[Model] Add K2 Horizon FP8 checkpoint support (#38033)

Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>

python/sglang/srt/models/xllm.py

@@ -661,10 +661,10 @@ def _validate_mova_config(
                 "the released checkpoints persist float32 dtype metadata but "
                 "their weights and validated runtime contract are BF16."
             )
-        if quant_config is not None:
+        if quant_config is not None and quant_config.get_name() != "compressed_tensors":
             raise ValueError(
-                "Native xLLM/K2 Horizon serving does not support quantized "
+                "Native xLLM/K2 Horizon serving supports only "
-                "model weights"
+                "compressed-tensors quantized model weights"
             )
 
         runtime = get_exec()
@@ -1686,6 +1686,14 @@ class XllmModel(nn.Module):
 class XllmForCausalLM(nn.Module):
     fall_back_to_pt_during_load = False
 
+    # Quantized checkpoints store these projections separately. This mapping
+    # lets quantization configs resolve fused runtime modules and their ignore
+    # lists consistently.
+    packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+
     def __init__(
         self,
         config: PretrainedConfig,

test/registered/unit/models/test_xllm.py

@@ -266,15 +266,35 @@ def test_native_xllm_requires_bfloat16(monkeypatch):
         _validate_mova_config(config, quant_config=None)
 
 
-def test_native_xllm_rejects_quantized_weights(monkeypatch):
+def test_native_xllm_accepts_compressed_tensors_quantization(monkeypatch):
     config = XllmConfig(num_values=0, num_experts=0)
+    quant_config = SimpleNamespace(get_name=lambda: "compressed_tensors")
+    monkeypatch.setattr(torch, "get_default_dtype", lambda: torch.bfloat16)
+
+    with get_context().override_server_args(**_native_runtime_config()):
+        _validate_mova_config(config, quant_config=quant_config)
+
+
+def test_native_xllm_rejects_other_quantization(monkeypatch):
+    config = XllmConfig(num_values=0, num_experts=0)
+    quant_config = SimpleNamespace(get_name=lambda: "awq")
     monkeypatch.setattr(torch, "get_default_dtype", lambda: torch.bfloat16)
 
     with (
         get_context().override_server_args(**_native_runtime_config()),
-        pytest.raises(ValueError, match="does not support quantized"),
+        pytest.raises(ValueError, match="supports only compressed-tensors"),
     ):
-        _validate_mova_config(config, quant_config=object())
+        _validate_mova_config(config, quant_config=quant_config)
+
+
+def test_native_xllm_declares_quantized_fused_module_mapping():
+    assert XllmForCausalLM.packed_modules_mapping == {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"],
+    }
+    assert K2HorizonForCausalLM.packed_modules_mapping == (
+        XllmForCausalLM.packed_modules_mapping
+    )
 
 
 def test_native_xllm_accepts_bfloat16_without_expert_remapping(monkeypatch):