dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

fix(build): std::optional -> c10::optional for corex torch compatibility

Browse Source
This commit is contained in:
claude
2026-08-11 07:37:20 +00:00
parent e440207977
commit f944ef912b
2 changed files with 41 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
ex_engine/csrc/ilu/ilu_ops_api.h
View File

@@ -7,7 +7,7 @@
#pragma once
#include <torch/all.h>
#include <optional>
// #include <optional> // use c10::optional instead
#include <iostream>
#include <stdexcept>
@@ -44,23 +44,23 @@ void act_and_mul(torch::Tensor out,
void reshape_paged_cache(
torch::Tensor& key,
std::optional<torch::Tensor>& value,
c10::optional<torch::Tensor>& value,
torch::Tensor& key_cache,
std::optional<torch::Tensor>& value_cache,
c10::optional<torch::Tensor>& value_cache,
torch::Tensor& slot_mapping);
void batch_prefill(torch::Tensor& query,
const torch::Tensor& key,
const std::optional<torch::Tensor>& value,
const c10::optional<torch::Tensor>& value,
torch::Tensor& output,
std::optional<torch::Tensor>& output_lse,
const std::optional<torch::Tensor>& q_cu_seq_lens,
const std::optional<torch::Tensor>& kv_cu_seq_lens,
const std::optional<torch::Tensor>& alibi_slope,
const std::optional<torch::Tensor>& attn_bias,
const std::optional<torch::Tensor>& q_quant_scale,
const std::optional<torch::Tensor>& k_quant_scale,
const std::optional<torch::Tensor>& v_quant_scale,
c10::optional<torch::Tensor>& output_lse,
const c10::optional<torch::Tensor>& q_cu_seq_lens,
const c10::optional<torch::Tensor>& kv_cu_seq_lens,
const c10::optional<torch::Tensor>& alibi_slope,
const c10::optional<torch::Tensor>& attn_bias,
const c10::optional<torch::Tensor>& q_quant_scale,
const c10::optional<torch::Tensor>& k_quant_scale,
const c10::optional<torch::Tensor>& v_quant_scale,
const torch::Tensor& block_tables,
int64_t max_query_len,
int64_t max_seq_len,
@@ -76,14 +76,14 @@ void batch_decode(torch::Tensor& query,
torch::Tensor& output,
const torch::Tensor& block_table,
const torch::Tensor& seq_lens,
const std::optional<torch::Tensor>& v_cache,
std::optional<torch::Tensor>& output_lse,
const std::optional<torch::Tensor>& q_quant_scale,
const std::optional<torch::Tensor>& k_cache_quant_scale,
const std::optional<torch::Tensor>& v_cache_quant_scale,
const std::optional<torch::Tensor>& out_quant_scale,
const std::optional<torch::Tensor>& alibi_slope,
const std::optional<torch::Tensor>& mask,
const c10::optional<torch::Tensor>& v_cache,
c10::optional<torch::Tensor>& output_lse,
const c10::optional<torch::Tensor>& q_quant_scale,
const c10::optional<torch::Tensor>& k_cache_quant_scale,
const c10::optional<torch::Tensor>& v_cache_quant_scale,
const c10::optional<torch::Tensor>& out_quant_scale,
const c10::optional<torch::Tensor>& alibi_slope,
const c10::optional<torch::Tensor>& mask,
const std::string& compute_dtype,
int64_t max_seq_len,
int64_t window_size_left,
@@ -95,10 +95,10 @@ void batch_decode(torch::Tensor& query,
void residual_layer_norm(torch::Tensor& input,
torch::Tensor& output,
std::optional<torch::Tensor>& residual,
c10::optional<torch::Tensor>& residual,
torch::Tensor& weight,
std::optional<torch::Tensor>& bias,
std::optional<torch::Tensor>& residual_out,
c10::optional<torch::Tensor>& bias,
c10::optional<torch::Tensor>& residual_out,
double eps);
void rms_norm(torch::Tensor& output,
@@ -108,7 +108,7 @@ void rms_norm(torch::Tensor& output,
torch::Tensor matmul(torch::Tensor a,
torch::Tensor b,
std::optional<torch::Tensor> bias);
c10::optional<torch::Tensor> bias);
std::tuple<torch::Tensor, torch::Tensor> moe_active_topk(
const torch::Tensor& input,
@@ -116,11 +116,11 @@ std::tuple<torch::Tensor, torch::Tensor> moe_active_topk(
int64_t num_expert_group,
int64_t topk_group,
bool normalize,
const std::optional<torch::Tensor>& mask,
const c10::optional<torch::Tensor>& mask,
const std::string& normed_by,
const std::string& scoring_func,
double route_scale,
const std::optional<torch::Tensor>& e_score_correction_bias);
const c10::optional<torch::Tensor>& e_score_correction_bias);
std::vector<torch::Tensor> moe_gen_idx(torch::Tensor& expert_id,
int64_t expert_num);
@@ -133,7 +133,7 @@ torch::Tensor moe_expand_input(const torch::Tensor& input,
torch::Tensor group_gemm(torch::Tensor& input,
torch::Tensor& weight,
torch::Tensor& tokens_per_experts,
const std::optional<torch::Tensor>& dst_to_src,
const c10::optional<torch::Tensor>& dst_to_src,
torch::Tensor& output);
torch::Tensor moe_combine_result(torch::Tensor& input, torch::Tensor& weight);

28
ex_engine/csrc/ilu/ix_unified_bridge.cpp
View File

@@ -41,7 +41,7 @@ static torch::Tensor py_silu_and_mul(torch::Tensor input) {
// --- Norm ---
static void py_rms_norm(torch::Tensor output, torch::Tensor input,
torch::Tensor weight, double eps) {
std::optional<torch::Tensor> bias = std::nullopt;
c10::optional<torch::Tensor> bias = c10::nullopt;
infer::rms_norm(input, weight, output, bias, eps);
}
@@ -49,7 +49,7 @@ static void py_fused_add_rms_norm(torch::Tensor input, torch::Tensor residual,
torch::Tensor weight, double eps) {
auto output = torch::empty_like(input);
auto residual_out = torch::empty_like(input);
std::optional<torch::Tensor> bias = std::nullopt;
c10::optional<torch::Tensor> bias = c10::nullopt;
infer::residual_rms_norm(input, residual, weight, output, residual_out,
bias, /*alpha=*/1.0, eps, /*is_post=*/false);
// Copy back in-place
@@ -109,9 +109,9 @@ static torch::Tensor py_flash_attn_prefill(
int64_t wl = -1, wr = -1;
double softcap = 0.0;
bool sqrt_alibi = false;
std::optional<torch::Tensor> alibi = std::nullopt;
std::optional<torch::Tensor> sinks = std::nullopt;
std::optional<torch::Tensor> lse = std::nullopt;
c10::optional<torch::Tensor> alibi = c10::nullopt;
c10::optional<torch::Tensor> sinks = c10::nullopt;
c10::optional<torch::Tensor> lse = c10::nullopt;
return infer::ixinfer_flash_attn_unpad_with_block_tables(
query, key_cache, value_cache, output, block_tables,
cu_seq_q, cu_seq_k, max_seq_q, max_seq_k,
@@ -126,13 +126,13 @@ static torch::Tensor py_paged_attention(
int64_t num_kv_heads, double scale,
torch::Tensor block_tables, torch::Tensor context_lens,
int64_t block_size, int64_t max_context_len) {
std::optional<torch::Tensor> alibi = std::nullopt;
c10::optional<torch::Tensor> alibi = c10::nullopt;
bool causal = true;
int32_t wl = -1, wr = -1;
double softcap = 0.0;
bool enable_cuda_graph = false;
bool sqrt_alibi = false;
std::optional<torch::Tensor> sinks = std::nullopt;
c10::optional<torch::Tensor> sinks = c10::nullopt;
return infer::xllm_paged_attention(
output, query, key_cache, value_cache,
num_kv_heads, scale, block_tables, context_lens,
@@ -170,9 +170,9 @@ static std::vector<torch::Tensor> py_moe_gen_idx(
infer::moe_compute_token_index_api(
expert_ids, src_dst, dst_src, expert_sizes,
/*expert_mask=*/std::nullopt,
/*expert_sizes_cpu=*/std::nullopt,
/*expand_tokens_gpu=*/std::nullopt,
/*expert_mask=*/c10::nullopt,
/*expert_sizes_cpu=*/c10::nullopt,
/*expand_tokens_gpu=*/c10::nullopt,
/*start_expert_id=*/0,
/*end_expert_id=*/num_experts,
/*num_experts=*/num_experts);
@@ -200,8 +200,8 @@ static torch::Tensor py_moe_group_gemm(
auto output = input.new_empty({input.size(0), out_features});
infer::moe_w16a16_group_gemm(
output, input, weight, tokens_per_experts,
/*dst_to_src=*/std::nullopt,
/*bias=*/std::nullopt,
/*dst_to_src=*/c10::nullopt,
/*bias=*/c10::nullopt,
/*format=*/"TN",
/*persistent=*/0,
/*output_n=*/input.size(0));
@@ -216,8 +216,8 @@ static torch::Tensor py_moe_combine_result(
auto output = input.new_empty({inp_3d.size(0), inp_3d.size(2)});
infer::moe_output_reduce_sum(
output, inp_3d, weights,
/*mask=*/std::nullopt,
/*extra_residual=*/std::nullopt,
/*mask=*/c10::nullopt,
/*extra_residual=*/c10::nullopt,
/*scaling_factor=*/1.0);
return output;
}