musa: disable MUL_MAT_ID (q2_k × f32) due to precision issues

Signed-off-by: Xiaodong Ye <[email protected]>

Author: yeahdongcn

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -3070,6 +3070,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                             a->type == GGML_TYPE_F16 && b->type == GGML_TYPE_F16) {
                         return false;
                     }
+                    if (GGML_CUDA_CC_IS_QY2(cc) && op->op == GGML_OP_MUL_MAT_ID &&
+                            a->type == GGML_TYPE_Q2_K && b->type == GGML_TYPE_F32) {
+                        return false;
+                    }
                 }
 #endif // GGML_USE_MUSA
                 switch (a->type) {

ggml/src/ggml-musa/mudnn.cuh

@@ -1,7 +1,7 @@
 #pragma once
 
-#include "../include/ggml.h"
-#include "../ggml-cuda/common.cuh"
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
 
 // Asynchronously copies data from src tensor to dst tensor using the provided context.
 // Returns a musaError_t indicating success or failure.