EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

moe_gating_top_k (#5271)

Browse Source 
1. What this PR does / why we need it?
This PR supports the moe_gating_top_k operator, which enables
post-positioned renormalization (renorm) on the basis of softmax.
2. Does this PR introduce any user-facing change?
No user-facing changes are required.
3. How was this patch tested?
This patch was tested with the test_npu_moe_gating_top_k test case.
vLLM version: release/v0.13.0
vLLM main:
ad32e3e19c

---------

Signed-off-by: ZCG12345 <2097562023@qq.com>
Signed-off-by: zzzzwwjj <34335947+zzzzwwjj@users.noreply.github.com>
Co-authored-by: zzzzwwjj <34335947+zzzzwwjj@users.noreply.github.com>
This commit is contained in:
ZCG12345
2025-12-30 09:28:01 +08:00
committed by GitHub
parent 15d73f248e
commit 45c3c279e2
34 changed files with 4791 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
csrc/torch_binding_meta.cpp
View File

@@ -283,6 +283,42 @@ std::tuple<at::Tensor, at::Tensor> matmul_allreduce_add_rmsnorm_meta(
return {output, add_out};
}
std::tuple<at::Tensor,at::Tensor, at::Tensor> moe_gating_top_k_meta(
const at::Tensor& x,
int64_t k,
int64_t kGroup,
int64_t groupCount,
int64_t groupSelectMode,
int64_t renorm,
int64_t normType,
bool outFlag,
double routedScalingFactor,
double eps,
const c10::optional<at::Tensor>& biasOptional)
{
TORCH_CHECK(x.dim() == 2, "The x should be 2D");
TORCH_CHECK(
x.scalar_type() == at::kHalf || x.scalar_type() == at::kFloat || x.scalar_type() == at::kBFloat16,
"float16、float32 or bfloat16 tensor expected but got a tensor with dtype: ",
x.scalar_type());
auto x_size = x.sizes();
auto rows = x_size[0];
auto expert_num = x_size[1];
const at::Tensor &bias = c10::value_or_else(biasOptional, [] { return at::Tensor(); });
if (bias.defined()) {
TORCH_CHECK(x.scalar_type() == bias.scalar_type(), "The dtype of x and bias should be same");
TORCH_CHECK(bias.dim() == 1, "The bias should be 1D");
auto bias_size = bias.sizes();
TORCH_CHECK(bias_size[0] == expert_num, "The bias first dim should be same as x second dim");
}
at::Tensor yOut = at::empty({rows, k}, x.options());
at::Tensor expertIdxOut = at::empty({rows, k}, x.options().dtype(at::kInt));
at::Tensor outOut = at::empty({rows, expert_num}, x.options().dtype(at::kFloat));
return std::tuple<at::Tensor, at::Tensor, at::Tensor>(outOut,expertIdxOut, yOut);
}
std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> npu_moe_init_routing_custom_meta(
const at::Tensor &x, const at::Tensor &expert_idx,
const c10::optional<at::Tensor> &scale, const c10::optional<at::Tensor> &offset, int64_t active_num,
@@ -367,12 +403,15 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor> npu_moe_init_routing_
}
} // namespace meta
} // namespace vllm_ascend
namespace {
// Register the meta implementations of the custom kernels for symbolic tracing, this will also
// the custom kernel been captured into aclgraph
TORCH_LIBRARY_IMPL_EXPAND(CONCAT(_C, _ascend), Meta, ops) {
// Moe_gating_top_k
ops.impl("moe_gating_top_k", &vllm_ascend::meta::moe_gating_top_k_meta);
// Rotary embedding meta implementation
ops.impl("rotary_embedding", &vllm_ascend::meta::rotary_embedding_meta);
// Masked input and mask meta implementation