Skip to content

fix(ep): keep D2H handles in device memory #7

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
fergusfinn wants to merge 1 commit into from
Closed

fix(ep): keep D2H handles in device memory #7

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
52 changes: 18 additions & 34 deletions ep/src/uccl_ep.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -385,12 +385,16 @@ class Buffer {
device_index, low_latency_mode);
num_d2h_channel_addrs = static_cast<int>(host_addrs.size());
if (num_d2h_channel_addrs > 0) {
CUDA_CHECK(cudaMallocManaged(
&d_handle_objs, num_d2h_channel_addrs * sizeof(d2hq::D2HHandle)));

CUDA_CHECK(cudaMallocManaged(
&d_handles, num_d2h_channel_addrs * sizeof(uint64_t)));

size_t const handle_objs_bytes =
num_d2h_channel_addrs * sizeof(d2hq::D2HHandle);
size_t const handle_ptrs_bytes =
num_d2h_channel_addrs * sizeof(uint64_t);
CUDA_CHECK(cudaMalloc(&d_handle_objs, handle_objs_bytes));
CUDA_CHECK(cudaMalloc(&d_handles, handle_ptrs_bytes));

std::vector<d2hq::D2HHandle> host_handle_objs(
num_d2h_channel_addrs);
std::vector<uint64_t> host_handles(num_d2h_channel_addrs);
for (int i = 0; i < num_d2h_channel_addrs; ++i) {
#ifndef USE_MSCCLPP_FIFO_BACKEND
void* host_ptr = reinterpret_cast<void*>(host_addrs[i]);
Expand All @@ -401,40 +405,20 @@ class Buffer {
#else
dev_ptr = host_ptr;
#endif
d_handle_objs[i].init_from_dev_ptr(dev_ptr);
d_handles[i] = reinterpret_cast<uint64_t>(&d_handle_objs[i]);
host_handle_objs[i].init_from_dev_ptr(dev_ptr);
host_handles[i] = reinterpret_cast<uint64_t>(&d_handle_objs[i]);
#else
Comment on lines +408 to 410
auto* fifo = reinterpret_cast<mscclpp::Fifo*>(host_addrs[i]);
mscclpp::FifoDeviceHandle h = fifo->deviceHandle();
d_handle_objs[i].init_from_host_value(h);
d_handles[i] = reinterpret_cast<uint64_t>(d_handle_objs + i);
host_handle_objs[i].init_from_host_value(h);
host_handles[i] = reinterpret_cast<uint64_t>(d_handle_objs + i);
#endif
Comment on lines +413 to 415
}

// Prefetch so the device immediately sees initialized contents
#if !defined(__HIP_PLATFORM_AMD__) && !defined(__HIPCC__) && \
CUDA_VERSION >= 12000
// CUDA 12+: cudaMemPrefetchAsync(ptr, count, cudaMemLocation, flags,
// stream)
cudaMemLocation loc;
loc.type = cudaMemLocationTypeDevice;
loc.id = device_index;
CUDA_CHECK(cudaMemPrefetchAsync(
d_handle_objs, num_d2h_channel_addrs * sizeof(d2hq::D2HHandle),
loc, 0));
CUDA_CHECK(cudaMemPrefetchAsync(
d_handles, num_d2h_channel_addrs * sizeof(uint64_t), loc, 0));
#else
// CUDA 11.x / HIP: cudaMemPrefetchAsync(ptr, count, dstDevice,
// stream)
CUDA_CHECK(cudaMemPrefetchAsync(
d_handle_objs, num_d2h_channel_addrs * sizeof(d2hq::D2HHandle),
device_index, 0));
CUDA_CHECK(cudaMemPrefetchAsync(
d_handles, num_d2h_channel_addrs * sizeof(uint64_t),
device_index));
#endif
CUDA_CHECK(cudaDeviceSynchronize());
CUDA_CHECK(cudaMemcpy(d_handle_objs, host_handle_objs.data(),
handle_objs_bytes, cudaMemcpyHostToDevice));
CUDA_CHECK(cudaMemcpy(d_handles, host_handles.data(),
handle_ptrs_bytes, cudaMemcpyHostToDevice));
}
// Allocate device memory for IPC base pointers
CUDA_CHECK(
Expand Down