[CI Patch] Qwen 2.5 CI Fixes for Intel XPU (#3083)

Signed-off-by: Daniel Huang <daniel1.huang@intel.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

.buildkite/pipeline-intel.yaml

@@ -13,4 +13,4 @@ steps:
           VLLM_VERSION: "v0.20.0"
         priority: 100
         timeout_in_minutes: 60
-        soft_fail: true
+        soft_fail: false

tests/helpers/stage_config.py

@@ -306,7 +306,11 @@ _CI_OVERLAYS: dict[str, dict[str, Any]] = {
                         "max_num_batched_tokens": 16384,
                         "max_model_len": 16384,
                     },
-                    {"stage_id": 1, "gpu_memory_utilization": 0.5},
+                    {
+                        "stage_id": 1,
+                        "gpu_memory_utilization": 0.5,
+                        "default_sampling_params": {"max_tokens": 2048},
+                    },
                     {
                         "stage_id": 2,
                         "gpu_memory_utilization": 0.3,

vllm_omni/config/stage_config.py

@@ -677,6 +677,15 @@ def _apply_platform_overrides(
             base.devices = devices
         for key, val in overrides.items():
             if hasattr(base, key):
+                # Deep-merge dict-valued fields listed in _DEEP_MERGE_KEYS so
+                # platform overlays don't silently clobber sibling keys (e.g.
+                # setting default_sampling_params={max_tokens: 2048} must not
+                # drop temperature / top_p / top_k from the base stage).
+                if key in _DEEP_MERGE_KEYS and isinstance(val, dict):
+                    base_val = getattr(base, key, None)
+                    if isinstance(base_val, dict):
+                        setattr(base, key, {**base_val, **val})
+                        continue
                 setattr(base, key, val)
             else:
                 base.engine_extras[key] = val