Enable CPU device on SGLang (#2806)

python/sglang/srt/models/deepseek_v2.py

@@ -49,6 +49,7 @@ from sglang.srt.layers.quantization.fp8_utils import (
     normalize_e4m3fn_to_e4m3fnuz,
 )
 from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.rotary_embedding import get_rope_wrapper
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
@@ -271,13 +272,14 @@ class DeepseekV2Attention(nn.Module):
             quant_config=quant_config,
         )
         rope_scaling["rope_type"] = "deepseek_yarn"
-        self.rotary_emb = get_rope(
+        self.rotary_emb = get_rope_wrapper(
             qk_rope_head_dim,
             rotary_dim=qk_rope_head_dim,
             max_position=max_position_embeddings,
             base=rope_theta,
             rope_scaling=rope_scaling,
             is_neox_style=False,
+            device=global_server_args_dict["device"],
         )
 
         if rope_scaling: