Rename InputMetadata -> ForwardBatch (#1543)

python/sglang/srt/models/llama.py

@@ -43,7 +43,7 @@ from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.torchao_utils import apply_torchao_config_
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.model_executor.forward_batch_info import InputMetadata
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 
 
 class LlamaMLP(nn.Module):
@@ -162,12 +162,12 @@ class LlamaAttention(nn.Module):
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
-        input_metadata: InputMetadata,
+        forward_batch: ForwardBatch,
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
         q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
         q, k = self.rotary_emb(positions, q, k)
-        attn_output = self.attn(q, k, v, input_metadata)
+        attn_output = self.attn(q, k, v, forward_batch)
         output, _ = self.o_proj(attn_output)
         return output
 
@@ -221,7 +221,7 @@ class LlamaDecoderLayer(nn.Module):
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
-        input_metadata: InputMetadata,
+        forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         # Self Attention
@@ -233,7 +233,7 @@ class LlamaDecoderLayer(nn.Module):
         hidden_states = self.self_attn(
             positions=positions,
             hidden_states=hidden_states,
-            input_metadata=input_metadata,
+            forward_batch=forward_batch,
         )
 
         # Fully Connected
@@ -270,7 +270,7 @@ class LlamaModel(nn.Module):
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        input_metadata: InputMetadata,
+        forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
         if input_embeds is None:
@@ -283,7 +283,7 @@ class LlamaModel(nn.Module):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,
-                input_metadata,
+                forward_batch,
                 residual,
             )
         hidden_states, _ = self.norm(hidden_states, residual)
@@ -310,12 +310,12 @@ class LlamaForCausalLM(nn.Module):
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        input_metadata: InputMetadata,
+        forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
     ) -> LogitsProcessorOutput:
-        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
+        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
         return self.logits_processor(
-            input_ids, hidden_states, self.lm_head.weight, input_metadata
+            input_ids, hidden_states, self.lm_head.weight, forward_batch
         )
 
     def get_hidden_dim(self, module_name):