[Kernel][UX] Add --linear-backend arg for linear kernel selection (#39538)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-05-15 17:07:39 -07:00
committed by GitHub
parent d0921bafef
commit b2a27b82d9
4 changed files with 266 additions and 34 deletions
+7 -1
View File
@@ -71,7 +71,7 @@ from vllm.config.cache import (
PrefixCachingHashAlgo,
)
from vllm.config.device import Device
from vllm.config.kernel import IrOpPriorityConfig, MoEBackend
from vllm.config.kernel import IrOpPriorityConfig, LinearBackend, MoEBackend
from vllm.config.lora import MaxLoRARanks
from vllm.config.mamba import MambaBackendEnum
from vllm.config.model import (
@@ -477,6 +477,7 @@ class EngineArgs:
enable_expert_parallel: bool = ParallelConfig.enable_expert_parallel
enable_ep_weight_filter: bool = ParallelConfig.enable_ep_weight_filter
moe_backend: MoEBackend = KernelConfig.moe_backend
linear_backend: LinearBackend = KernelConfig.linear_backend
all2all_backend: All2AllBackend = ParallelConfig.all2all_backend
enable_elastic_ep: bool = ParallelConfig.enable_elastic_ep
enable_dbo: bool = ParallelConfig.enable_dbo
@@ -1412,6 +1413,9 @@ class EngineArgs:
moe_backend_kwargs = kernel_kwargs["moe_backend"]
moe_backend_kwargs["type"] = lambda s: s.lower().replace("-", "_")
kernel_group.add_argument("--moe-backend", **moe_backend_kwargs)
linear_backend_kwargs = kernel_kwargs["linear_backend"]
linear_backend_kwargs["type"] = lambda s: s.lower().replace("-", "_")
kernel_group.add_argument("--linear-backend", **linear_backend_kwargs)
# vLLM arguments
vllm_kwargs = get_kwargs(VllmConfig)
@@ -2085,6 +2089,8 @@ class EngineArgs:
kernel_config.enable_flashinfer_autotune = self.enable_flashinfer_autotune
if self.moe_backend != "auto":
kernel_config.moe_backend = self.moe_backend
if self.linear_backend != "auto":
kernel_config.linear_backend = self.linear_backend
# Transfer top-level ir_op_priority into KernelConfig.ir_op_priority
for op_name, op_priority in asdict(self.ir_op_priority).items():