A high-performance inference engine for AI models
Fix classifier (#494)
uuuvn71aca00b4e24
diff --git a/crates/backend-uzu/src/classifier/classifier.rs b/crates/backend-uzu/src/classifier/classifier.rsindex ac8b2a066..83ffc881b 100644--- a/crates/backend-uzu/src/classifier/classifier.rs+++ b/crates/backend-uzu/src/classifier/classifier.rs@@ -4,7 +4,7 @@ use std::{collections::HashMap, path::Path, time::Instant};use super::ActivationTrace;use super::{ClassificationOutput, ClassificationStats, ClassifierContext};use crate::{- backends::common::{Allocation, AsBufferRangeRef, Backend, Encoder},+ backends::common::{Allocation, AllocationType, AsBufferRangeRef, Backend, Context, Encoder},config::model::classifier_model::ClassifierModelConfig,data_type::DataType,encodable_block::{LayerArguments, PrecalculatedRope},@@ -199,19 +199,26 @@ impl<B: Backend> Classifier<B> {encoder.encode_copy(&logits, .., trace.logits.allocation_mut(), ..);}+ let mut logits_output = self+ .context+ .context+ .create_allocation(logits.as_buffer_range_ref().range().len(), AllocationType::Global)+ .map_err(|e| Error::EncodeFailed(Box::new(e)))?;+ encoder.encode_copy(&logits, .., &mut logits_output, ..);++ drop(logits);+ drop(main);+ drop(shortcut);+ drop(ropes);+let completed = encoder.end_encoding().submit().wait_until_completed().map_err(|e| Error::CommandBufferFailed(Box::new(e)))?;-- let logits_result = self.copy_logits_from_allocation(&logits);- drop(logits);- drop(main);- drop(shortcut);drop(completed);- logits_result+ self.copy_logits_from_allocation(&logits_output)}#[cfg(not(feature = "tracing"))]