strands-sglang 0.3.2

SGLang model provider for Strands Agents SDK with Token-in/Token-out support for agentic RL training.

Strands-SGLang

SGLang model provider for Strands Agents SDK with Token-in/Token-out rollouts for on-policy agentic RL training (no retokenization drift).

Features

This package is designed to make the serving-oriented agent scaffold Strands Agents SDK training-ready by exposing end-to-end, token-level rollouts from SGLang while reusing Strands’ customizable agent loop.

• Token-In/Token-Out rollouts (token IDs + logprobs/masks): no retokenization drift
• Strict, on-policy tool-call parsing: no heuristic repair or post-processing; tool calls are parsed exactly as generated by models
• Native SGLang /generate: high-throughput, non-streaming rollouts

For RL environment integration, please refer to strands-env

Requirements

• Python 3.10+
• Strands Agents SDK
• SGLang server running with your model
• HuggingFace tokenizer for the model

Installation

pip install strands-sglang strands-agents-tools

Or install from source with development dependencies:

git clone https://github.com/horizon-rl/strands-sglang.git
cd strands-sglang
pip install -e ".[dev]"

Quick Start

1. Start SGLang Server

python -m sglang.launch_server \
    --model-path Qwen/Qwen3-4B-Instruct-2507 \
    --port 30000 \
    --host 0.0.0.0

2. Basic Agent

import asyncio
from transformers import AutoTokenizer
from strands import Agent
from strands_tools import calculator
from strands_sglang import SGLangClient, SGLangModel

async def main():
    client = SGLangClient(base_url="http://localhost:30000")
    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
    model = SGLangModel(client=client, tokenizer=tokenizer)
    agent = Agent(model=model, tools=[calculator])

    result = await agent.invoke_async("What is 25 * 17?")
    print(result)

    # Access token data for RL training
    print(f"Tokens: {model.token_manager.token_ids}")
    print(f"Loss mask: {model.token_manager.loss_mask}")
    print(f"Logprobs: {model.token_manager.logprobs}")

asyncio.run(main())

Training with slime

For RL training with slime, SGLangModel eliminates the retokenization step, see an concrete example at slime/examples/strands_sglang:

import logging
from strands import Agent, tool
from strands_sglang import SGLangModel, ToolLimiter, get_client_from_slime_args
from strands_sglang.tool_parsers import HermesToolParser
from slime.rollout.sglang_rollout import GenerateState
from slime.utils.types import Sample

SYSTEM_PROMPT = "..."
MAX_TOOL_ITERS = 5
MAX_TOOL_CALLS = None  # No limit


@tool
def execute_python_code(code: str):
    """Execute Python code and return the output."""
    ...


async def generate(args, sample: Sample, sampling_params) -> Sample:
    """Generate with tokens captured during generation, no retokenization."""
    assert not args.partial_rollout, "Partial rollout not supported."

    state = GenerateState(args)
    model = SGLangModel(
        tokenizer=state.tokenizer,
        client=get_client_from_slime_args(args),  # this is lru-cached client
        tool_parser=HermesToolParser(),  # tool parsing for wrapped JSON tool calls
        sampling_params=sampling_params,
    )

    tool_limiter = ToolLimiter(max_tool_iters=MAX_TOOL_ITERS, max_tool_calls=MAX_TOOL_CALLS)
    agent = Agent(
        model=model,
        tools=[execute_python_code],
        hooks=[tool_limiter],
        callback_handler=None,
        system_prompt=SYSTEM_PROMPT,
    )

    prompt = sample.prompt if isinstance(sample.prompt, str) else sample.prompt[0]["content"]

    try:
        await agent.invoke_async(prompt)
        sample.status = Sample.Status.COMPLETED
    except Exception as e:
        # Always use TRUNCATED instead of ABORTED because slime doesn't properly
        # handle ABORTED samples in reward processing. See: https://github.com/THUDM/slime/issues/200
        sample.status = Sample.Status.TRUNCATED
        logger.warning(f"TRUNCATED: {type(e).__name__}: {e}")

    # Extract token trajectory from token_manager
    tm = model.token_manager
    prompt_len = len(tm.segments[0])  # system + user are first segment
    sample.tokens = tm.token_ids
    sample.loss_mask = tm.loss_mask[prompt_len:]
    sample.rollout_log_probs = tm.logprobs[prompt_len:]
    sample.response_length = len(sample.tokens) - prompt_len
    sample.response = model.tokenizer.decode(sample.tokens[prompt_len:], skip_special_tokens=False)

    # Record tool call stats for reward computation if needed
    # Multiple parallel tool calls count as one tool_iter
    sample.tool_iters = tool_limiter.tool_iter_count
    sample.tool_calls = tool_limiter.tool_call_count

    model.reset()
    agent.cleanup()
    return sample

Testing

# Unit tests
pytest tests/unit/ -v

# Integration tests (requires SGLang server)
pytest tests/integration/ -v --sglang-base-url=http://localhost:30000

Contributing

Contributions welcome! Install pre-commit hooks for code style and commit message validation:

pip install -e ".[dev]"
pre-commit install -t pre-commit -t commit-msg

This project uses Conventional Commits. Commit messages must follow the format:

<type>(<scope>): <description>

# Examples:
feat(client): add retry backoff configuration
fix(sglang): handle empty response from server
docs: update usage examples

Allowed types: feat, fix, docs, style, refactor, perf, test, build, ci, chore, revert

Related Projects

• strands-vllm - Community vLLM provider for Strands Agents SDK

License

Apache License 2.0 - see LICENSE.