Struct PerTensorFP8Linear 

pub struct PerTensorFP8Linear { /* private fields */ }

Per-tensor FP8 Linear layer with static activation scaling.

This is used for models that have per-tensor FP8 quantization (weight_block_size = null) with static activation scales. Each linear layer has:

• <layer>.weight (FP8 E4M3)
• <layer>.weight_scale_inv (F32 scalar) - dequantization scale for weights
• <layer>.activation_scale (F32 scalar) - quantization scale for activations

Trait Implementations

impl Debug for PerTensorFP8Linear

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl QuantMethod for PerTensorFP8Linear

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

where Self: Sized,

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

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

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

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

If a quantized method, return the activation dtype.

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.

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

Weight dtype and device

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

If the quant is backed by a qmatmul.

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

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

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

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

fn gather_forward(&self, _a: &Tensor, _indices: &Tensor) -> Result<Tensor>

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

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

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

Begin tracking stats into an ImatrixLayerStats

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

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

fn is_distributed(&self) -> Option<DistributedKind>

impl QuantizedSerde for PerTensorFP8Linear

fn isq_serde_supported(&self) -> bool

fn name(&self) -> &'static str

fn serialize(&self) -> Result<Cow<'_, [u8]>>

fn serialize_with_bias(&self, bias: Option<Tensor>) -> Result<Cow<'_, [u8]>>

NOT meant for external calling

fn deserialize( _data: Cow<'_, [u8]>, _device: &Device, _comm: &Arc<Comm>, _guard: QuantizeOntoGuard, ) -> Result<Arc<dyn QuantMethod>>

where Self: Sized,

fn deserialize_ext_bias( _data: Cow<'_, [u8]>, _device: &Device, _guard: QuantizeOntoGuard, ) -> Result<(Arc<dyn QuantMethod>, Option<Tensor>)>

where Self: Sized,