Update ggml/src/ggml-cuda/solve_tri.cu

wsbagnsv1 · JohannesGaessler · web-flow · commit 2fd926483529 · 2025-12-04T20:11:05.000+01:00
Co-authored-by: Johannes Gäßler &lt;johannesg@5d6.de&gt;
diff --git a/ggml/src/ggml-cuda/solve_tri.cu b/ggml/src/ggml-cuda/solve_tri.cu
@@ -70,7 +70,7 @@ static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
     for (int row = 0; row < nrows_low; ++row) {
         float sum = 0.0f;
         if (lane < row) {
-            sum = fmaf(sA[row * n + lane], x_low, sum);
+            sum += sA[row * n + lane] * x_low;
         }
         sum = warp_reduce_sum(sum);
 

Original file line number	Diff line number	Diff line change
`@@ -70,7 +70,7 @@ static __global__ void solve_tri_f32_fast(const float * __restrict__ A,`
`70`	`70`	`for (int row = 0; row < nrows_low; ++row) {`
`71`	`71`	`float sum = 0.0f;`
`72`	`72`	`if (lane < row) {`
`73`		`- sum = fmaf(sA[row * n + lane], x_low, sum);`
	`73`	`+ sum += sA[row * n + lane] * x_low;`
`74`	`74`	`}`
`75`	`75`	`sum = warp_reduce_sum(sum);`
`76`	`76`