mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 32 Packages Projects Releases Wiki Activity

ggml-virtgpu: make the code thread safe (#19204)

Browse Source 
* ggml-virtgpu: regenerate_remoting.py: add the ability to deprecate a function

* ggml-virtgpu: deprecate buffer_type is_host remoting

not necessary

* ggml-virtgpu: stop using static vars as cache

The static init isn't thread safe.

* ggml-virtgpu: protect the use of the shared memory to transfer data

* ggml-virtgpu: make the remote calls thread-safe

* ggml-virtgpu: backend: don't continue if couldn't allocate the tensor memory

* ggml-virtgpu: add a cleanup function for consistency

* ggml-virtgpu: backend: don't crash if buft->iface.get_max_size is missing

* fix style and ordering

* Remove the static variable in apir_device_get_count

* ggml-virtgpu: improve the logging

* fix review minor formatting changes
This commit is contained in:
Kevin Pouget
2026-02-04 03:46:18 +01:00
committed by GitHub
parent 2ceda3f662
commit 015deb9048
27 changed files with 397 additions and 237 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp
+22 -6
View File
@@ -36,13 +36,18 @@ void apir_buffer_set_tensor(virtgpu * gpu,
virtgpu_shmem temp_shmem; // Local storage for large buffers
virtgpu_shmem * shmem = &temp_shmem;
bool using_shared_shmem = false;
if (size <= gpu->data_shmem.mmap_size) {
// prefer the init-time allocated page, if large enough
// Lock mutex before using shared data_shmem buffer
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
}
using_shared_shmem = true;
shmem = &gpu->data_shmem;
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
}
memcpy(shmem->mmap_ptr, data, size);
@@ -55,7 +60,10 @@ void apir_buffer_set_tensor(virtgpu * gpu,
remote_call_finish(gpu, encoder, decoder);
if (shmem != &gpu->data_shmem) {
// Unlock mutex before cleanup
if (using_shared_shmem) {
mtx_unlock(&gpu->data_shmem_mutex);
} else {
virtgpu_shmem_destroy(gpu, shmem);
}
@@ -79,13 +87,18 @@ void apir_buffer_get_tensor(virtgpu * gpu,
virtgpu_shmem temp_shmem; // Local storage for large buffers
virtgpu_shmem * shmem = &temp_shmem;
bool using_shared_shmem = false;
if (size <= gpu->data_shmem.mmap_size) {
// prefer the init-time allocated page, if large enough
// Lock mutex before using shared data_shmem buffer
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
GGML_ABORT(GGML_VIRTGPU "%s: Failed to lock data_shmem mutex", __func__);
}
using_shared_shmem = true;
shmem = &gpu->data_shmem;
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
GGML_ABORT(GGML_VIRTGPU "%s: Couldn't allocate the guest-host shared buffer", __func__);
}
apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
@@ -98,7 +111,10 @@ void apir_buffer_get_tensor(virtgpu * gpu,
remote_call_finish(gpu, encoder, decoder);
if (shmem != &gpu->data_shmem) {
// Unlock mutex before cleanup
if (using_shared_shmem) {
mtx_unlock(&gpu->data_shmem_mutex);
} else {
virtgpu_shmem_destroy(gpu, shmem);
}
}