upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-28 08:57:40 -05:00
Code Issues Packages Projects Releases Wiki Activity

webgpu: Handle unaligned writes in ggml_backend_webgpu_buffer_set_tensor (#29471)

Browse Source 
* Fix:  Handle unaligned writes in ggml_backend_webgpu_buffer_set_tensor

* Clang formatting
This commit is contained in:
jbooth
2026-09-28 16:38:10 +03:00
committed by GitHub
parent d77dd0806d
commit f00a64c147
1 changed files with 22 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/ggml-webgpu.cpp
+22 -1
View File
@@ -3774,7 +3774,28 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
size_t total_offset = ggml_webgpu_tensor_offset(tensor) + offset;
buf_ctx->global_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
// WriteBuffer needs the offset and the size to be multiples of 4.
// Write the misaligned head bytes using compute memset, then increment total_offset
// and data pointer so that they are 4-aligned.
if (total_offset % 4 != 0) {
size_t lane = total_offset % 4; // in-word lane the head starts at (the tail below always starts at 0)
size_t head = std::min<size_t>(4 - lane, size);
// Pack head bytes into a uint32_t
uint32_t head_val = 0;
for (size_t i = 0; i < head; i++) {
((uint8_t *) &head_val)[lane + i] = ((const uint8_t *) data)[i];
}
ggml_backend_webgpu_buffer_memset(buf_ctx->global_ctx, buf_ctx->buffer, head_val, total_offset, head);
total_offset += head;
size -= head;
data = (const uint8_t *) data + head;
}
if (size > 0) {
buf_ctx->global_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
}
if (size % 4 != 0) {
// If size is not a multiple of 4, we need to memset the remaining bytes