A high-performance inference engine for AI models
Fix QuantInput packed-weight allocation overflow (#762)
CCYehd8095eea629d
diff --git a/crates/uzu-engine/unit/common/matmul/quant.rs b/crates/uzu-engine/unit/common/matmul/quant.rsindex 4ff4f5ffe..b0772b60e 100644--- a/crates/uzu-engine/unit/common/matmul/quant.rs+++ b/crates/uzu-engine/unit/common/matmul/quant.rs@@ -67,7 +67,8 @@ impl<T: ArrayElement + Float> QuantInput<T> {let num_groups_k = k.div_ceil(group_size);let mut rng = SmallRng::seed_from_u64(seed);- let w_packed: Vec<u32> = (0..(n * k * bits / 32) as usize).map(|_| rng.random_range(0..u32::MAX)).collect();+ let w_packed: Vec<u32> =+ (0..(n as usize * k as usize * bits as usize / 32)).map(|_| rng.random_range(0..u32::MAX)).collect();let scales: Vec<T> =(0..(n * num_groups_k) as usize).map(|_| T::from(rng.random_range(0.01f32..0.3f32)).unwrap()).collect();let x: Vec<T> = (0..(m * k) as usize).map(|_| T::from(rng.random_range(-0.3f32..0.3f32)).unwrap()).collect();