fix(kimi-k3): preserve dense ModelSlim MLA weights (#36603)

test/registered/expert_pack/test_kimi_k3_gguf.py

@@ -10,18 +10,29 @@ from sglang.srt.layers.quantization.gguf import (
     GGUFLinearMethod,
     _ordered_gguf_shard_ids,
 )
+from sglang.srt.layers.quantization.modelslim.modelslim import ModelSlimConfig
 from sglang.srt.model_loader.kimi_k3_gguf import (
     _kda_a_log_target_value,
     _residual_target_value,
     kimi_k3_checkpoint_targets,
     routed_expert_tensor,
 )
+from sglang.srt.models.kimi_k3 import _uses_split_gguf_kv_b
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=10, suite="base-a-test-cpu")
 
 
 class TestKimiK3GGUFMapping(unittest.TestCase):
+    def test_split_kv_capability_is_expert_pack_specific(self) -> None:
+        self.assertTrue(ModelSlimConfig.supports_kimi_k3_quantized_latent_projections)
+        self.assertFalse(_uses_split_gguf_kv_b(ModelSlimConfig))
+        self.assertTrue(
+            _uses_split_gguf_kv_b(
+                SimpleNamespace(supports_kimi_k3_split_gguf_kv_b=True)
+            )
+        )
+
     def test_maps_dense_kda_mla_moe_and_residual_tensors(self) -> None:
         cases = {
             "token_embd.weight": ("model.embed_tokens.weight",),