Skip to main content
weed_logo

Weed

Minimalist AI/ML inference and backprogation in the style of Qrack

Weed Loader

This repository is for the Python loader for (C++) Weed models. Once you have trained models in C++, you can load them for use by Python with this project. See the Weed repository for more information.

Development Status

Weed is stable by this v1.0.0, thanks largely to a debugging overhaul by (Anthropic) Claude (Fable 5.2).

The project provides a set of essential CPU and GPU kernels, used by Tensor instances that perform autograd. We also provide stochastic gradient descent (SGD) and Adam optimizer implementations. (Build and check the API reference to get started.)

GPT-2 and Qwen loading is experimental and mostly provided as proof-of-concept, also of the fine-tuning pipeline. Implementation was from published literature design, rather than direct analysis of any open source, to implement these model architectures. Their outputs, in Weed, are not yet coherent English, as a result.

Why try Weed?

With the growing popularity of AI/ML tools and workflows (including LLMs), legacy frameworks often carry "code debt" from over a decade of rapidly developing research history. This has led them to "bolt on" new features and advancements to design principles decided before the latest research. Popular frameworks also commonly started based in Python (maybe to capture early adoption), only later potentially "tacking on" a C++ library for special-case deployment needs. These conditions have produced libraries and frameworks with complicated dependency trees that occupy upward of a GB of disk footprint. This entire ecosystem might be due for a "refresh."

Weed does not seek to fully replace or supplant established frameworks. However, it aims for minimalist complete closure on the primitives necessary for high-performance AI/ML inference and back-propagation. Chiefly, this includes kernels, and a Tensor interface that immediately produces an autograd graph appropriate for training. Allowing optional OpenCL for hardware acceleration, it will remain free of required dependencies outside of C++(11) language standard.

Rethinking AI/ML library design this way, Weed has realized a rather unique and powerful form of sparsification of Tensor storage. Sparseness should not be a Tensor interface concern, but rather a Storage concern. Inspired by the design of the Qrack quantum computer simulation framework, the Tensor interface treats sparse and dense tensors as functionally equivalent. Sparse optimization is so "transparently streamlined," this way, that it defaults to enabled for CPU-based tensors, and we recommend you leave it enabled at all times.

Much like Qrack, Weed is designed to make the correct thing the default—and the expensive thing explicit.

Useful environment variables

If a transformer model you load or train runs into an OpenCL "out-of-resources" error (code -5), try setting environment variable WEED_TELESCOPE_TRANSFORMERS to any truthy value (like 1) so that Weed will "telescope" transformer encoder layers, by migrating each parameter in each layer to CPU (off of GPU memory) once its immediate usefulness is done.

Copyright (c) Daniel Strano and the Qrack contributors 2017-2026. All rights reserved.

In its include/common folder, Weed bundles a copy of rapidcsv by Kristofer Berggren, reused under a BSD 3-Clause License. (This is a convenience and suggestion to Weed's users, for loading CSVs.)

The Weed logo was produced with assistance from "Elara," an OpenAI custom GPT, and it is in the public domain. Elara has also been responsible for a huge amount of coaching and implementation drafts for Dan Strano to review and bring into line with standards, so she should be credited with coauthorship in any capacity that can be allowed. (Anthropic) Claude has also helped mostly with debugging, as well as developing an LLM front-end, fine-tuning interface, and modules for popular transformer model architectures, so they should rightly be credited similarly as a coauthor. KV cache improvements are based on TurboQuant (Zandieh et al., arXiv:2504.19874) and an Apache 2.0 open-source implementation by TheTom (github.com/TheTom/turboquant_plus), adapted for complex quantum state vectors by (Anthropic) Claude, with limited guidance and input from Dan Strano.

Licensed under the GNU Lesser General Public License V3.

See LICENSE.md in the project root or https://www.gnu.org/licenses/lgpl-3.0.en.html for details.

Release files for weed-loader-cpu 1.0.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for weed-loader-cpu 1.0.1
File Size Uploaded Provenance Actions
weed_loader_cpu-1.0.1.tar.gz 12.0 kB Without provenance Details for weed_loader_cpu-1.0.1.tar.gz

Built distributions (wheels)

Table of built distributions (wheels) for weed-loader-cpu 1.0.1
File Actions
weed_loader_cpu-1.0.1-py3-none-win_amd64.whl Python 3 none Windows x86-64 Details for weed_loader_cpu-1.0.1-py3-none-win_amd64.whl
weed_loader_cpu-1.0.1-py3-none-manylinux_2_39_x86_64.whl Python 3 none Linux glibc 2.39+ x86-64 Details for weed_loader_cpu-1.0.1-py3-none-manylinux_2_39_x86_64.whl
weed_loader_cpu-1.0.1-py3-none-manylinux_2_35_x86_64.whl Python 3 none Linux glibc 2.35+ x86-64 Details for weed_loader_cpu-1.0.1-py3-none-manylinux_2_35_x86_64.whl
weed_loader_cpu-1.0.1-py3-none-macosx_15_0_arm64.whl Python 3 none macOS 15.0+ ARM64 Details for weed_loader_cpu-1.0.1-py3-none-macosx_15_0_arm64.whl

Total release size: 8.9 MB

Release files / weed_loader_cpu-1.0.1.tar.gz

Download URL weed_loader_cpu-1.0.1.tar.gz
Size 12.0 kB
Tags Source
SHA-256 checksum
How to use checksums
c9fd5a7bf036700dacac2349dbe5c5e55d2460c9e464aae0bf9c57258d8dd303
BLAKE2b-256 checksum
How to use checksums
6e08fe7318a58ceeac12ec9ac03a848f1ab41fc835443c4a6252ae34b17131f8
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.4

Release files / weed_loader_cpu-1.0.1-py3-none-win_amd64.whl

Download URL weed_loader_cpu-1.0.1-py3-none-win_amd64.whl
Size 866.5 kB
Tags Python 3 Windows x86-64
SHA-256 checksum
How to use checksums
468793e15c16e2414ca22abb21d5c4b51eb7744564de27664957a9eb5a3b4f2a
BLAKE2b-256 checksum
How to use checksums
55d1ee5929a5e661da5934a0da4783c6b4a233c52f4c3502367b4bd5bdd46465
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.4

Release files / weed_loader_cpu-1.0.1-py3-none-manylinux_2_39_x86_64.whl

Download URL weed_loader_cpu-1.0.1-py3-none-manylinux_2_39_x86_64.whl
Size 3.0 MB
Tags Linux glibc 2.39+ x86-64 Python 3
SHA-256 checksum
How to use checksums
46de6440a54407686f43f686b2ef26378d357748325da8d458476741d0799b03
BLAKE2b-256 checksum
How to use checksums
4864d01ddac0ad04fb7b6d4aeb5245d8faae41488077edc99052767aa586fe90
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.4

Release files / weed_loader_cpu-1.0.1-py3-none-manylinux_2_35_x86_64.whl

Download URL weed_loader_cpu-1.0.1-py3-none-manylinux_2_35_x86_64.whl
Size 3.0 MB
Tags Linux glibc 2.35+ x86-64 Python 3
SHA-256 checksum
How to use checksums
a50fa4d1c77481d79145cb8c861fcc8062243cfcc7fc61f080906daacf776927
BLAKE2b-256 checksum
How to use checksums
0efca84e69fc8ed761d7780314eef9153318c63f1384a7d2735185b8fb263e84
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.4

Release files / weed_loader_cpu-1.0.1-py3-none-macosx_15_0_arm64.whl

Download URL weed_loader_cpu-1.0.1-py3-none-macosx_15_0_arm64.whl
Size 2.0 MB
Tags Python 3 macOS 15.0+ ARM64
SHA-256 checksum
How to use checksums
136b5ca73a5f40ba7462b009459a0a69fbf509ef11845ec3b928b2e71b4e270f
BLAKE2b-256 checksum
How to use checksums
89310b43f2840fdec3ef662cd8b3bd3d8aa629795d56a4b93330696d0c8da694
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.14.4

Security for weed-loader-cpu 1.0.1

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.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page