Comments (5)
I think I understand this use case but just to confirm is there a reason that this cannot be part of the prompt?
from vllm.
There is a difference between instructing the model with a prompt like Start all your responses with "Sure Thing!"
and the model having Sure Thing!
as the initial tokens that follow the begin-response token. The former is an instruction given to the model each time, while the latter means the model is pre-configured to always start responses with 'Sure Thing!'
Here is an example for my use case:
You can also think of it this way Start all your responses with "[" and create a json with the following information that I provide
. I am telling the prompt to start as a Json but it is not a guarantee and all too often it will start off saying Here is your Json: [{
. But if you have the initial token following the begin response token as [
then it will basically guarantee returning a Json.
from vllm.
Why not just using the system prompt for this one?
Also, where have you seen this before? I have never encountered such feature
from vllm.
@Etelis What I described was for the system prompt.
from vllm.
I think anthropic's api has similar feature.
https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/prefill-claudes-response
I'm not familiar but may it achievable with outlines?
from vllm.
Related Issues (20)
- [Bug]: RuntimeError with tensor_parallel_size > 1 in Process Bootstrapping Phase HOT 5
- [Bug]: error when requesting max-num-seqs with speculation
- [Installation]: vllm on NVIDIA jetson AGX orin HOT 2
- [Bug]: Error while running command-r in high load for a long period in OpenBLAS on 2 GPUs
- [Bug]: multiprocessing KeyError from `cache[rtype].remove(name)` HOT 4
- IfEval Metrics not consistent with different vLLM versions HOT 3
- [Bug]: OOT models not included in ModelRegistry.get_supported_archs() HOT 4
- [Bug]: Ray distributed backend does not support out-of-tree models via ModelRegistry APIs
- [Usage]: Running Cohere command R+ 104b using VLLM 16bf, getting 5 tokens per second very slow
- éćććŻćšćŻćšæ„é
- [Usage]: vllm0.5.0.post error when using openai completion interface HOT 1
- [Bug]: Qwen2-72B-Instruct-gptq-int4 Repetitive issues HOT 1
- [Usage]: Does class LLM support inference quantization on CPU?
- [Bug]: Eabling Prefix-Caching doesn't speed up inference HOT 2
- [Bug]: Option for preemption_mode uses underscore (_) instead of dash (-) HOT 1
- [Installation]: Build from source: Could NOT find Python. Could not build wheels for vllm. HOT 4
- [Installation]: pip install -e failed HOT 3
- ValueError: The input size is not aligned with the quantized weight shape. This can be caused by too large tensor parallel size.[Bug]: HOT 2
- [Usage]: Use quantization=gptq_marlin for faster inference HOT 2
- [Bug]: `flash_attn_cuda.varlen_fwd` may output a bad result when enabling prefix caching HOT 2
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
đ Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. đđđ
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google â€ïž Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from vllm.