A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 20:45 UTC
[Model Runner V2] Do not error on attention backends (#32820)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Woosuk Kwon committed
Jan 21, 2026 at 17:02 UTC
5e00b561cddd2cecf8be7a341cb49d446613d6ef
1 file changed
-10
vllm/v1/worker/gpu/model_runner.py
-10
@@ -247,16 +247,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
247
self.block_tables,
248
)
249
250
- # TODO(woosuk): Support other backends.
251
- supported_backends = ("FLASH_ATTN", "FLASHINFER", "FLASHINFER_MLA")
252
- for backend in self.attn_backends.values():
253
- backend_name = backend.get_name()
254
- if backend_name not in supported_backends:
255
- raise NotImplementedError(
256
- f"The {backend_name} attention backend is not supported yet. "
257
- f"Supported backends are: {supported_backends}."
258
- )
259
-
250
self.kv_caches: list[torch.Tensor] = []
251
init_kv_cache(
252
self.kv_caches,