friendliai/tokasaurus

friendliai/tokasaurus

License: Apache-2.0

Stars: 0

Forks: 0

Open issues: 0

Created: 2025-06-06T10:14:59Z

Pushed: 2025-06-05T18:02:45Z

Default branch: main

Fork: yes

Parent repository: ScalingIntelligence/tokasaurus

Archived: no

README:

Tokasaurus: The Little (LLM) Engine That Could!

Table of Contents

• [What is This?](#what-is-this)
• [Installation](#installation)
• [Quickstart](#quickstart)
• [Walkthrough of CLI Flags](#walkthrough-of-cli-flags)
• [System Design](#system-design)

What is This?

Tokasaurus is an LLM inference engine designed for high-throughput workloads. Features include:

• OpenAI chat, completions, and batch APIs.
• Data, pipeline, and tensor parallelism (with support for AsyncTP).
• Support for Llama3 and Qwen2 architectures.
• Paged KV caching with prefix caching.
• Efficient attention over shared prefixes with Hydragen, with automatic detection of shared prefixes across groups of sequences.
• End-to-end torch compile with dynamic shapes.
• CUDA graphs.
• Very low CPU overhead (important for small models/fast GPUs).
• A scheduler that can simulate the number of available KV cache blocks thousands of steps in the future, allowing us to aggressively onboard new sequences and keep our batch size as large as possible.
• No OOMs or recompiles in production: on engine startup, we launch a series of warmup inputs that trigger all torch recompiles ahead-of-time (torch will recompile whenever a tensor has an input dimension is 0 or 1) and make check for OOMs using the largest configured batch size.

NOTE: as a new project, expect some potentially rough edges :).

Installation

Tokasaurus has been tested on Python >= 3.10. To install from PyPI, run:

pip install tokasaurus

Alternatively, clone the repo and run:

pip install -e .

Quickstart

Once installed, you can launch the engine with:

# launch engine for Llama 1B (by default on port 10210).
toka model=meta-llama/Llama-3.2-1B-Instruct

# make a request to the engine (this command just wraps the OpenAI client)
toka ping prompt='tell me a joke' max_tokens=256 chat=True

# launch a 70B model with pipeline parallelism across 8 gpus
toka model=meta-llama/Llama-3.1-70B-Instruct kv_cache_num_tokens='(512 * 1024)' pp_size=8

To ping the engine once it's been launched, you can use the OpenAI client:

from openai import OpenAI
client = OpenAI(
api_key='fake-key',
base_url="http://0.0.0.0:10210/v1"
)
response = client.completions.create(
model="default",
prompt="On a dark desert highway, cool wind in my hair, warm smell of colitas, rising up through the air, up ahead in the distance, I saw a shimmering light, my head grew heavy and my sight grew dim, I had to stop for the night",
temperature=0,
n=2,
max_tokens=100,
)

LM Eval Harness

Since the engine supports the OpenAI API, you can plug it into the EleutherAI LM Eval harness using their local completions feature. First spin up an engine (see above) and then run evals on it with:

lm_eval --model local-completions --tasks gsm8k --model_args model=MODEL,base_url=http://0.0.0.0:PORT/v1/completions,num_concurrent=256,max_retries=3,tokenized_requests=False

Walkthrough of CLI Flags

The tokasaurus CLI uses Pydra, which uses a key=value format to set config flags. It also allows for boolean shorthands (e.g. key=T is equivalent to key=True) and allows for Python expression evaluation between parentheses (e.g. key='(2 * 1024)' is equivalent to key=2048).

The Basics

The only required parameter to launch an engine is the model field, which can point to a repo on HF or a local directory where a model is stored in HF format (just like when calling from_pretrained on a HF model). By default, the tokenizer will also be loaded using using the model flag. This can be overridden by setting the tokenizer flag yourself:

toka model=meta-llama/Llama-3.2-1B-Instruct

# e.g. if you want to load a fine-tuned model you saved to disk
toka model=my_local_dir tokenizer=meta-llama/Llama-3.2-1B-Instruct

Leveraging Multiple GPUs

By default, the engine will only use a single GPU to serve the model. You can change this with the dp_size, pp_size, and tp_size flags to control data, pipeline, and tensor parallelism, respectively. These flags are composable: for example, dp_size=2 and pp_size=4 will use 8 GPUs in total by creating two data-parallel replicas that each contain 4 GPUs in a pipeline:

Managing GPU Memory with KV Cache Limits and Concurrency Controls

The total amount of GPU memory used by the engine is the sum of GPU memory used to store the model weights, the activations, and the KV cache. While the model's GPU memory is fixed for a given model, we can control the size of the KV cache and the amount of activation memory we use.

The KV cache size is controlled with kv_cache_size_num_tokens, and we can cap activation memory with the flags max_tokens_per_forward and max_seqs_per_forward. With max_tokens_per_forward, you directly control the number of tokens being sent through the model in a single forward pass, which can include tokens from sequences running either prefill or decode. With max_seqs_per_forward, we control the total number of sequences that can be running (i.e. that are in prefill or in decode) at a given time. Importantly, this limits the number of tokens per forward pass that can ever be sent through the language modeling head of the model, which can have a disproportionately large impact on activation memory. Prefill tokens don't run through the LM head (since we don't need to decode anything from them), so they take less activation memory.

How should you tune these flags? Well, one of the most important factors for achieving high throughput is making the batch size as large as possible. A common bottleneck that limits the batch size in practice is the size of the KV cache - once your KV cache is full, you can't run any more sequences concurrently. Therefore, we want to make the KV cache as large as possible. However, in order to benefit from a large KV cache that can fit many sequences, we also must increase max_seqs_per_forward and…