From 8d4fcf2f7bff1cca95ad17cc7520343004a22328 Mon Sep 17 00:00:00 2001
From: jianan-gu <jianan.gu@intel.com>
Date: Thu, 19 Mar 2026 13:57:55 +0800
Subject: [PATCH] [CPU] Fix MoE layer support for DeepSeek-OCR models (#12555)

---
 python/sglang/srt/models/deepseek.py     | 40 ++++++++++++++++++------
 python/sglang/srt/models/deepseek_ocr.py | 35 ++++++++++++++++++++-
 2 files changed, 65 insertions(+), 10 deletions(-)

diff --git a/python/sglang/srt/models/deepseek.py b/python/sglang/srt/models/deepseek.py
index 675bb1673..39ce8749c 100644
--- a/python/sglang/srt/models/deepseek.py
+++ b/python/sglang/srt/models/deepseek.py
@@ -48,7 +48,12 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 )
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
-from sglang.srt.utils import add_prefix
+from sglang.srt.utils import add_prefix, cpu_has_amx_support, is_cpu
+
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
+if _is_cpu and _is_cpu_amx_available:
+    import sgl_kernel  # noqa: F401
 
 
 class DeepseekMLP(nn.Module):
@@ -176,14 +181,31 @@ class DeepseekMoE(nn.Module):
         # router_logits: (num_tokens, n_experts)
         router_logits, _ = self.gate(hidden_states)
         topk_output = self.topk(hidden_states, router_logits)
-        final_hidden_states = fused_moe.fused_moe(
-            hidden_states,
-            w1=self.w1,
-            w2=self.w2,
-            topk_output=topk_output,
-            moe_runner_config=MoeRunnerConfig(inplace=True),
-        )
-
+        if _is_cpu and _is_cpu_amx_available:
+            topk_weights, topk_ids, _ = topk_output
+            final_hidden_states = torch.ops.sgl_kernel.fused_experts_cpu(
+                hidden_states,
+                self.w1,
+                self.w2,
+                topk_weights,
+                topk_ids,
+                False,  # inplace # See [Note] inplace should be False in fused_experts.
+                0,  # CPUQuantMethod.UNQUANT,
+                None,  # w1_scale
+                None,  # w2_scale
+                None,  # w1_zp
+                None,  # w2_zp
+                None,  # block_size
+                True,  # is_vnni
+            )
+        else:
+            final_hidden_states = fused_moe.fused_moe(
+                hidden_states,
+                w1=self.w1,
+                w2=self.w2,
+                topk_output=topk_output,
+                moe_runner_config=MoeRunnerConfig(inplace=True),
+            )
         if self.config.n_shared_experts is not None:
             final_hidden_states = final_hidden_states + shared_output
         final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
diff --git a/python/sglang/srt/models/deepseek_ocr.py b/python/sglang/srt/models/deepseek_ocr.py
index f87c06d32..02ccfb4d6 100644
--- a/python/sglang/srt/models/deepseek_ocr.py
+++ b/python/sglang/srt/models/deepseek_ocr.py
@@ -41,6 +41,10 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.deepseek import DeepseekForCausalLM
 from sglang.srt.models.deepseek_v2 import DeepseekV2ForCausalLM, DeepseekV3ForCausalLM
 from sglang.srt.models.transformers import maybe_prefix
+from sglang.srt.utils import cpu_has_amx_support, is_cpu
+
+_is_cpu_amx_available = cpu_has_amx_support()
+_is_cpu = is_cpu()
 
 NestedTensors: TypeAlias = Union[
     list["NestedTensors"],
@@ -1772,7 +1776,6 @@ class DeepseekOCRForCausalLM(nn.Module):
 
         params_dict = dict(self.named_parameters())
         loaded_params: Set[str] = set()
-
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
@@ -1852,6 +1855,36 @@ class DeepseekOCRForCausalLM(nn.Module):
             raise RuntimeError(
                 f"Some weights are not initialized from checkpoints: {unloaded_params}"
             )
+        self.post_load_weights()
+
+    def post_load_weights(self):
+        if _is_cpu and _is_cpu_amx_available:
+            from sglang.srt.layers.amx_utils import _amx_process_weight_after_loading
+
+            layer_ids = int(self.config.num_hidden_layers)
+            first_k_dense_replace_id = (
+                self.config.first_k_dense_replace
+                if hasattr(self.config, "first_k_dense_replace")
+                else -1
+            )
+            moe_layer_freq_id = (
+                self.config.moe_layer_freq
+                if hasattr(self.config, "moe_layer_freq")
+                else 1
+            )
+            for layer_id in range(0, layer_ids):
+                if (
+                    layer_id >= first_k_dense_replace_id
+                    and layer_id % moe_layer_freq_id == 0
+                ):
+                    if (
+                        hasattr(self.model, "model")
+                        and hasattr(self.model.model, "layers")
+                        and hasattr(self.model.model.layers[layer_id], "mlp")
+                    ):
+                        self_moe = self.model.model.layers[layer_id].mlp
+                        if hasattr(self_moe, "w1") and hasattr(self_moe, "w2"):
+                            _amx_process_weight_after_loading(self_moe, ["w1", "w2"])
 
 
 EntryClass = [DeepseekOCRForCausalLM]