openhands

nemo_gym/global_config.py

@@ -489,4 +489,4 @@ def format_almost_server_warning(server_name: str, error: ValidationError) -> st
 {error_str}
 
     This server will NOT be started.
-    """
+    """

nemo_gym/openai_utils.py

@@ -25,6 +25,7 @@
     Union,
 )
 
+import orjson
 from openai.types.chat import (
     ChatCompletion,
     ChatCompletionAssistantMessageParam,
@@ -474,33 +475,44 @@ async def _raise_for_status(self, response: ClientResponse, request_kwargs: Dict
     async def create_chat_completion(self, **kwargs):
         request_kwargs = dict(
             url=f"{self.base_url}/chat/completions",
-            json=kwargs,
-            headers={"Authorization": f"Bearer {self.api_key}"},
+            data=orjson.dumps(kwargs),
+            headers={
+                "Authorization": f"Bearer {self.api_key}",
+                "Content-Type": "application/json",
+            },
         )
         response = await self._request(method="POST", **request_kwargs)
 
         await self._raise_for_status(response, request_kwargs)
-        return await response.json()
+        response_dict = orjson.loads(await response.read())
+        return response_dict
 
     async def create_response(self, **kwargs):
         request_kwargs = dict(
             url=f"{self.base_url}/responses",
-            json=kwargs,
-            headers={"Authorization": f"Bearer {self.api_key}"},
+            data=orjson.dumps(kwargs),
+            headers={
+                "Authorization": f"Bearer {self.api_key}",
+                "Content-Type": "application/json",
+            },
         )
         response = await self._request(method="POST", **request_kwargs)
 
-        await self._raise_for_status(response, request_kwargs)
-        return await response.json()
+        response_dict = orjson.loads(await response.read())
+        return response_dict
 
     async def create_tokenize(self, **kwargs):
         base_url = self.base_url.removesuffix("/v1")
         request_kwargs = dict(
             url=f"{base_url}/tokenize",
-            json=kwargs,
-            headers={"Authorization": f"Bearer {self.api_key}"},
+            data=orjson.dumps(kwargs),
+            headers={
+                "Authorization": f"Bearer {self.api_key}",
+                "Content-Type": "application/json",
+            },
         )
         response = await self._request(method="POST", **request_kwargs)
 
         await self._raise_for_status(response, request_kwargs)
-        return await response.json()
+        response_dict = orjson.loads(await response.read())
+        return response_dict

nemo_gym/rollout_collection.py

@@ -154,4 +154,4 @@ def collect_rollouts():  # pragma: no cover
     config = RolloutCollectionConfig.model_validate(get_global_config_dict())
     rch = RolloutCollectionHelper()
 
-    asyncio.run(rch.run_from_config(config))
+    asyncio.run(rch.run_from_config(config))

pyproject.toml

@@ -158,6 +158,11 @@ dependencies = [
     # Updated Thu Dec 04, 2025 with datasets==4.4.1
     # License: Apache 2.0 https://github.com/huggingface/datasets/blob/main/LICENSE
     "datasets",
+
+    # orjson: Fast, correct Python JSON library supporting dataclasses, datetimes, and numpy
+    # Updated: Thu Jan 08, 2026 with orjson==3.11.3
+    # License: Apache 2.0 https://github.com/ijl/orjson/blob/master/LICENSE
+    "orjson",
 ]
 
 [dependency-groups]

responses_api_agents/swe_agents/.gitignore

@@ -0,0 +1,4 @@
+swe_openhands_setup
+swe_swebench_setup
+swe_r2e_gym_setup
+swebench_results_*

responses_api_agents/swe_agents/README.md

@@ -0,0 +1,229 @@
+# Quick Start: Running SWE Agents
+
+This guide shows how to run the SWE agents that use OpenAI GPT-4.1 (or any other model) to solve real-world GitHub issues.
+
+## Prerequisites
+
+1. **Install Apptainer** (for container execution):
+```bash
+# Install Apptainer on Ubuntu/Debian
+apt install -y wget && \
+    cd /tmp && \
+    wget https://github.com/apptainer/apptainer/releases/download/v1.4.1/apptainer_1.4.1_amd64.deb && \
+    apt install -y ./apptainer_1.4.1_amd64.deb
+
+# Verify installation
+apptainer --version
+```
+
+
+## Step 1: Configure Your API Key
+
+Create or update your `env.yaml` file in the NeMo-Gym root directory:
+
+```yaml
+# For OpenAI models
+policy_base_url: https://api.openai.com/v1
+policy_api_key: {your OpenAI API key}
+policy_model_name: gpt-4.1-2025-04-14
+```
+
+You can also host a vLLM model.
+
+Start VLLM server (in separate terminal):
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct \
+  --max-model-len 131072 \
+  --enable-expert-parallel \
+  --tensor-parallel-size 4 \
+  --enable-auto-tool-choice \
+  --tool-call-parser qwen3_coder \
+  --port 8000 \
+  --enforce-eager
+```
+Then set
+```yaml
+policy_base_url: http://localhost:8000/v1
+policy_api_key: dummy
+policy_model_name: Qwen/Qwen3-Coder-30B-A3B-Instruct
+```
+
+
+## Step 2: Run the SWE Agents
+
+Start the servers with SWE-agent configuration:
+
+```bash
+# Define config paths
+# OpenAI model
+config_paths="responses_api_agents/swe_agents/configs/swebench_swe_agent.yaml,\
+responses_api_models/openai_model/configs/openai_model.yaml"
+
+or
+# vLLM model
+config_paths="responses_api_agents/swe_agents/configs/swebench_swe_agent.yaml,\
+responses_api_models/vllm_model/configs/vllm_model.yaml"
+
+# Run the servers
+# If you have pre-downloaded images, you can set the path with container_formatter, e.g.
+ng_run "+config_paths=[$config_paths]" \
+     +swe_agents.responses_api_agents.swe_agents.container_formatter=/lustre/xxx/images/swe-bench/swebench_sweb.eval.x86_64.\{instance_id\}.sif \
+     +swe_agents.responses_api_agents.swe_agents.model_server.name=vllm_model 
+
+```
+
+To run OpenHands server, simply replace the SWE-agent config path to OpenHands config 
+```bash
+responses_api_agents/swe_agents/configs/swebench_openhands.yaml
+```
+
+For how to download images and convert to .sif, you can refer to https://github.com/NVIDIA/NeMo-Skills/blob/main/nemo_skills/dataset/swe-bench/dump_images.py
+
+
+You should see output like:
+```
+INFO:     Started server process [1815588]
+INFO:     Uvicorn running on http://127.0.0.1:25347 (Press CTRL+C to quit)
+INFO:     Started server process [1815587]
+INFO:     Uvicorn running on http://127.0.0.1:56809 (Press CTRL+C to quit)
+```
+
+## Step 3: Query the Agent
+
+In a new terminal, run the client script:
+
+```bash
+python responses_api_agents/swe_agents/client.py
+```
+
+
+## Advanced usage: Run Batch  Evaluation/Data Collection
+
+For multiple problems, use rollout collection:
+
+```
+# Collect rollouts
+ng_collect_rollouts +agent_name=swe_agents \
+    +input_jsonl_fpath=swebench-verified-converted.jsonl \
+    +output_jsonl_fpath=swebench-verified.openhands.qwen3-30b-coder.jsonl \
+    +model=Qwen/Qwen3-Coder-30B-A3B-Instruct \
+    +temperature=0.7 \
+    +top_p=0.8
+```
+By default, the concurrency of ng_collect_rollouts is 100. You may want to adjust it based on your hardware configuration accordingly. 
+
+## Step 6: View Results
+
+View the collected results:
+
+```bash
+ng_viewer +jsonl_fpath=swebench-verified.openhands.qwen3-30b-coder.jsonl
+```
+
+
+## Expected Output
+
+A successful run will show:
+```json
+{
+  "responses_create_params": {
+    "background": null,
+    "include": null,
+    "input": [
+      {
+        "content": "You are OpenHands agent, a helpful AI assistant...",
+        "role": "system",
+        "type": "message"
+      },
+      {
+        "content": "I've uploaded a python code repository...",
+        "role": "user", 
+        "type": "message"
+      }
+    ],
+    "instructions": null,
+    "max_output_tokens": null,
+    "max_tool_calls": null,
+    "metadata": {
+      "instance_id": "astropy__astropy-12907",
+      "base_commit": "d16bfe05a744909de4b27f5875fe0d4ed41ce607",
+      "dataset_name": "princeton-nlp/SWE-bench_Verified",
+      "split": "test",
+      "problem_statement": "Modeling's `separability_matrix` does not compute separability correctly for nested CompoundModels\nConsider the following model:\n\n```python\nfrom astropy.modeling import models as m\nfrom astropy.modeling.separable import separability_matrix\n\ncm = m.Linear1D(10) & m.Linear1D(5)\n```\n\nIt's separability matrix as you might expect is a diagonal:\n\n```python\n>>> separability_matrix(cm)\narray([[ True, False],\n[False, True]])\n```\n\nIf I make the model more complex:\n```python\n>>> separability_matrix(m.Pix2Sky_TAN() & m.Linear1D(10) & m.Linear1D(5))\narray([[ True, True, False, False],\n[ True, True, False, False],\n[False, False, True, False],\n[False, False, False, True]])\n```\n\nThe output matrix is again, as expected, the outputs and inputs to the linear models are separable and independent of each other.\n\nIf however, I nest these compound models:\n```python\n>>> separability_matrix(m.Pix2Sky_TAN() & cm)\narray([[ True, True, False, False],\n[ True, True, False, False],\n[False, False, True, True],\n[False, False, True, True]])\n```\nSuddenly the inputs and outputs are no longer separable?\n\nThis feels like a bug to me, but I might be missing something?"
+    },
+    "model": "Qwen/Qwen/Qwen3-Coder-30B-A3B-Instruct",
+    "parallel_tool_calls": true,
+    "previous_response_id": null,
+    "prompt": null,
+    "reasoning": null,
+    "service_tier": null,
+    "store": null,
+    "temperature": 0.7,
+    "text": null,
+    "tool_choice": "auto",
+    "tools": [...]
+  },
+  "response": {
+    "id": "swebench-astropy__astropy-12907",
+    "created_at": 1757366053,
+    "error": null,
+    "incomplete_details": null,
+    "instructions": null,
+    "metadata": null,
+    "model": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
+    "object": "response",
+    "output": [
+      {
+        "id": "msg-2",
+        "content": [
+          {
+            "annotations": [],
+            "text": "I'll help you implement the necessary changes...",
+            "type": "output_text",
+            "logprobs": null
+          }
+        ],
+        "role": "assistant",
+        "status": "completed",
+        "type": "message"
+      }
+    ],
+    "parallel_tool_calls": true,
+    "temperature": null,
+    "tool_choice": "auto",
+    "tools": [...],
+    "top_p": null,
+    "background": null,
+    "max_output_tokens": null,
+    "max_tool_calls": null,
+    "previous_response_id": null,
+    "prompt": null,
+    "reasoning": null,
+    "service_tier": null,
+    "status": null,
+    "text": null,
+    "top_logprobs": null,
+    "truncation": null,
+    "usage": null,
+    "user": null
+  },
+  "reward": 1.0,
+  "swebench_metrics": {
+    "patch_is_None": false,
+    "patch_exists": true,
+    "patch_successfully_applied": true,
+    "resolved": true,
+    },
+  "resolved": 1,
+  "patch_exists": 1,
+  "patch_successfully_applied": 1,
+  "metadata": {
+    "instance_id": "astropy__astropy-12907",
+    "agent_framework": "openhands",
+    "patch_exists": true,
+    "patch_successfully_applied": true,
+    "resolved": true
+  }
+}
+```

responses_api_agents/swe_agents/__init__.py

@@ -0,0 +1,22 @@
+"""SWE-bench wrapper agent for NeMo-Gym.
+
+This module provides integration between NeMo-Skills' SWE-bench evaluation
+capabilities and NeMo-Gym's agent framework.
+"""
+
+from .app import (
+    SWEBenchRunRequest,
+    SWEBenchVerifyRequest,
+    SWEBenchVerifyResponse,
+    SWEBenchWrapper,
+    SWEBenchWrapperConfig,
+)
+
+
+__all__ = [
+    "SWEBenchWrapper",
+    "SWEBenchWrapperConfig",
+    "SWEBenchRunRequest",
+    "SWEBenchVerifyRequest",
+    "SWEBenchVerifyResponse",
+]