fix: enhance human_readable_int function (#30337)

Signed-off-by: Andy Xie <andy.xning@gmail.com>

tests/engine/test_arg_utils.py

@@ -350,21 +350,35 @@ def test_human_readable_model_len():
     assert args.max_model_len == 1_000_000
     args = parser.parse_args(["--max-model-len", "10k"])
     assert args.max_model_len == 10_000
+    args = parser.parse_args(["--max-model-len", "2g"])
+    assert args.max_model_len == 2_000_000_000
+    args = parser.parse_args(["--max-model-len", "2t"])
+    assert args.max_model_len == 2_000_000_000_000
 
     # Capital
     args = parser.parse_args(["--max-model-len", "3K"])
-    assert args.max_model_len == 1024 * 3
+    assert args.max_model_len == 2**10 * 3
     args = parser.parse_args(["--max-model-len", "10M"])
     assert args.max_model_len == 2**20 * 10
+    args = parser.parse_args(["--max-model-len", "4G"])
+    assert args.max_model_len == 2**30 * 4
+    args = parser.parse_args(["--max-model-len", "4T"])
+    assert args.max_model_len == 2**40 * 4
 
     # Decimal values
     args = parser.parse_args(["--max-model-len", "10.2k"])
     assert args.max_model_len == 10200
     # ..truncated to the nearest int
-    args = parser.parse_args(["--max-model-len", "10.212345k"])
+    args = parser.parse_args(["--max-model-len", "10.2123451234567k"])
     assert args.max_model_len == 10212
+    args = parser.parse_args(["--max-model-len", "10.2123451234567m"])
+    assert args.max_model_len == 10212345
+    args = parser.parse_args(["--max-model-len", "10.2123451234567g"])
+    assert args.max_model_len == 10212345123
+    args = parser.parse_args(["--max-model-len", "10.2123451234567t"])
+    assert args.max_model_len == 10212345123456
 
     # Invalid (do not allow decimals with binary multipliers)
-    for invalid in ["1a", "pwd", "10.24", "1.23M"]:
+    for invalid in ["1a", "pwd", "10.24", "1.23M", "1.22T"]:
         with pytest.raises(ArgumentError):
-            args = parser.parse_args(["--max-model-len", invalid])
+            parser.parse_args(["--max-model-len", invalid])

vllm/engine/arg_utils.py

@@ -1783,6 +1783,7 @@ class EngineArgs:
         except Exception:
             # This is only used to set default_max_num_batched_tokens
             device_memory = 0
+            device_name = ""
 
         # NOTE(Kuntai): Setting large `max_num_batched_tokens` for A100 reduces
         # throughput, see PR #17885 for more details.
@@ -2042,11 +2043,13 @@ def human_readable_int(value):
             "k": 10**3,
             "m": 10**6,
             "g": 10**9,
+            "t": 10**12,
         }
         binary_multiplier = {
             "K": 2**10,
             "M": 2**20,
             "G": 2**30,
+            "T": 2**40,
         }
 
         number, suffix = match.groups()