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, 23:00 UTC
[Model Runner V2] Support FLASHINFER_MLA backend (#32709)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Woosuk Kwon committed
Jan 20, 2026 at 10:26 UTC
9ab4388cd3dc6bb2069125c1a84feaa1aa193e0b
2 files changed
+2
-2
vllm/v1/worker/gpu/attn_utils.py
+1
-1
@@ -57,7 +57,7 @@ def init_attn_backend(
57
)
58
attn_metadata_builders.append(attn_metadata_builder) # type: ignore
59
60
- if "FLASHINFER" in attn_backend.get_name():
60
+ if attn_backend.get_name() == "FLASHINFER":
61
if flashinfer_workspace is None:
62
flashinfer_workspace = attn_metadata_builder._get_workspace_buffer()
63
else:
vllm/v1/worker/gpu/model_runner.py
+1
-1
@@ -248,7 +248,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
248
)
249
250
# TODO(woosuk): Support other backends.
251
- supported_backends = ("FLASH_ATTN", "FLASHINFER")
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: