[XPU] Enable QK Norm + RoPE fusion pass on XPU (#49394)

Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

.buildkite/intel_jobs/misc_intel.yaml

@@ -262,3 +262,25 @@ steps:
     pytest -v -s detokenizer &&
     pytest -v -s -m "not cpu_test" ./multimodal &&
     pytest -v -s utils_ --ignore=utils_/test_mem_utils.py'
+
+- label: Fusion Unit Tests
+  timeout_in_minutes: 30
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/compilation/
+    - tests/compile/passes/test_qk_norm_rope_fusion.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s compile/passes/test_qk_norm_rope_fusion.py'

vllm/compilation/passes/pass_manager.py

@@ -30,19 +30,19 @@ if rocm_aiter_ops.is_enabled():
     )
 
 if current_platform.is_cuda_alike() or current_platform.is_xpu():
+    from .fusion.qk_norm_rope_fusion import QKNormRoPEFusionPass
     from .fusion.sequence_parallelism import SequenceParallelismPass
+    from .utility.split_coalescing import SplitCoalescingPass
 
 if current_platform.is_cuda_alike():
     from .fusion.act_quant_fusion import ActivationQuantFusionPass
     from .fusion.attn_quant_fusion import AttnQuantFusionPass
     from .fusion.mla_attn_quant_fusion import MLAAttnQuantFusionPass
     from .fusion.mla_rope_kvcache_cat_fusion import MLARoPEKVCacheCatFusionPass
-    from .fusion.qk_norm_rope_fusion import QKNormRoPEFusionPass
     from .fusion.qk_norm_rope_kvcache_fusion import QkNormRopeKvCacheFusionPass
     from .fusion.rms_quant_fusion import RMSNormQuantFusionPass
     from .fusion.rope_kvcache_fusion import RopeKVCacheFusionPass
     from .utility.scatter_split_replace import ScatterSplitReplacementPass
-    from .utility.split_coalescing import SplitCoalescingPass
 
 if current_platform.is_cuda():
     from .fusion.allreduce_rms_fusion import AllReduceFusionPass

vllm/platforms/xpu.py

@@ -304,7 +304,6 @@ class XPUPlatform(Platform):
             "fuse_act_padding": "Activation + padding fusion",
             "fuse_rope_kvcache": "RoPE + KV cache fusion",
             "fuse_rope_kvcache_cat_mla": "RoPE + KV cache + MLA fusion",
-            "enable_qk_norm_rope_fusion": "QK Norm + RoPE fusion",
         }
         if compilation_config.mode != CompilationMode.NONE:
             for flag, feature_name in fusion_passes_to_disable.items():