[Intel XPU] Enable fused_moe_triton tuning on XPU and add tuned DeepSeek-OCR-2 configs (#28723)

test/registered/unit/layers/moe/test_fused_moe_common_utils.py

@@ -29,6 +29,8 @@ def test_get_model_config_supports_kimi_vl():
         n_routed_experts=64,
         num_experts_per_tok=6,
         moe_intermediate_size=1408,
+        num_hidden_layers=27,
+        first_k_dense_replace=1,
         torch_dtype=torch.bfloat16,
     )
     model_config = SimpleNamespace(
@@ -42,6 +44,8 @@ def test_get_model_config_supports_kimi_vl():
             "moonshotai/Kimi-VL-A3B-Instruct", tp_size=4, ep_size=4
         )
 
+    # The layer counts come from the text config, not the outer one, which has
+    # neither field.
     assert tuned_config == {
         "num_experts": 16,
         "topk": 6,
@@ -50,6 +54,48 @@ def test_get_model_config_supports_kimi_vl():
         "dtype": torch.bfloat16,
         "block_shape": None,
         "architecture": "KimiVLForConditionalGeneration",
+        "num_layers": 27,
+        "dense_layers": 1,
+    }
+
+
+def test_get_model_config_reports_deepseek_ocr_layer_layout():
+    """``load_topk_ids`` indexes recorded topk_ids per MoE layer.
+
+    It derives the MoE layer count as ``num_layers - dense_layers``, so a config
+    that reports either as 0 makes the tuner read the wrong (or no) recording.
+    """
+    common_utils = _load_common_utils()
+    text_config = SimpleNamespace(
+        hidden_size=1280,
+        n_routed_experts=64,
+        num_experts_per_tok=6,
+        moe_intermediate_size=896,
+        num_hidden_layers=12,
+        first_k_dense_replace=1,
+        torch_dtype=torch.bfloat16,
+    )
+    model_config = SimpleNamespace(
+        architectures=["DeepseekOCRForCausalLM"],
+        text_config=text_config,
+        get_text_config=lambda: text_config,
+    )
+
+    with patch.object(common_utils, "get_config", return_value=model_config):
+        tuned_config = common_utils.get_model_config(
+            "deepseek-ai/DeepSeek-OCR-2", tp_size=1, ep_size=1
+        )
+
+    assert tuned_config == {
+        "num_experts": 64,
+        "topk": 6,
+        "hidden_size": 1280,
+        "shard_intermediate_size": 1792,
+        "dtype": torch.bfloat16,
+        "block_shape": None,
+        "architecture": "DeepseekOCRForCausalLM",
+        "num_layers": 12,
+        "dense_layers": 1,
     }