Update dynamic_quantize_gpu_kv_cache.cl

openvinotoolkit · Jan 15, 2025 · 1fbd98f · 1fbd98f
1 parent 317638e
commit 1fbd98f
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/dynamic_quantize_gpu_kv_cache.cl b/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/dynamic_quantize_gpu_kv_cache.cl
@@ -89,7 +89,7 @@ KERNEL(dynamic_quantize_gpu_kv_cache)(
     max_value = work_group_reduce_max(max_value);
 
     // If the range of input data is zero, it is adjusted to the minimum value(0.001).
-    half diff_value = max_value == min_value ? (grp_max) : (max_value - min_value);
+    ACCUMULATOR_TYPE diff_value = max_value == min_value ? (grp_max) : (max_value - min_value);
     ACCUMULATOR_TYPE scale_tmp = (ACCUMULATOR_TYPE)((CHAR_MAX - CHAR_MIN) / diff_value);
     ACCUMULATOR_TYPE zp_tmp = (ACCUMULATOR_TYPE)(-min_value * scale_tmp) + CHAR_MIN;
     OUTPUT1_TYPE scale = (OUTPUT1_TYPE)(scale_tmp);