ggml-zdnn: add more safeguards in matmul

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-07-28 16:44:39 +08:00
parent f239bbb02d
commit cf0e190c40

View File

@@ -48,6 +48,7 @@ static bool ggml_zdnn_op_mul_mat(struct ggml_backend_zdnn_context * ctx, const g
const int64_t output_cols = ne0;
// have to do this because weights apparently do not go through set_tensor
if (&weights_extra->ztensor.is_transformed) zdnn_reset_ztensor(&weights_extra->ztensor);
zdnn_init_pre_transformed_desc(
ZDNN_2D,
FP32,
@@ -55,6 +56,20 @@ static bool ggml_zdnn_op_mul_mat(struct ggml_backend_zdnn_context * ctx, const g
weights->ne[1], weights->ne[0]
);
ZDNN_CHECK(zdnn_transform_ztensor(&weights_extra->ztensor, weights->data));
ZDNN_CHECK(zdnn_generate_transformed_desc(&weights_extra->pre_tfm_desc, &weights_extra->tfm_desc));
ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&weights_extra->pre_tfm_desc, &weights_extra->tfm_desc, &weights_extra->ztensor));
// have to do this here because although it was transformed, the shape is wrong
if (&inputs_extra->ztensor.is_transformed) zdnn_reset_ztensor(&inputs_extra->ztensor);
zdnn_init_pre_transformed_desc(
ZDNN_2D,
FP32,
&inputs_extra->pre_tfm_desc,
inputs->ne[1], inputs->ne[0]
);
ZDNN_CHECK(zdnn_transform_ztensor(&inputs_extra->ztensor, inputs->data));
ZDNN_CHECK(zdnn_generate_transformed_desc(&inputs_extra->pre_tfm_desc, &inputs_extra->tfm_desc));
ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&inputs_extra->pre_tfm_desc, &inputs_extra->tfm_desc, &inputs_extra->ztensor));
// have to transform the bias ztensor here because only GGML_OP_NONE goes through set_tensor
ZDNN_CHECK(zdnn_transform_ztensor(&bias_extra->ztensor, bias_extra->data));