[CPU] Fix macOS/Apple Silicon hang by enabling OpenMP in the build (#46769)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
parent
3d3b96488f
commit
c6554f321c
@@ -124,7 +124,7 @@ struct AttentionMetadata {
|
||||
workitem_group_num(workitem_group_num),
|
||||
reduction_item_num(reduction_item_num),
|
||||
reduction_split_num(reduction_split_num),
|
||||
thread_num(omp_get_max_threads()),
|
||||
thread_num(cpu_utils::get_max_threads()),
|
||||
effective_thread_num(thread_num),
|
||||
split_kv_q_token_num_threshold(split_kv_q_token_num_threshold),
|
||||
attention_scratchpad_size_per_thread(0),
|
||||
@@ -405,7 +405,7 @@ class AttentionScheduler {
|
||||
torch::Tensor schedule(const ScheduleInput& input) const {
|
||||
const bool causal = input.causal;
|
||||
const bool is_dynamic_causal = input.dynamic_causal != nullptr;
|
||||
const int32_t thread_num = omp_get_max_threads();
|
||||
const int32_t thread_num = cpu_utils::get_max_threads();
|
||||
const int64_t cache_size = cpu_utils::get_available_l2_size();
|
||||
const int32_t max_num_q_per_iter = input.max_num_q_per_iter;
|
||||
const int32_t kv_len_alignment = input.kv_block_alignment;
|
||||
@@ -1423,7 +1423,7 @@ class AttentionMainLoop {
|
||||
|
||||
public:
|
||||
void operator()(const AttentionInput* input) {
|
||||
const int thread_num = omp_get_max_threads();
|
||||
const int thread_num = cpu_utils::get_max_threads();
|
||||
TORCH_CHECK_EQ(input->metadata->thread_num, thread_num);
|
||||
std::atomic<int32_t> guard_counter(0);
|
||||
std::atomic<int32_t>* guard_counter_ptr = &guard_counter;
|
||||
|
||||
@@ -267,7 +267,7 @@ void fused_moe_impl(scalar_t* __restrict__ output, scalar_t* __restrict__ input,
|
||||
TORCH_CHECK_EQ(output_size_2 % gemm_n_tile_size, 0);
|
||||
TORCH_CHECK_EQ(output_size_13 / 2, input_size_2);
|
||||
|
||||
const int32_t thread_num = omp_get_max_threads();
|
||||
const int32_t thread_num = cpu_utils::get_max_threads();
|
||||
|
||||
const int32_t w13_input_buffer_size = cpu_utils::round_up<64>(
|
||||
gemm_m_tile_size * input_size_13 * sizeof(scalar_t));
|
||||
|
||||
@@ -25,4 +25,20 @@
|
||||
#include <omp.h>
|
||||
#endif
|
||||
|
||||
#include <c10/util/Exception.h>
|
||||
|
||||
namespace cpu_utils {
|
||||
// Without OpenMP the omp pragmas compile to serial loops, so report 1: kernels
|
||||
// that barrier on the thread count would otherwise deadlock.
|
||||
inline int get_max_threads() {
|
||||
#ifdef _OPENMP
|
||||
return omp_get_max_threads();
|
||||
#else
|
||||
TORCH_WARN_ONCE(
|
||||
"vLLM CPU was built without OpenMP; running single-threaded.");
|
||||
return 1;
|
||||
#endif
|
||||
}
|
||||
} // namespace cpu_utils
|
||||
|
||||
#endif
|
||||
@@ -155,7 +155,7 @@ void cpu_gemm_wna16_impl(
|
||||
constexpr int32_t gemm_m_tile_size = gemm_t::MaxMSize;
|
||||
constexpr int32_t n_block_size = 16;
|
||||
static_assert(gemm_n_tile_size % n_block_size == 0);
|
||||
const int32_t thread_num = omp_get_max_threads();
|
||||
const int32_t thread_num = cpu_utils::get_max_threads();
|
||||
|
||||
// a simple schedule policy, just to hold more B tiles in L2 and make sure
|
||||
// each thread has tasks
|
||||
|
||||
@@ -202,7 +202,7 @@ void dynamic_quant_epilogue(const float* input, scalar_t* output,
|
||||
using cvt_vec_t = typename KernelVecType<scalar_t>::cvt_vec_type;
|
||||
constexpr int vec_elem_num = load_vec_t::VEC_ELEM_NUM;
|
||||
|
||||
const int64_t thread_num = omp_get_max_threads();
|
||||
const int64_t thread_num = cpu_utils::get_max_threads();
|
||||
if (num_tokens > thread_num) {
|
||||
#pragma omp parallel for
|
||||
for (int64_t i = 0; i < num_tokens; ++i) {
|
||||
|
||||
@@ -251,7 +251,7 @@ void mla_decode_kvcache_cpu_impl(
|
||||
constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
|
||||
|
||||
// shared across threads
|
||||
const int max_threads = omp_get_max_threads();
|
||||
const int max_threads = cpu_utils::get_max_threads();
|
||||
const int acc_out_nbytes =
|
||||
max_threads * num_heads * V_HEAD_DIM * sizeof(float);
|
||||
float* acc_out = static_cast<float*>(std::aligned_alloc(64, acc_out_nbytes));
|
||||
|
||||
Reference in New Issue
Block a user