An efficient Linear Attention Decoding package
1. installation
conda create -n leetDecoding python==3.9
conda activate leetDecoding
pip install leetDecoding
The code has been test under the following environment:
triton>=2.1.0
torch>=2.1.0
pycuda
pynvml
numpy<2
You can use the following command to install:
pip install triton==2.1.0
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install pycuda
pip install pynvml
pip install numpy
2. usage
import torch
from leetDecoding.efficient_linear_decoding import causal_linear_decoder
# Create input tensor
Q = torch.randn(2,32,1024,128,device='cuda:0')
K = torch.randn(2,32,1024,128,device='cuda:0')
V = torch.randn(2,32,1024,128,device='cuda:0')
# Inference using causal_linear_decoder
output = causal_linear_decoder(Q,K,V)
# If you want to input a mask with weight that values are exp(-gamma), set the is_mask_weight: True and is_need_exp:True
gamma = torch.full((32,),0.5,device='cuda:0')
output = causal_linear_decoder(Q,K,V,is_mask_weight=True,gamma=gamma,is_need_exp=True)
# If you just want to input a mask with weight, set the is_mask_weight: True and is_need_exp:False
gamma = torch.full((32,),0.5,device='cuda:0')
output = causal_linear_decoder(Q,K,V,is_mask_weight=True,gamma=gamma,is_need_exp=False)
# If you want to use a specified methods, such as FleetAttention, set the attn-method: 'FleetAttention'
gamma = torch.full((32,),0.5,device='cuda:0')
output = causal_linear_decoder(Q,K,V,is_mask_weight=False,attn_method='FleetAttention')
3. acknowledgement
| method | Title | Paper | Code |
|---|---|---|---|
| causal_dot_product | Fast Transformers with Clustered Attention | arxiv | code |
| Lighting Attention-2 | Lightning Attention-2: A Free Lunch for Handling Unlimited Sequence Lengths in Large Language Models | arxiv | code |
Metadata
Release files for leetDecoding 0.0.3
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| leetDecoding-0.0.3.tar.gz | 29.1 kB | Without provenance | Details for leetDecoding-0.0.3.tar.gz |
Release files / leetDecoding-0.0.3.tar.gz
| Download URL | leetDecoding-0.0.3.tar.gz |
|---|---|
| Size | 29.1 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
0c380ecc61fe9444efbfcff0668b85963b8ff2ce54ed2dc17f42189aacd140a7
|
|
BLAKE2b-256 checksum How to use checksums |
67bbf25e4d311fc662de4eba1d2fd20c288df7ec48b18e7e28953c29975c9065
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/5.1.1 CPython/3.11.4
|
Security for leetDecoding 0.0.3
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.