# sgl-project/sglang · Reposen Release & Migration Intelligence

> Upgrade DeepSeek-V4 chat encoding and default SWA bounded replay. Update API configurations. Migration: Moderate.

## Metadata
- **Repository URL**: https://github.com/sgl-project/sglang
- **Primary Package**: `sglang` (PYPI)
- **Categories**: `LOCAL-AI`, `RUNTIME`
- **Latest Version**: `v0.5.23`
- **Semantic Status**: `BREAKING`
- **Bundle Size Impact (unavailable)**: `Size Unavailable`
- **Atom Ingestion Feed**: https://github.com/sgl-project/sglang/releases.atom

## The Human Context (Plain-English Pitch)
SGLang is a high-performance serving framework for large language models and diffusion models, designed to maximize throughput and minimize latency through advanced Radix Attention, tensor parallelism, and custom kernel fusions. It addresses the high compute costs and memory bottlenecks of modern generative AI inference by optimizing KV caching and memory layouts across hardware accelerators.

## Stage 1 Executive Summary ([Core Change] + [Required Action] + [Migration Effort])
Upgrade DeepSeek-V4 chat encoding and default SWA bounded replay. Update API configurations. Migration: Moderate.

## Breaking Changes & Required Actions
- [BREAKING] DeepSeek-V4/V4.1 chat requests now use `deepseek-recipe`; custom `encoding_dsv4.py` endpoints are removed.
- [BREAKING] Numeric `reasoning_effort` on V4 inputs returns a 400 error.
- [BREAKING] The `tokenspeed_mla` attention backend and `--speculative-skip-dp-mlp-sync` flag are completely removed.
- [BREAKING] Unsafe pickle fallback (`weights_only=False`) for `.pt` checkpoints is disabled.
- DeepSeek-V4 and DeepSeek-V4.1 chat API requests now route through official `deepseek-recipe` encoding.
- Enabled DeepSeek-V4.1 decoder SWA bounded replay by default, reducing first-token latency.
- Removed the deprecated `tokenspeed_mla` attention backend and `--speculative-skip-dp-mlp-sync` flag.
- Dropped unsafe `weights_only=False` pickle loading support for legacy `.pt` model checkpoints.

## Migration Snippet
```ts
uv pip install --prerelease=allow sglang==0.5.23
```

## Ecosystem Compatibility Matrix & Graph Warnings
- **Python >=3.10** [`COMPATIBLE`]: Requires Python runtime compatibility for underlying server and plugin bindings.
- **PyTorch >=2.5** [`COMPATIBLE`]: Required for native tensor operations, custom operators, and compiled CUDA graphs.

## Zero-Noise Release Timeline
- **v0.5.23** (2026-11-02) [BREAKING]: Upgrade DeepSeek-V4 chat encoding and default SWA bounded replay. Update API configurations. Migration: Moderate.

## Upgrade & Migration Q&A (GEO Direct Answers)
### Is sgl-project/sglang v0.5.23 safe to upgrade in production?
sgl-project/sglang v0.5.23 is classified as BREAKING (MINOR release). Upgrade DeepSeek-V4 chat encoding and default SWA bounded replay. Update API configurations. Migration: Moderate.

### What breaking changes or compatibility risks exist in sgl-project/sglang v0.5.23?
DeepSeek-V4/V4.1 chat requests now use `deepseek-recipe`; custom `encoding_dsv4.py` endpoints are removed. Numeric `reasoning_effort` on V4 inputs returns a 400 error. The `tokenspeed_mla` attention backend and `--speculative-skip-dp-mlp-sync` flag are completely removed. Unsafe pickle fallback (`weights_only=False`) for `.pt` checkpoints is disabled.

### What is the bundle size impact of sglang v0.5.23?
Measured bundle size telemetry is currently unavailable for sglang v0.5.23; verify build output locally before deploying.

### How do I upgrade sglang to v0.5.23?
Upgrade sglang to v0.5.23 by running "pip install sglang==0.5.23" and applying the migration steps: uv pip install --prerelease=allow sglang==0.5.23
