Quantized GGUF style (#1523)

* 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>

candle-transformers/src/quantized_var_builder.rs

@@ -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(),
         })
     }