Add low-latency kernel PCIe usage flag (#195)

* Add low-latency kernel usage flag

* Update comments

csrc/deep_ep.cpp

@@ -76,6 +76,13 @@ Buffer::Buffer(int rank, int num_ranks, int64_t num_nvl_bytes, int64_t num_rdma_
         CUDA_CHECK(cudaHostGetDevicePointer(&moe_recv_rdma_counter_mapped, const_cast<int*>(moe_recv_rdma_counter), 0));
         *moe_recv_rdma_counter = -1;
     }
+
+    // Low-latency kernels' usage flag
+    if (low_latency_mode) {
+        CUDA_CHECK(cudaMallocHost(&low_latency_usage_flag, sizeof(int), cudaHostAllocMapped));
+        CUDA_CHECK(cudaHostGetDevicePointer(&low_latency_usage_flag_mapped, const_cast<int*>(low_latency_usage_flag), 0));
+        *low_latency_usage_flag = 0;
+    }
 }
 
 Buffer::~Buffer() noexcept(false) {
@@ -997,6 +1004,11 @@ Buffer::internode_combine(const torch::Tensor& x, const std::optional<torch::Ten
     return {combined_x, combined_topk_weights, event};
 }
 
+uint64_t Buffer::get_low_latency_usage_flag() const {
+    EP_HOST_ASSERT(low_latency_usage_flag != nullptr);
+    return reinterpret_cast<uint64_t>(low_latency_usage_flag);
+}
+
 void Buffer::clean_low_latency_buffer(int num_max_dispatch_tokens_per_rank, int hidden, int num_experts) {
     EP_HOST_ASSERT(low_latency_mode);
 
@@ -1078,7 +1090,7 @@ Buffer::low_latency_dispatch(const torch::Tensor& x, const torch::Tensor& topk_i
                                next_clean_meta.first, next_clean_meta.second,
                                num_tokens, hidden, num_max_dispatch_tokens_per_rank,
                                num_topk, num_experts, rank, num_ranks, use_fp8,
-                               workspace, launch_stream, phases);
+                               workspace, low_latency_usage_flag_mapped, launch_stream, phases);
     };
     launcher(return_recv_hook ? LOW_LATENCY_SEND_PHASE : (LOW_LATENCY_SEND_PHASE | LOW_LATENCY_RECV_PHASE));
 
@@ -1165,7 +1177,7 @@ Buffer::low_latency_combine(const torch::Tensor& x, const torch::Tensor& topk_id
                               next_clean_meta.first, next_clean_meta.second,
                               num_combined_tokens, hidden, num_max_dispatch_tokens_per_rank,
                               num_topk, num_experts, rank, num_ranks,
-                              workspace, launch_stream,
+                              workspace, low_latency_usage_flag_mapped, launch_stream,
                               phases, zero_copy);
     };
     launcher(return_recv_hook ? LOW_LATENCY_SEND_PHASE : (LOW_LATENCY_SEND_PHASE | LOW_LATENCY_RECV_PHASE));
@@ -1238,6 +1250,7 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
         .def("intranode_combine", &deep_ep::Buffer::intranode_combine)
         .def("internode_dispatch", &deep_ep::Buffer::internode_dispatch)
         .def("internode_combine", &deep_ep::Buffer::internode_combine)
+        .def("get_low_latency_usage_flag", &deep_ep::Buffer::get_low_latency_usage_flag)
         .def("clean_low_latency_buffer", &deep_ep::Buffer::clean_low_latency_buffer)
         .def("low_latency_dispatch", &deep_ep::Buffer::low_latency_dispatch)
         .def("low_latency_combine", &deep_ep::Buffer::low_latency_combine)

csrc/deep_ep.hpp

@@ -71,6 +71,10 @@ private:
     volatile int* moe_recv_rdma_counter = nullptr;
     int* moe_recv_rdma_counter_mapped = nullptr;
 
+    // Host-side low-latency kernels' usages
+    volatile int* low_latency_usage_flag = nullptr;
+    int* low_latency_usage_flag_mapped = nullptr;
+
 private:
     void move_fifo_slots(int num_slots = 1);
 
@@ -132,6 +136,8 @@ public:
                       const torch::Tensor& combined_rdma_head, const torch::Tensor& combined_nvl_head,
                       const Config& config, std::optional<EventHandle>& previous_event, bool async, bool allocate_on_comm_stream);
 
+    uint64_t get_low_latency_usage_flag() const;
+
     void clean_low_latency_buffer(int num_max_dispatch_tokens_per_rank, int hidden, int num_experts);
 
     std::tuple<torch::Tensor, std::optional<torch::Tensor>, torch::Tensor, torch::Tensor, torch::Tensor, std::optional<EventHandle>, std::optional<std::function<void()>>>

csrc/kernels/api.cuh

@@ -138,7 +138,8 @@ void dispatch(void* packed_recv_x, float* packed_recv_x_scales,
               int* next_clean, int num_next_clean_int,
               int num_tokens, int hidden, int num_max_dispatch_tokens_per_rank,
               int num_topk, int num_experts, int rank, int num_ranks, bool use_fp8,
-              void* workspace, cudaStream_t stream, int phases);
+              void* workspace, int* usage_flag,
+              cudaStream_t stream, int phases);
 
 void combine(void* combined_x,
              void* rdma_recv_x, int* rdma_recv_flag, void* rdma_send_x,
@@ -147,8 +148,8 @@ void combine(void* combined_x,
              int* next_clean, int num_next_clean_int,
              int num_combined_tokens, int hidden, int num_max_dispatch_tokens_per_rank,
              int num_topk, int num_experts, int rank, int num_ranks,
-             void* workspace, cudaStream_t stream,
-             int phases, bool zero_copy);
+             void* workspace, int* usage_flag,
+             cudaStream_t stream, int phases, bool zero_copy);
 
 } // namespace internode_ll
 

csrc/kernels/internode_ll.cu

@@ -47,7 +47,7 @@ dispatch(void* packed_recv_x, float* packed_recv_x_scales,
          int* next_clean, int num_next_clean_int,
          int num_tokens, int num_max_dispatch_tokens_per_rank,
          int num_topk, int num_experts, int rank, int num_ranks,
-         int phases) {
+         int* usage_flag, int phases) {
     const auto sm_id = static_cast<int>(blockIdx.x);
     const auto thread_id = static_cast<int>(threadIdx.x);
     const auto warp_id = thread_id / 32, lane_id = get_lane_id();
@@ -180,6 +180,10 @@ dispatch(void* packed_recv_x, float* packed_recv_x_scales,
             #pragma unroll
             for (int i = lane_id; i < num_experts; i += 32)
                 atomic_add_release_global(atomic_finish_counter_per_expert + i, FINISHED_SUM_TAG);
+        } else if (sm_id == 1) {
+            // The second SM is also responsible for notifying PCIe usage
+            if (lane_id == 0)
+                atomicAdd_system(usage_flag, 1);
         }
 
         // This SM should be responsible for some destination experts, read `topk_idx` for them
@@ -311,7 +315,8 @@ void dispatch(void* packed_recv_x, float* packed_recv_x_scales,
               int* next_clean, int num_next_clean_int,
               int num_tokens, int hidden, int num_max_dispatch_tokens_per_rank,
               int num_topk, int num_experts, int rank, int num_ranks, bool use_fp8,
-              void* workspace, cudaStream_t stream, int phases) {
+              void* workspace, int* usage_flag,
+              cudaStream_t stream, int phases) {
     constexpr int kNumMaxTopK = 9;
     constexpr int kNumWarpsPerGroup = 10;
     constexpr int kNumWarpGroups = 3;
@@ -338,7 +343,8 @@ LAUNCH_KERNEL(&cfg, dispatch_func, \
               atomic_counter_per_expert, atomic_finish_counter_per_expert, \
               next_clean, num_next_clean_int, \
               num_tokens, num_max_dispatch_tokens_per_rank, \
-              num_topk, num_experts, rank, num_ranks, phases); } break
+              num_topk, num_experts, rank, num_ranks, \
+              usage_flag, phases); } break
 
     SETUP_LAUNCH_CONFIG(num_sms, num_warps * 32, stream);
     SWITCH_HIDDEN(DISPATCH_LAUNCH_CASE);
@@ -356,7 +362,7 @@ combine(void* combined_x,
         int num_combined_tokens, int hidden, int num_topk,
         int num_max_dispatch_tokens_per_rank,
         int num_experts, int rank, int num_ranks,
-        int phases, bool zero_copy) {
+        int* usage_flag, int phases, bool zero_copy) {
     const auto sm_id = static_cast<int>(blockIdx.x);
     const auto num_sms = static_cast<int>(gridDim.x);
     const auto thread_id = static_cast<int>(threadIdx.x);
@@ -451,11 +457,14 @@ combine(void* combined_x,
     if ((phases & LOW_LATENCY_RECV_PHASE) == 0)
         return;
 
-    // Wait all ranks to arrive
+    // Wait all ranks to arrive and notify usages
     if (responsible_expert_idx < num_experts) {
         EP_STATIC_ASSERT(kNumWarpsPerGroup > 1, "Invalid number of warps per group");
-        if (sub_warp_id == 0 and lane_id == 0)
+        if (sub_warp_id == 0 and lane_id == 0) {
             while (ld_acquire_sys_global(rdma_recv_flag + responsible_expert_idx) == 0);
+        } else if (sm_id == 0 and sub_warp_id == 1 and lane_id == 0) {
+            atomicAdd_system(usage_flag, 1);
+        }
     }
     cg::this_grid().sync();
 
@@ -506,8 +515,8 @@ void combine(void* combined_x,
              int* next_clean, int num_next_clean_int,
              int num_combined_tokens, int hidden, int num_max_dispatch_tokens_per_rank,
              int num_topk, int num_experts, int rank, int num_ranks,
-             void* workspace, cudaStream_t stream,
-             int phases, bool zero_copy) {
+             void* workspace, int* usage_flag,
+             cudaStream_t stream, int phases, bool zero_copy) {
     constexpr int kNumWarpsPerGroup = 10;
     constexpr int kNumWarpGroups = 3;
     constexpr int kNumMaxTopk = 9;
@@ -531,6 +540,7 @@ LAUNCH_KERNEL(&cfg, combine_func, \
               num_combined_tokens, hidden, num_topk, \
               num_max_dispatch_tokens_per_rank, \
               num_experts, rank, num_ranks, \
+              usage_flag, \
               phases, zero_copy); } break
 
     SETUP_LAUNCH_CONFIG(num_sms, num_warps * 32, stream);

deep_ep/buffer.py

@@ -443,6 +443,19 @@ class Buffer:
             async_finish, allocate_on_comm_stream)
         return combined_x, combined_topk_weights, EventOverlap(event)
 
+    def get_low_latency_usage_flag(self):
+        """
+        Return a host-side integer flag, which indicates the stages of low-latency kernels.
+        The initial value is 0, the low-latency dispatch will add 1 before communication, the low-latency combine
+            will add 1 after communication.
+        This is useful when there is no two-batch overlap, and you want to overlap H2D/D2H transfer with attention layers.
+
+        Returns:
+            flag: the host-side integer flag pointer. The value is in `int`, but returns a `uint64_t` pointer. Please
+                `reinterpret_cast` the returned value into `int*`.
+        """
+        return self.runtime.get_low_latency_usage_flag()
+
     def clean_low_latency_buffer(self, num_max_dispatch_tokens_per_rank: int, hidden: int, num_experts: int) -> None:
         """
         As low-latency kernels require part of the buffer to be zero-initialized, so it is vital to clean the buffer

tests/test_low_latency.py

@@ -155,6 +155,7 @@ def test_loop(local_rank: int, num_local_ranks: int):
         print(f'Allocating buffer size: {num_rdma_bytes / 1e6} MB ...', flush=True)
     buffer = deep_ep.Buffer(group, num_rdma_bytes=num_rdma_bytes, low_latency_mode=True,
                             num_qps_per_rank=num_experts // num_ranks)
+    buffer.get_low_latency_usage_flag()
     test_main(num_tokens, hidden, num_experts, num_topk, rank, num_ranks, group, buffer, seed=1)
 
     do_pressure_test = False