pub enum Nvfp4Act {
Unstamped,
A4,
Wide,
}Expand description
2026-09-28: The activation format the policy set for one NVFP4 weight, which the loader
stamps on it (QuantizedWeight::act).
Variants§
Unstamped
2026-09-28: No stamp: the loader does not consult the policy, or the tier is nvfp4.
Decode follows the published tier’s lever, and prefill may run FP4 activations (the
dense-FFN MMQ arm), as before the plan existed.
A4
2026-09-28: The checkpoint declares FP4 input activations (W4A4): decode and prefill run them FP4.
Wide
2026-09-28: The checkpoint declares wider activations (A8 or A16): no FP4 activations in decode or prefill.
Implementations§
Source§impl Nvfp4Act
impl Nvfp4Act
Sourcepub fn combine(parts: impl IntoIterator<Item = Nvfp4Act>) -> Nvfp4Act
pub fn combine(parts: impl IntoIterator<Item = Nvfp4Act>) -> Nvfp4Act
2026-09-28: The stamp of a weight built from parts (a fused concatenation): FP4 only
when every part is, Wide when any part is, else unstamped.
Sourcepub fn allows_fp4_prefill(self) -> bool
pub fn allows_fp4_prefill(self) -> bool
2026-09-28: Whether prefill may quantize this weight’s activations to FP4.
Trait Implementations§
impl Copy for Nvfp4Act
impl Eq for Nvfp4Act
impl StructuralPartialEq for Nvfp4Act
Auto Trait Implementations§
impl Freeze for Nvfp4Act
impl RefUnwindSafe for Nvfp4Act
impl Send for Nvfp4Act
impl Sync for Nvfp4Act
impl Unpin for Nvfp4Act
impl UnwindSafe for Nvfp4Act
Blanket Implementations§
Source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
Source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
Source§impl<T> CloneToUninit for Twhere
T: Clone,
impl<T> CloneToUninit for Twhere
T: Clone,
§impl<Q, K> Equivalent<K> for Q
impl<Q, K> Equivalent<K> for Q
§fn equivalent(&self, key: &K) -> bool
fn equivalent(&self, key: &K) -> bool
§impl<Q, K> Equivalent<K> for Q
impl<Q, K> Equivalent<K> for Q
§fn equivalent(&self, key: &K) -> bool
fn equivalent(&self, key: &K) -> bool
key and return true if they are equal.