[CPU] [BF16] Call fused_experts_cpu, weight_packed_linear and bmm_cpu kernel in DeepSeek model (#6641)

Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>

python/sglang/srt/layers/linear.py

@@ -30,7 +30,12 @@ from sglang.srt.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,
 )
-from sglang.srt.utils import set_weight_attrs
+from sglang.srt.utils import (
+    _process_weight_after_loading,
+    cpu_has_amx_support,
+    is_cpu,
+    set_weight_attrs,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -52,6 +57,9 @@ WEIGHT_LOADER_V2_SUPPORTED = [
     "IPEXAWQLinearMethod",
 ]
 
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
+
 
 def adjust_marlin_shard(param, shard_size, shard_offset):
     marlin_tile_size = getattr(param, "marlin_tile_size", None)
@@ -165,6 +173,10 @@ class UnquantizedLinearMethod(LinearMethodBase):
         layer.register_parameter("weight", weight)
         set_weight_attrs(weight, extra_weight_attrs)
 
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        if _is_cpu and _is_cpu_amx_available:
+            _process_weight_after_loading(layer, ["weight"])
+
     def apply(
         self,
         layer: torch.nn.Module,
@@ -172,6 +184,11 @@ class UnquantizedLinearMethod(LinearMethodBase):
         bias: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
 
+        if getattr(layer, "use_intel_amx_backend", False):
+            return torch.ops.sgl_kernel.weight_packed_linear(
+                x, layer.weight, bias, True  # is_vnni
+            )
+
         return F.linear(x, layer.weight, bias)