[CPU] Fix argument issues in qkv_proj_with_rope_fused_weight and bmm… (#21367)

Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

python/sglang/srt/model_executor/cpu_graph_runner.py

@@ -222,6 +222,7 @@ def register_fake_ops():
         use_fp8_w8a16,
         qkv_a_proj_scale,
         q_b_proj_scale,
+        w_scale,
         is_vnni,
         block_size,
         q_lora_rank,

python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_fused_rope_cpu.py

@@ -100,7 +100,7 @@ class DeepseekMLACpuForwardMixin:
                         else None
                     )
                 ),
-                self.w_scale,
+                self.w_scale if self.qkv_proj_with_rope_is_fp8 else None,
                 True,  # is_vnni
                 self.weight_block_size,
                 self.q_lora_rank,
@@ -145,7 +145,7 @@ class DeepseekMLACpuForwardMixin:
             attn_output.transpose(0, 1),
             self.w_vc,
             True,  # is_vnni
-            self.w_scale,  # scale
+            self.w_scale if self.qkv_proj_with_rope_is_fp8 else None,  # scale
         )
         attn_output = output
         output, _ = self.o_proj(attn_output)