backend,
head_size,
dtype,
kv_cache_dtype,
block_size,
is_attention_free,
is_blocksparse,
use_mla,
use_sparse=False,
**kwargs,
) -> str:
```

This is the standard forward-compatibility pattern used by other vLLM platform backends (TPU, XPU) when the base class gains new parameters.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions