vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

cuda : ignore peer access already enabled errors

Browse source
This commit is contained in:
slaren 2024-02-19 21:37:49 +01:00
parent 9d679f0fcc
commit 3b82c9c7b0
1 changed files with 13 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

20
ggml-cuda.cu
View file

@ -9325,9 +9325,15 @@ static void ggml_cuda_set_peer_access(const int n_tokens) {
CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other)); CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
if (can_access_peer) { if (can_access_peer) {
if (enable_peer_access) { if (enable_peer_access) {
CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0)); cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
if (err != cudaErrorPeerAccessAlreadyEnabled) {
CUDA_CHECK(err);
}
} else { } else {
CUDA_CHECK(cudaDeviceDisablePeerAccess(id_other)); cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
if (err != cudaErrorPeerAccessNotEnabled) {
CUDA_CHECK(err);
}
} }
} }
} }
@ -10999,10 +11005,10 @@ GGML_CALL static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backe
UNUSED(buffer); UNUSED(buffer);
} }
// unused at the moment static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
//static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) { return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
// return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name; UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
//} }
GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) { GGML_CALL static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
@ -11390,7 +11396,7 @@ GGML_CALL static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, gg
for (int j = 0; j < GGML_MAX_SRC; j++) { for (int j = 0; j < GGML_MAX_SRC; j++) {
if (node->src[j] != nullptr) { if (node->src[j] != nullptr) {
assert(node->src[j]->backend == GGML_BACKEND_GPU || node->src[j]->backend == GGML_BACKEND_GPU_SPLIT); assert(node->src[j]->backend == GGML_BACKEND_GPU || node->src[j]->backend == GGML_BACKEND_GPU_SPLIT);
assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device)); assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
assert(node->src[j]->extra != nullptr); assert(node->src[j]->extra != nullptr);
} }
} }