[Misc] Move dynamic seed initialization to EngineArgs (#29165)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/engine/arg_utils.py

@@ -367,7 +367,7 @@ class EngineArgs:
     config_format: str = ModelConfig.config_format
     dtype: ModelDType = ModelConfig.dtype
     kv_cache_dtype: CacheDType = CacheConfig.cache_dtype
-    seed: int | None = ModelConfig.seed
+    seed: int | None = None
     max_model_len: int | None = ModelConfig.max_model_len
     cuda_graph_sizes: list[int] | None = CompilationConfig.cudagraph_capture_sizes
     cudagraph_capture_sizes: list[int] | None = (
@@ -1188,6 +1188,20 @@ class EngineArgs:
         if check_gguf_file(self.model):
             self.quantization = self.load_format = "gguf"
 
+        # NOTE(woosuk): In V1, we use separate processes for workers (unless
+        # VLLM_ENABLE_V1_MULTIPROCESSING=0), so setting a seed here
+        # doesn't affect the user process.
+        if self.seed is None:
+            self.seed = 0
+            if not envs.VLLM_ENABLE_V1_MULTIPROCESSING:
+                logger.warning(
+                    "The global random seed is set to %d. Since "
+                    "VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may "
+                    "affect the random state of the Python process that "
+                    "launched vLLM.",
+                    self.seed,
+                )
+
         if self.disable_mm_preprocessor_cache:
             logger.warning(
                 "`--disable-mm-preprocessor-cache` is deprecated "