[Frontend] Dynamic RoPE scaling (#4638)

vllm/engine/arg_utils.py

@@ -1,5 +1,6 @@
 import argparse
 import dataclasses
+import json
 from dataclasses import dataclass
 from typing import List, Optional, Tuple, Union
 
@@ -49,6 +50,7 @@ class EngineArgs:
     disable_log_stats: bool = False
     revision: Optional[str] = None
     code_revision: Optional[str] = None
+    rope_scaling: Optional[dict] = None
     tokenizer_revision: Optional[str] = None
     quantization: Optional[str] = None
     enforce_eager: bool = False
@@ -330,6 +332,11 @@ class EngineArgs:
                             'None, we assume the model weights are not '
                             'quantized and use `dtype` to determine the data '
                             'type of the weights.')
+        parser.add_argument('--rope-scaling',
+                            default=None,
+                            type=json.loads,
+                            help='RoPE scaling configuration in JSON format. '
+                            'For example, {"type":"dynamic","factor":2.0}')
         parser.add_argument('--enforce-eager',
                             action='store_true',
                             help='Always use eager-mode PyTorch. If False, '
@@ -548,11 +555,12 @@ class EngineArgs:
         model_config = ModelConfig(
             self.model, self.tokenizer, self.tokenizer_mode,
             self.trust_remote_code, self.dtype, self.seed, self.revision,
-            self.code_revision, self.tokenizer_revision, self.max_model_len,
-            self.quantization, self.quantization_param_path,
-            self.enforce_eager, self.max_context_len_to_capture,
-            self.max_seq_len_to_capture, self.max_logprobs,
-            self.skip_tokenizer_init, self.served_model_name)
+            self.code_revision, self.rope_scaling, self.tokenizer_revision,
+            self.max_model_len, self.quantization,
+            self.quantization_param_path, self.enforce_eager,
+            self.max_context_len_to_capture, self.max_seq_len_to_capture,
+            self.max_logprobs, self.skip_tokenizer_init,
+            self.served_model_name)
         cache_config = CacheConfig(self.block_size,
                                    self.gpu_memory_utilization,
                                    self.swap_space, self.kv_cache_dtype,