Skip to content

armgit / uzu

Public Imported from GitHub
0

A high-performance inference engine for AI models

work around llvm regalloc segfault when compiling attn gemm (#729)

uuuvn9d02ea0cb6d8
diff --git a/crates/backend-uzu/src/backends/metal/kernel/attention/attention_gemm.metal b/crates/backend-uzu/src/backends/metal/kernel/attention/attention_gemm.metal
index b19f733f1..3010a678a 100644
--- a/crates/backend-uzu/src/backends/metal/kernel/attention/attention_gemm.metal
+++ b/crates/backend-uzu/src/backends/metal/kernel/attention/attention_gemm.metal
@@ -112,7 +112,7 @@ KERNEL(AttentionGemm)(
using ScoreFragment = Fragment<AccumType, QUERY_ROW_FRAGMENTS, KEY_COL_FRAGMENTS, Ops>;
QueryFragment query_frags[QUERY_FRAGMENT_COUNT];
ScoreFragment score_fragment;
- constexpr int VALUE_COL_FRAGMENTS = USE_MXU ? 2 : HEAD_DIM_FRAGMENTS;
+ constexpr int VALUE_COL_FRAGMENTS = USE_MXU ? 2 : (HEAD_DIM_FRAGMENTS < 16 ? HEAD_DIM_FRAGMENTS : 16);
static_assert(HEAD_DIM_FRAGMENTS % VALUE_COL_FRAGMENTS == 0, "head-dim must split into VALUE_COL_FRAGMENTS frags");
static_assert(!USE_MXU || VALUE_COL_FRAGMENTS % 2 == 0, "MXU PV needs even N (VALUE_COL_FRAGMENTS)");
constexpr int OUTPUT_CHUNKS = HEAD_DIM_FRAGMENTS / VALUE_COL_FRAGMENTS;