CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-09 10:17:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml-zdnn: rework init_tensor to create new buffers

Browse Source 
Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo
2025-07-24 01:03:53 +08:00
parent ee0ed78d54
commit b7f4b6fde3
1 changed files with 22 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
ggml/src/ggml-zdnn/ggml-zdnn-rewrite.cpp
View File

@@ -244,32 +244,32 @@ 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;
// Find which buffer entry this tensor belongs to // Create a dedicated buffer entry for this tensor
const int64_t tsize = ggml_nbytes(tensor); const int64_t tsize = ggml_nbytes(tensor);
size_t offset = 0;
int buffer_idx = -1;
for (int i = 0; i < ctx->n_buffers; ++i) { if (ctx->n_buffers >= 64) {
const int64_t ioffs = (int64_t)tensor->data - (int64_t)ctx->buffers[i].data; GGML_LOG_ERROR("%s: too many tensors, maximum 64 supported\n", __func__);
if (ioffs >= 0 && ioffs + tsize <= (int64_t)ctx->buffers[i].size) { return GGML_STATUS_FAILED;
buffer_idx = i;
offset = (size_t)ioffs;
break;
}
} }
if (buffer_idx >= 0) { int buffer_idx = ctx->n_buffers;
// Initialize the buffer entry for this specific tensor
ctx->buffers[buffer_idx].data = tensor->data;
ctx->buffers[buffer_idx].size = tsize;
// Copy tensor name to buffer for logging // Copy tensor name to buffer for logging
strncpy(ctx->buffers[buffer_idx].name, tensor->name, sizeof(ctx->buffers[buffer_idx].name) - 1); strncpy(ctx->buffers[buffer_idx].name, tensor->name, sizeof(ctx->buffers[buffer_idx].name) - 1);
ctx->buffers[buffer_idx].name[sizeof(ctx->buffers[buffer_idx].name) - 1] = '\0'; ctx->buffers[buffer_idx].name[sizeof(ctx->buffers[buffer_idx].name) - 1] = '\0';
GGML_LOG_INFO("%s: initialized tensor '%s' in buffer %d, size = %8.2f MiB, offset = %zu\n", // Initialize tensor descriptors for this tensor
// TODO: Set up pre_tfm_desc, tfm_desc, and ztensor based on tensor properties
ctx->n_buffers++;
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[buffer_idx].name, buffer_idx,
(float)tsize / (1024.0f * 1024.0f), offset); (float)tsize / (1024.0f * 1024.0f));
} else {
GGML_LOG_WARN("%s: could not find buffer for tensor '%s'\n", __func__, tensor->name);
return GGML_STATUS_FAILED;
}
tensor->extra = &ctx->buffers[buffer_idx]; tensor->extra = &ctx->buffers[buffer_idx];