[Kernel] GLM5 Router GEMM (#46385)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-06-23 22:54:50 -07:00
committed by GitHub
parent 4c5bc41ba6
commit 9d6fdc2901
4 changed files with 158 additions and 32 deletions
@@ -14,7 +14,7 @@ from vllm.utils.torch_utils import direct_register_custom_op
class GateLinear(ReplicatedLinear):
"""MoE gate linear layer with multi-tier GEMM dispatch:
1. DSV3 specialized kernel (SM90+, fp32 out, M<=16, H=7168, E=256/384)
1. DSV3 specialized kernel (SM90+, M<=16, H=7168 E=256/384, H=6144 E=256)
2. fp32 specialized kernel (SM90+, bf16/fp32 in, fp32 out,
M<=32, H=3072, E=256)
3. cuBLAS bf16×bf16→fp32 (SM90+ + bf16 weight + fp32 out_dtype)
@@ -25,9 +25,14 @@ class GateLinear(ReplicatedLinear):
method which is only known later).
"""
# Dimensions supported by the DSV3 specialized kernel
# Dimensions supported by the DSV3 specialized kernel.
# Valid (hidden_size, num_experts) combinations:
# (7168, 256) -> DeepSeek-V3, (7168, 384) -> Kimi-K2,
# (6144, 256) -> GLM-5
DSV3_SUPPORTED_NUM_EXPERTS = [256, 384]
DSV3_SUPPORTED_HIDDEN_SIZES = [7168]
DSV3_SUPPORTED_HIDDEN_SIZES = [7168, 6144]
# num_experts=384 is only instantiated for hidden_size=7168.
DSV3_UNSUPPORTED_SHAPES = {(6144, 384)}
# (hidden_size, num_experts) pairs with an instantiated fp32 kernel:
# (3072, 256) -> MiniMax-M2/M2.5, (6144, 128) -> MiniMax-M3
@@ -71,6 +76,7 @@ class GateLinear(ReplicatedLinear):
self.allow_specialized_router_gemm
and output_size in self.DSV3_SUPPORTED_NUM_EXPERTS
and input_size in self.DSV3_SUPPORTED_HIDDEN_SIZES
and (input_size, output_size) not in self.DSV3_UNSUPPORTED_SHAPES
)
# See https://github.com/vllm-project/vllm/pull/44217
# for more details.