[Misc] Set default value of seed to None (#14274)

Signed-off-by: மனோஜ்குமார் பழனிச்சாமி <smartmanoj42857@gmail.com>

tests/distributed/test_torchrun_example.py

@@ -25,7 +25,8 @@ llm = LLM(model="facebook/opt-125m",
           tensor_parallel_size=2,
           distributed_executor_backend="external_launcher",
           gpu_memory_utilization=random.uniform(0.7, 0.9),
-          swap_space=random.randint(1, 4))
+          swap_space=random.randint(1, 4),
+          seed=0)
 
 outputs = llm.generate(prompts, sampling_params)
 

tests/entrypoints/llm/test_encode.py

@@ -34,7 +34,8 @@ def llm():
               max_num_batched_tokens=32768,
               tensor_parallel_size=1,
               gpu_memory_utilization=0.75,
-              enforce_eager=True)
+              enforce_eager=True,
+              seed=0)
 
     with llm.deprecate_legacy_api():
         yield weakref.proxy(llm)

tests/entrypoints/llm/test_guided_generate.py

@@ -21,7 +21,7 @@ GUIDED_DECODING_BACKENDS = ["outlines", "lm-format-enforcer", "xgrammar"]
 def llm():
     # pytest caches the fixture so we use weakref.proxy to
     # enable garbage collection
-    llm = LLM(model=MODEL_NAME, max_model_len=1024)
+    llm = LLM(model=MODEL_NAME, max_model_len=1024, seed=0)
 
     with llm.deprecate_legacy_api():
         yield weakref.proxy(llm)

tests/entrypoints/openai/test_chat_echo.py

@@ -24,6 +24,8 @@ def server():
         "4080",
         "--chat-template",
         DUMMY_CHAT_TEMPLATE,
+        "--seed",
+        "0",
     ]
 
     with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:

tests/entrypoints/openai/test_metrics.py

@@ -47,6 +47,8 @@ def default_server_args():
         "--enforce-eager",
         "--max-num-seqs",
         "128",
+        "--seed",
+        "0",
     ]
 
 

tests/entrypoints/openai/test_root_path.py

@@ -30,6 +30,8 @@ def server():
         "/" + ROOT_PATH,
         "--chat-template",
         DUMMY_CHAT_TEMPLATE,
+        "--seed",
+        "0",
     ]
     envs = os.environ.copy()
 

vllm/engine/arg_utils.py

@@ -104,7 +104,7 @@ class EngineArgs:
     config_format: ConfigFormat = ConfigFormat.AUTO
     dtype: str = 'auto'
     kv_cache_dtype: str = 'auto'
-    seed: int = 0
+    seed: Optional[int] = None
     max_model_len: Optional[int] = None
     # Note: Specifying a custom executor backend by passing a class
     # is intended for expert use only. The API may change without

vllm/entrypoints/llm.py

@@ -169,7 +169,7 @@ class LLM:
         quantization: Optional[str] = None,
         revision: Optional[str] = None,
         tokenizer_revision: Optional[str] = None,
-        seed: int = 0,
+        seed: Optional[int] = None,
         gpu_memory_utilization: float = 0.9,
         swap_space: float = 4,
         cpu_offload_gb: float = 0,

vllm/utils.py

@@ -644,7 +644,7 @@ def create_kv_caches_with_random_flash(
     head_size: int,
     cache_dtype: Optional[Union[str, torch.dtype]],
     model_dtype: Optional[Union[str, torch.dtype]] = None,
-    seed: int = 0,
+    seed: Optional[int] = None,
     device: Optional[str] = "cuda",
 ) -> tuple[list[torch.Tensor], list[torch.Tensor]]:
     from vllm.platforms import current_platform
@@ -681,7 +681,7 @@ def create_kv_caches_with_random(
     head_size: int,
     cache_dtype: Optional[Union[str, torch.dtype]],
     model_dtype: Optional[Union[str, torch.dtype]] = None,
-    seed: int = 0,
+    seed: Optional[int] = None,
     device: Optional[str] = "cuda",
 ) -> tuple[list[torch.Tensor], list[torch.Tensor]]: