Comments (3)
How do you send requests? can you share the code here?
Also note that when your batch size is large, enough it will reach to compute bound. And once it hits the compute bound, increasing batch doesn't improve much performance.
Other possibility is you don't have enough kv caches to batch all requests. In this case, although you max num seqs is 256, it may never reach that batch size because you cannot batch requests more than your available kv caches.
from vllm.
@rkooo567 The demo codes are attched. I changed the .py files to .txt since it dosen't support the upload of .py files. I used threadpool to send requests in order to mimic the behavior of concurrent requests. DeepseekCoder is used as the engine model, after awq, loading model weights took 3.7GB. So i guess the compute bound rather than the kv cache might be the reason.
api_server.txt
github_demo.txt
multi_8192.json
from vllm.
one thing you can try is to set disable_log_stats=False
, and it can also show you the # of running requests. If it is close to max num seqs, I think it is the compute bound case. if it is too low, maybe a code bug (since you don't use much memory for model weights)
from vllm.
Related Issues (20)
- [Bug]: UnboundLocalError: local variable 'lora_b_k' referenced before assignment HOT 3
- [Bug]: Mistral 7b inst v0.3 fails to run HOT 1
- [Bug]: HOT 2
- [Usage]: I use llama3. I found that one token is 'Ġor' in tokenizer.get_vocab(). But when I use vllm server, I got ' or' in response. HOT 1
- [Bug]: Command-R incorrect output contains `<EOS_TOKEN>` and seems to do text prediction rather than conversation
- [Misc]: LLM is responding with advertisement HOT 2
- [Bug]: 英伟达最新驱动555.85,vllm运行报错 HOT 2
- [Feature]: Additional metrics to enable better autoscaling / load balancing of vLLM servers in Kubernetes HOT 4
- [Misc]: Understanding Batching Mechanism in Prefill and Decode Phases HOT 1
- [Installation]:
- [Feature]: Add num_requests_preempted metric HOT 1
- Running Vllm on ray cluster, logging stuck at loading
- [Feature]: multi-steps model_runner? HOT 1
- [Bug]: Cannot build cpu docker image
- [Bug]: vllm.engine.async_llm_engine.AsyncEngineDeadError: Background loop has errored already. HOT 4
- [Usage]: not support for mistralai/Mistral-7B-Instruct-v0.3 HOT 3
- [Bug]: When load model weights, there are infinite loading
- [Misc]: How to use guided decoding and regex as well? HOT 2
- [Feature]: Integration of transformers past_key_values into the vllm kvcache Function HOT 4
- [Bug]: The VRAM usage of calculating log_probs is not considered in profile run HOT 5
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 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. 📊📈🎉
-
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 ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from vllm.