Add Dynamo renderer transport

pyproject.toml

@@ -126,10 +126,9 @@ prime-sandboxes = false
 renderers = false
 
 [tool.uv.sources]
-# Pinned to renderers main until the next PyPI release lands; drop after.
-# fe67f9f = renderers main: PR #4 squash-merge — construction-time
-# preserve_*_thinking flags on create_renderer / create_renderer_pool.
-renderers = { git = "https://github.com/PrimeIntellect-ai/renderers.git", rev = "fe67f9f" }
+# Pinned to renderers PR #11 until the next PyPI release lands; drop after.
+# 1f3de65 = Dynamo chat nvext transport for token-in /chat/completions.
+renderers = { git = "https://github.com/PrimeIntellect-ai/renderers.git", rev = "7ca1ab3" }
 
 [tool.uv.extra-build-dependencies]
 flash-attn = [{ requirement = "torch", match-runtime = true }]

tests/test_renderer_client.py

@@ -155,6 +155,48 @@ async def test_renderer_client_rejects_empty_dict_native_response():
         await client.raise_from_native_response({})
 
 
+@pytest.mark.asyncio
+async def test_get_native_response_passes_renderer_transport_to_generate(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    client = object.__new__(RendererClient)
+    client._client = object()
+    client._config = vf.ClientConfig(
+        client_type="renderer",
+        renderer_transport="dynamo_chat_nvext",
+    )
+    renderer = _BridgeRenderer()
+    calls = []
+
+    client._get_renderer_or_pool = lambda model: renderer  # type: ignore[method-assign]
+
+    async def fake_get_incremental_prompt_ids(**kwargs):
+        return [10, 20]
+
+    async def fake_generate(**kwargs):
+        calls.append(kwargs)
+        return {"content": "ok"}
+
+    monkeypatch.setattr(
+        "verifiers.clients.renderer_client._get_incremental_prompt_ids",
+        fake_get_incremental_prompt_ids,
+    )
+    monkeypatch.setattr("verifiers.clients.renderer_client.generate", fake_generate)
+
+    response = await client.get_native_response(
+        prompt=[{"role": "user", "content": "hi"}],
+        model="test-model",
+        sampling_args={"temperature": 0.1},
+        tools=None,
+        state={"trajectory": []},
+    )
+
+    assert response == {"content": "ok"}
+    assert len(calls) == 1
+    assert calls[0]["transport"] == "dynamo_chat_nvext"
+    assert calls[0]["prompt_ids"] == [10, 20]
+
+
 @pytest.mark.asyncio
 async def test_from_native_response_uses_request_id_and_token_lengths():
     """vLLM's /inference/v1/generate returns ``request_id`` (not ``id``) and

verifiers/clients/renderer_client.py

@@ -530,6 +530,9 @@ async def get_native_response(
             or sampling_params.pop("cache_salt", None),
             priority=args.get("priority") or sampling_params.pop("priority", None),
             extra_headers=args.get("extra_headers"),
+            transport=getattr(
+                self._config, "renderer_transport", "prime_vllm_generate"
+            ),
         )
 
     async def raise_from_native_response(self, response: dict[str, Any]) -> None:

verifiers/types.py

@@ -45,6 +45,7 @@
     "anthropic_messages",
     "nemorl_chat_completions",
 ]
+RendererTransport = Literal["prime_vllm_generate", "dynamo_chat_nvext"]
 MessageType = Literal["chat", "completion"]  # deprecated
 
 
@@ -581,6 +582,7 @@ class ClientConfig(BaseModel):
     renderer: str = "auto"
     renderer_model_name: str | None = None
     renderer_pool_size: int | None = None
+    renderer_transport: RendererTransport = "prime_vllm_generate"
     tool_parser: str | None = None
     reasoning_parser: str | None = None
     preserve_all_thinking: bool = False