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 4, 21:15 UTC
[NIXL][XPU]Fix nixl import on XPU (#40430)
Signed-off-by: Soila Kavulya <soila.p.kavulya.intel.com>
Soila Kavulya committed
Apr 21, 2026 at 18:26 UTC
f90aa446629f74d4d7049771fe938e04b5c2d3ef
1 file changed
+3
-3
vllm/distributed/nixl_utils.py
+3
-3
@@ -24,7 +24,7 @@ if "UCX_RCACHE_MAX_UNRELEASED" not in os.environ:
24
os.environ["UCX_RCACHE_MAX_UNRELEASED"] = "1024"
25
26
try:
27
- if current_platform.is_cuda():
27
+ if not current_platform.is_rocm():
28
from nixl._api import nixl_agent as NixlWrapper
29
else:
30
from rixl._api import nixl_agent as NixlWrapper
@@ -35,7 +35,7 @@ except ImportError:
35
NixlWrapper = None # type: ignore[assignment, misc]
36
37
try:
38
- if current_platform.is_cuda():
38
+ if not current_platform.is_rocm():
39
from nixl._api import nixl_agent_config
40
else:
41
from rixl._api import nixl_agent_config
@@ -44,7 +44,7 @@ except ImportError:
44
logger.warning_once("NIXL agent config is not available")
45
46
try:
47
- if current_platform.is_cuda():
47
+ if not current_platform.is_rocm():
48
from nixl._bindings import nixlXferTelemetry
49
else:
50
from rixl._bindings import nixlXferTelemetry