[DYNAMO] pass renderer transport to verifiers

packages/prime-rl-configs/src/prime_rl/configs/orchestrator.py

@@ -1339,5 +1339,4 @@ def resolve_env_config(self):
             if is_vllm:
                 env.sampling.extra_body.setdefault("top_k", -1)
                 env.sampling.extra_body.setdefault("min_p", 0.0)
-                env.sampling.extra_body.setdefault("return_token_ids", True)
         return self

pyproject.toml

@@ -29,6 +29,7 @@ dependencies = [
     "uvloop>=0.21.0",
     "torchtitan",
     "verifiers",
+    "renderers",
     "dion",
     "tilelang>=0.1.8",
     "flash-linear-attention",
@@ -150,6 +151,7 @@ prime-tunnel = false
 prime-evals = false
 # Trusted git/URL sources pinned by rev or wheel URL (defensive: belt-and-suspenders)
 verifiers = false
+renderers = false
 vllm-router = false
 dion = false
 pydantic-config = false
@@ -160,7 +162,8 @@ nixl-cu12 = false
 [tool.uv.sources]
 prime-rl-configs = { workspace = true }
 torch = { index = "pytorch-cu128" }
-verifiers = { git = "https://github.com/PrimeIntellect-ai/verifiers.git", rev = "3b77145" }
+verifiers = { git = "https://github.com/PrimeIntellect-ai/verifiers.git", rev = "9485194f" }
+renderers = { git = "https://github.com/PrimeIntellect-ai/renderers.git", rev = "17005dd" }
 torchtitan = { git = "https://github.com/pytorch/torchtitan", rev = "a1fdd7e" }
 dion = { git = "https://github.com/samsja/dion.git", rev = "d891eeb" }
 transformers = { git = "https://github.com/huggingface/transformers.git", rev = "c1c3424" }

src/prime_rl/utils/client.py

@@ -310,6 +310,7 @@ def setup_clients(
 ) -> list[vf.ClientConfig]:
     clients = []
     client_idx = 0
+    renderer_transport = "dynamo" if client_type == "renderer" and client_config.backend == "dynamo" else "vllm"
     for base_url in client_config.base_url:
         for dp_rank in range(client_config.dp_rank_count):
             headers = client_config.headers.copy()
@@ -321,6 +322,7 @@ def setup_clients(
                     client_type=client_type,
                     renderer=renderer_name,
                     renderer_model_name=renderer_model_name,
+                    renderer_transport=renderer_transport,
                     renderer_pool_size=renderer_pool_size,
                     tool_parser=tool_parser,
                     reasoning_parser=reasoning_parser,

src/prime_rl/utils/elastic.py

@@ -203,6 +203,7 @@ def _rebuild_clients(self) -> None:
                 base_url=urls,
                 api_key_var=self.client_config.api_key_var,
                 headers=self.client_config.headers,
+                backend=self.client_config.backend,
                 dp_rank_count=self.client_config.dp_rank_count,
                 extra_headers_from_state=self.client_config.extra_headers_from_state,
             )

tests/unit/test_configs.py

@@ -156,6 +156,15 @@ def test_removed_fused_lm_head_chunk_size_field_is_rejected():
         TrainerModelConfig.model_validate({"fused_lm_head_chunk_size": "auto"})
 
 
+def test_orchestrator_env_defaults_do_not_force_return_token_ids():
+    config = OrchestratorConfig()
+
+    extra_body = config.train.env[0].sampling.extra_body
+    assert extra_body["top_k"] == -1
+    assert extra_body["min_p"] == 0.0
+    assert "return_token_ids" not in extra_body
+
+
 def test_selective_activation_checkpointing_requires_custom_impl():
     with pytest.raises(ValidationError, match="Selective activation checkpointing requires model.impl='custom'"):
         TrainerModelConfig.model_validate({"impl": "hf", "ac": {"mode": "selective"}})

tests/unit/utils/test_client.py

@@ -89,6 +89,18 @@ def test_setup_clients_assigns_renderer_model_name():
     assert clients[0].renderer_model_name == "Qwen/Qwen3-VL-4B-Instruct"
 
 
+def test_setup_clients_uses_dynamo_transport_for_dynamo_renderer():
+    client_config = ClientConfig(
+        base_url=["http://worker-a:8000/v1"],
+        api_key_var="PRIME_API_KEY",
+        backend="dynamo",
+    )
+
+    clients = setup_clients(client_config, client_type="renderer")
+
+    assert clients[0].renderer_transport == "dynamo"
+
+
 def test_setup_clients_preserves_chat_client_defaults():
     client_config = ClientConfig(
         base_url=["http://worker-a:8000/v1"],

tests/unit/utils/test_elastic.py

@@ -417,6 +417,7 @@ def test_elastic_clients_preserve_renderer_model_name_when_model_name_updates():
         client_config.headers = {}
         client_config.extra_headers_from_state = {}
         client_config.dp_rank_count = 1
+        client_config.backend = "vllm"
 
         pool = ElasticInferencePool(
             client_config=client_config,