v0.2.2: Fix logprob regression for RL training

What's Changed

Bug Fixes

• fix(sglang): regression on logprob computation - Fixed a regression introduced in v0.2.0 where logprob_start_len=0 was
  accidentally removed from the SGLang /generate request. This caused input_token_logprobs to not be returned for tool result
  tokens, resulting in None values in the token trajectory.

Impact

This bug affected all token-in/token-out training workflows with tool-calling agents. Users would see:
TypeError: must be real number, not NoneType
when training frameworks (e.g., slime) attempted to create tensors from the logprobs.

Upgrade

pip install --upgrade strands-sglang

v0.2.1 will be yanked

Full Changelog: v0.2.1...v0.2.2

v0.2.1: Smoother SGLang API and Tool Registry

Compared to v0.1.0, the v0.2.1 release mainly brings:

• Supporting tool parser registry and three kinds of tool parsers
• Significant code refactoring to align with SGLang's native API while removing redundant code

Full Changelog: v0.1.0...v0.2.1

v0.1.3 - QwenXMLToolParser & Renamed Tool Parser API

Added

• QwenXMLToolParser: New parser for Qwen3-Coder XML tool call format
  (<function=name><parameter=key>value</parameter></function>)

Changed

• [Breaking] Renamed tool parser classes for consistency:
  • ToolCallParser → ToolParser
  • ToolCallParseResult → ToolParseResult
  • HermesToolCallParser → HermesToolParser
  • tool_call_parser parameter → tool_parser
• Decorator-based parser registry: New parsers self-register via @register_tool_parser("name") decorator
• Reorganized tool_parsers module: Extracted into tool_parsers/ module with base.py, hermes.py, qwen_xml.py

Removed

• CHANGELOG.md (tracking releases via GitHub releases instead)

Full Changelog: v0.1.2...v0.1.3

v0.1.2: Align SGLangModel with other Strands-Agents Model Providers

This release has breaking changes related to aligning SGLangModel with other model providers in Strands-Agents and remove redundancy like ephemeral client.

What's Changed

Breaking Changes

• SGLangModel requires client parameter: SGLangModel now takes a
  required client: SGLangClient instead of base_url/timeout config.
  Ephemeral client creation is removed. All parameters are keyword-only.
• SGLangConfig no longer contains base_url or timeout: These
  belong to SGLangClient. SGLangConfig now only has model_id,
  params, return_logprobs, and enable_thinking.
• SGLangClient base_url is keyword-only:
  SGLangClient(base_url="http://...") instead of positional.

Migration

# Before (0.1.1)                                                          
model = SGLangModel(tokenizer=tokenizer,                                  
base_url="http://localhost:30000")                                      
                                                                          
# After (0.1.2)                                                           
client = SGLangClient(base_url="http://localhost:30000")                  
model = SGLangModel(tokenizer=tokenizer, client=client)

Added

• TokenManager.initial_prompt: Property to access the first segment
  (initial prompt tokens) directly, instead of indexing segments[0].

Changed

• SGLangModel.client is a public attribute: Stored as self.client
  (previously self._client).
• TokenManager.add_response() guard: Raises RuntimeError if called before
  any add_prompt(), enforcing that the first segment is always a prompt.
• TokenManager.segments: No longer returns defensive copies; returns
  internal segment lists directly.

Removed

• SGLangModel._get_client(): Removed along with ephemeral client
  creation/cleanup logic.

Full Changelog:
v0.1.1...v0.1.2

v0.1.1: Enhance Tool Parsing Logic

Added

• Model Info Getter: Native get_model_info() for SGLang client.

• Think Block Exclusion: HermesToolCallParser excludes tool calls inside <think> blocks by default, preventing parsing of draft tool calls from reasoning models (Qwen3, DeepSeek-R1). Configurable via think_start_token/think_end_token.

• Tool Result Ordering: Tool results are now sorted by sequential IDs (call_0000, call_0001, ...) before tokenization. Fixes ordering issues when Strands executes tools concurrently and returns results in completion order.

Full Changelog: v0.1.0...v0.1.1

v0.1.0: Strands-SGLang Beta Release

Upon testing on various agentic RL tasks including search and coding agents, we announce beta release of strands-sglang.

Full Changelog: v0.0.3...v0.1.0

v0.0.3 - Effectively Support Qwen3 Series Agentic RL Training

Experiments

strands-sglang has been proven working with Slime on Qwen3 series models for agentic RL, providing customizable agent loops and more stabilized training compared with using OpenAIModel.

New Features

enable_thinking Config Option: Control Qwen3 hybrid models' internal reasoning mode

 # Enable thinking for reasoning tasks
 model = SGLangModel(tokenizer=tokenizer, enable_thinking=True)

 # Disable thinking for faster non-reasoning tasks  
 model = SGLangModel(tokenizer=tokenizer, enable_thinking=False)

Bug Fixes

• SLIME-Aligned Retry for Local Servers: 400 errors are now retried (transient during weight reloading, memory pressure). Only truly non-retryable: 401, 403, 404, and 400 with context length patterns.
• Improved Context Length Detection: Expanded patterns ("exceed", "too long", "max model len", etc.) to properly raise ContextWindowOverflowException

Changes

• Simplified TokenManager API: Removed unused tokenizer parameter and decode() method
• Message Formatting Methods: Converted to @classmethod for clearer intent

Installation

pip install strands-sglang==0.0.3

Full Changelog: v0.0.2...v0.0.3

v0.0.2 - Align SGLang client with Slime for RL Training

🚀 Highlights

This release introduces a non-streaming architecture aligned with Slime's http_utils.py, providing ~20x better parallelism for RL training at scale.

⚠️ Breaking Changes

• SGLangClient.generate() now returns dict[str, Any] directly instead of AsyncGenerator
• Removed stream config option from SGLangModel

✨ What's New

SGLangClient

• High-level async HTTP client with connection pooling (1000 max connections)
• Aggressive retry: 60 attempts with 1s delay (aligned with Slime)
• Infinite timeout by default for long generations
• Non-streaming POST for optimal parallelism

Factory Method

client = SGLangClient.from_slime_args(args)

model = SGLangModel(tokenizer=tokenizer, client=client)

Improved Retry Logic

• Retries all 5xx server errors
• Retries 408 (Request Timeout) and 429 (Rate Limit)
• Retries connection errors (ConnectError, PoolTimeout, ReadTimeout)
• Only non-retryable: permanent client errors (400, 401, 403, 404, 422)

Developer Experience

• Added conventional commits enforcement via pre-commit hook
• Default port changed to 30000 (SGLang default)

📦 Installation

pip install strands-sglang==0.0.2

📖 Full Changelog

See CHANGELOG.md for complete details.

Full Changelog: v0.0.1...v0.0.2

v0.0.1

strands-sglang Initial Release

strands-sglang is a SGLang model provider for Strands Agents SDK with Token-In/Token-Out (TITO) support for agentic RL training.

Features

• SGLang Native API: Uses SGLang's native /generate endpoint for efficient token-level generation
• TITO Support: Tracks complete token trajectories with logprobs for RL training - no retokenization drift (see examples/retokenization_drift/)
• Tool Call Parsing: Customizable tool parsing aligned with model chat templates (Hermes/Qwen format)
• Iteration Limiting: Built-in hook to limit tool iterations with clean trajectory truncation