CUDA: fix compilation with GGML_CUDA_F16 (ggml-org#14837)

JohannesGaessler · web-flow · commit b284197df426 · 2025-07-23T18:22:30.000+02:00
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
@@ -31,8 +31,8 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     dequantize_kernel(vx, ib, iqs, v);
 
     const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = v.x;
-    y[iy0 + y_offset] = v.y;
+    y[iy0 + 0]        = float(v.x);
+    y[iy0 + y_offset] = float(v.y);
 }
 
 template <bool need_check>

Original file line number	Diff line number	Diff line change
`@@ -31,8 +31,8 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __`
`31`	`31`	`dequantize_kernel(vx, ib, iqs, v);`
`32`	`32`
`33`	`33`	`const int64_t iy0 = ((i03ne02 + i02)ne01 + i01)*ne00 + iybs + iqs;`
`34`		`- y[iy0 + 0] = v.x;`
`35`		`- y[iy0 + y_offset] = v.y;`
	`34`	`+ y[iy0 + 0] = float(v.x);`
	`35`	`+ y[iy0 + y_offset] = float(v.y);`
`36`	`36`	`}`
`37`	`37`
`38`	`38`	`template <bool need_check>`