The sample generated

The global state for the generation process.

Configuration for a single server group. Attributes: worker_type: One of "regular", "prefill", "decode", "placeholder",

A tiny char-level tokenizer that models the Qwen3.5 assistant formatting rule. The critical behavior we need for these tests is: 1. `add_gene

Configuration for a single model deployment. Attributes: name: Unique name for this model (e.g. "actor", "reward"). model_path: H

Configuration for SGLang engine deployment. Loaded from ``--sglang-config`` YAML file. **Config format**:: sglang: - name

OpenAI format assistant message structure

A base class for our AI agents.

Simplified asynchronous rollout worker, using threads instead of processes Supports continuous running, independent of rollout function lifec

Minimal config class (simulating your argparse) Replace this with your real arguments or load them dynamically.

Micro-batch iterator over rollout dicts. Supports either fixed contiguous micro-batches or an explicit per-step index schedule (for dynamic b

Configuration for a single evaluation dataset.

Minimal interaction environment for multi-turn geo3k with a scoring tool. The model is expected to emit a <tool_call>{...}</tool_call> paylo

GLM-4.6V vision-language model for Megatron training. Wraps an HF vision encoder (only on first PP stage) together with a standard Megatron C

Adapter class that converts sglang tool call parsing results to OpenAI compatible format This class encapsulates existing tool call parsing

OpenAI format tool call structure

Python code sandbox, provides safe code execution environment

A group of ray actors Functions start with 'async' should return list of object refs Args: args (Namespace): Arguments for the a

The agent responsible for rewriting solutions.

Health monitor for rollout engines. The monitor runs continuously once started, but can be paused/resumed based on whether the engines are of

A model served behind a shared router, with one or more server groups. Each RolloutServer represents one model deployed behind a single router.

The agent responsible for selecting solutions.

A group of homogeneous SGLang engines with the same configuration. All engines in a group share the same tp_size / nodes_per_engine / pg. A R

The agent responsible for generating and improving solutions.

Tool registry, manages available tools and their execution

A trainable version of ToolCallingAgent that uses sglang rollout for training. This agent combines the original ToolCallingAgent functionali

Base class that defines the explicit contract for interaction environments.

Message processor for DeepSeek-R1-Distill-Qwen/Llama series models. Only supports single-turn conversation, no tool calls. Prompt format

Bridge implementation for Qwen2 models. This class extends LLMBridge to provide specific configurations and optimizations for Qwen2 mode

Bridge implementation for Qwen2 models. This class extends LLMBridge to provide specific configurations and optimizations for Qwen2 mode

Bridge for GLM-4.7-Flash (glm4_moe_lite) models. Extends DeepseekV3Bridge with: - Dynamic MTP layer indexing (parent hardcodes layer 61

Bridge between HuggingFace GLM-4.6V and the Megatron VL model.

Provider that creates Glm4vMoeVLModel. Inherits from Qwen3MoEModelProvider to reuse MoE + TransformerConfig infra. Defined at module level (n

Bridge implementation for GPT-OSS models. Handles weight conversion between preprocessed GPT-OSS HF format (BF16 per-expert) and Megatro