CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-zdnn: add bias init_tensor

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-07-28 15:41:56 +08:00
parent f263f5d9ae
commit aef93b3908
2 changed files with 46 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
ggml/src/ggml-zdnn/ggml-zdnn-impl.h
View File

@@ -65,6 +65,9 @@ struct ggml_backend_zdnn_buffer {
zdnn_ztensor ztensor; zdnn_ztensor ztensor;
char name[GGML_MAX_NAME]; char name[GGML_MAX_NAME];
struct ggml_backend_zdnn_buffer * extra; // for bias etc.
ggml_backend_zdnn_buffer() : extra(nullptr) {}
}; };
#endif // GGML_ZDNN_IMPL #endif // GGML_ZDNN_IMPL

52
ggml/src/ggml-zdnn/ggml-zdnn-rewrite.cpp
View File

@@ -281,6 +281,18 @@ static enum ggml_status ggml_zdnn_graph_compute(ggml_backend_t backend, struct g
return GGML_STATUS_SUCCESS; return GGML_STATUS_SUCCESS;
} }
static void ggml_zdnn_init_bias_tensor(struct ggml_backend_zdnn_buffer * buffer, struct ggml_tensor * tensor) {
zdnn_init_pre_transformed_desc(
ZDNN_1D,
FP32,
&buffer->pre_tfm_desc,
tensor->ne[0], 1, 1, 1
);
ZDNN_CHECK(zdnn_generate_transformed_desc(&buffer->pre_tfm_desc, &buffer->tfm_desc));
ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&buffer->pre_tfm_desc, &buffer->tfm_desc, &buffer->ztensor));
}
static void ggml_zdnn_init_tensor(struct ggml_backend_zdnn_buffer * buffer, struct ggml_tensor * tensor) { static void ggml_zdnn_init_tensor(struct ggml_backend_zdnn_buffer * buffer, struct ggml_tensor * tensor) {
switch (tensor->op) { switch (tensor->op) {
case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT:
@@ -341,23 +353,45 @@ static enum ggml_status ggml_backend_zdnn_buffer_init_tensor(ggml_backend_buffer
struct ggml_backend_zdnn_buffer_context * ctx = (struct ggml_backend_zdnn_buffer_context *)buffer->context; struct ggml_backend_zdnn_buffer_context * ctx = (struct ggml_backend_zdnn_buffer_context *)buffer->context;
// Create a dedicated buffer entry for this tensor // Create a dedicated buffer entry for this tensor
int tensor_buffer_idx;
int bias_buffer_idx;
const int64_t tsize = ggml_nbytes(tensor); const int64_t tsize = ggml_nbytes(tensor);
int buffer_idx = ctx->n_buffers;
ctx->buffers[buffer_idx].data = tensor->data; struct ggml_backend_zdnn_buffer * tensor_buffer;
ctx->buffers[buffer_idx].size = tsize; tensor_buffer_idx = ctx->n_buffers;
strncpy(ctx->buffers[buffer_idx].name, tensor->name, sizeof(ctx->buffers[buffer_idx].name) - 1); tensor_buffer = &ctx->buffers[tensor_buffer_idx];
ctx->buffers[buffer_idx].name[sizeof(ctx->buffers[buffer_idx].name) - 1] = '\0'; tensor_buffer->data = tensor->data;
tensor_buffer->size = tsize;
ggml_zdnn_init_tensor(&ctx->buffers[buffer_idx], tensor); strncpy(tensor_buffer->name, tensor->name, sizeof(tensor_buffer->name) - 1);
tensor_buffer->name[sizeof(tensor_buffer->name) - 1] = '\0';
ggml_zdnn_init_tensor(tensor_buffer, tensor);
ctx->n_buffers++; ctx->n_buffers++;
if (tensor->op == GGML_OP_MUL_MAT) {
struct ggml_backend_zdnn_buffer * bias_buffer;
bias_buffer_idx = tensor_buffer_idx + 1;
bias_buffer = &ctx->buffers[bias_buffer_idx];
bias_buffer->data = calloc(tensor->ne[0], tensor->ne[0] * sizeof(float));
bias_buffer->size = tensor->ne[0] * sizeof(float);
strncpy(bias_buffer->name, "bias", sizeof(bias_buffer->name) - 1);
bias_buffer->name[sizeof(bias_buffer->name) - 1] = '\0';
ggml_zdnn_init_bias_tensor(bias_buffer, tensor);
ctx->n_buffers++;
tensor_buffer->extra = bias_buffer;
GGML_LOG_INFO("%s: initialized bias tensor '%s' in buffer %d, size = %8.2f MiB\n",
__func__, bias_buffer->name, bias_buffer_idx,
(float)bias_buffer->size / (1024.0f * 1024.0f));
}
GGML_LOG_INFO("%s: initialized tensor '%s' in buffer %d, size = %8.2f MiB\n", GGML_LOG_INFO("%s: initialized tensor '%s' in buffer %d, size = %8.2f MiB\n",
__func__, ctx->buffers[buffer_idx].name, buffer_idx, __func__, ctx->buffers[tensor_buffer_idx].name, tensor_buffer_idx,
(float)tsize / (1024.0f * 1024.0f)); (float)tsize / (1024.0f * 1024.0f));
tensor->extra = &ctx->buffers[buffer_idx]; tensor->extra = &ctx->buffers[tensor_buffer_idx];
return GGML_STATUS_SUCCESS; return GGML_STATUS_SUCCESS;
} }