huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-19 19:58:35 +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

41
candle-transformers/src/models/quantized_llama.rs
View File

@ -356,6 +356,7 @@ impl ModelWeights {
pub fn from_gguf<R: std::io::Seek + std::io::Read>(
ct: gguf_file::Content,
reader: &mut R,
device: &Device,
) -> Result<Self> {
let cpu = &Device::Cpu;
let md_get = |s: &str| match ct.metadata.get(s) {
@ -383,21 +384,28 @@ impl ModelWeights {
.unwrap_or(10000f32);
let (cos, sin) = precomput_freqs_cis(rope_dim, rope_freq_base)?;
let tok_embeddings = ct.tensor(reader, "token_embd.weight")?;
let tok_embeddings = ct.tensor(reader, "token_embd.weight", device)?;
let tok_embeddings = tok_embeddings.dequantize(cpu)?;
let norm = RmsNorm::new(ct.tensor(reader, "output_norm.weight")?, rms_norm_eps)?;
let output = ct.tensor(reader, "output.weight")?;
let norm = RmsNorm::new(
ct.tensor(reader, "output_norm.weight", device)?,
rms_norm_eps,
)?;
let output = ct.tensor(reader, "output.weight", device)?;
let mut layers = Vec::with_capacity(block_count);
for layer_idx in 0..block_count {
let prefix = format!("blk.{layer_idx}");
let attention_wq = ct.tensor(reader, &format!("{prefix}.attn_q.weight"))?;
let attention_wk = ct.tensor(reader, &format!("{prefix}.attn_k.weight"))?;
let attention_wv = ct.tensor(reader, &format!("{prefix}.attn_v.weight"))?;
let attention_wo = ct.tensor(reader, &format!("{prefix}.attn_output.weight"))?;
let attention_wq = ct.tensor(reader, &format!("{prefix}.attn_q.weight"), device)?;
let attention_wk = ct.tensor(reader, &format!("{prefix}.attn_k.weight"), device)?;
let attention_wv = ct.tensor(reader, &format!("{prefix}.attn_v.weight"), device)?;
let attention_wo =
ct.tensor(reader, &format!("{prefix}.attn_output.weight"), device)?;
let mlp_or_moe = if n_expert <= 1 {
let feed_forward_w1 = ct.tensor(reader, &format!("{prefix}.ffn_gate.weight"))?;
let feed_forward_w2 = ct.tensor(reader, &format!("{prefix}.ffn_down.weight"))?;
let feed_forward_w3 = ct.tensor(reader, &format!("{prefix}.ffn_up.weight"))?;
let feed_forward_w1 =
ct.tensor(reader, &format!("{prefix}.ffn_gate.weight"), device)?;
let feed_forward_w2 =
ct.tensor(reader, &format!("{prefix}.ffn_down.weight"), device)?;
let feed_forward_w3 =
ct.tensor(reader, &format!("{prefix}.ffn_up.weight"), device)?;
MlpOrMoe::Mlp(Mlp {
feed_forward_w1: QMatMul::from_qtensor(feed_forward_w1)?,
feed_forward_w2: QMatMul::from_qtensor(feed_forward_w2)?,
@ -405,15 +413,15 @@ impl ModelWeights {
})
} else {
let feed_forward_gate_inp =
ct.tensor(reader, &format!("{prefix}.ffn_gate_inp.weight"))?;
ct.tensor(reader, &format!("{prefix}.ffn_gate_inp.weight"), device)?;
let mut experts = Vec::with_capacity(n_expert);
for i in 0..n_expert {
let feed_forward_w1 =
ct.tensor(reader, &format!("{prefix}.ffn_gate.{i}.weight"))?;
ct.tensor(reader, &format!("{prefix}.ffn_gate.{i}.weight"), device)?;
let feed_forward_w2 =
ct.tensor(reader, &format!("{prefix}.ffn_down.{i}.weight"))?;
ct.tensor(reader, &format!("{prefix}.ffn_down.{i}.weight"), device)?;
let feed_forward_w3 =
ct.tensor(reader, &format!("{prefix}.ffn_up.{i}.weight"))?;
ct.tensor(reader, &format!("{prefix}.ffn_up.{i}.weight"), device)?;
experts.push(Mlp {
feed_forward_w1: QMatMul::from_qtensor(feed_forward_w1)?,
feed_forward_w2: QMatMul::from_qtensor(feed_forward_w2)?,
@ -426,8 +434,9 @@ impl ModelWeights {
experts,
}
};
let attention_norm = ct.tensor(reader, &format!("{prefix}.attn_norm.weight"))?;
let ffn_norm = ct.tensor(reader, &format!("{prefix}.ffn_norm.weight"))?;
let attention_norm =
ct.tensor(reader, &format!("{prefix}.attn_norm.weight"), device)?;
let ffn_norm = ct.tensor(reader, &format!("{prefix}.ffn_norm.weight"), device)?;
let span_attn = tracing::span!(tracing::Level::TRACE, "attn");
let span_rot = tracing::span!(tracing::Level::TRACE, "attn-rot");
let span_mlp = tracing::span!(tracing::Level::TRACE, "attn-mlp");

4
candle-transformers/src/models/quantized_mixformer.rs
View File

@ -311,7 +311,7 @@ impl MixFormerSequentialForCausalLM {
let mut blocks = Vec::new();
for i in 0..cfg.n_layer {
let block = ParallelBlock::new(cfg, vb.pp(i + 1))?;
blocks.push(block)
blocks.push(block);
}
let head = CausalLMHead::new(cfg, vb.pp(cfg.n_layer + 1))?;
Ok(Self {
@ -332,7 +332,7 @@ impl MixFormerSequentialForCausalLM {
Some(get_mask(seq_len, xs.device())?)
};
for block in self.blocks.iter_mut() {
xs = block.forward(&xs, mask.as_ref())?
xs = block.forward(&xs, mask.as_ref())?;
}
xs.narrow(1, seq_len - 1, 1)?.apply(&self.head)?.squeeze(1)
}

12
candle-transformers/src/quantized_var_builder.rs
View File

@ -10,33 +10,33 @@ pub struct VarBuilder {
}
impl VarBuilder {
pub fn from_gguf<P: AsRef<std::path::Path>>(p: P) -> Result<Self> {
pub fn from_gguf<P: AsRef<std::path::Path>>(p: P, device: &Device) -> Result<Self> {
let mut file = std::fs::File::open(p)?;
let content = candle::quantized::gguf_file::Content::read(&mut file)?;
let mut data = std::collections::HashMap::new();
for tensor_name in content.tensor_infos.keys() {
let tensor = content.tensor(&mut file, tensor_name)?;
let tensor = content.tensor(&mut file, tensor_name, device)?;
data.insert(tensor_name.to_string(), Arc::new(tensor));
}
Ok(Self {
data: Arc::new(data),
path: Vec::new(),
device: Device::Cpu,
device: device.clone(),
})
}
pub fn from_gguf_buffer(buffer: &[u8]) -> Result<Self> {
pub fn from_gguf_buffer(buffer: &[u8], device: &Device) -> Result<Self> {
let mut cursor = std::io::Cursor::new(buffer);
let content = candle::quantized::gguf_file::Content::read(&mut cursor)?;
let mut data = std::collections::HashMap::new();
for tensor_name in content.tensor_infos.keys() {
let tensor = content.tensor(&mut cursor, tensor_name)?;
let tensor = content.tensor(&mut cursor, tensor_name, device)?;
data.insert(tensor_name.to_string(), Arc::new(tensor));
}
Ok(Self {
data: Arc::new(data),
path: Vec::new(),
device: Device::Cpu,
device: device.clone(),
})
}