Skip to content

armgit / uzu

Public Imported from GitHub
0

A high-performance inference engine for AI models

remove unused vars (#206)

eugenea8a31b8e850d
diff --git a/crates/uzu/src/backends/metal/kernel/moe/experts_two_pass_prefill.metal b/crates/uzu/src/backends/metal/kernel/moe/experts_two_pass_prefill.metal
index 9f28ba119..6ff575bdc 100644
--- a/crates/uzu/src/backends/metal/kernel/moe/experts_two_pass_prefill.metal
+++ b/crates/uzu/src/backends/metal/kernel/moe/experts_two_pass_prefill.metal
@@ -469,8 +469,7 @@ KERNEL(MoeExpertsPrefillPassB)(
// Cooperative load config - vec8 loads for better bandwidth
constexpr uint H_VEC = 8;
- constexpr uint H_TCOLS = Bk / H_VEC; // 64/8 = 8
- constexpr uint H_TROWS = TGP_SIZE / H_TCOLS; // 128/8 = 16
+ constexpr uint H_TCOLS = Bk / H_VEC; // 64/8 = 8
const uint h_bi = lin / H_TCOLS;
const uint h_bj = (lin % H_TCOLS) * H_VEC;
@@ -590,4 +589,4 @@ KERNEL(MoeExpertsPrefillPassB)(
output[out_row * (ulong)d_model + out_col] = T(val);
}
}
-}
\ No newline at end of file
+}
diff --git a/crates/uzu/src/backends/metal/kernel/quant_matmul/quant_matmul.metal b/crates/uzu/src/backends/metal/kernel/quant_matmul/quant_matmul.metal
index c2408b656..ed923674d 100644
--- a/crates/uzu/src/backends/metal/kernel/quant_matmul/quant_matmul.metal
+++ b/crates/uzu/src/backends/metal/kernel/quant_matmul/quant_matmul.metal
@@ -1096,7 +1096,6 @@ void qmv_impl_dispatch(
const int in_vec_size_g =
(K + group_size - 1) / group_size; // ceil(K / group_size)
const device T* scales_base = scales;
- const device uint8_t* zero_points_base = zero_points;
const int out_row = tid.y * (num_simdgroups * results_per_simdgroup) +
simd_gid * results_per_simdgroup;
const int used_out_row = min(N - results_per_simdgroup, out_row);
diff --git a/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.metal b/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.metal
index bf1898172..f8517a6ba 100644
--- a/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.metal
+++ b/crates/uzu/src/backends/metal/kernel/ssm/ssd_prefill.metal
@@ -31,7 +31,6 @@ KERNEL(SSDPrefill64)(
const uint lane_idx THREADS(32)
) {
const uint simd_width = 32;
- const int state_dim = state_size;
const uint h_idx = pair_idx / head_dim;
const uint dh_idx = pair_idx % head_dim;
const uint safe_group = uint(max(group_size, 1));
@@ -275,4 +274,4 @@ KERNEL(SSDPrefillSequential)(
acc *= this_z;
y[x_idx] = acc;
}
-}
\ No newline at end of file
+}