mistralrs_quant

Trait QuantMethod

Source 
pub trait QuantMethod:
    Send
    + Sync
    + Debug
    + QuantizedSerde {

Show 13 methods    // Required methods
    fn new(method: QuantMethodConfig) -> Result<Self>
       where Self: Sized;
    fn dequantize_w(&self) -> Result<Tensor>;
    fn forward(&self, a: &Tensor) -> Result<Tensor>;
    fn quantized_act_type(&self) -> Option<DType>;
    fn dtype_and_device(&self) -> (DType, Device);
    fn add_delta_w(&self, delta: &Tensor) -> Result<Arc<dyn QuantMethod>>;
    fn apply_isq(
        self: Arc<Self>,
        dtype: Option<IsqType>,
        device: Device,
        n_quantized: &AtomicUsize,
        imatrix_weight: Option<Vec<f32>>,
    ) -> Result<Arc<dyn QuantMethod>>;
    fn get_max_isq_cpu_threads(&self, dtype: IsqType) -> Option<NonZeroUsize>;

    // Provided methods
    fn forward_autocast(&self, a: &Tensor) -> Result<Tensor> { ... }
    fn forward_via_half(&self, a: &Tensor) -> Result<Tensor> { ... }
    fn unquant_weight_bias(&self) -> Option<(Tensor, Option<Tensor>)> { ... }
    fn begin_track_stats(&mut self) -> Result<()> { ... }
    fn end_track_stats(&self) -> Result<Tensor> { ... }

}
Expand description

Quantized method for a quantized matmul.

Required Methods§

Source

fn new(method: QuantMethodConfig) -> Result<Self>
where Self: Sized,

Source

fn dequantize_w(&self) -> Result<Tensor>

Source

fn forward(&self, a: &Tensor) -> Result<Tensor>

Compute matmul of self and a. self should contain the weights.

Source

fn quantized_act_type(&self) -> Option<DType>

If a quantized method, return the activation dtype.

Source

fn dtype_and_device(&self) -> (DType, Device)

Weight dtype and device

Source

fn add_delta_w(&self, delta: &Tensor) -> Result<Arc<dyn QuantMethod>>

Add a delta weight from LoRA to the weights. This should be prescaled with alpha.

Source

fn apply_isq( self: Arc<Self>, dtype: Option<IsqType>, device: Device, n_quantized: &AtomicUsize, imatrix_weight: Option<Vec<f32>>, ) -> Result<Arc<dyn QuantMethod>>

If the quant is backed by a qmatmul.

Source

fn get_max_isq_cpu_threads(&self, dtype: IsqType) -> Option<NonZeroUsize>

Provided Methods§

Source

fn forward_autocast(&self, a: &Tensor) -> Result<Tensor>

Compute matmul of self and a. self should contain the weights. Automatically cast to required quantization actiation type and back

Source

fn forward_via_half(&self, a: &Tensor) -> Result<Tensor>

Compute matmul of self and a. self should contain the weights. This may go via half precision if it is supported.

Source

fn unquant_weight_bias(&self) -> Option<(Tensor, Option<Tensor>)>

Source

fn begin_track_stats(&mut self) -> Result<()>

Begin tracking stats into an ImatrixLayerStats

Source

fn end_track_stats(&self) -> Result<Tensor>

End tracking stats into an ImatrixLayerStats. Returns the computed imatrix.

Trait Implementations§

Source§

impl Module for dyn QuantMethod

Source§

fn forward(&self, xs: &Tensor) -> Result<Tensor>

Implementors§