llmanim 0.1.0

Reusable Manim components for creating LLM and Transformer explanation videos.

Reusable Manim components for creating LLM and Transformer explanation videos.

A Python library of pre-built visual components, animations, and utilities for creating educational videos about Large Language Models and Transformer architectures using Manim.

Installation · Quick Start · Components · Examples · Contributing

---

Installation

pip install llmanim

Prerequisites

Manim requires a few system dependencies before you install:

# macOS
brew install cairo pango ffmpeg
brew install --cask mactex          # LaTeX for MathTex

# Ubuntu / Debian
sudo apt install libcairo2-dev libpango1.0-dev ffmpeg texlive-full

# Windows — install MiKTeX from https://miktex.org, then:
# pip install manim

Full Manim installation guide: docs.manim.community

Install from source

git clone https://github.com/rishabhbhartiya/LLManim.git
cd LLManim
pip install -e ".[dev]"

---

Quick Start

from manim import *
from llmanim.base.shapes import MatrixBox, VectorBar, TokenBox
from llmanim.base.animations import highlight_sequence, data_flow_arrow
from llmanim.base.utils import apply_dark_theme, make_attention_weights
from llmanim.tokenization.token_box import TokenRow
from llmanim.styles.colors import TOKEN_COLOR, HIGHLIGHT_COLOR

class MyTransformerScene(Scene):
    def construct(self):
        apply_dark_theme(self)

        # 1. Show a token sequence
        row = TokenRow(["the", "cat", "sat"], token_ids=[1, 2, 3])
        self.play(row.appear())

        # 2. Show an embedding matrix
        mat = MatrixBox(rows=4, cols=6, color=TOKEN_COLOR).shift(DOWN * 2)
        self.play(mat.fill_anim())

        # 3. Connect them with an animated arrow
        arrow, anim = data_flow_arrow(row, mat, label="embed")
        self.add(arrow)
        self.play(anim)

        # 4. Highlight components
        self.play(highlight_sequence(row.boxes))
        self.play(mat.highlight_row(2))
        self.wait(1)

Render it:

manim -pql my_scene.py MyTransformerScene   # fast preview
manim -pqh my_scene.py MyTransformerScene   # 1080p final

---

Example Videos

Five complete video scripts are included in examples/:

#	File	Scene	Topic
1	01_what_is_a_token.py	WhatIsAToken	Tokenization basics
2	02_token_embeddings.py	TokenEmbeddings	Embedding lookup
3	03_attention_mechanism.py	AttentionMechanism	Q, K, V & softmax
4	04_multi_head_attention.py	MultiHeadAttention	Multiple attention heads
5	05_full_forward_pass.py	FullForwardPass	Complete forward pass

Video 1 — What is a Token?

Video 2 — Token Embeddings

Video 3 — Attention Mechanism

Video 4 — Multi-Head Attention

Video 5 — Full Forward Pass

# Clone to access examples (they're not installed with pip)
git clone https://github.com/rishabhbhartiya/LLManim.git
cd LLManim

manim -pql examples/01_what_is_a_token.py WhatIsAToken     # preview
manim -pqh examples/05_full_forward_pass.py FullForwardPass # 1080p
manim -pqk examples/05_full_forward_pass.py FullForwardPass # 4K

---

Components

Base Shapes — llmanim.base.shapes

Class	Description	Key Methods
MatrixBox(rows, cols)	Colored grid for weight matrices	.highlight_row(r), .highlight_col(c), .fill_anim()
VectorBar(dim, values)	Colored strip for embedding vectors	.pulse(), .highlight_dim(i), .add_value_labels()
TokenBox(text, token_id)	Rounded rectangle for a single token	.highlight(), .shake(), .pop_in(), .pop_out()
LabeledBlock(label)	Named rectangular block	.highlight(), .expand_to(mob)
SoftmaxCurve(logits)	Probability bar chart	.set_temperature(T), .show_values()
MathLabel(latex)	LaTeX equation label	.write(), .unwrite()
ConnectionLine(srcs, tgts)	Fully connected layer lines	.draw(), .activate(i,j), .pulse_all()

Animations — llmanim.base.animations

Function	Description
flow_through(mob, path)	Animate an object along a path
matrix_multiply_anim(A, B, C, scene)	Step-by-step matrix multiplication
highlight_sequence(boxes)	Flash boxes one by one
attention_flow(tokens, weights)	Weighted attention arrows between tokens
data_flow_arrow(A, B, label)	Animated labeled arrow between mobjects
equation_reveal(terms, scene)	Build an equation term by term
pulse_glow(mob)	Expanding glow effect
zoom_into(mob, scene)	Zoom in and back out
forward_pass_pulse(blocks)	Light pulse travelling through layers

Utilities — llmanim.base.utils

Function	Description
softmax(x, temperature)	Numerically stable softmax
scaled_dot_product(Q, K, V, mask)	Full attention computation
causal_mask(seq_len)	Upper-triangular causal mask
sinusoidal_pe(seq_len, d_model)	Sinusoidal positional encoding
make_attention_weights(seq_len, n_heads)	Generate demo attention weights
make_token_sequence(words)	Create a demo token sequence
value_to_color(v)	Map a float to a Manim color
attention_palette(weights)	Color palette from attention weights
apply_dark_theme(scene)	Set dark background on a Scene

Tokenization — llmanim.tokenization.token_box

Class	Description
TokenRow(tokens, token_ids)	A horizontal row of token boxes
TokenBox(text, token_id)	Individual token visualization
SpecialTokenBox(text)	Special tokens like [CLS], [SEP]
BPESequence	Byte-pair encoding visualization
VocabTable	Vocabulary lookup table
ContextWindowBar	Context window capacity indicator

---

Color System

All colors are in llmanim.styles.colors:

Constant	Hex	Used for
TOKEN_COLOR	#4CAF50	Tokens, raw input
EMBEDDING_COLOR	#2196F3	Embeddings, weight matrices
ATTENTION_COLOR	#9C27B0	Attention mechanisms
FFN_COLOR	#FF9800	Feed-forward networks
NORM_COLOR	#00BCD4	Layer normalization
OUTPUT_COLOR	#F44336	Output logits
QUERY_COLOR	#E91E63	Query (Q) vectors
KEY_COLOR	#3F51B5	Key (K) vectors
VALUE_COLOR	#009688	Value (V) vectors
HIGHLIGHT_COLOR	#FFEB3B	Highlights, focus
DIM_COLOR	#555555	Dimmed / background elements

---

Project Structure

LLManim/
├── llmanim/           # Installed package (import as llmanim)
│   ├── attention/               # QKV projection, scores, output, multi-head
│   ├── base/                    # shapes.py, animations.py, utils.py
│   ├── blocks/                  # transformer_block.py
│   ├── embeddings/              # embedding_lookup.py, positional_encoding.py
│   ├── feedforward/             # ffn.py
│   ├── normalization/           # layer_norm.py
│   ├── output/                  # logits_sampling.py
│   ├── styles/                  # colors.py, constants.py, fonts.py
│   ├── tokenization/            # token_box.py
│   └── full_model.py
├── examples/                    # Standalone video scripts (not installed)
├── videos/                      # Output GIFs for all example scenes
└── tests/

---

Video Production Workflow

1. Plan — write your script outline and narration
2. Import — pick components from llmanim
3. Sketch — plan scene layout on paper
4. Code — build scenes incrementally, preview with -pql
5. Render — final export with -pqh or -pqk
6. Voiceover — record audio to match animation timing
7. Edit — combine in your video editor of choice

---

Contributing

Contributions are welcome — new components, bug fixes, example videos, or docs improvements. See CONTRIBUTING.md to get started.

---

License

MIT © Rishabh Bhartiya

---

Built with Manim Community Edition