[V1] Fully Transparent Implementation of CPU Offloading (#15354)

Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2025-03-31 20:22:34 +08:00
committed by GitHub
parent e7ae3bf3d6
commit 555aa21905
12 changed files with 148 additions and 25 deletions
+39
View File
@@ -0,0 +1,39 @@
#include <torch/all.h>
#include <torch/cuda.h>
#include <cuda_runtime.h>
// This function assumes that `cpu_tensor` is a CPU tensor allocated with pinned
// memory, and that UVA (Unified Virtual Addressing) is enabled.
torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor) {
TORCH_CHECK(cpu_tensor.device().is_cpu(), "Input tensor must be on CPU");
// Get raw host pointer from CPU tensor
void* host_ptr = cpu_tensor.data_ptr();
// Get a device pointer corresponding to the pinned host memory
void* device_ptr = nullptr;
cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
TORCH_CHECK(err == cudaSuccess,
"cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
// We'll use the same sizes, strides, and dtype as the CPU tensor.
// TODO: check if layout is respected.
auto sizes = cpu_tensor.sizes();
auto strides = cpu_tensor.strides();
auto options = cpu_tensor.options().device(torch::kCUDA);
// from_blob signature: from_blob(void *data, IntArrayRef sizes, ..., Deleter,
// const TensorOptions &) Provide a no-op deleter. The CPU tensor holds the
// memory, so we don't free it here.
auto deleter = [](void*) {
// no-op, since the memory is owned by the original CPU tensor
};
torch::Tensor cuda_tensor =
torch::from_blob(device_ptr, sizes, strides, deleter, options);
TORCH_CHECK(cuda_tensor.device().is_cuda(),
"Resulting tensor is not on CUDA device");
return cuda_tensor;
}
+2
View File
@@ -119,6 +119,8 @@ void advance_step_flashinfer(
torch::Tensor& paged_kv_indices, torch::Tensor& paged_kv_indptr,
torch::Tensor& paged_kv_last_page_len, torch::Tensor& block_table_bounds);
torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor);
#ifndef USE_ROCM
torch::Tensor aqlm_gemm(const torch::Tensor& input, const torch::Tensor& codes,
const torch::Tensor& codebooks,
+4
View File
@@ -31,6 +31,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
ops.def("weak_ref_tensor(Tensor input) -> Tensor");
ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);
ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
&get_cuda_view_from_cpu_tensor);
// Attention ops
// Compute the attention between an input query and the cached
// keys/values using PagedAttention.