gadget-ml 0.4.6

Model creation in Python powered by GGML

gadget is a Python library for model creation using the GGML compute framework. It provides Python bindings for most low-level GGML functions, a Python interface for reading/writing GGUF files, and a high-level interface for creating and executing models.

Here's a minimal example of how to use gadget to create a model and run inference on the CPU:

import numpy as np
from gadget import GgmlModel, Tensor
from gadget.ggml import ggml_mul_mat

# simple model interface
class LinearModel(GgmlModel):
    weight: Tensor('F32', ('input_dim', 'output_dim'))
    inputs: Tensor('F32', ('input_dim', 'batch_size'))

    def forward(self):
        ctx = self.ctx_graph
        w, x = self.tensors['weight', 'inputs']
        return ggml_mul_mat(ctx, w, x)

# generate random weights and input data (note reverse order of dimensions)
input_dim, output_dim, batch_size = 256, 32, 16
weight_np = np.random.randn(output_dim, input_dim).astype(np.float32)
inputs_np = np.random.randn(batch_size, input_dim).astype(np.float32)

# create model and execute
model = LinearModel.from_values(
    dict(weight=weight_np), input_dim=input_dim,
    output_dim=output_dim, batch_size=batch_size
)
output_np = model(inputs=inputs_np)

To run on the GPU, you can pass backend='cuda' to the GgmlModel constructor and pass the weights and data as torch tensors on the GPU. In practice, you'll likely be loading the weights for these models from GGUF files, in which case you can use the from_gguf and from_path constructors for GgmlModel.

For more examples, see the test_* functions in model.py and compute.py, or the full fledged implementations in gadget/models. One can implement fairly complex models with a relatively small amount of code. Though adding in Python as a dependency is a deal-breaker for some projects, some advantages of gadget are:

• tokenization is hell, let Huggingface handle it!
• can do rapid prototyping and experimentation without having to compile anything
• no need for round-trips to and from the GPU (could be important for embeddings?)
• easier to rapidly integrate things like novel sampling methods (entropix??)

Install

To install with pip run:

pip install gadget-ml

You can pass arguments to cmake using the CMAKE_ARGS environment variable. For example, to add CUDA support:

CMAKE_ARGS="-DGGML_CUDA=ON"

You can install locally after cloning this repository with:

pip install -e .

To build the shared libraries for local testing, you can use cmake directly

cmake -B build .
cmake --build build -j

Usage

gadget comes with built-in support for popular models, such as Llama for text generation and BERT for embeddings, which actually covers a lot of cases. To do a simple completion with Llama, you can run something like

from gadget import TextGen
model = TextGen(path_to_gguf, huggingface_model_id)
reply = model.generate(prompt)

For a conversation interface, you can use the TextChat class with

from gadget import TextChat
model = TextChat(path_to_gguf, huggingface_model_id, system=system_prompt)
reply = model.generate_chat(prompt)

To do streaming generation, use the respective stream and stream_chat methods. You can control the maximum length of the generated text with the max_gen argument. As for BERT embeddings, you can run the following on a string or a list of strings texts:

from gadget import EmbedTorch
model = EmbedTorch(path_to_gguf, huggingface_model_id)
vecs = model.embed(texts)

In all of the above, path_to_gguf is the path to the GGUF file and huggingface_model_id is the full name of the model on Huggingface. The default backend is cpu. To run on the GPU, use backend='cuda'. Metal is currently not supported, as I don't have anything to test it on.

Internals

GgmlCompute

The lowest level interface is the GgmlCompute class, which takes care of creating, setting, and getting tensors, as well as graph creation and execution. The constructor takes three arguments:

• params: a dictionary of parameter names and values (name: value)
• tensors: a dictionary of tensors specifications (name: (dtype, shape))
• model: a function that takes fields and tensors as inputs and returns an output tensor

The model function should have the signature model(context, params, tensors) and return an output tensor. There are some simple usage examples at the end of compute.py. It has the following methods:

• create_graph(model) — creates the computational graph for a model function
• destroy_graph() — deallocates the computational graph
• get_input(name) — retrieves input tensor value by name
• set_input(name, array, offset=None) — sets input tensor value by name
• get_named_node(name) — retrieves tensor values for a graph node by name
• compute() — executes the computational graph
• __call__(**values) — sets input values, computes, and returns the output

GgmlModel

In most cases, however, you'll want to use the higher level GgmlModel interface. This takes a cue from the JAX library equinox and uses class-level type hints to dictate which tensors should be created. Additionally, it takes a GGUF file (via GgufFile) as input and loads tensors from that. There are three types of metadata that can be included:

• Parameter: values that can be set on object creation like batch_size
• Tensor: tensors (input or working) that can be provided by the user
• State: runtime variables whose mutation will trigger a graph recompile (like n_tokens)

This handles loading directly from GGUF objects or paths, and will automatically recompile the compute graph whenever self.state is modified. To use this class, subclass it with the appropriate type hints and implement the forward method. You can then initialize the model using the from_values or from_gguf/from_path methods. It also has the following methods:

• rebuild_graph() — rebuilds the graph
• __call__(**kwargs) — rebuilds the graph (if self.state has changed) and calls GgmlCompute.__call__

LlamaModel

The class LlamaModel does single sequence logit computation with KV caching. The context_length parameter controls the size of the KV cache, and the batch_size parameter controls the maximum number of tokens that can be passed to the model in a single call.

BertModel

The class BertModel implements BERT embeddings, which covers a very wide variety of embedding models today. Currently, pooling is done outside of ggml in numpy or torch, but you could imagine subclassing BertModel and wrapping the forward function to perform pooling inside the model.

Conventions

Matrix shape and order: tensors are row-major, meaning elements in a row are stored in contiguous order. However, the way in which dimensions are reported is reversed from numpy. The first number in the shape is the number of columns, the second is the number of rows, and so on. The logic here is that the first number denotes the number of elements in a row, the second denotes the number of columns and so on. This makes computing strides for element access much easier.

Tensor shapes are reported using the ggml convention. But you still have to keep this in mind that when setting tensor values with set_input, you need to use the reverse order from what ggml reports. So something that is shape (k, n, m) in ggml should be shape (m, n, k) in numpy.

Matrix multiplication: for ggml_mul_mat and others, the GGML-style shape of inputs a and b and output c should be

a ~ (k, n, i, j)
b ~ (k, m, i, j)
c ~ (n, m, i, j)

In other words, the matmul is performed over the first two dimensions according to c.T = b.T @ a (or c = a.T @ b) and is batched over the remaining dimensions.

If we think about shapes numpy-style, then this would read

a ~ (n, k, i, j)
b ~ (m, k, i, j)
c ~ (m, n, i, j)

And we would write the matrix multiplication as c = b @ a.T. So basically the same thing, but everything is transposed to undo the shape reversed notation.