A high-performance inference engine for AI models
Causal swa dflash (#709)
uuuvn47c67a4fca38
diff --git a/crates/backend-uzu/src/encodable_block/mixer/attention/mode.rs b/crates/backend-uzu/src/encodable_block/mixer/attention/mode.rsindex e2b205bc5..c0f3a4b51 100644--- a/crates/backend-uzu/src/encodable_block/mixer/attention/mode.rs+++ b/crates/backend-uzu/src/encodable_block/mixer/attention/mode.rs@@ -8,12 +8,15 @@ use crate::{encodable_block::{batch_topology::BatchTopology,linear::Linear,- mixer::attention::{- Attention,- core::AttentionCoreEncodeArguments,- qkv_norm::QKVNorm,- rope::PrecalculatedRoPE,- state::{AttentionState, AttentionStateType},+ mixer::{+ MixerState,+ attention::{+ Attention,+ core::AttentionCoreEncodeArguments,+ qkv_norm::QKVNorm,+ rope::PrecalculatedRoPE,+ state::{AttentionState, AttentionStateType},+ },},},utils::maybe_mut::MaybeMut,@@ -166,7 +169,7 @@ impl<B: Backend> Attention<B> {batch_dim,encoder,)?;- state.append_full(batch_dim);+ state.encode_accept(&(0..batch_dim).collect::<Box<[usize]>>(), encoder)?;Ok(())}diff --git a/crates/backend-uzu/src/encodable_block/mixer/attention/state.rs b/crates/backend-uzu/src/encodable_block/mixer/attention/state.rsindex 3b482ae36..6fb9fc64e 100644--- a/crates/backend-uzu/src/encodable_block/mixer/attention/state.rs+++ b/crates/backend-uzu/src/encodable_block/mixer/attention/state.rs@@ -134,20 +134,6 @@ impl<B: Backend> AttentionState<B> {kv_cache_update,})}-- pub(super) fn append_full(- &mut self,- length: usize,- ) {- let AttentionStateType::Full {- length: state_length,- } = &mut self.state_type- else {- panic!("append_full requires full attention state");- };- *state_length += length;- self.cur_context_length += length;- }}impl<B: Backend> MixerState<B> for AttentionState<B> {