[ROCm][CI] Fix `rlhf_async_new_apis` Example On ROCm (#46895)
Signed-off-by: Micah Williamson <micah.williamson@amd.com> Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com> Co-authored-by: Matthew Wong <Matthew.Wong2@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Micah Williamson committed
Jun 28, 2026 at 12:50 UTC
c2127a25c787492fea657b867a6c668a317166fb
1 file changed
+3
-4
examples/rl/rlhf_async_new_apis.py
+3
-4
@@ -190,12 +190,11 @@ class TrainModel:
190
191
192
# Build platform-specific env vars for Ray
193
-ray_env_vars = {
194
- # Prevent Ray from setting CUDA_VISIBLE_DEVICES
195
- "RAY_EXPERIMENTAL_NOSET_CUDA_ENV_VAR": "1",
196
-}
193
+ray_env_vars = {}
194
195
if current_platform.is_rocm():
196
+ # Workaround for RCCL bug. See https://github.com/ROCm/rocm-systems/issues/5756
197
+ ray_env_vars["RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES"] = "1"
198
# For ROCm, BATCH_INVARIANT vllm is not supported
199
ray_env_vars["VLLM_ROCM_USE_SKINNY_GEMM"] = "0"
200
else: