[Feature] support deepseek v3/r1/v3.2 (#78)

* [Feature] support deepseek v3/r1/v3.2

* fix gpt_oss

* update readme

* update readme

---------

Co-authored-by: hanhaowen <hanhaowen@baidu.com>

vllm_kunlun/models/__init__.py

@@ -80,6 +80,14 @@ def register_model():
     ModelRegistry.register_model(
         "GptOssForCausalLM",
         "vllm_kunlun.models.gpt_oss:GptOssForCausalLM")   
-    
+
+    ModelRegistry.register_model(
+        "DeepseekV3ForCausalLM",
+        "vllm_kunlun.models.deepseek_v2:DeepseekV3ForCausalLM")
+
+    ModelRegistry.register_model(
+        "DeepseekV32ForCausalLM",
+        "vllm_kunlun.models.deepseek_v2:DeepseekV3ForCausalLM")
+
 def register_quant_method():
     """to do"""

vllm_kunlun/models/gpt_oss.py

@@ -16,7 +16,7 @@ from vllm.distributed import (get_ep_group, get_pp_group,
                               get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size,
                               tensor_model_parallel_all_gather)
-from vllm_kunlun.ops.fused_moe.layer import FusedMoE
+from vllm.model_executor.layers.fused_moe import FusedMoE
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (QKVParallelLinear,
                                                RowParallelLinear)
@@ -176,7 +176,7 @@ class MLPBlock(torch.nn.Module):
             x = sequence_parallel_chunk(x)
 
         g = self.router(x)
-        x = self.experts(hidden_states=x, router_logits=g, linear_weights=self.router.weight)
+        x = self.experts(hidden_states=x, router_logits=g)
 
         if self.is_sequence_parallel:
             x = tensor_model_parallel_all_gather(x.contiguous(), 0)

vllm_kunlun/models/mimo_v2_flash.py

@@ -21,7 +21,7 @@ from vllm.distributed import (
     tensor_model_parallel_all_gather,
 )
 from vllm.logger import init_logger
-from vllm_kunlun.ops.fused_moe.layer import FusedMoE
+from vllm.model_executor.layers.fused_moe import FusedMoE
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (
     MergedColumnParallelLinear,
@@ -185,8 +185,7 @@ class MiMoV2MoE(nn.Module):
             gate_input = hidden_states
         router_logits = self.gate(gate_input)
         final_hidden_states = self.experts(
-            hidden_states=hidden_states, router_logits=router_logits, linear_weights=self.gate.weight
-        )
+            hidden_states=hidden_states, router_logits=router_logits)
 
         return final_hidden_states.squeeze(0) if is_input_1d else final_hidden_states