[XPU] update warning of XPU Graph (#50236)
Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
liuzhenwei committed
Aug 6, 2026 at 14:33 UTC
8c0f029ac7718af4400d7542bb99417cb901d220
2 files changed
+31
-7
.buildkite/intel_jobs/test-intel.yaml
+29
@@ -196,3 +196,32 @@ steps:
196
bash .buildkite/scripts/hardware_ci/run-intel-test.sh
197
'cd tests &&
198
pytest -v -s quantization/test_compressed_tensors.py::test_compressed_tensors_fp8'
199
+ - label: "XPU Graph"
200
+ depends_on:
201
+ - image-build-xpu
202
+ timeout_in_minutes: 60
203
+ device: intel_gpu
204
+ agent_tags:
205
+ label: production
206
+ gpu: 1+
207
+ mem: 32+
208
+ no_plugin: true
209
+ env:
210
+ REGISTRY: "public.ecr.aws/q9t5s3a7"
211
+ REPO: "vllm-ci-test-repo"
212
+ VLLM_TEST_DEVICE: "xpu"
213
+ source_file_dependencies:
214
+ - vllm/compilation/
215
+ - vllm/platforms/xpu.py
216
+ - vllm/v1/worker/xpu_model_runner.py
217
+ - vllm/v1/worker/xpu_worker.py
218
+ - .buildkite/intel_jobs/test-intel.yaml
219
+ commands:
220
+ - >-
221
+ bash .buildkite/scripts/hardware_ci/run-intel-test.sh
222
+ 'export VLLM_XPU_ENABLE_XPU_GRAPH=1 &&
223
+ python3 examples/basic/offline_inference/generate.py --model Qwen/Qwen3-0.6B &&
224
+ python3 examples/basic/offline_inference/generate.py --model Qwen/Qwen3-0.6B --kv-cache-dtype fp8 &&
225
+ python3 examples/basic/offline_inference/generate.py --model INCModel/Qwen3-30B-A3B-Instruct-2507-MXFP4-CT-AutoRound --max-model-len 4096'
226
+ # Temporarily disabled (multi-GPU/MXFP8 not yet supported with XPU Graph):
227
+ # python3 examples/basic/offline_inference/generate.py --model INCModel/Qwen3-30B-A3B-Instruct-2507-MXFP8-CT-AutoRound -tp 2 --max-model-len 4096
vllm/platforms/xpu.py
+2
-7
@@ -301,13 +301,8 @@ class XPUPlatform(Platform):
301
)
302
else:
303
logger.warning_once(
304
- "XPU Graph support is experimental and has known limitations: "
305
- "(1) only single-GPU execution is supported; "
306
- "(2) FLASH_ATTN supports PIECEWISE mode only; use TRITON_ATTN "
307
- "for FULL mode; "
308
- "(3) XPU Graph may increase device memory usage, "
309
- "potentially causing OOM errors or leaving less memory "
310
- "for the KV cache and reducing performance."
304
+ "XPU Graph support is experimental and currently only supports "
305
+ "single-GPU execution."
306
)
307
308
# Disable fusion passes not yet supported on XPU.