huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-15 18:28:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

Quantized GGUF style (#1523)

Browse Source 
* Metal quantized modifications proposal.

- Add a device param, wherever needed.
- Create new QMetal storage thing that implements QuantizedType.
- Update everywhere needed.

Fix Python.

Fixing examples.

Fix: fmt + clippy + stub.

Moving everything around.

Only missing the actual implems.

Fixing everything + adding dequantized kernels.

More work.

Fixing matmul.

Fmt + Clippy

Some clippy fixes.

Working state.

Q2K Metal -> Bugged (also present in GGML).
Q4K CPU -> Bugged (present previously, new test catch it).
Q5K CPU -> Bugged (present previously).
Q8_1 Both -> Never really implemented it seems
Q8K metal -> Never implemented in metal

Fixing Q2K bug (present in ggml).

* Cleanup.

* Fix the rebase.

* Removing the fences speeds everything up and *is* correct this time...

* Cleanup the fence.

* After rebase.

* Bad code removal.

* Rebase after phi2 merge + fix replit default to CPU.

* Making the CI happy.

* More happy tests.

---------

Co-authored-by: Nicolas Patry <nicolas@Nicolass-MacBook-Pro.local>
This commit is contained in:
Nicolas Patry
2024-01-17 10:27:58 +01:00
committed by GitHub
parent 5270224f40
commit 403680f17d
31 changed files with 6446 additions and 515 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
candle-examples/examples/quantized/main.rs
View File

@ -9,7 +9,7 @@ use std::io::Write;
use tokenizers::Tokenizer;
use candle::quantized::{ggml_file, gguf_file};
use candle::{Device, Tensor};
use candle::Tensor;
use candle_transformers::generation::LogitsProcessor;
use candle_examples::token_output_stream::TokenOutputStream;
@ -361,6 +361,7 @@ fn main() -> anyhow::Result<()> {
let model_path = args.model()?;
let mut file = std::fs::File::open(&model_path)?;
let start = std::time::Instant::now();
let device = candle_examples::device(false)?;
let mut model = match model_path.extension().and_then(|v| v.to_str()) {
Some("gguf") => {
@ -369,7 +370,7 @@ fn main() -> anyhow::Result<()> {
for (_, tensor) in model.tensor_infos.iter() {
let elem_count = tensor.shape.elem_count();
total_size_in_bytes +=
elem_count * tensor.ggml_dtype.type_size() / tensor.ggml_dtype.blck_size();
elem_count * tensor.ggml_dtype.type_size() / tensor.ggml_dtype.block_size();
}
println!(
"loaded {:?} tensors ({}) in {:.2}s",
@ -377,15 +378,16 @@ fn main() -> anyhow::Result<()> {
&format_size(total_size_in_bytes),
start.elapsed().as_secs_f32(),
);
ModelWeights::from_gguf(model, &mut file)?
ModelWeights::from_gguf(model, &mut file, &device)?
}
Some("ggml" | "bin") | Some(_) | None => {
let model = ggml_file::Content::read(&mut file).map_err(|e| e.with_path(model_path))?;
let model = ggml_file::Content::read(&mut file, &device)
.map_err(|e| e.with_path(model_path))?;
let mut total_size_in_bytes = 0;
for (_, tensor) in model.tensors.iter() {
let elem_count = tensor.shape().elem_count();
total_size_in_bytes +=
elem_count * tensor.dtype().type_size() / tensor.dtype().blck_size();
elem_count * tensor.dtype().type_size() / tensor.dtype().block_size();
}
println!(
"loaded {:?} tensors ({}) in {:.2}s",
@ -486,7 +488,7 @@ fn main() -> anyhow::Result<()> {
let start_prompt_processing = std::time::Instant::now();
let mut next_token = {
let input = Tensor::new(prompt_tokens.as_slice(), &Device::Cpu)?.unsqueeze(0)?;
let input = Tensor::new(prompt_tokens.as_slice(), &device)?.unsqueeze(0)?;
let logits = model.forward(&input, 0)?;
let logits = logits.squeeze(0)?;
logits_processor.sample(&logits)?
@ -507,7 +509,7 @@ fn main() -> anyhow::Result<()> {
let start_post_prompt = std::time::Instant::now();
let mut sampled = 0;
for index in 0..to_sample {
let input = Tensor::new(&[next_token], &Device::Cpu)?.unsqueeze(0)?;
let input = Tensor::new(&[next_token], &device)?.unsqueeze(0)?;
let logits = model.forward(&input, prompt_tokens.len() + index)?;
let logits = logits.squeeze(0)?;
let logits = if args.repeat_penalty == 1. {