[XPU][NIXL] Add GPUDirect RDMA support for XPU (#35270)

Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
liuzhenwei
2026-03-03 08:42:49 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent c8b678e53e
commit 9dd656f0ea
3 changed files with 62 additions and 5 deletions
@@ -135,7 +135,10 @@ _NIXL_SUPPORTED_DEVICE = {
"cpu",
),
"tpu": ("cpu",),
"xpu": ("cpu",),
"xpu": (
"cpu",
"xpu",
),
"cpu": ("cpu",),
}
# support for oot platform by providing mapping in current_platform
@@ -945,7 +948,7 @@ class NixlConnectorWorker:
# type based on kv_buffer_device
nixl_memory_type = current_platform.get_nixl_memory_type()
if nixl_memory_type is None:
if self.kv_buffer_device == "cuda":
if self.kv_buffer_device in ["cuda", "xpu"]:
nixl_memory_type = "VRAM"
elif self.kv_buffer_device == "cpu":
nixl_memory_type = "DRAM"
+6
View File
@@ -221,6 +221,12 @@ class XPUPlatform(Platform):
vllm_config.scheduler_config.DEFAULT_MAX_NUM_BATCHED_TOKENS,
)
# In some cases, the internal memory type cache can misdetect GPU
# memory as host memory, also leading to invalid memory access.
# This cache can be disabled by setting UCX_MEMTYPE_CACHE=n.
# ref. https://openucx.readthedocs.io/en/master/faq.html
os.environ["UCX_MEMTYPE_CACHE"] = "n"
@classmethod
def support_hybrid_kv_cache(cls) -> bool:
return True