CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

CANN: Add 310P operator support check (#12962)

Browse Source
This commit is contained in:
Chenguang Li
2025-04-16 16:21:05 +08:00
committed by GitHub
parent 80f19b4186
commit b43d89e311
2 changed files with 18 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@@ -625,6 +625,10 @@ static void ggml_cann_avg_pool2d(ggml_backend_cann_context& ctx,
bool count_include_pad = true; bool count_include_pad = true;
int64_t divisor_override = 0; int64_t divisor_override = 0;
int8_t cube_math_type = 0; int8_t cube_math_type = 0;
#ifdef ASCEND_310P
cube_math_type = 1;
#endif
GGML_CANN_CALL_ACLNN_OP(AvgPool2d, acl_src, kernel_size, strides, paddings_avg, GGML_CANN_CALL_ACLNN_OP(AvgPool2d, acl_src, kernel_size, strides, paddings_avg,
ceil_mode, count_include_pad, divisor_override, ceil_mode, count_include_pad, divisor_override,
cube_math_type, acl_dst); cube_math_type, acl_dst);
@@ -2590,6 +2594,10 @@ void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* ds
int64_t groups = 1; int64_t groups = 1;
int8_t cubeMathType = 0; int8_t cubeMathType = 0;
#ifdef ASCEND_310P
cubeMathType = 1;
#endif
GGML_CANN_CALL_ACLNN_OP(Convolution, acl_input, acl_weight, nullptr, stride, GGML_CANN_CALL_ACLNN_OP(Convolution, acl_input, acl_weight, nullptr, stride,
padding, dilation, transposed, padding, groups, acl_dst, cubeMathType); padding, dilation, transposed, padding, groups, acl_dst, cubeMathType);

10
ggml/src/ggml-cann/ggml-cann.cpp
View File

@@ -2022,6 +2022,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
return true; return true;
case GGML_TYPE_Q8_0: case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_0:
#ifdef ASCEND_310P
// Q4 && Q8 per group is not suppor on 310p device
return false;
#endif
// only support contiguous for quantized types. // only support contiguous for quantized types.
return ggml_is_contiguous(op->src[0]) && return ggml_is_contiguous(op->src[0]) &&
ggml_is_contiguous(op->src[1]); ggml_is_contiguous(op->src[1]);
@@ -2107,6 +2111,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
} }
case GGML_OP_POOL_2D: { case GGML_OP_POOL_2D: {
const int32_t * opts = (const int32_t *) op->op_params; const int32_t * opts = (const int32_t *) op->op_params;
#ifdef ASCEND_310P
enum ggml_op_pool opt = static_cast<ggml_op_pool>(opts[0]);
if(opt == GGML_OP_POOL_MAX){
return false;
}
#endif
const int k0 = opts[1]; const int k0 = opts[1];
const int k1 = opts[2]; const int k1 = opts[2];
const int p0 = opts[5]; const int p0 = opts[5];