[Core] Loading model from S3 using RunAI Model Streamer as optional loader (#10192)

Signed-off-by: OmerD <omer@run.ai>

vllm/config.py

@@ -29,6 +29,7 @@ from vllm.transformers_utils.config import (
     get_hf_text_config, get_pooling_config,
     get_sentence_transformer_tokenizer_config, is_encoder_decoder,
     try_get_generation_config, uses_mrope)
+from vllm.transformers_utils.utils import is_s3
 from vllm.utils import (GiB_bytes, LayerBlockType, cuda_device_count_stateless,
                         get_cpu_memory, print_warning_once, random_uuid,
                         resolve_obj_by_qualname)
@@ -256,6 +257,8 @@ class ModelConfig:
                    f"'Please instead use `--hf-overrides '{hf_override!r}'`")
             warnings.warn(DeprecationWarning(msg), stacklevel=2)
 
+        self.maybe_pull_model_tokenizer_for_s3(model, tokenizer)
+
         # The tokenizer version is consistent with the model version by default.
         if tokenizer_revision is None:
             self.tokenizer_revision = revision
@@ -357,6 +360,39 @@ class ModelConfig:
         self._verify_cuda_graph()
         self._verify_bnb_config()
 
+    def maybe_pull_model_tokenizer_for_s3(self, model: str,
+                                          tokenizer: str) -> None:
+        """
+        Pull the model config or tokenizer to a temporary 
+        directory in case of S3.
+
+        Args:
+            model: The model name or path.
+            tokenizer: The tokenizer name or path.
+
+        """
+        if is_s3(model) or is_s3(tokenizer):
+            try:
+                from vllm.transformers_utils.s3_utils import S3Model
+            except ImportError as err:
+                raise ImportError(
+                    "Please install Run:ai optional dependency "
+                    "to use the S3 capabilities. "
+                    "You can install it with: pip install vllm[runai]"
+                ) from err
+
+            if is_s3(model):
+                self.s3_model = S3Model()
+                self.s3_model.pull_files(model, allow_pattern=["*config.json"])
+                self.model_weights = self.model
+                self.model = self.s3_model.dir
+
+            if is_s3(tokenizer):
+                self.s3_tokenizer = S3Model()
+                self.s3_tokenizer.pull_files(
+                    model, ignore_pattern=["*.pt", "*.safetensors", "*.bin"])
+                self.tokenizer = self.s3_tokenizer.dir
+
     def _init_multimodal_config(
         self, limit_mm_per_prompt: Optional[Mapping[str, int]]
     ) -> Optional["MultiModalConfig"]:
@@ -1099,6 +1135,7 @@ class LoadFormat(str, enum.Enum):
     GGUF = "gguf"
     BITSANDBYTES = "bitsandbytes"
     MISTRAL = "mistral"
+    RUNAI_STREAMER = "runai_streamer"
 
 
 @dataclass