Comments (10)
from vllm.
command:
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
python3 -m vllm.entrypoints.openai.api_server \
--model Qwen1.5-72B-Chat \
--tensor-parallel-size 8 \
--max-model-len 8192 \
--trust-remote-code \
--disable-custom-all-reduce \
--enable-prefix-caching \
--tokenizer-mode slow \
--enforce-eager \
--gpu-memory-utilization 0.9 \
--port 8861
from vllm.
If you can biset to find out the commit that leads to the degradation, that would be helpful. Otherwise, it is very difficult to answer a generic report of a performance regression.
from vllm.
If you can biset to find out the commit that leads to the degradation, that would be helpful. Otherwise, it is very difficult to answer a generic report of a performance regression.
vLLM seems to have been under reconstruction recently. I'm not quite sure what's causing TPOT to be slow
from vllm.
most commits should be runnable since they pass the ci tests. it's not related with the refactoring. just bisect to find out the potentially commit that leads to this regression.
from vllm.
I have test on L20, not sure the device is same as in CI.
from vllm.
with cuda graph: 35.7ms -> 36.7ms, without cuda graph: 39ms -> 45ms
from vllm.
maybe relate to #5207
from vllm.
feel free to test if #5207 solves your problem here.
from vllm.
@youkaichao close, seems the latest vllm (up to #5410) has fixed this problem. (TP0T 45ms v0.4.2 -> 39ms v0.5, eager mode)
[I][2024-06-11 16:31:36][ 1/20][ 1/20 5%] session:0 turn:0 req:0 output: 50 prompt: 1024(+[1024]) first(ms): 426.7(+[ 426.7]) other(ms): 39.3 latency(s): 2.35
[I][2024-06-11 16:31:46][ 2/20][ 2/20 10%] session:1 turn:0 req:1 output: 50 prompt: 1024(+[1024]) first(ms): 427.0(+[ 427.0]) other(ms): 40.6 latency(s): 2.42
[I][2024-06-11 16:31:56][ 3/20][ 3/20 15%] session:2 turn:0 req:2 output: 50 prompt: 1024(+[1024]) first(ms): 427.8(+[ 427.8]) other(ms): 39.2 latency(s): 2.35
[I][2024-06-11 16:32:06][ 4/20][ 4/20 20%] session:3 turn:0 req:3 output: 50 prompt: 1024(+[1024]) first(ms): 428.5(+[ 428.5]) other(ms): 39.2 latency(s): 2.35
[I][2024-06-11 16:32:16][ 5/20][ 5/20 25%] session:4 turn:0 req:4 output: 50 prompt: 1024(+[1024]) first(ms): 427.6(+[ 427.6]) other(ms): 39.2 latency(s): 2.35
[I][2024-06-11 16:32:26][ 6/20][ 6/20 30%] session:5 turn:0 req:5 output: 50 prompt: 1024(+[1024]) first(ms): 429.1(+[ 429.1]) other(ms): 39.0 latency(s): 2.34
[I][2024-06-11 16:32:36][ 7/20][ 7/20 35%] session:6 turn:0 req:6 output: 50 prompt: 1024(+[1024]) first(ms): 427.0(+[ 427.0]) other(ms): 40.5 latency(s): 2.41
[I][2024-06-11 16:32:46][ 8/20][ 8/20 40%] session:7 turn:0 req:7 output: 50 prompt: 1024(+[1024]) first(ms): 427.0(+[ 427.0]) other(ms): 39.1 latency(s): 2.34
[I][2024-06-11 16:32:56][ 9/20][ 9/20 45%] session:8 turn:0 req:8 output: 50 prompt: 1024(+[1024]) first(ms): 425.3(+[ 425.3]) other(ms): 38.8 latency(s): 2.33
[I][2024-06-11 16:33:06][10/20][10/20 50%] session:9 turn:0 req:9 output: 50 prompt: 1024(+[1024]) first(ms): 426.2(+[ 426.2]) other(ms): 39.1 latency(s): 2.34
[I][2024-06-11 16:33:16][11/20][11/20 55%] session:10 turn:0 req:10 output: 50 prompt: 1024(+[1024]) first(ms): 428.2(+[ 428.2]) other(ms): 38.9 latency(s): 2.33
[I][2024-06-11 16:33:26][12/20][12/20 60%] session:11 turn:0 req:11 output: 50 prompt: 1024(+[1024]) first(ms): 426.4(+[ 426.4]) other(ms): 39.0 latency(s): 2.34
[I][2024-06-11 16:33:36][13/20][13/20 65%] session:12 turn:0 req:12 output: 50 prompt: 1024(+[1024]) first(ms): 425.6(+[ 425.6]) other(ms): 39.6 latency(s): 2.37
[I][2024-06-11 16:33:46][14/20][14/20 70%] session:13 turn:0 req:13 output: 50 prompt: 1024(+[1024]) first(ms): 425.5(+[ 425.5]) other(ms): 39.2 latency(s): 2.35
[I][2024-06-11 16:33:56][15/20][15/20 75%] session:14 turn:0 req:14 output: 50 prompt: 1024(+[1024]) first(ms): 425.2(+[ 425.2]) other(ms): 39.9 latency(s): 2.38
from vllm.
Related Issues (20)
- [Bug]: tensor parallel processes not working in vllm_cpu HOT 2
- [Bug]: use cpu_offload_gb in gguf failed. HOT 1
- [Usage]: Question about dequantization HOT 1
- [Usage]: how to acquire logits in vllm HOT 2
- [Usage]: Total generated tokens in benchmarking script HOT 5
- [Usage]: output were empty HOT 6
- [Usage]: output were empty HOT 1
- [Bug]: output is empty HOT 1
- [Bug]: LLMEngine cannot be pickled error vllm 0.6.1.post2 HOT 2
- vLLM's V2 Engine Architecture HOT 3
- [RFC]: Add Goodput Metric to Benchmark Serving
- [Bug]: Decode n tokens gives different output for first seq position compared to decode 1 token
- [Bug]: Disabling Marlin by setting --quantization gptq doesn't work when using a draft model
- [Tracking Issue][Help Wanted]: FlashInfer backend improvements
- [Installation]: Installing vLLM on ROCm - Distro:Gentoo
- [Bug]: Port binding failure when using pp > 1 after commit 7c7714d856eee6fa94aade729b67f00584f72a4c HOT 8
- [Doc]: Is Qwen2.5's long context YARN handled? HOT 9
- > > Specify the local folder you have the model in instead of a HF model ID. If you have all the necessary files and the model is using a supported architecture, then it will work. HOT 2
- [Bug]: Bug in the convert_fp8 function, a function for testing.
- [Bug]: Loading a model with bitsandbytes 8bit quantization HOT 1
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.