Development Roadmap

[Ying1123]

Function Calling

• Frontend
  • Add tools argument in sgl.gen. See also guidance tools
• Backend
  • OpenAI: Translate to their function calling API (https://platform.openai.com/docs/guides/function-calling).
    • Function calling for OpenAI backend #573
  • Local Models (SGLang)
    1. Use SGLang primitives (regex, select) and constrained decoding to implement a workflow
    2. Directly use models that support function calling (e.g., Gorilla OpenFunctions, https://huggingface.co/jondurbin/bagel-dpo-7b-v0.4#prompting-strategies)
  • Local Models (OpenAI-compatible API)

High-level Pythonic Interface

• LLM integration with normal programming patterns or, a high level sglang interface #39

Inference Optimizations

• Speculative decoding for local models
• Speculative execution for OpenAI Chat API
  • support speculative execution for openai API #48

Structured Decoding

• Support parallel JSON decoding SGLang Integration varunshenoy/super-json-mode#8
• Support auto parallel decoding https://arxiv.org/abs/2401.06761

Compiler

• Support tracing and compiling sgl.fork
• Support sending a full serialized SGL program to the server

LoRA Support

• Port multi-LoRA batching and unified memory from S-LoRA

Model Coverage

• Vision Langauge Models. Support top-performing models from https://github.com/open-compass/VLMEvalKit
• Language Models. Port the implementation of popular models from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/models. (help)

Device Coverage

• AMD support. Investigate AMD support in Trion and FlashInfer.
• CPU support. This is better done by adding a llama.cpp backend.

[AriMKatz]

Are there still plans for a high level pythonic interface? #39 (comment)

[Ying1123]

Are there still plans for a high level pythonic interface? #39 (comment)

Hi @AriMKatz, thanks for the reference. This is very important, I just added it.

[nivibilla]

For the vision models support, is it possible to align with the openai gpt4v API?
https://platform.openai.com/docs/guides/vision

[aliencaocao]

Are there plans for loading models in 8bit or 4bit?

[Ying1123]

For the vision models support, is it possible to align with the openai gpt4v API? https://platform.openai.com/docs/guides/vision

@nivibilla Yes, it is already aligned with the openai gpt4v API, see here.
You can also find a runnable example of serving it with Sky Serve here.

[Ying1123]

Are there plans for loading models in 8bit or 4bit?

@aliencaocao Thanks for the question! The AWQ and GPTQ are already supported. But we do not support an automatic dtype translation yet. You are welcome to submit a PR for that.

[aliencaocao]

Are there plans for loading models in 8bit or 4bit?

@aliencaocao Thanks for the question! The AWQ and GPTQ are already supported. But we do not support an automatic dtype translation yet. You are welcome to submit a PR for that.

I'm looking to load llava 1.6 in 8bit, but it does not seem that llava series has AWQ or GPTQ quants, or did I miss out anything here?

EDIT: I saw 1.5 has but not 1.6 yet. Perhaps its just too new and no one did a calibration yet.

[qeternity]

Hi all - is anyone working on the S-LoRA integration currently? I see the branch, but it looks a few months old.

Would love to see this, happy to pick up from existing work or start fresh.

[Ying1123]

Hi all - is anyone working on the S-LoRA integration currently? I see the branch, but it looks a few months old.

Would love to see this, happy to pick up from existing work or start fresh.

Hi @qeternity, I was working on it but have been blocked by other affairs. You are welcome to contribute, either continue on the branch or start fresh! I'll be happy to review and collaborate.

[Bit0r]

Tools support is very important, which is necessary for many use cases.

[omri-sap]

Is TinyLlama supported? TinyLlama/TinyLlama-1.1B-Chat-v1.0
generation seems a bit slow...

[wille-x]

I see llama.cpp integration is on the roadmap. When will this feature be delivered? It would be very nice to have it , since it will support running local LLMs, such as llama models, on Mac computers and experiment them with the powerful and expressive SGLang.

[Gintasz]

I'd request to include support for Phi-3-mini

[binarycrayon]

Hi all - is anyone working on the S-LoRA integration currently? I see the branch, but it looks a few months old.
Would love to see this, happy to pick up from existing work or start fresh.

Hi @qeternity, I was working on it but have been blocked by other affairs. You are welcome to contribute, either continue on the branch or start fresh! I'll be happy to review and collaborate.

Hi which branch is it? looks like better start fresh