[Model]: Support internlm3 (#12037)

docs/source/models/supported_models.md

@@ -216,6 +216,11 @@ See [this page](#generative-models) for more information on how to use generativ
   - `internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc.
   - ✅︎
   - ✅︎
+* - `InternLM3ForCausalLM`
+  - InternLM3
+  - `internlm/internlm3-8b-instruct`, etc.
+  - ✅︎
+  - ✅︎
 * - `JAISLMHeadModel`
   - Jais
   - `inceptionai/jais-13b`, `inceptionai/jais-13b-chat`, `inceptionai/jais-30b-v3`, `inceptionai/jais-30b-chat-v3`, etc.

tests/models/registry.py

@@ -85,6 +85,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
                                             trust_remote_code=True),
     "InternLM2VEForCausalLM": _HfExamplesInfo("OpenGVLab/Mono-InternVL-2B",
                                               trust_remote_code=True),
+    "InternLM3ForCausalLM": _HfExamplesInfo("internlm/internlm3-8b-instruct",
+                                            trust_remote_code=True),
     "JAISLMHeadModel": _HfExamplesInfo("inceptionai/jais-13b-chat"),
     "JambaForCausalLM": _HfExamplesInfo("ai21labs/AI21-Jamba-1.5-Mini"),
     "LlamaForCausalLM": _HfExamplesInfo("meta-llama/Meta-Llama-3-8B"),

vllm/model_executor/models/llama.py

@@ -97,20 +97,19 @@ class LlamaMLP(nn.Module):
 
 class LlamaAttention(nn.Module):
 
-    def __init__(
-        self,
-        config: LlamaConfig,
-        hidden_size: int,
-        num_heads: int,
-        num_kv_heads: int,
-        rope_theta: float = 10000,
-        rope_scaling: Optional[Dict[str, Any]] = None,
-        max_position_embeddings: int = 8192,
-        quant_config: Optional[QuantizationConfig] = None,
-        bias: bool = False,
-        cache_config: Optional[CacheConfig] = None,
-        prefix: str = "",
-    ) -> None:
+    def __init__(self,
+                 config: LlamaConfig,
+                 hidden_size: int,
+                 num_heads: int,
+                 num_kv_heads: int,
+                 rope_theta: float = 10000,
+                 rope_scaling: Optional[Dict[str, Any]] = None,
+                 max_position_embeddings: int = 8192,
+                 quant_config: Optional[QuantizationConfig] = None,
+                 bias: bool = False,
+                 cache_config: Optional[CacheConfig] = None,
+                 prefix: str = "",
+                 bias_o_proj: bool = False) -> None:
         super().__init__()
         layer_idx = extract_layer_index(prefix)
         self.hidden_size = hidden_size
@@ -150,7 +149,7 @@ class LlamaAttention(nn.Module):
         self.o_proj = RowParallelLinear(
             input_size=self.total_num_heads * self.head_dim,
             output_size=hidden_size,
-            bias=bias,
+            bias=bias_o_proj,
             quant_config=quant_config,
             prefix=f"{prefix}.o_proj",
         )
@@ -231,6 +230,11 @@ class LlamaDecoderLayer(nn.Module):
         # Support internlm/internlm-7b with bias
         attention_bias = getattr(config, "attention_bias", False) or getattr(
             config, "bias", False)
+        bias_o_proj = attention_bias
+        # support internlm/internlm3-8b with qkv_bias
+        if hasattr(config, 'qkv_bias'):
+            attention_bias = config.qkv_bias
+
         self.self_attn = LlamaAttention(
             config=config,
             hidden_size=self.hidden_size,
@@ -242,6 +246,7 @@ class LlamaDecoderLayer(nn.Module):
             max_position_embeddings=max_position_embeddings,
             quant_config=quant_config,
             bias=attention_bias,
+            bias_o_proj=bias_o_proj,
             cache_config=cache_config,
             prefix=f"{prefix}.self_attn",
         )

vllm/model_executor/models/registry.py

@@ -60,6 +60,7 @@ _TEXT_GENERATION_MODELS = {
     "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
     "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
     "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
+    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
     "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
     "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
     "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),