A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 23:00 UTC
[CI][amd] Revert NIXL connector change to avoid crash (#32570)
Signed-off-by: Qiang Li <qiang.li2@amd.com> Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
qli88 committed
Jan 19, 2026 at 12:39 UTC
a0490be8f1f232ef1350234c899bc5159c1491cd
3 files changed
+8
-8
.buildkite/test-amd.yaml
+2
-2
@@ -1451,7 +1451,7 @@ steps:
1451
- bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt
1452
1453
- label: NixlConnector PD accuracy tests (Distributed) # 30min
1454
- mirror_hardwares: [amdexperimental]
1454
+ mirror_hardwares: [amdexperimental, amdproduction]
1455
agent_pool: mi325_4
1456
# grade: Blocking
1457
timeout_in_minutes: 30
@@ -1465,7 +1465,7 @@ steps:
1465
- VLLM_ATTENTION_BACKEND=ROCM_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1466
1467
- label: DP EP NixlConnector PD accuracy tests (Distributed) # 15min
1468
- mirror_hardwares: [amdexperimental]
1468
+ mirror_hardwares: [amdexperimental, amdproduction]
1469
agent_pool: mi325_4
1470
# grade: Blocking
1471
timeout_in_minutes: 15
docker/Dockerfile.rocm
+1
-1
@@ -385,5 +385,5 @@ RUN echo "VLLM_BASE_IMAGE=${BASE_IMAGE}" >> ${COMMON_WORKDIR}/versions.txt
385
CMD ["/bin/bash"]
386
387
#Set entrypoint for vllm-openai official images
388
-FROM final As vllm-openai
388
+FROM final AS vllm-openai
389
ENTRYPOINT ["vllm", "serve"]
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+5
-5
@@ -89,20 +89,20 @@ logger = init_logger(__name__)
89
90
# Lazy import nixl_wrapper to avoid loading nixl_bindings if nixl is not used
91
try:
92
- if "UCX_MEM_MMAP_HOOK_MODE" not in os.environ:
92
+ if "UCX_RCACHE_MAX_UNRELEASED" not in os.environ:
93
# avoid a memory leak in UCX when using NIXL on some models
94
# see: https://github.com/vllm-project/vllm/issues/24264
95
if "nixl" in sys.modules or "rixl" in sys.modules:
96
logger.warning(
97
- "NIXL was already imported, we can't disable UCX mmap hooks. "
98
- "Please set UCX_MEM_MMAP_HOOK_MODE to 'none' manually."
97
+ "NIXL was already imported, we can't reset UCX_RCACHE_MAX_UNRELEASED. "
98
+ "Please set it to '1024' manually."
99
)
100
else:
101
logger.info(
102
- "Setting UCX_MEM_MMAP_HOOK_MODE to 'none' to avoid a rare "
102
+ "Setting UCX_RCACHE_MAX_UNRELEASED to '1024' to avoid a rare "
103
"memory leak in UCX when using NIXL."
104
)
105
- os.environ["UCX_MEM_MMAP_HOOK_MODE"] = "none"
105
+ os.environ["UCX_RCACHE_MAX_UNRELEASED"] = "1024"
106
107
if not current_platform.is_rocm():
108
from nixl._api import nixl_agent as NixlWrapper