Struct BlockwiseFP8Linear 

pub struct BlockwiseFP8Linear { /* private fields */ }

Trait Implementations

impl Debug for BlockwiseFP8Linear

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

Formats the value using the given formatter.

impl QuantMethod for BlockwiseFP8Linear

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

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

If a is (n_tokens, 1, cols), self weights are (n_experts, rows, cols), then the indices are (n_tokens, n_experts_per_tok).

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

where Self: Sized,

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

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

Raw matmul without dtype casting. Implementors override this. Callers should use forward instead.

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

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

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

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

fn afq_inner(&self) -> Option<AfqInner<'_>>

If this is an AFQ layer, return its (w_q, scales, biases, bits, group_size). Used by Metal fused QKV / gate-up paths.

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

fn has_bias(&self) -> bool

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>

fn dummy_info(&self) -> Option<&DummyLayerInfo>

impl QuantizedSerde for BlockwiseFP8Linear

fn isq_serde_supported(&self) -> bool

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

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

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,

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

NOT meant for external calling