[Bugfix] Fix prefix strings for quantized VLMs (#9772)

vllm/model_executor/models/internlm2.py

@@ -30,7 +30,8 @@ from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP
 from .utils import (is_pp_missing_parameter,
-                    make_empty_intermediate_tensors_factory, make_layers)
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
 
 
 class InternLM2MLP(nn.Module):
@@ -41,16 +42,23 @@ class InternLM2MLP(nn.Module):
         intermediate_size: int,
         hidden_act: str,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.gate_up_proj = MergedColumnParallelLinear(
-            hidden_size, [intermediate_size] * 2,
+            hidden_size,
+            [intermediate_size] * 2,
             bias=False,
-            quant_config=quant_config)
-        self.w2 = RowParallelLinear(intermediate_size,
-                                    hidden_size,
-                                    bias=False,
-                                    quant_config=quant_config)
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj",
+        )
+        self.w2 = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.w2",
+        )
         if hidden_act != "silu":
             raise ValueError(f"Unsupported activation: {hidden_act}. "
                              "Only silu is supported for now.")
@@ -75,6 +83,7 @@ class InternLM2Attention(nn.Module):
         max_position_embeddings: int = 8192,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.hidden_size = hidden_size
@@ -108,12 +117,14 @@ class InternLM2Attention(nn.Module):
             self.total_num_kv_heads,
             bias=False,
             quant_config=quant_config,
+            prefix=f"{prefix}.wqkv",
         )
         self.wo = RowParallelLinear(
             self.total_num_heads * self.head_dim,
             hidden_size,
             bias=False,
             quant_config=quant_config,
+            prefix=f"{prefix}.wo",
         )
 
         self.rotary_emb = get_rope(
@@ -123,12 +134,15 @@ class InternLM2Attention(nn.Module):
             base=rope_theta,
             rope_scaling=rope_scaling,
         )
-        self.attn = Attention(self.num_heads,
-                              self.head_dim,
-                              self.scaling,
-                              num_kv_heads=self.num_kv_heads,
-                              cache_config=cache_config,
-                              quant_config=quant_config)
+        self.attn = Attention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
+        )
 
     def split_qkv(self, qkv: torch.Tensor):
         seq_len = qkv.shape[0]
@@ -176,6 +190,7 @@ class InternLMDecoderLayer(nn.Module):
         config: PretrainedConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
@@ -192,12 +207,14 @@ class InternLMDecoderLayer(nn.Module):
             max_position_embeddings=max_position_embeddings,
             cache_config=cache_config,
             quant_config=quant_config,
+            prefix=f"{prefix}.attention",
         )
         self.feed_forward = InternLM2MLP(
             hidden_size=self.hidden_size,
             intermediate_size=config.intermediate_size,
             hidden_act=config.hidden_act,
             quant_config=quant_config,
+            prefix=f"{prefix}.feed_forward",
         )
         self.attention_norm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
@@ -251,8 +268,8 @@ class InternLM2Model(nn.Module):
         )
         self.start_layer, self.end_layer, self.layers = make_layers(
             config.num_hidden_layers,
-            lambda prefix: InternLMDecoderLayer(config, cache_config,
-                                                quant_config),
+            lambda prefix: InternLMDecoderLayer(
+                config, cache_config, quant_config, prefix=prefix),
             prefix=f"{prefix}.layers")
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.make_empty_intermediate_tensors = (
@@ -306,14 +323,19 @@ class InternLM2ForCausalLM(nn.Module, SupportsPP):
         config: PretrainedConfig,
         cache_config: Optional[CacheConfig] = None,
         quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
     ) -> None:
         super().__init__()
         self.config = config
         self.quant_config = quant_config
-        self.model = InternLM2Model(config, cache_config, quant_config)
+        self.model = InternLM2Model(config,
+                                    cache_config,
+                                    quant_config,
+                                    prefix=maybe_prefix(prefix, "model"))
         self.output = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
-                                     quant_config=quant_config)
+                                     quant_config=quant_config,
+                                     prefix=maybe_prefix(prefix, "output"))
         if self.config.tie_word_embeddings:
             self.output.weight = self.model.tok_embeddings.weight
         self.logits_processor = LogitsProcessor(config.vocab_size)