[XPU] add warning for xpu graph limitations (#49419)
Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com>
liuzhenwei committed
Jul 25, 2026 at 10:26 UTC
318b527cc2d1f672683407be05ea26a2cf1f3ea6
1 file changed
+10
vllm/platforms/xpu.py
+10
@@ -282,6 +282,16 @@ class XPUPlatform(Platform):
282
"XPU Graph is disabled by environment variable, "
283
"please set VLLM_XPU_ENABLE_XPU_GRAPH=1 to enable it."
284
)
285
+ else:
286
+ logger.warning_once(
287
+ "XPU Graph support is experimental and has known limitations: "
288
+ "(1) only single-GPU execution is supported; "
289
+ "(2) FLASH_ATTN supports PIECEWISE mode only; use TRITON_ATTN "
290
+ "for FULL mode; "
291
+ "(3) XPU Graph may increase device memory usage, "
292
+ "potentially causing OOM errors or leaving less memory "
293
+ "for the KV cache and reducing performance."
294
+ )
295
296
# Disable fusion passes not yet supported on XPU.
297
from vllm.config.compilation import CompilationMode