PD send cache via storage & Refine swap_cache_layout op

Author: juncaipeng

custom_ops/gpu_ops/swap_cache_layout.cu

@@ -15,90 +15,254 @@
 #include "helper.h"
 #include "paddle/extension.h"
 
-// #define SWAP_DEBUG
+// D2H: Each thread block handles ALL layers for one swap block.
+// This produces perfectly contiguous host writes (1 block × all layers),
+// maximizing write-combining efficiency.
+template <typename T>
+__global__ void swap_d2h_kernel(T** __restrict__ layer_ptrs,
+                                T* __restrict__ cpu_buffer,
+                                const int64_t* __restrict__ gpu_block_ids,
+                                int n_blocks,
+                                int layer_num,
+                                int64_t block_stride) {
+  int block_idx = blockIdx.x;
+  if (block_idx >= n_blocks) return;
+
+  int64_t gpu_block = gpu_block_ids[block_idx];
+  int64_t num_vec_per_layer = (block_stride * sizeof(T)) / sizeof(float4);
+
+  T* dst_base = cpu_buffer + (int64_t)block_idx * layer_num * block_stride;
+
+  for (int layer_idx = 0; layer_idx < layer_num; layer_idx++) {
+    const T* src = layer_ptrs[layer_idx] + gpu_block * block_stride;
+    float4* dst4 =
+        reinterpret_cast<float4*>(dst_base + layer_idx * block_stride);
+    const float4* src4 = reinterpret_cast<const float4*>(src);
+
+    for (int64_t i = threadIdx.x; i < num_vec_per_layer; i += blockDim.x) {
+      dst4[i] = src4[i];
+    }
+  }
+}
+
+// H2D: scatter from contiguous staging buffer to scattered GPU layer tensors
+template <typename T>
+__global__ void scatter_blocks_kernel(T** __restrict__ layer_ptrs,
+                                      const T* __restrict__ staging,
+                                      const int64_t* __restrict__ gpu_block_ids,
+                                      int n_blocks,
+                                      int layer_num,
+                                      int64_t block_stride) {
+  int pair_idx = blockIdx.x;
+  int block_idx = pair_idx / layer_num;
+  int layer_idx = pair_idx % layer_num;
+
+  if (block_idx >= n_blocks) return;
+
+  int64_t gpu_block = gpu_block_ids[block_idx];
+  const T* src = staging + (int64_t)block_idx * layer_num * block_stride +
+                 layer_idx * block_stride;
+  T* dst = layer_ptrs[layer_idx] + gpu_block * block_stride;
+
+  int64_t num_vec = (block_stride * sizeof(T)) / sizeof(float4);
+  const float4* src4 = reinterpret_cast<const float4*>(src);
+  float4* dst4 = reinterpret_cast<float4*>(dst);
+
+  for (int64_t i = threadIdx.x; i < num_vec; i += blockDim.x) {
+    dst4[i] = src4[i];
+  }
+}
+
+static void* g_staging_buffer = nullptr;
+static size_t g_staging_buffer_size = 0;
+static void* g_device_block_ids = nullptr;
+static size_t g_device_block_ids_size = 0;
+static void* g_device_layer_ptrs = nullptr;
+static size_t g_device_layer_ptrs_size = 0;
+
+static void ensure_staging_buffer(size_t required_size) {
+  if (g_staging_buffer_size < required_size) {
+    if (g_staging_buffer) cudaFree(g_staging_buffer);
+    cudaError_t err = cudaMalloc(&g_staging_buffer, required_size);
+    PADDLE_ENFORCE_EQ(
+        err,
+        cudaSuccess,
+        phi::errors::External("cudaMalloc staging buffer failed: %s",
+                              cudaGetErrorString(err)));
+    g_staging_buffer_size = required_size;
+  }
+}
+
+static void ensure_device_block_ids(size_t required_size) {
+  if (g_device_block_ids_size < required_size) {
+    if (g_device_block_ids) cudaFree(g_device_block_ids);
+    cudaError_t err = cudaMalloc(&g_device_block_ids, required_size);
+    PADDLE_ENFORCE_EQ(
+        err,
+        cudaSuccess,
+        phi::errors::External("cudaMalloc device block_ids failed: %s",
+                              cudaGetErrorString(err)));
+    g_device_block_ids_size = required_size;
+  }
+}
+
+static void ensure_device_layer_ptrs(size_t required_size) {
+  if (g_device_layer_ptrs_size < required_size) {
+    if (g_device_layer_ptrs) cudaFree(g_device_layer_ptrs);
+    cudaError_t err = cudaMalloc(&g_device_layer_ptrs, required_size);
+    PADDLE_ENFORCE_EQ(
+        err,
+        cudaSuccess,
+        phi::errors::External("cudaMalloc device layer_ptrs failed: %s",
+                              cudaGetErrorString(err)));
+    g_device_layer_ptrs_size = required_size;
+  }
+}
+
+static bool is_cpu_block_ids_sequential(
+    const std::vector<int64_t>& cpu_block_ids) {
+  if (cpu_block_ids.empty()) return true;
+  int64_t start = cpu_block_ids[0];
+  for (size_t i = 1; i < cpu_block_ids.size(); i++) {
+    if (cpu_block_ids[i] != start + static_cast<int64_t>(i)) return false;
+  }
+  return true;
+}
 
 template <paddle::DataType D>
-void SwapCacheImpLayout(
-    const std::vector<paddle::Tensor>& cache_gpu_tensors,  // gpu
-    const int64_t& cache_cpu_pointer,                      // cpu
-    const std::vector<int64_t>& cache_shape,
-    const std::vector<int64_t>& gpu_block_ids,
-    const std::vector<int64_t>& cpu_block_ids,
-    int mode) {
-  /*
-  mode is 0: gpu to cpu; 1: cpu to gpu
-
-  cache layout: layer_num * [block_num, head_num, block_size, head_dim]
-  scale layout: layer_num * [block_num, head_num, block_size]
-  cache buffer layout: [block_num, layer_num, head_num, block_size, head_dim]
-  scale buffer layout: [block_num, layer_num, head_num, block_size]
-  */
+void SwapCacheImpLayout(const std::vector<paddle::Tensor>& cache_gpu_tensors,
+                        const int64_t& cache_cpu_pointer,
+                        const std::vector<int64_t>& cache_shape,
+                        const std::vector<int64_t>& gpu_block_ids,
+                        const std::vector<int64_t>& cpu_block_ids,
+                        int mode) {
   typedef PDTraits<D> traits_;
   typedef typename traits_::DataType DataType_;
   typedef typename traits_::data_t data_t;
 
   const int64_t layer_number = cache_gpu_tensors.size();
   int64_t cache_block_stride = 1;
-  for (int i = 1; i < cache_shape.size(); i++) {
+  for (size_t i = 1; i < cache_shape.size(); i++) {
     cache_block_stride *= cache_shape[i];
   }
 
+  const int n_blocks = gpu_block_ids.size();
+  if (n_blocks == 0) return;
+
   auto stream = cache_gpu_tensors[0].stream();
-  const cudaMemcpyKind copy_kind =
-      (mode == 0) ? cudaMemcpyDeviceToHost : cudaMemcpyHostToDevice;
-
-  for (int layer_idx = 0; layer_idx < cache_gpu_tensors.size(); layer_idx++) {
-    const paddle::Tensor& cache_gpu = cache_gpu_tensors[layer_idx];
-    data_t* cache_gpu_ptr = const_cast<data_t*>(cache_gpu.data<data_t>());
-    auto* cache_cpu_ptr = reinterpret_cast<data_t*>(cache_cpu_pointer);
-
-    for (int block_idx = 0; block_idx < gpu_block_ids.size(); block_idx++) {
-      auto cur_gpu_block_id = gpu_block_ids[block_idx];
-      auto cur_cpu_block_id = cpu_block_ids[block_idx];
-      auto* cache_gpu_ptr_now =
-          cache_gpu_ptr + cur_gpu_block_id * cache_block_stride;
-      auto* cache_cpu_ptr_now =
-          cache_cpu_ptr + cur_cpu_block_id * cache_block_stride * layer_number +
-          layer_idx * cache_block_stride;
-
-      cudaError_t status = cudaMemcpyAsync(
-          (copy_kind == cudaMemcpyDeviceToHost) ? cache_cpu_ptr_now
-                                                : cache_gpu_ptr_now,
-          (copy_kind == cudaMemcpyDeviceToHost) ? cache_gpu_ptr_now
-                                                : cache_cpu_ptr_now,
-          cache_block_stride * sizeof(DataType_),
-          copy_kind,
-          stream);
+  const size_t block_bytes = cache_block_stride * sizeof(DataType_);
+  const size_t total_bytes = (size_t)n_blocks * layer_number * block_bytes;
+
+  bool use_optimized = is_cpu_block_ids_sequential(cpu_block_ids);
+
+  if (use_optimized) {
+    ensure_device_block_ids(n_blocks * sizeof(int64_t));
+    ensure_device_layer_ptrs(layer_number * sizeof(DataType_*));
 
+    cudaMemcpyAsync(g_device_block_ids,
+                    gpu_block_ids.data(),
+                    n_blocks * sizeof(int64_t),
+                    cudaMemcpyHostToDevice,
+                    stream);
+
+    std::vector<DataType_*> h_layer_ptrs(layer_number);
+    for (int64_t i = 0; i < layer_number; i++) {
+      h_layer_ptrs[i] = reinterpret_cast<DataType_*>(
+          const_cast<data_t*>(cache_gpu_tensors[i].data<data_t>()));
+    }
+    cudaMemcpyAsync(g_device_layer_ptrs,
+                    h_layer_ptrs.data(),
+                    layer_number * sizeof(DataType_*),
+                    cudaMemcpyHostToDevice,
+                    stream);
+
+    int64_t cpu_start_block = cpu_block_ids[0];
+    auto* cache_cpu_base = reinterpret_cast<DataType_*>(cache_cpu_pointer) +
+                           cpu_start_block * layer_number * cache_block_stride;
+
+    int grid_size = n_blocks * layer_number;
+
+    if (mode == 0) {
+      // GPU→CPU: direct kernel write to pinned host memory
+      // Multi-layer kernel: each block handles all layers for one swap block
+      swap_d2h_kernel<DataType_><<<n_blocks, 512, 0, stream>>>(
+          reinterpret_cast<DataType_**>(g_device_layer_ptrs),
+          cache_cpu_base,
+          reinterpret_cast<int64_t*>(g_device_block_ids),
+          n_blocks,
+          layer_number,
+          cache_block_stride);
+    } else {
+      // CPU→GPU: DMA memcpy to staging then scatter kernel
+      ensure_staging_buffer(total_bytes);
+
+      cudaError_t status = cudaMemcpyAsync(g_staging_buffer,
+                                           cache_cpu_base,
+                                           total_bytes,
+                                           cudaMemcpyHostToDevice,
+                                           stream);
       PADDLE_ENFORCE_EQ(status,
                         cudaSuccess,
-                        phi::errors::External("cudaMemcpyAsync failed: %s",
+                        phi::errors::External("cudaMemcpyAsync H2D failed: %s",
                                               cudaGetErrorString(status)));
 
-#ifdef SWAP_DEBUG
-      cudaStreamSynchronize(stream);
-      std::cout << "mode:" << mode << ", layer_idx:" << layer_idx
-                << ", block_idx:" << block_idx << ", cache_cpu_ptr_now data:"
-                << static_cast<float>(*cache_cpu_ptr_now) << std::endl;
-#endif
+      scatter_blocks_kernel<DataType_><<<grid_size, 256, 0, stream>>>(
+          reinterpret_cast<DataType_**>(g_device_layer_ptrs),
+          reinterpret_cast<const DataType_*>(g_staging_buffer),
+          reinterpret_cast<int64_t*>(g_device_block_ids),
+          n_blocks,
+          layer_number,
+          cache_block_stride);
+    }
+  } else {
+    const cudaMemcpyKind copy_kind =
+        (mode == 0) ? cudaMemcpyDeviceToHost : cudaMemcpyHostToDevice;
+    for (int64_t layer_idx = 0; layer_idx < layer_number; layer_idx++) {
+      const paddle::Tensor& cache_gpu = cache_gpu_tensors[layer_idx];
+      data_t* cache_gpu_ptr = const_cast<data_t*>(cache_gpu.data<data_t>());
+      auto* cache_cpu_ptr = reinterpret_cast<data_t*>(cache_cpu_pointer);
+
+      for (int block_idx = 0; block_idx < n_blocks; block_idx++) {
+        auto cur_gpu_block_id = gpu_block_ids[block_idx];
+        auto cur_cpu_block_id = cpu_block_ids[block_idx];
+        auto* cache_gpu_ptr_now =
+            cache_gpu_ptr + cur_gpu_block_id * cache_block_stride;
+        auto* cache_cpu_ptr_now =
+            cache_cpu_ptr +
+            cur_cpu_block_id * cache_block_stride * layer_number +
+            layer_idx * cache_block_stride;
+
+        cudaError_t status = cudaMemcpyAsync(
+            (copy_kind == cudaMemcpyDeviceToHost) ? cache_cpu_ptr_now
+                                                  : cache_gpu_ptr_now,
+            (copy_kind == cudaMemcpyDeviceToHost) ? cache_gpu_ptr_now
+                                                  : cache_cpu_ptr_now,
+            block_bytes,
+            copy_kind,
+            stream);
+        PADDLE_ENFORCE_EQ(status,
+                          cudaSuccess,
+                          phi::errors::External("cudaMemcpyAsync failed: %s",
+                                                cudaGetErrorString(status)));
+      }
     }
   }
+
   cudaError_t sync_status = cudaStreamSynchronize(stream);
   PADDLE_ENFORCE_EQ(sync_status,
                     cudaSuccess,
                     phi::errors::External("cudaStreamSynchronize failed: %s",
                                           cudaGetErrorString(sync_status)));
 }
 
-void SwapCacheLayout(
-    const std::vector<paddle::Tensor>& cache_gpu_tensors,  // gpu
-    const int64_t& cache_cpu_ptrs,                         // cpu memory pointer
-    const std::vector<int64_t>& cache_shape,
-    const std::vector<int64_t>& gpu_block_ids,
-    const std::vector<int64_t>& cpu_block_ids,
-    int rank,
-    int mode) {
-  cudaSetDevice(rank);  // used for distributed launch
+void SwapCacheLayout(const std::vector<paddle::Tensor>& cache_gpu_tensors,
+                     const int64_t& cache_cpu_ptrs,
+                     const std::vector<int64_t>& cache_shape,
+                     const std::vector<int64_t>& gpu_block_ids,
+                     const std::vector<int64_t>& cpu_block_ids,
+                     int rank,
+                     int mode) {
+  cudaSetDevice(rank);
   assert(cache_gpu_tensors.size() > 0);
   switch (cache_gpu_tensors[0].dtype()) {
     case paddle::DataType::BFLOAT16:

examples/cache_storage/run_03b_pd_storage.sh

@@ -18,7 +18,7 @@ metadata_port=15002
 
 export MOONCAKE_MASTER_SERVER_ADDR="${master_ip}:${master_port}"
 export MOONCAKE_METADATA_SERVER="http://${master_ip}:${metadata_port}/metadata"
-export MOONCAKE_GLOBAL_SEGMENT_SIZE="50000000000"
+export MOONCAKE_GLOBAL_SEGMENT_SIZE="200000000000"
 # export MOONCAKE_PROTOCOL="tcp"
 export MOONCAKE_PROTOCOL="rdma"
 # export MOONCAKE_RDMA_DEVICES="mlx5_0"

fastdeploy/cache_manager/cache_messager.py

@@ -682,6 +682,24 @@ def prefill_layerwise_send_cache_thread(self):
             try:
                 batch_engine_signals = self.cache_prefilled_engine_ids_queue.get()
                 self.engine_worker_queue.begin_send_cache_barrier.wait()
+
+                # Storage pool mode: skip RDMA/IPC transfer, immediately notify completion
+                if envs.FD_PD_TRANSFER_VIA_STORAGE:
+                    with self.engine_cache_task_thread_lock:
+                        for engine_idx, _ in batch_engine_signals:
+                            self._maybe_wait_for_cache_task(engine_idx)
+                            task = self.idx_cache_task_dict[engine_idx]
+                            task["status"] = "finished"
+                            logger.info(
+                                f"[PD Storage] Skip RDMA transfer, mark as finished, " f"req_id: {task['request_id']}"
+                            )
+                            self.engine_worker_queue.finish_send_cache_barrier.wait()
+                            self.engine_worker_queue.put_finished_req([[task["request_id"], task["status"]]])
+                            self.engine_cache_tasks[task["current_id"]] = dict()
+                            del self.cache_info[task["request_id"]]
+                            del self.idx_cache_task_dict[task["current_id"]]
+                    continue
+
                 block_start_end_list = []
                 current_prefilled_token_num_list = []
                 for engine_index, current_step_prefilled_token_num in batch_engine_signals: