[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
@@ -286,3 +286,52 @@ template void invokeRouterGemmBf16Output<__nv_bfloat16, 15, 384, 7168>(
template void invokeRouterGemmBf16Output<__nv_bfloat16, 16, 384, 7168>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
// Template instantiations for GLM-5 (DEFAULT_NUM_EXPERTS, hidden_dim=6144)
template void invokeRouterGemmBf16Output<__nv_bfloat16, 1, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 2, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 3, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 4, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 5, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 6, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 7, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 8, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 9, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 10, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 11, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 12, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 13, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 14, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 15, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmBf16Output<__nv_bfloat16, 16, 256, 6144>(
__nv_bfloat16*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
@@ -41,6 +41,7 @@ inline int getSMVersion() {
static constexpr int DEFAULT_NUM_EXPERTS = 256;
static constexpr int KIMI_K2_NUM_EXPERTS = 384;
static constexpr int DEFAULT_HIDDEN_DIM = 7168;
static constexpr int GLM_5_HIDDEN_DIM = 6144;
template <typename T, int kNumTokens, int kNumExperts, int kHiddenDim>
void invokeRouterGemmFloatOutput(float* output, T const* mat_a, T const* mat_b,
@@ -121,14 +122,21 @@ void dsv3_router_gemm(
STD_TORCH_CHECK(mat_a.size(1) == mat_b.size(1),
"mat_a and mat_b must have the same hidden_dim");
STD_TORCH_CHECK(hidden_dim == DEFAULT_HIDDEN_DIM,
"Expected hidden_dim=", DEFAULT_HIDDEN_DIM,
", but got hidden_dim=", hidden_dim);
STD_TORCH_CHECK(
hidden_dim == DEFAULT_HIDDEN_DIM || hidden_dim == GLM_5_HIDDEN_DIM,
"Expected hidden_dim=", DEFAULT_HIDDEN_DIM,
" or hidden_dim=", GLM_5_HIDDEN_DIM, ", but got hidden_dim=", hidden_dim);
STD_TORCH_CHECK(
num_experts == DEFAULT_NUM_EXPERTS || num_experts == KIMI_K2_NUM_EXPERTS,
"Expected num_experts=", DEFAULT_NUM_EXPERTS,
" or num_experts=", KIMI_K2_NUM_EXPERTS,
", but got num_experts=", num_experts);
// KIMI_K2_NUM_EXPERTS is only instantiated for the default hidden_dim.
STD_TORCH_CHECK(
hidden_dim == DEFAULT_HIDDEN_DIM || num_experts == DEFAULT_NUM_EXPERTS,
"hidden_dim=", GLM_5_HIDDEN_DIM,
" only supports num_experts=", DEFAULT_NUM_EXPERTS,
", but got num_experts=", num_experts);
STD_TORCH_CHECK(num_tokens >= 1 && num_tokens <= 16,
"currently num_tokens must be less than or equal to 16 for "
"router_gemm");
@@ -148,35 +156,49 @@ void dsv3_router_gemm(
const cudaStream_t stream = get_current_cuda_stream(mat_a.get_device_index());
__nv_bfloat16 const* a_ptr =
reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr());
__nv_bfloat16 const* b_ptr =
reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr());
if (output.scalar_type() == torch::headeronly::ScalarType::Float) {
if (num_experts == DEFAULT_NUM_EXPERTS) {
LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
unroll_float_output(
num_tokens, reinterpret_cast<float*>(output.mutable_data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
} else if (num_experts == KIMI_K2_NUM_EXPERTS) {
LoopUnroller<1, 16, KIMI_K2_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
unroll_float_output(
num_tokens, reinterpret_cast<float*>(output.mutable_data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
float* out_ptr = reinterpret_cast<float*>(output.mutable_data_ptr());
if (hidden_dim == DEFAULT_HIDDEN_DIM) {
if (num_experts == DEFAULT_NUM_EXPERTS) {
LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS,
DEFAULT_HIDDEN_DIM>::unroll_float_output(num_tokens,
out_ptr, a_ptr,
b_ptr, stream);
} else {
LoopUnroller<1, 16, KIMI_K2_NUM_EXPERTS,
DEFAULT_HIDDEN_DIM>::unroll_float_output(num_tokens,
out_ptr, a_ptr,
b_ptr, stream);
}
} else { // GLM_5_HIDDEN_DIM
LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS,
GLM_5_HIDDEN_DIM>::unroll_float_output(num_tokens, out_ptr,
a_ptr, b_ptr, stream);
}
} else if (output.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
if (num_experts == DEFAULT_NUM_EXPERTS) {
LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
unroll_bf16_output(
num_tokens,
reinterpret_cast<__nv_bfloat16*>(output.mutable_data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
} else if (num_experts == KIMI_K2_NUM_EXPERTS) {
LoopUnroller<1, 16, KIMI_K2_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
unroll_bf16_output(
num_tokens,
reinterpret_cast<__nv_bfloat16*>(output.mutable_data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
__nv_bfloat16* out_ptr =
reinterpret_cast<__nv_bfloat16*>(output.mutable_data_ptr());
if (hidden_dim == DEFAULT_HIDDEN_DIM) {
if (num_experts == DEFAULT_NUM_EXPERTS) {
LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS,
DEFAULT_HIDDEN_DIM>::unroll_bf16_output(num_tokens,
out_ptr, a_ptr,
b_ptr, stream);
} else {
LoopUnroller<1, 16, KIMI_K2_NUM_EXPERTS,
DEFAULT_HIDDEN_DIM>::unroll_bf16_output(num_tokens,
out_ptr, a_ptr,
b_ptr, stream);
}
} else { // GLM_5_HIDDEN_DIM
LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS,
GLM_5_HIDDEN_DIM>::unroll_bf16_output(num_tokens, out_ptr,
a_ptr, b_ptr, stream);
}
}
}
@@ -286,3 +286,52 @@ template void invokeRouterGemmFloatOutput<__nv_bfloat16, 15, 384, 7168>(
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 16, 384, 7168>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
// Template instantiations for GLM-5 (DEFAULT_NUM_EXPERTS, hidden_dim=6144)
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 1, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 2, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 3, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 4, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 5, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 6, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 7, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 8, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 9, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 10, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 11, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 12, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 13, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 14, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 15, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
template void invokeRouterGemmFloatOutput<__nv_bfloat16, 16, 256, 6144>(
float*, __nv_bfloat16 const*, __nv_bfloat16 const*, cudaStream_t);
@@ -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.