nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

ggml: add env var GGML_OP_OFFLOAD_MIN_BATCH (#18535)

Browse Source 
* ggml: add env var GGML_OP_OFFLOAD_MIN_BATCH
* makes the min_batch_size for triggering op offload configurable via env var, defaulting to the prior hardcoded value of 32

* ggml: read GGML_OP_OFFLOAD_MIN_BATCH once and store to dev ctx

* cann: forward declaration of device context struct

* cann: move offload op check after device context declaration

* cuda: fix whitespace

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

---------

Co-authored-by: Aman Gupta <amangupta052@gmail.com>
This commit is contained in:
Doctor Shotgun
2026-01-08 11:03:21 +02:00
committed by GitHub
co-authored by Aman Gupta
parent 9c142e3a2a
commit 9a5724dee2
7 changed files with 45 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cann/ggml-cann.cpp
+23 -21
View File
@@ -2541,27 +2541,6 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
}
/**
* @brief Determines if a tensor operation should be offloaded to the CANN
* backend.
*
* This function checks if a given tensor operation should be offloaded to the
* CANN backend based on the operation type and the size of the tensor. It
* returns true if the second dimension (ne[1]) of the tensor is greater than or
* equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
*
* @param backend Pointer to the CANN backend.
* @param op Pointer to the tensor operation to check.
* @return bool Returns true if the operation should be offloaded, otherwise
* false.
*/
static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
const int min_batch_size = 32;
GGML_UNUSED(dev);
return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
}
/**
* @brief Records an event on the CANN backend stream.
*
@@ -2637,6 +2616,7 @@ struct ggml_backend_cann_device_context {
int device;
std::string name;
std::string description;
int op_offload_min_batch_size;
};
static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
@@ -2713,6 +2693,26 @@ static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(
return ggml_backend_cann_host_buffer_type();
}
/**
* @brief Determines if a tensor operation should be offloaded to the CANN
* backend.
*
* This function checks if a given tensor operation should be offloaded to the
* CANN backend based on the operation type and the size of the tensor. It
* returns true if the second dimension (ne[1]) of the tensor is greater than or
* equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
*
* @param backend Pointer to the CANN backend.
* @param op Pointer to the tensor operation to check.
* @return bool Returns true if the operation should be offloaded, otherwise
* false.
*/
static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
return op->ne[1] >= dev_ctx->op_offload_min_batch_size && op->op != GGML_OP_GET_ROWS;
}
/**
* @brief Creates a new event for the CANN backend device.
*
@@ -2829,12 +2829,14 @@ ggml_backend_reg_t ggml_backend_cann_reg() {
if (!initialized) {
aclInit(nullptr);
ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
for (int i = 0; i < ggml_cann_info().device_count; i++) {
ggml_backend_cann_device_context * dev_ctx = new ggml_backend_cann_device_context();
dev_ctx->description = aclrtGetSocName();
dev_ctx->device = i;
dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
dev_ctx->op_offload_min_batch_size = min_batch_size;
ggml_cann_set_device(i);
ggml_backend_dev_t dev = new ggml_backend_device{ /* .iface = */ ggml_backend_cann_device_interface,
/* .reg = */ &reg,