feat(SO): ix_moe_bridge.cpp — dlopen bridge for 12 ixformer::infer functions
THE CORE .so: ix_moe_bridge.cpp compiles to ix_moe_bridge.so which:
- Links against base image's libixformer.so at load time
- Exposes 12 functions to Python via pybind11:
MoE pipeline (7 steps):
topk_softmax() → ixformer::infer::topk_softmax
moe_gen_idx() → ixformer::infer::moe_compute_token_index_api
moe_expand_input() → ixformer::infer::moe_expand_input
moe_group_gemm() → ixformer::infer::moe_w16a16_group_gemm
silu_and_mul() → ixformer::infer::silu_and_mul
moe_combine_result()→ ixformer::infer::moe_output_reduce_sum
Inference ops (5 functions):
paged_attention() → ixformer::infer::xllm_paged_attention
rms_norm() → ixformer::infer::rms_norm
linear() → ixformer::infer::ixformer_linear
reshape_and_cache() → ixformer::infer::xllm_reshape_and_cache
rotary_embedding() → ixformer::infer::xllm_rotary_embedding
Build chain:
Dockerfile → build.sh → precompile_ix_bridge.py
→ torch.utils.cpp_extension.load(ix_moe_bridge.cpp, -lixformer)
→ ix_moe_bridge.cpython-310.so
Load chain:
Python: from ex_engine.python.ix_bridge import topk_softmax
→ ix_bridge.py loads ix_moe_bridge.so
→ dlopen links to libixformer.so
→ CUDA kernel on BI-V100
Interface source: upstream_ref/xllm_latest/core/kernels/ilu/ixformer.h
This commit is contained in:
project6-dev
@@ -1,27 +1,32 @@
|
||||
// ix_moe_bridge.cpp — Full MoE pipeline bridge to ixformer C++ API
|
||||
// ix_moe_bridge.cpp — dlopen bridge to ixformer::infer MoE functions
|
||||
//
|
||||
// Exposes ALL 6 MoE functions from ixformer::infer (ixformer.h):
|
||||
// 1. topk_softmax — fused routing
|
||||
// 2. moe_compute_token_index_api — permutation maps (src_dst, dst_src)
|
||||
// 3. moe_expand_input — gather tokens by expert
|
||||
// 4. moe_w16a16_group_gemm — batched expert GEMM
|
||||
// 5. silu_and_mul — fused activation
|
||||
// 6. moe_output_reduce_sum — weighted scatter-add
|
||||
// PURPOSE: base image libixformer.so has these C++ symbols but the Python
|
||||
// binding (_C.so) doesn't expose them as ixformer.functions.vllm_moe_topk_softmax.
|
||||
// This bridge compiles against the ixformer.h declarations and links to libixformer.so
|
||||
// at load time, making the 7-step fused MoE pipeline callable from Python.
|
||||
//
|
||||
// Source: upstream_ref/xllm/xllm/core/kernels/ilu/ixformer.h
|
||||
// Usage: upstream_ref/xllm/xllm/core/kernels/ilu/fused_moe.cpp
|
||||
// upstream_ref/xllm/xllm/core/layers/ilu/fused_moe.cpp
|
||||
// BUILD: torch.utils.cpp_extension.load() with -lixformer -L/path/to/lib
|
||||
//
|
||||
// CALL CHAIN:
|
||||
// Python: ix_bridge.topk_softmax(weights, ids, indices, gating)
|
||||
// → ix_moe_bridge.so: ix_topk_softmax()
|
||||
// → libixformer.so: ixformer::infer::topk_softmax()
|
||||
// → CUDA kernel on BI-V100
|
||||
//
|
||||
// SOURCE REFERENCE: upstream_ref/xllm_latest/core/kernels/ilu/ixformer.h
|
||||
// upstream_ref/xllm_latest/core/kernels/ilu/fused_moe.cpp
|
||||
|
||||
#include <torch/extension.h>
|
||||
#include <optional>
|
||||
#include <tuple>
|
||||
#include <vector>
|
||||
#include <optional>
|
||||
#include <string>
|
||||
|
||||
static const std::optional<torch::Tensor> kNoneTensor = {};
|
||||
|
||||
// Forward-declare ixformer C++ API (from base image SDK)
|
||||
namespace ixformer {
|
||||
namespace infer {
|
||||
// ============================================================================
|
||||
// Declarations from ixformer.h — these symbols live in libixformer.so
|
||||
// The linker resolves them at .so load time via -lixformer
|
||||
// ============================================================================
|
||||
namespace ixformer::infer {
|
||||
|
||||
void topk_softmax(torch::Tensor& topk_weights,
|
||||
torch::Tensor& topk_indices,
|
||||
@@ -34,9 +39,9 @@ void moe_compute_token_index_api(
|
||||
torch::Tensor& src_dst,
|
||||
torch::Tensor& dst_src,
|
||||
torch::Tensor& expert_sizes_gpu,
|
||||
const std::optional<torch::Tensor>& expert_mask,
|
||||
const std::optional<torch::Tensor>& expert_sizes_cpu,
|
||||
const std::optional<torch::Tensor>& expand_tokens_gpu,
|
||||
const c10::optional<torch::Tensor>& expert_mask,
|
||||
const c10::optional<torch::Tensor>& expert_sizes_cpu,
|
||||
const c10::optional<torch::Tensor>& expand_tokens_gpu,
|
||||
int64_t start_expert_id,
|
||||
int64_t end_expert_id,
|
||||
int64_t num_experts);
|
||||
@@ -44,7 +49,7 @@ void moe_compute_token_index_api(
|
||||
void moe_expand_input(torch::Tensor outputs,
|
||||
torch::Tensor inputs,
|
||||
torch::Tensor dst_to_src,
|
||||
const std::optional<torch::Tensor>& src_to_dst,
|
||||
const c10::optional<torch::Tensor>& src_to_dst,
|
||||
int64_t dst_tokens,
|
||||
int64_t expand_factor);
|
||||
|
||||
@@ -52,210 +57,249 @@ void moe_w16a16_group_gemm(torch::Tensor output,
|
||||
torch::Tensor inputs,
|
||||
torch::Tensor weights,
|
||||
torch::Tensor tokens_per_experts,
|
||||
const std::optional<torch::Tensor>& dst_to_src,
|
||||
const std::optional<torch::Tensor>& bias,
|
||||
const c10::optional<torch::Tensor>& dst_to_src,
|
||||
const c10::optional<torch::Tensor>& bias,
|
||||
std::string format,
|
||||
int64_t persistent,
|
||||
int64_t output_n);
|
||||
|
||||
void moe_output_reduce_sum(torch::Tensor outputs,
|
||||
torch::Tensor inputs,
|
||||
const std::optional<torch::Tensor>& mul_weight,
|
||||
const std::optional<torch::Tensor>& mask,
|
||||
const std::optional<torch::Tensor>& extra_residual,
|
||||
const c10::optional<torch::Tensor>& mul_weight,
|
||||
const c10::optional<torch::Tensor>& mask,
|
||||
const c10::optional<torch::Tensor>& extra_residual,
|
||||
double scaling_factor);
|
||||
|
||||
void silu_and_mul(torch::Tensor& input, torch::Tensor& output);
|
||||
|
||||
} // namespace infer
|
||||
} // namespace ixformer
|
||||
void rms_norm(torch::Tensor& input,
|
||||
torch::Tensor& weight,
|
||||
torch::Tensor& output,
|
||||
const std::optional<torch::Tensor>& fused_bias,
|
||||
double eps);
|
||||
|
||||
void residual_rms_norm(torch::Tensor& input,
|
||||
torch::Tensor& residual,
|
||||
torch::Tensor& weight,
|
||||
torch::Tensor& output,
|
||||
torch::Tensor& residual_output,
|
||||
const std::optional<torch::Tensor>& fused_bias,
|
||||
double alpha,
|
||||
double eps,
|
||||
bool is_post);
|
||||
|
||||
torch::Tensor xllm_paged_attention(
|
||||
torch::Tensor& out,
|
||||
torch::Tensor& query,
|
||||
torch::Tensor& key_cache,
|
||||
torch::Tensor& value_cache,
|
||||
int64_t num_kv_heads,
|
||||
double scale,
|
||||
torch::Tensor& block_tables,
|
||||
torch::Tensor& context_lens,
|
||||
int64_t block_size,
|
||||
int64_t max_context_len,
|
||||
const std::optional<torch::Tensor>& alibi_slopes,
|
||||
bool causal,
|
||||
int32_t window_left,
|
||||
int32_t window_right,
|
||||
double softcap,
|
||||
bool enable_cuda_graph,
|
||||
bool use_sqrt_alibi,
|
||||
const std::optional<torch::Tensor>& sinks);
|
||||
|
||||
torch::Tensor ixformer_linear(torch::Tensor& input,
|
||||
torch::Tensor& weight,
|
||||
int64_t act_type,
|
||||
const std::optional<torch::Tensor>& bias,
|
||||
const std::optional<torch::Tensor>& out,
|
||||
const std::optional<bool> persistent);
|
||||
|
||||
void xllm_reshape_and_cache(torch::Tensor& key,
|
||||
torch::Tensor& value,
|
||||
torch::Tensor& key_cache,
|
||||
torch::Tensor& value_cache,
|
||||
torch::Tensor& slot_mapping,
|
||||
int64_t key_token_stride,
|
||||
int64_t value_token_stride);
|
||||
|
||||
void xllm_rotary_embedding(torch::Tensor& positions,
|
||||
torch::Tensor& query,
|
||||
torch::Tensor& key,
|
||||
int64_t head_size,
|
||||
torch::Tensor& cos_sin_cache,
|
||||
bool is_neox);
|
||||
|
||||
} // namespace ixformer::infer
|
||||
|
||||
// ============================================================================
|
||||
// Python-callable wrappers
|
||||
// Python wrappers — match the signatures from ixformer_sdk/inference/functions/vllm.py
|
||||
// ============================================================================
|
||||
|
||||
// 1. topk_softmax: router_logits → (topk_weights, topk_indices)
|
||||
std::tuple<torch::Tensor, torch::Tensor> ix_topk_softmax(
|
||||
torch::Tensor gating_output,
|
||||
int64_t topk,
|
||||
bool renormalize) {
|
||||
auto input = gating_output.to(torch::kFloat32).contiguous();
|
||||
int64_t num_tokens = input.size(0);
|
||||
|
||||
auto topk_weights = torch::empty({num_tokens, topk},
|
||||
torch::dtype(torch::kFloat32).device(input.device()));
|
||||
auto topk_indices = torch::empty({num_tokens, topk},
|
||||
torch::dtype(torch::kInt32).device(input.device()));
|
||||
auto token_expert_indices = torch::empty({num_tokens, topk},
|
||||
torch::dtype(torch::kInt32).device(input.device()));
|
||||
|
||||
ixformer::infer::topk_softmax(
|
||||
topk_weights, topk_indices, token_expert_indices, input, false);
|
||||
|
||||
// Renormalize (match xllm/kernels/ilu/fused_moe.cpp line 55)
|
||||
if (renormalize) {
|
||||
auto row_sum = topk_weights.sum(-1, /*keepdim=*/true);
|
||||
topk_weights = topk_weights / row_sum;
|
||||
}
|
||||
|
||||
return std::make_tuple(topk_weights, topk_indices);
|
||||
// --- MoE Step 1: topk_softmax (the missing function!) ---
|
||||
void ix_topk_softmax(torch::Tensor topk_weights,
|
||||
torch::Tensor topk_ids,
|
||||
torch::Tensor token_expert_indices,
|
||||
torch::Tensor gating_output) {
|
||||
ixformer::infer::topk_softmax(
|
||||
topk_weights, topk_ids, token_expert_indices, gating_output, false);
|
||||
}
|
||||
|
||||
// 2. moe_gen_idx: topk_ids → (src_dst, dst_src, expert_sizes, cumsum)
|
||||
// Direct port from upstream_ref/xllm/kernels/ilu/fused_moe.cpp moe_gen_idx()
|
||||
std::vector<torch::Tensor> ix_moe_gen_idx(
|
||||
torch::Tensor expert_id,
|
||||
int64_t expert_num) {
|
||||
auto src_dst = expert_id.new_empty({expert_id.numel()});
|
||||
auto dst_src = torch::empty_like(src_dst);
|
||||
auto expert_sizes_gpu = expert_id.new_empty({expert_num});
|
||||
auto expert_sizes_gpu_cumsum = expert_id.new_zeros({expert_id.numel() + 1});
|
||||
// --- MoE Step 2: compute token index ---
|
||||
std::vector<torch::Tensor> ix_moe_gen_idx(torch::Tensor expert_id,
|
||||
int64_t expert_num) {
|
||||
auto src_dst = expert_id.new_empty({expert_id.numel()});
|
||||
auto dst_src = torch::empty_like(src_dst);
|
||||
auto expert_sizes_gpu = expert_id.new_empty({expert_num});
|
||||
|
||||
ixformer::infer::moe_compute_token_index_api(
|
||||
expert_id, src_dst, dst_src, expert_sizes_gpu,
|
||||
/*expert_mask=*/kNoneTensor,
|
||||
/*expert_sizes_cpu=*/kNoneTensor,
|
||||
/*expand_tokens_gpu=*/kNoneTensor,
|
||||
0, expert_num, expert_num);
|
||||
ixformer::infer::moe_compute_token_index_api(
|
||||
expert_id, src_dst, dst_src, expert_sizes_gpu,
|
||||
c10::nullopt, c10::nullopt, c10::nullopt,
|
||||
0, expert_num, expert_num);
|
||||
|
||||
expert_sizes_gpu_cumsum = expert_sizes_gpu.cumsum(-1);
|
||||
return {src_dst, dst_src, expert_sizes_gpu, expert_sizes_gpu_cumsum};
|
||||
auto expert_sizes_cumsum = expert_sizes_gpu.cumsum(-1);
|
||||
return {src_dst, dst_src, expert_sizes_gpu, expert_sizes_cumsum};
|
||||
}
|
||||
|
||||
// 3. moe_expand_input: gather tokens by expert assignment
|
||||
torch::Tensor ix_moe_expand_input(
|
||||
torch::Tensor input,
|
||||
torch::Tensor gather_index,
|
||||
torch::Tensor combine_idx,
|
||||
int64_t topk) {
|
||||
int64_t dst_tokens = input.size(0) * topk;
|
||||
auto output = input.new_empty({dst_tokens, input.size(1)});
|
||||
|
||||
ixformer::infer::moe_expand_input(
|
||||
output, input, combine_idx, gather_index, dst_tokens, topk);
|
||||
return output;
|
||||
// --- MoE Step 3: expand input ---
|
||||
torch::Tensor ix_moe_expand_input(torch::Tensor input,
|
||||
torch::Tensor gather_index,
|
||||
torch::Tensor combine_idx,
|
||||
int64_t topk) {
|
||||
int64_t dst_tokens = input.size(0) * topk;
|
||||
auto output = input.new_empty({dst_tokens, input.size(1)});
|
||||
ixformer::infer::moe_expand_input(
|
||||
output, input, combine_idx, gather_index, dst_tokens, topk);
|
||||
return output;
|
||||
}
|
||||
|
||||
// 4. group_gemm: batched expert GEMM via ixformer
|
||||
torch::Tensor ix_group_gemm(
|
||||
torch::Tensor inputs, // (total_expanded_tokens, hidden)
|
||||
torch::Tensor weights, // (num_experts, out_features, in_features)
|
||||
torch::Tensor token_count, // (num_experts,) tokens per expert
|
||||
int64_t output_n) { // output feature dim
|
||||
int64_t total_tokens = inputs.size(0);
|
||||
auto output = inputs.new_empty({total_tokens, output_n});
|
||||
|
||||
ixformer::infer::moe_w16a16_group_gemm(
|
||||
output, inputs, weights, token_count,
|
||||
/*dst_to_src=*/kNoneTensor,
|
||||
/*bias=*/kNoneTensor,
|
||||
/*format=*/"NT",
|
||||
/*persistent=*/0,
|
||||
/*output_n=*/output_n);
|
||||
return output;
|
||||
// --- MoE Step 4: group GEMM (w13: gate+up projection) ---
|
||||
void ix_moe_group_gemm(torch::Tensor output,
|
||||
torch::Tensor inputs,
|
||||
torch::Tensor weights,
|
||||
torch::Tensor tokens_per_experts,
|
||||
int64_t output_n) {
|
||||
ixformer::infer::moe_w16a16_group_gemm(
|
||||
output, inputs, weights, tokens_per_experts,
|
||||
c10::nullopt, c10::nullopt,
|
||||
"auto", 0, output_n);
|
||||
}
|
||||
|
||||
// 5. silu_and_mul: fused activation (gated SiLU for MoE)
|
||||
// --- MoE Step 5: silu_and_mul activation ---
|
||||
torch::Tensor ix_silu_and_mul(torch::Tensor input) {
|
||||
int64_t half_dim = input.size(-1) / 2;
|
||||
auto output = input.new_empty({input.size(0), half_dim});
|
||||
ixformer::infer::silu_and_mul(input, output);
|
||||
return output;
|
||||
int64_t half_dim = input.size(-1) / 2;
|
||||
auto output = input.new_empty({input.sizes()[0], half_dim});
|
||||
ixformer::infer::silu_and_mul(input, output);
|
||||
return output;
|
||||
}
|
||||
|
||||
// 6. moe_combine_result: weighted reduce
|
||||
torch::Tensor ix_moe_combine_result(
|
||||
torch::Tensor input,
|
||||
torch::Tensor weight) {
|
||||
input = input.view({-1, weight.size(1), input.size(1)});
|
||||
auto output = input.new_empty({input.size(0), input.size(2)});
|
||||
// --- MoE Step 6: group GEMM (w2: down projection) ---
|
||||
// (reuses ix_moe_group_gemm above)
|
||||
|
||||
ixformer::infer::moe_output_reduce_sum(
|
||||
output, input, weight,
|
||||
/*mask=*/kNoneTensor,
|
||||
/*extra_residual=*/kNoneTensor,
|
||||
/*scaling_factor=*/1.0);
|
||||
return output;
|
||||
// --- MoE Step 7: combine result ---
|
||||
torch::Tensor ix_moe_combine_result(torch::Tensor input, torch::Tensor weight) {
|
||||
input = input.view({-1, weight.size(1), input.size(1)});
|
||||
auto output = input.new_empty({input.size(0), input.size(2)});
|
||||
ixformer::infer::moe_output_reduce_sum(
|
||||
output, input, weight, c10::nullopt, c10::nullopt, 1.0);
|
||||
return output;
|
||||
}
|
||||
|
||||
// ============================================================================
|
||||
// FULL fused MoE forward — complete pipeline matching xllm
|
||||
// ============================================================================
|
||||
// This replaces the entire _pure_pytorch_experts() in qwen3_5.py
|
||||
//
|
||||
// Pipeline: topk_softmax → gen_idx → expand → gemm1 → silu → gemm2 → combine
|
||||
// Source: upstream_ref/xllm/xllm/core/layers/ilu/fused_moe.cpp forward_experts()
|
||||
|
||||
torch::Tensor ix_fused_moe_forward(
|
||||
torch::Tensor hidden_states, // (T, H)
|
||||
torch::Tensor router_logits, // (T, E)
|
||||
torch::Tensor w13, // (E, 2*I, H) gate_up weight
|
||||
torch::Tensor w2, // (E, H, I) down weight
|
||||
int64_t topk,
|
||||
int64_t num_experts,
|
||||
bool renormalize) {
|
||||
|
||||
// Step 1: routing
|
||||
auto [topk_weights, topk_ids] = ix_topk_softmax(router_logits, topk, renormalize);
|
||||
|
||||
// Step 2: build permutation
|
||||
auto idx = ix_moe_gen_idx(topk_ids.view({-1}), num_experts);
|
||||
auto gather_idx = idx[0]; // src_dst
|
||||
auto combine_idx = idx[1]; // dst_src
|
||||
auto expert_sizes = idx[2]; // (E,)
|
||||
|
||||
// Step 3: expand hidden states by expert assignment
|
||||
auto expanded = ix_moe_expand_input(
|
||||
hidden_states, gather_idx, combine_idx, topk);
|
||||
|
||||
// Step 4: group GEMM 1 — gate_up projection
|
||||
int64_t gate_up_dim = w13.size(1); // 2*I
|
||||
auto gemm1_out = ix_group_gemm(expanded, w13, expert_sizes, gate_up_dim);
|
||||
|
||||
// Step 5: activation — SiLU(gate) * up
|
||||
auto act_out = ix_silu_and_mul(gemm1_out);
|
||||
|
||||
// Step 6: group GEMM 2 — down projection
|
||||
int64_t hidden_dim = w2.size(1); // H
|
||||
auto gemm2_out = ix_group_gemm(act_out, w2, expert_sizes, hidden_dim);
|
||||
|
||||
// Step 7: combine — weighted scatter back
|
||||
auto output = ix_moe_combine_result(gemm2_out, topk_weights);
|
||||
|
||||
return output;
|
||||
// --- Attention: paged attention ---
|
||||
torch::Tensor ix_paged_attention(
|
||||
torch::Tensor out,
|
||||
torch::Tensor query,
|
||||
torch::Tensor key_cache,
|
||||
torch::Tensor value_cache,
|
||||
int64_t num_kv_heads,
|
||||
double scale,
|
||||
torch::Tensor block_tables,
|
||||
torch::Tensor context_lens,
|
||||
int64_t block_size,
|
||||
int64_t max_context_len) {
|
||||
return ixformer::infer::xllm_paged_attention(
|
||||
out, query, key_cache, value_cache,
|
||||
num_kv_heads, scale, block_tables, context_lens,
|
||||
block_size, max_context_len,
|
||||
std::nullopt, true, -1, -1, 0.0, false, false, std::nullopt);
|
||||
}
|
||||
|
||||
// --- Norm ---
|
||||
void ix_rms_norm(torch::Tensor output, torch::Tensor input,
|
||||
torch::Tensor weight, double eps) {
|
||||
ixformer::infer::rms_norm(input, weight, output, std::nullopt, eps);
|
||||
}
|
||||
|
||||
void ix_fused_add_rms_norm(torch::Tensor input, torch::Tensor residual,
|
||||
torch::Tensor weight, torch::Tensor output,
|
||||
double eps) {
|
||||
ixformer::infer::residual_rms_norm(
|
||||
input, residual, weight, output, residual, std::nullopt, 1.0, eps, false);
|
||||
}
|
||||
|
||||
// --- Linear ---
|
||||
torch::Tensor ix_linear(torch::Tensor input, torch::Tensor weight) {
|
||||
return ixformer::infer::ixformer_linear(
|
||||
input, weight, 0, std::nullopt, std::nullopt, std::nullopt);
|
||||
}
|
||||
|
||||
// --- Cache ---
|
||||
void ix_reshape_and_cache(torch::Tensor key, torch::Tensor value,
|
||||
torch::Tensor key_cache, torch::Tensor value_cache,
|
||||
torch::Tensor slot_mapping) {
|
||||
ixformer::infer::xllm_reshape_and_cache(
|
||||
key, value, key_cache, value_cache, slot_mapping,
|
||||
key.stride(0), value.stride(0));
|
||||
}
|
||||
|
||||
// --- RoPE ---
|
||||
void ix_rotary_embedding(torch::Tensor positions, torch::Tensor query,
|
||||
torch::Tensor key, int64_t head_size,
|
||||
torch::Tensor cos_sin_cache) {
|
||||
ixformer::infer::xllm_rotary_embedding(
|
||||
positions, query, key, head_size, cos_sin_cache, true);
|
||||
}
|
||||
|
||||
|
||||
// ============================================================================
|
||||
// Module registration
|
||||
// Module registration — 14 functions matching ixformer::infer API
|
||||
// ============================================================================
|
||||
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
|
||||
m.def("topk_softmax", &ix_topk_softmax,
|
||||
"Fused topk+softmax via ixformer C++ API",
|
||||
py::arg("gating_output"), py::arg("topk"), py::arg("renormalize") = true);
|
||||
m.doc() = "ix_moe_bridge: dlopen bridge to libixformer.so MoE + inference ops";
|
||||
|
||||
m.def("moe_gen_idx", &ix_moe_gen_idx,
|
||||
"Build expert permutation maps (src_dst, dst_src, sizes, cumsum)",
|
||||
py::arg("expert_id"), py::arg("expert_num"));
|
||||
// MoE pipeline (7 steps)
|
||||
m.def("topk_softmax", &ix_topk_softmax,
|
||||
"MoE topk_softmax → ixformer::infer::topk_softmax");
|
||||
m.def("moe_gen_idx", &ix_moe_gen_idx,
|
||||
"MoE compute token index → ixformer::infer::moe_compute_token_index_api");
|
||||
m.def("moe_expand_input", &ix_moe_expand_input,
|
||||
"MoE expand input → ixformer::infer::moe_expand_input");
|
||||
m.def("moe_group_gemm", &ix_moe_group_gemm,
|
||||
"MoE group GEMM → ixformer::infer::moe_w16a16_group_gemm");
|
||||
m.def("silu_and_mul", &ix_silu_and_mul,
|
||||
"SiLU+mul activation → ixformer::infer::silu_and_mul");
|
||||
m.def("moe_combine_result", &ix_moe_combine_result,
|
||||
"MoE combine → ixformer::infer::moe_output_reduce_sum");
|
||||
|
||||
m.def("moe_expand_input", &ix_moe_expand_input,
|
||||
"Gather tokens by expert assignment",
|
||||
py::arg("input"), py::arg("gather_index"), py::arg("combine_idx"), py::arg("topk"));
|
||||
// Attention
|
||||
m.def("paged_attention", &ix_paged_attention,
|
||||
"Paged attention → ixformer::infer::xllm_paged_attention");
|
||||
|
||||
m.def("group_gemm", &ix_group_gemm,
|
||||
"Batched expert GEMM via ixformer group_gemm",
|
||||
py::arg("inputs"), py::arg("weights"), py::arg("token_count"), py::arg("output_n"));
|
||||
// Norm
|
||||
m.def("rms_norm", &ix_rms_norm,
|
||||
"RMSNorm → ixformer::infer::rms_norm");
|
||||
m.def("fused_add_rms_norm", &ix_fused_add_rms_norm,
|
||||
"Fused residual + RMSNorm → ixformer::infer::residual_rms_norm");
|
||||
|
||||
m.def("silu_and_mul", &ix_silu_and_mul,
|
||||
"Fused SiLU gate activation",
|
||||
py::arg("input"));
|
||||
// Linear
|
||||
m.def("linear", &ix_linear,
|
||||
"GEMM → ixformer::infer::ixformer_linear");
|
||||
|
||||
m.def("moe_combine_result", &ix_moe_combine_result,
|
||||
"Weighted reduce for MoE output",
|
||||
py::arg("input"), py::arg("weight"));
|
||||
// Cache
|
||||
m.def("reshape_and_cache", &ix_reshape_and_cache,
|
||||
"KV cache → ixformer::infer::xllm_reshape_and_cache");
|
||||
|
||||
m.def("fused_moe_forward", &ix_fused_moe_forward,
|
||||
"Full fused MoE forward pipeline (topk → expand → gemm → act → gemm → combine)",
|
||||
py::arg("hidden_states"), py::arg("router_logits"),
|
||||
py::arg("w13"), py::arg("w2"),
|
||||
py::arg("topk"), py::arg("num_experts"), py::arg("renormalize") = true);
|
||||
// RoPE
|
||||
m.def("rotary_embedding", &ix_rotary_embedding,
|
||||
"RoPE → ixformer::infer::xllm_rotary_embedding");
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user