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

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

vllm/config/model.py

@@ -146,9 +146,12 @@ class ModelConfig:
     - "bfloat16" for a balance between precision and range.\n
     - "float" is shorthand for FP32 precision.\n
     - "float32" for FP32 precision."""
-    seed: int | None = None
-    """Random seed for reproducibility. Initialized to None in V0, but
-    initialized to 0 in V1."""
+    seed: int = 0
+    """Random seed for reproducibility.
+
+    We must set the global seed because otherwise,
+    different tensor parallel workers would sample different tokens,
+    leading to inconsistent results."""
     hf_config: PretrainedConfig = field(init=False)
     """The Hugging Face config of the model."""
     hf_text_config: PretrainedConfig = field(init=False)
@@ -415,7 +418,7 @@ class ModelConfig:
     def __post_init__(
         self,
         # Multimodal config init vars
-        limit_mm_per_prompt: dict[str, int] | None,
+        limit_mm_per_prompt: dict[str, int | dict[str, int]] | None,
         enable_mm_embeds: bool | None,
         media_io_kwargs: dict[str, dict[str, Any]] | None,
         mm_processor_kwargs: dict[str, Any] | None,
@@ -428,23 +431,6 @@ class ModelConfig:
         skip_mm_profiling: bool | None,
         video_pruning_rate: float | None,
     ) -> None:
-        # Set the default seed to 0 in V1.
-        # 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. However, without a consistent seed,
-        # different tensor parallel workers would sample different tokens,
-        # leading to inconsistent results.
-        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,
-                )
-
         # Keep set served_model_name before maybe_model_redirect(self.model)
         self.served_model_name = get_served_model_name(
             self.model, self.served_model_name
@@ -1151,12 +1137,6 @@ class ModelConfig:
         self,
         parallel_config: ParallelConfig,
     ) -> None:
-        if parallel_config.distributed_executor_backend == "external_launcher":
-            assert self.seed is not None, (
-                "Seed must be set when using external launcher backend to "
-                "make sure sampling results are the same across workers."
-            )
-
         total_num_attention_heads = getattr(
             self.hf_text_config, "num_attention_heads", 0
         )

vllm/config/speculative.py

@@ -9,6 +9,7 @@ from pydantic import Field, SkipValidation, model_validator
 from pydantic.dataclasses import dataclass
 from typing_extensions import Self
 
+from vllm.config.model import ModelConfig
 from vllm.config.parallel import ParallelConfig
 from vllm.config.utils import config
 from vllm.logger import init_logger
@@ -18,10 +19,8 @@ if TYPE_CHECKING:
     from transformers import PretrainedConfig
 
     import vllm.model_executor.layers.quantization as me_quant
-    from vllm.config import ModelConfig
 else:
     PretrainedConfig = Any
-    ModelConfig = Any
 
     me_quant = LazyLoader(
         "model_executor", globals(), "vllm.model_executor.layers.quantization"
@@ -316,10 +315,6 @@ class SpeculativeConfig:
             self.prompt_lookup_min = 0
 
             if self.model is not None:
-                # TODO: Move this import to the top once `ModelConfig`
-                # lives in `vllm.config.model`.
-                from vllm.config import ModelConfig
-
                 self.draft_model_config = ModelConfig(
                     model=self.model,
                     runner="draft",

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 "

vllm/v1/worker/tpu_worker.py

@@ -106,9 +106,6 @@ class TPUWorker:
                 "Profiling enabled. Traces will be saved to: %s", self.profile_dir
             )
 
-        if self.model_config.seed is None:
-            self.model_config.seed = 0
-
     def initialize_cache(self, num_gpu_blocks: int, num_cpu_blocks: int) -> None:
         self.cache_config.num_gpu_blocks = num_gpu_blocks
         self.cache_config.num_cpu_blocks = num_cpu_blocks