c10d/gloo: add ibverbs backend (pytorch#153015)

d4l3k · pytorchmergebot · commit d900c68ea656 · 2025-05-08T18:26:29.000Z
Summary: X-link: pytorch/gloo#437 This provides a new "UnboundBuffer" implementation for Gloo ibverbs backend so it can be used with PyTorch. This currently is passing basic tests such as `reduce_test` and `send_recv_test` but there are a number of failures. Putting this up for review so the follow up fixes are less of a mega PR and also so we can start doing some initial testing with this E2E with PyTorch. Known issues: * support recv from any is not supported * AllreduceBcubeBase2 is failing Test Plan: ``` buck2 run mode/dbgo //gloo/test:send_recv_test_ibverbs buck2 test //gloo/test: GLOO_DEVICE_TRANSPORT=IBVERBS buck2 run @//mode/opt //caffe2/test/distributed:c10d -- -r '.*gloo.*' -f ``` We can't run any of the gloo tests in CI since none of our CI machines have ibverbs so they're disabled by default and need to be manually run. Differential Revision: D73291471 Pull Request resolved: pytorch#153015 Approved by: https://github.com/fduwjj
diff --git a/torch/csrc/distributed/c10d/GlooDeviceFactory.cpp b/torch/csrc/distributed/c10d/GlooDeviceFactory.cpp
@@ -1,5 +1,7 @@
 #include <torch/csrc/distributed/c10d/GlooDeviceFactory.hpp>
 
+#include <torch/csrc/distributed/c10d/Utils.hpp>
+
 #ifdef USE_C10D_GLOO
 
 #include <cstdlib>
@@ -19,6 +21,10 @@
 #include <gloo/transport/uv/device.h>
 #endif
 
+#if GLOO_HAVE_TRANSPORT_IBVERBS
+#include <gloo/transport/ibverbs/device.h>
+#endif
+
 // On Linux, check that the tcp transport is available.
 #ifdef __linux__
 #if !GLOO_HAVE_TRANSPORT_TCP
@@ -140,6 +146,45 @@ C10_REGISTER_CREATOR(GlooDeviceRegistry, WIN32, makeUVDevice)
 C10_REGISTER_CREATOR(GlooDeviceRegistry, UV, makeUVDevice)
 #endif
 
+#if GLOO_HAVE_TRANSPORT_IBVERBS
+static std::shared_ptr<::gloo::transport::Device> makeIBVerbsDevice(
+    const std::string& interface,
+    const std::string& hostname,
+    bool lazyInit) {
+  TORCH_CHECK(hostname.empty(), "ibverbs transport does not support hostname");
+
+  TORCH_CHECK(!lazyInit, "transport does not support lazy init");
+
+  ::gloo::transport::ibverbs::attr attr;
+  attr.name = getCvarString(
+      {
+          "TORCH_GLOO_IBV_NAME",
+      },
+      "");
+  attr.port = getCvarInt(
+      {
+          "TORCH_GLOO_IBV_PORT",
+      },
+      1);
+  attr.index = getCvarInt(
+      {
+          "TORCH_GLOO_IBV_INDEX",
+      },
+      0);
+
+  if (!interface.empty()) {
+    attr.name = interface;
+  }
+
+  // use global port
+  attr.port = 1;
+
+  return ::gloo::transport::ibverbs::CreateDevice(attr);
+}
+
+C10_REGISTER_CREATOR(GlooDeviceRegistry, IBVERBS, makeIBVerbsDevice)
+#endif
+
 namespace {
 std::shared_ptr<::gloo::transport::Device> makeGlooDevice(
     const std::string& interfaceName,
