From 5c8455d6ea87712cf711863c3a466059bea779e8 Mon Sep 17 00:00:00 2001
From: QingshuChen <qingshu.chen714@gmail.com>
Date: Wed, 3 Feb 2021 15:52:59 +0800
Subject: [PATCH] try again if kunlun memory malloc failed (#30855)

* try again if kunlun memory malloc failed

* minor
---
 paddle/fluid/memory/allocation/naive_best_fit_allocator.cc | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc b/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc
index dbea74e7e0f..6306ad5ffc8 100644
--- a/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc
+++ b/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc
@@ -135,6 +135,11 @@ void *Alloc<platform::XPUPlace>(const platform::XPUPlace &place, size_t size) {
                         "Baidu Kunlun Card is properly installed.",
                         ret));
   ret = xpu_malloc(reinterpret_cast<void **>(&p), size);
+  if (ret != XPU_SUCCESS) {
+    std::cout << "xpu memory malloc(" << size << ") failed, try again\n";
+    xpu_wait();
+    ret = xpu_malloc(reinterpret_cast<void **>(&p), size);
+  }
   PADDLE_ENFORCE_EQ(
       ret, XPU_SUCCESS,
       platform::errors::External(
-- 
GitLab