[DSA][MLA] Tiny refactor on DeepSeek to make it reusable for different backends (#26656)

Signed-off-by: MengqingCao <cmq0113@163.com>

vllm/model_executor/models/deepseek_mtp.py

@@ -17,9 +17,13 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding,
 )
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 
-from .deepseek_v2 import DeepseekV2DecoderLayer, get_spec_layer_idx_from_weight_name
+from .deepseek_v2 import (
+    DeepseekV2DecoderLayer,
+    get_spec_layer_idx_from_weight_name,
+)
 from .interfaces import SupportsPP
 from .utils import maybe_prefix
 
@@ -56,6 +60,8 @@ class DeepSeekMultiTokenPredictorLayer(nn.Module):
         self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
 
+        self.device = current_platform.device_type
+
         self.is_v32 = hasattr(config, "index_topk")
         if self.is_v32:
             topk_tokens = config.index_topk
@@ -63,7 +69,7 @@ class DeepSeekMultiTokenPredictorLayer(nn.Module):
                 vllm_config.scheduler_config.max_num_batched_tokens,
                 topk_tokens,
                 dtype=torch.int32,
-                device="cuda",
+                device=self.device,
             )
         else:
             topk_indices_buffer = None