From: Shawn Gu <redacted>
Date: Tue, 23 Jun 2026 05:25:21 +0000 (-0700)
Subject: opencl: q8_0 gemv precision improvement (#24923)
X-Git-Tag: upstream/0.0.10438~674
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=23ee8797e11b1f9502f547a9250bfa0f5c36a63f;p=pkg%2Fggml%2Fsources%2Fllama.cpp

opencl: q8_0 gemv precision improvement (#24923)
---

diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl
index 9703b693e..f5c6fb3e8 100644
--- a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl
@@ -174,7 +174,7 @@ __kernel void kernel_gemv_noshuffle_q8_0_f32(
         regA.s6 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
         regA.s7 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
 
-        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, regS, regB);
+        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, convert_float(regS), regB);
     }
 
     // reduction in local memory, assumes #wave=4