Support minicpm3-4b (#2465)

* Support minicpm3-4b

* update doc and fix typo

* comments

* sum -> max

* add TODO

* use pytorch engine's rotary embedding

README.md

@@ -140,6 +140,7 @@ For detailed inference benchmarks in more devices and more settings, please refe
   <li>Phi-3-mini (3.8B)</li>
   <li>Phi-3.5-mini (3.8B)</li>
   <li>Phi-3.5-MoE (16x3.8B)</li>
+  <li>MiniCPM3 (4B)</li>
 </ul>
 </td>
 <td>

README_ja.md

@@ -139,6 +139,7 @@ LMDeploy TurboMindエンジンは卓越した推論能力を持ち、さまざ
   <li>Phi-3-mini (3.8B)</li>
   <li>Phi-3.5-mini (3.8B)</li>
   <li>Phi-3.5-MoE (16x3.8B)</li>
+  <li>MiniCPM3 (4B)</li>
 </ul>
 </td>
 <td>

README_zh-CN.md

@@ -141,6 +141,7 @@ LMDeploy TurboMind 引擎拥有卓越的推理能力，在各种规模的模型
   <li>Phi-3-mini (3.8B)</li>
   <li>Phi-3.5-mini (3.8B)</li>
   <li>Phi-3.5-MoE (16x3.8B)</li>
+  <li>MiniCPM3 (4B)</li>
 </ul>
 </td>
 <td>

docs/en/supported_models/supported_models.md

@@ -64,6 +64,7 @@ The TurboMind engine doesn't support window attention. Therefore, for models tha
 |     QWen2      | 0.5B - 72B  | LLM  |    Yes    |   No    |  No  |  Yes  |
 |  DeepSeek-MoE  |     16B     | LLM  |    Yes    |   No    |  No  |  No   |
 |  DeepSeek-V2   |  16B, 236B  | LLM  |    Yes    |   No    |  No  |  No   |
+|    MiniCPM3    |     4B      | LLM  |    Yes    |   No    |  No  |  No   |
 |     Gemma      |    2B-7B    | LLM  |    Yes    |   No    |  No  |  No   |
 |      Dbrx      |    132B     | LLM  |    Yes    |   No    |  No  |  No   |
 |   StarCoder2   |   3B-15B    | LLM  |    Yes    |   No    |  No  |  No   |

docs/zh_cn/supported_models/supported_models.md

@@ -64,6 +64,7 @@ turbomind 引擎不支持 window attention。所以，对于应用了 window att
 |     QWen2      | 0.5B - 72B  | LLM  |    Yes    |   No    |  No  |  Yes  |
 |  DeepSeek-MoE  |     16B     | LLM  |    Yes    |   No    |  No  |  No   |
 |  DeepSeek-V2   |  16B, 236B  | LLM  |    Yes    |   No    |  No  |  No   |
+|    MiniCPM3    |     4B      | LLM  |    Yes    |   No    |  No  |  No   |
 |     Gemma      |    2B-7B    | LLM  |    Yes    |   No    |  No  |  No   |
 |      Dbrx      |    132B     | LLM  |    Yes    |   No    |  No  |  No   |
 |   StarCoder2   |   3B-15B    | LLM  |    Yes    |   No    |  No  |  No   |

lmdeploy/model.py

@@ -887,6 +887,7 @@ def match(cls, model_path: str) -> Optional[str]:
 
 
 @MODELS.register_module(name='minicpmv-2d6')
+@MODELS.register_module(name='minicpm3')
 @MODELS.register_module(name='qwen')
 class Qwen7BChat(BaseChatTemplate):
     """Chat template for Qwen-7B-Chat."""
@@ -924,6 +925,8 @@ def match(cls, model_path: str) -> Optional[str]:
             return 'qwen'
         if 'minicpm-v-2_6' in model_path.lower():
             return 'minicpmv-2d6'
+        if 'minicpm3-' in model_path.lower():
+            return 'minicpm3'
 
 
 @MODELS.register_module(name='codellama')

lmdeploy/pytorch/configurations/minicpm3.py

@@ -0,0 +1,32 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from lmdeploy.pytorch.config import ModelConfig
+
+from .builder import AutoModelConfigBuilder
+
+
+class MiniCPM3ModelConfigBuilder(AutoModelConfigBuilder):
+
+    @classmethod
+    def condition(cls, hf_config):
+        """config."""
+        return hf_config.architectures[0] in ['MiniCPM3ForCausalLM']
+
+    @classmethod
+    def build(cls, hf_config, model_path: str = None):
+        """build."""
+        head_dim = (hf_config.qk_nope_head_dim + hf_config.qk_rope_head_dim)
+        k_head_dim = head_dim
+        v_head_dim = head_dim
+        num_attention_heads = hf_config.num_attention_heads
+        num_key_value_heads = hf_config.num_key_value_heads
+        return ModelConfig(hidden_size=hf_config.hidden_size,
+                           num_layers=hf_config.num_hidden_layers,
+                           num_attention_heads=num_attention_heads,
+                           num_key_value_heads=num_key_value_heads,
+                           bos_token_id=hf_config.bos_token_id,
+                           eos_token_id=hf_config.eos_token_id,
+                           head_dim=head_dim,
+                           k_head_dim=k_head_dim,
+                           v_head_dim=v_head_dim,
+                           vocab_size=hf_config.vocab_size,
+                           multi_query_attention=False)