Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-20 01:31:31 +02:00
Code Issues Packages Projects Releases Wiki Activity

ggml-cuda: avoid ROCm_Host compute on HIP integrated GPUs

Browse Source
This commit is contained in:
Victor-Loos
2026-07-18 14:47:23 +02:00
committed by Piotr Wilkin
parent 5efd7400ac
commit 52ff274b82
1 changed files with 23 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/ggml-cuda.cu
+23 -2
View File
@@ -4946,6 +4946,20 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
: GGML_BACKEND_DEVICE_TYPE_GPU;
}
static bool ggml_backend_cuda_host_buffer_supported() {
return getenv("GGML_CUDA_NO_PINNED") == nullptr;
}
static bool ggml_backend_cuda_device_supports_cuda_host_buft(int device) {
#if defined(GGML_USE_HIP)
if (ggml_cuda_info().devices[device].integrated) {
return false;
}
#endif
return ggml_backend_cuda_host_buffer_supported();
}
static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
@@ -4955,7 +4969,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
props->device_id = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
bool host_buffer = ggml_backend_cuda_host_buffer_supported();
#ifdef GGML_CUDA_NO_PEER_COPY
bool events = false;
#else
@@ -4984,6 +4998,10 @@ static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_buffer_type(ggml_
static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_host_buffer_type(ggml_backend_dev_t dev) {
GGML_UNUSED(dev);
if (!ggml_backend_cuda_host_buffer_supported()) {
return nullptr;
}
return ggml_backend_cuda_host_buffer_type();
}
@@ -5444,7 +5462,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
const bool integrated = ggml_cuda_info().devices[dev_ctx->device].integrated;
return (ggml_backend_buft_is_cuda(buft) && buft->device == dev) || (integrated && ggml_backend_buft_is_cuda_host(buft));
return (ggml_backend_buft_is_cuda(buft) && buft->device == dev) ||
(integrated &&
ggml_backend_buft_is_cuda_host(buft) &&
ggml_backend_cuda_device_supports_cuda_host_buft(dev_ctx->device));
}
static int64_t get_op_batch_size(const ggml_tensor * op) {