deepseek-ai--deepep
wehub-resource-sync
19 行
542 B
Plaintext
19 行
542 B
Plaintext
#pragma once
|
|
|
|
#include <deep_ep/common/compiled.cuh>
|
|
|
|
#define LEGACY_NUM_MAX_NVL_PEERS 8
|
|
#define LEGACY_NUM_MAX_RDMA_PEERS 20
|
|
#define LEGACY_NUM_WORKSPACE_BYTES (32 * 1024 * 1024)
|
|
#define LEGACY_NUM_MAX_LOCAL_EXPERTS 1024
|
|
#define LEGACY_NUM_BUFFER_ALIGNMENT_BYTES 128
|
|
|
|
#define LEGACY_LOW_LATENCY_SEND_PHASE 1
|
|
#define LEGACY_LOW_LATENCY_RECV_PHASE 2
|
|
|
|
#define LEGACY_FINISHED_SUM_TAG 1024
|
|
#define LEGACY_NUM_WAIT_NANOSECONDS 500
|
|
|
|
#define LEGACY_NUM_CPU_TIMEOUT_SECS 100
|
|
#define LEGACY_NUM_TIMEOUT_CYCLES 200000000000ull // 200G cycles ~= 100s
|