121 releases (7 breaking)
0.8.19 | Sep 22, 2024 |
---|---|
0.8.17 | Jul 3, 2024 |
0.6.35 | Mar 25, 2024 |
0.4.9 | Dec 25, 2023 |
0.1.16 | Jul 30, 2023 |
#44 in Machine learning
89 downloads per month
Used in ai00-core
640KB
17K
SLoC
Web-RWKV
This is an inference engine for the language model of RWKV implemented in pure WebGPU.
Features
- No dependencies on CUDA/Python.
- Support Nvidia/AMD/Intel GPUs, including integrated GPUs.
- Vulkan/Dx12/OpenGL backends.
- WASM support (can run in browser).
- Batched inference.
- Int8 and NF4 quantization.
- Very fast.
- LoRA merging at loading time.
- Support RWKV V4, V5 and V6.
- Hooks to intervene the inference process at any point.
- Model (de)serialization.
Note that web-rwkv
is only an inference engine. It only provides the following functionalities:
- A tokenizer.
- Model loading.
- State creation and updating.
- Model implements
run
function that takes in prompt tokens and returns logits, and asoftmax
function that turns logits into predicted next token probabilities. Both of them are executed on GPU. - Model quantization and (de)serialization.
- WASM bindings.
It does not provide the following:
- OpenAI API or APIs of any kind.
- If you would like to deploy an API server, check AI00 RWKV Server which is a fully-functional OpenAI-compatible API server built upon
web-rwkv
. - You could also check the
web-rwkv-axum
project if you want some fancy inference pipelines, including Classifier-Free Guidance (CFG), Backus–Naur Form (BNF) guidance, and more.
- If you would like to deploy an API server, check AI00 RWKV Server which is a fully-functional OpenAI-compatible API server built upon
- Samplers, though in the examples a basic nucleus sampler is implemented, this is not included in the library itself.
- State caching or management system.
- Python bindings.
Compile
- Install Rust.
- Download the model from HuggingFace, and convert it using
convert_safetensors.py
. Put the.st
model underassets/models
. - Compile
$ cargo build --release --examples
Examples
Performance Test
The test generates 500 tokens and measure the time cost.
$ cargo run --release --example rt-gen
Chat Demo
To chat with the model, run
$ cargo run --release --example rt-chat
In this demo, type +
to retry last round's generation; type -
to exit.
-
To specify the location of your safetensors model, use
$ cargo run --release --example rt-chat -- --model /path/to/model
-
To load custom prompts for chat, use
$ cargo run --release --example rt-chat -- --prompt /path/to/prompt
See
assets/prompt.json
for details. -
To specify layer quantization, use
--quant <LAYERS>
or--quant-nf4 <LAYERS>
to quantize the first<LAYERS>
layers. For example, use$ cargo run --release --example rt-chat -- --quant 32
to quantize all 32 layers.
-
UseThe--turbo
flag to switch to alternativeGEMM
kernel when inferring long prompts.rt
examples enforcesturbo
.
Batched Inference
This demo showcases generation of 4 batches of text with various lengths simultaneously.
$ cargo run --release --example rt-batch
Inspector
The inspector demo is a guide to an advanced usage called hooks. Hooks allow user to inject any tensor ops into the model's inference process, fetching and modifying the contents of the runtime buffer, state, and even the model parameters. Hooks enable certain third-party implementations like dynamic LoRA, control net, and so on.
(De)serialization
All versions of models implements serde::ser::Serialize
and serde::de::DeserializeSeed<'de>
, which means that one can save quantized or lora-merged model into a file and load it afterwards.
Use in Your Project
To use in your own rust project, simply add web-rwkv = "0.8"
as a dependency in your Cargo.toml
.
Check examples on how to create the environment, the tokenizer and how to run the model.
Explanations
Inference Runtime
Since v0.7 there is a runtime
feature for the crate. When enabled, applications can use infrastructures of the asynchronous runtime
API.
In general, a runtime
is an asynchronous task that is driven by tokio
. It allows CPU and GPU to work in parallel, maximizing the utilization of GPU computing resource.
Check examples starting with rt
for more information, and compare the generation speed with their non-rt
counterparts.
Batched Inference
Since version v0.2.4, the engine supports batched inference, i.e., inference of a batch of prompts (with different length) in parallel.
This is achieved by a modified WKV
kernel.
When building the model, the user specifies token_chunk_size
(default: 32, but for powerful GPUs this could be much higher), which is the maximum number of tokens the engine could process in one run
call.
After creating the model, the user creates a ModelState
with num_batch
specified.
This means that there are num_batch
slots that could consume the inputs in parallel.
Before calling run()
, the user fills each slot with some tokens as prompt.
If a slot is empty, no inference will be run for it.
After calling run()
, some (but may not be all) input tokens are consumed, and logits
appears in their corresponding returned slots if the inference of that slot is finished during this run.
Since there are only token_chunk_size
tokens are processed during each run()
call, there may be none of logits
appearing in the results.
Hooks
Hooks are a very powerful tool for customizing model inference process.
The library provides with the Model::run_with_hooks
function, which takes into a HookMap
as a parameter.
A HookMap
is essentially a hashmap from Model::Hook
to functions.
A Model::Hook
defines a certain place the hook function can be injected into. A model generally has dozens of hooking points.
A hook function is a function of Fn(&Model<'_>, &ModelState, &Runtime) -> Result<TensorOp, TensorError>
, where you can create tensor ops that reads/writes all the tensors you get here.
An example that reads out every layer's output:
let info = model.info();
// create a buffer to store each layer's output
let buffer = Buffer::new(&context, &info);
let mut hooks = HookMap::default();
for layer in 0..info.num_layer {
let buffer = buffer.clone();
hooks.insert(
v5::Hook::PostFfn(layer),
Box::new(
move |_model, _state, runtime: &v5::Runtime| -> Result<TensorOp, TensorError> {
// figure out how many tokens this run has
let shape = runtime.ffn_x.shape();
let num_token = shape[1];
// "steal" the layer's output (activation), and put it into our buffer
TensorOp::blit(
runtime.ffn_x.view(.., num_token - 1, .., ..)?,
buffer.ffn_x.view(.., layer, .., ..)?,
)
},
),
);
}
let output = model.run_with_hooks(&mut tokens, &state, &hooks).await?;
Convert Models
You must download the model and put in assets/models
before running if you are building from source.
You can now download the converted models here.
You may download the official RWKV World series models from HuggingFace, and convert them via the provided convert_safetensors.py
.
$ python convert_safetensors.py --input /path/to/model.pth --output /path/to/model.st
If you don't have python installed or don't want to, there is a pure rust converter
.
You can clone that repo and run
$ cd /path/to/web-rwkv-converter
$ cargo run --release --example converter -- --input /path/to/model.pth --output /path/to/model.st
Troubleshoot
-
"thread 'main' panicked at 'called
Result::unwrap()
on anErr
value: HeaderTooLarge'"Your model is broken, mainly because you cloned the repo but did not set up git-lfs.Please download the model manually and overwrite that one in
assets/models
. -
"thread 'main' panicked at 'Error in Queue::submit: parent device is lost'"
Your GPU is not responding. Maybe you are running a model that is just too big for your device. If the model doesn't fit into your VRam, the driver needs to constantly swap and transfer the model parameters, causing it to be 10x slower. Try to quantize your model first.
Credits
- Tokenizer is implemented by @koute.
Dependencies
~10–45MB
~703K SLoC