dsalt

library0.4.34pypypiunverified

dsalt (Dynamic Sparse Attention with Landmark Tokens) is a high-performance Triton-based implementation of sparse attention for transformers. Version 0.4.34 supports PyTorch and provides fused kernels for landmark token selection and sparse attention computation, targeting long-context LLM inference and training. Released monthly.

ai-ml

pip install dsalt

github ↗homepage ↗

INSTALL

IMPORT

SIG · DSALT

dsalt

ai-mlpythonv0.4.34

harness data pending

Install & Compatibility

Where this runs

No compatibility data collected yet for this library.

Code

Verified usage

Verified import paths — ran on the pinned version, not inferred.

DynamicSparseAttention

✓ from dsalt import DynamicSparseAttention

✗ from dsalt.dynamic_sparse_attention import DynamicSparseAttention

Top-level import is the canonical way. Nested submodule path changed in 0.4.x.

Initialize the sparse attention module with model dimensions and run a forward pass.

import torch from dsalt import DynamicSparseAttention model = DynamicSparseAttention( d_model=1024, n_heads=8, landmark_ratio=0.1, sparse_topk=32 ).cuda() x = torch.randn(2, 512, 1024).cuda() # compute dynamic sparse attention output = model(x) print(output.shape)

Debug

Known issues

breakingIn dsalt 0.4.x, the module import path changed from `dsalt.dynamic_sparse_attention` to `dsalt`. Existing code using the old import will break.

fix

Update import to `from dsalt import DynamicSparseAttention`.

affects: <0.4.0

gotchadsalt requires GPU with compute capability >=7.0 (Volta) and Triton installed. Running on CPU will raise a RuntimeError during kernel compilation.

fix

Ensure a compatible NVIDIA GPU and install Triton via `pip install triton`.

affects: all

gotchaThe `landmark_ratio` parameter must be a float between 0 and 1. Values >=1 will silently fall back to dense attention, which may cause unexpected memory usage.

fix

Ensure `landmark_ratio` < 1.0, e.g., 0.1.

affects: all

deprecatedThe `num_landmarks` argument in the constructor is deprecated in 0.4.x and will be removed in 0.5. Use `landmark_ratio` instead.

fix

Replace `num_landmarks=64` with `landmark_ratio=64/seq_len`.

affects: >=0.4.0

Upgrade

Version history

0.4.34latest on PyPI · released Jun 6, 2026

Audit

Dependencies

torchrequiredCore dependency for tensor operations and autograd.

tritonrequiredRequired for GPU kernel compilation and execution.

Agent activity

23 hits · last 30 days

node

OpenAI (training)

Resources

githubhttps://github.com/LeonardoCofone/dsalt-library ↗

homepagehttps://github.com/LeonardoCofone/dsalt-library ↗