[Misc] Remove deprecated arg for cuda graph capture (#9864)

Signed-off-by: Roger Wang <ywang@roblox.com>

vllm/engine/arg_utils.py

@@ -126,7 +126,6 @@ class EngineArgs:
     tokenizer_revision: Optional[str] = None
     quantization: Optional[str] = None
     enforce_eager: Optional[bool] = None
-    max_context_len_to_capture: Optional[int] = None
     max_seq_len_to_capture: int = 8192
     disable_custom_all_reduce: bool = False
     tokenizer_pool_size: int = 0
@@ -504,14 +503,6 @@ class EngineArgs:
                             help='Always use eager-mode PyTorch. If False, '
                             'will use eager mode and CUDA graph in hybrid '
                             'for maximal performance and flexibility.')
-        parser.add_argument('--max-context-len-to-capture',
-                            type=int,
-                            default=EngineArgs.max_context_len_to_capture,
-                            help='Maximum context length covered by CUDA '
-                            'graphs. When a sequence has context length '
-                            'larger than this, we fall back to eager mode. '
-                            '(DEPRECATED. Use --max-seq-len-to-capture instead'
-                            ')')
         parser.add_argument('--max-seq-len-to-capture',
                             type=int,
                             default=EngineArgs.max_seq_len_to_capture,
@@ -939,7 +930,6 @@ class EngineArgs:
             quantization=self.quantization,
             quantization_param_path=self.quantization_param_path,
             enforce_eager=self.enforce_eager,
-            max_context_len_to_capture=self.max_context_len_to_capture,
             max_seq_len_to_capture=self.max_seq_len_to_capture,
             max_logprobs=self.max_logprobs,
             disable_sliding_window=self.disable_sliding_window,