Add support to vllm==0.11.0

pyproject.toml

@@ -98,7 +98,7 @@ nanotron = [
   "tensorboardX"
 ]
 tensorboardX = ["tensorboardX"]
-vllm = ["vllm>=0.10.0,<0.10.2", "ray", "more_itertools"]
+vllm = ["vllm>=0.10.0,<0.12.0", "ray", "more_itertools"]
 sglang = ["sglang"]
 quality = ["ruff>=v0.11.0","pre-commit"]
 tests = ["pytest>=7.4.0","deepdiff","pip>=25.2"]

src/lighteval/models/vllm/vllm_model.py

@@ -40,7 +40,6 @@
 from lighteval.utils.cache_management import SampleCache, cached
 from lighteval.utils.imports import is_package_available, requires
 
-
 logger = logging.getLogger(__name__)
 
 
@@ -52,6 +51,7 @@
         destroy_distributed_environment,
         destroy_model_parallel,
     )
+    from vllm.inputs import token_inputs
     from vllm.transformers_utils.tokenizer import get_tokenizer
     from vllm.v1.engine.async_llm import AsyncEngineArgs, AsyncLLM
 
@@ -437,7 +437,8 @@ def _generate(
             @ray.remote(num_gpus=self.tensor_parallel_size)
             def run_inference_one_model(model_args: dict, sampling_params: SamplingParams, requests):
                 llm = LLM(**model_args)
-                return llm.generate(prompt_token_ids=requests, sampling_params=sampling_params)
+                requests = [token_inputs(prompt_token_ids=request) for request in requests]
+                return llm.generate(prompts=requests, sampling_params=sampling_params)
 
             # dispatch requests to all self.data_parallel_size workers, in interleaved fashion
             # interleaved important to balance context lengths across workers
@@ -454,8 +455,9 @@ def run_inference_one_model(model_args: dict, sampling_params: SamplingParams, r
                 if x is not None
             ]
         else:
+            inputs = [token_inputs(prompt_token_ids=input) for input in inputs]
             outputs = self.model.generate(
-                prompt_token_ids=inputs,
+                prompts=inputs,
                 sampling_params=sampling_params,
                 use_tqdm=True,
             )