You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
- Supports automatic tuning in homogeneous scenarios via flagcxTuner.
9
+
- Integrates automated PyTorch API tests into CI/CD.
4
10
-**[2025/08]** Released [v0.4](https://github.com/FlagOpen/FlagCX/tree/release/v0.4):
5
11
- Supports heterogeneous training of ERNIE4.5 on Nvidia and Iluvatar GPUs with Paddle + FlagCX.
6
12
- Enables more robust and flexible deployments with full support of heterogeneous communication across arbitrary NIC configurations (bug fixes).
@@ -92,7 +98,7 @@ FlagCX also integrates with upper-layer applications such as PyTorch and PaddleP
92
98
2. Build the library with different flags targeting to different platforms:
93
99
```sh
94
100
cd FlagCX
95
-
make [USE_NVIDIA/USE_ILUVATAR_COREX/USE_CAMBRICON/USE_GLOO/USE_MPI/USE_METAX/USE_MUSA/USE_KUNLUNXIN/USE_DU/USE_ASCEND]=1
101
+
make [USE_NVIDIA/USE_ILUVATAR_COREX/USE_CAMBRICON/USE_GLOO/USE_MPI/USE_METAX/USE_MUSA/USE_KUNLUNXIN/USE_DU/USE_ASCEND/USE_AMD]=1
96
102
```
97
103
The default install path is set to `build/`, you can manually set`BUILDDIR` to specify the build path. You may also define `DEVICE_HOME` and `CCL_HOME` to indicate the install paths of device runtime and communication libraries.
98
104
@@ -117,6 +123,9 @@ All tests support the same set of arguments:
117
123
* Performance
118
124
*`-w, <warmup iteration count>` number of warmup iterations (not timed). Default: 5.
119
125
*`-n, <iteration count>` number of iterations. Default: 20.
126
+
* Test Operation
127
+
*`-R, <0/1>`enablelocal buffer registration on send/recv buffers. Default: 0.
128
+
*`-s, <OCT/DEC/HEX>` specify MPI communication split mode. Default: 0
0 commit comments