upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-07 02:20:48 -05:00
Code Issues Packages Projects Releases Wiki Activity

add messages for allreduce fallback

Browse Source
This commit is contained in:
Scott Cutler
2026-04-23 16:32:14 -07:00
parent 77c0e36ceb
commit 43743e5ade
1 changed files with 17 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -1351,6 +1351,7 @@ enum ggml_cuda_comm_allreduce_result {
static ggml_cuda_comm_allreduce_result ggml_backend_cuda_comm_try_allreduce_internal(
ggml_backend_cuda_comm_context * comm_ctx, struct ggml_tensor ** tensors) {
if (comm_ctx->ar_pipeline == nullptr) {
GGML_LOG_WARN("%s: internal unsupported: pipeline unavailable\n", __func__);
return GGML_CUDA_COMM_ALLREDUCE_UNSUPPORTED;
}
@@ -1362,10 +1363,12 @@ static ggml_cuda_comm_allreduce_result ggml_backend_cuda_comm_try_allreduce_inte
const ggml_type type = tensors[0]->type;
if (n_backends != 2) {
GGML_LOG_WARN("%s: internal unsupported: n_backends=%zu\n", __func__, n_backends);
return GGML_CUDA_COMM_ALLREDUCE_UNSUPPORTED;
}
if (type != GGML_TYPE_F32 && type != GGML_TYPE_F16 && type != GGML_TYPE_BF16) {
GGML_LOG_WARN("%s: internal unsupported: type=%d\n", __func__, (int) type);
return GGML_CUDA_COMM_ALLREDUCE_UNSUPPORTED;
}
@@ -1375,16 +1378,30 @@ static ggml_cuda_comm_allreduce_result ggml_backend_cuda_comm_try_allreduce_inte
const size_t bytes = (size_t) ne * ggml_type_size(type);
if (bytes > GGML_CUDA_AR_MAX_BYTES) {
GGML_LOG_WARN("%s: internal unsupported: ne=%" PRId64 " type=%d bytes=%zu max=%zu\n",
__func__, ne, (int) type, bytes, GGML_CUDA_AR_MAX_BYTES);
return GGML_CUDA_COMM_ALLREDUCE_UNSUPPORTED;
}
for (size_t i = 0; i < n_backends; ++i) {
if (tensors[i] == nullptr) {
GGML_LOG_ERROR("%s: internal failed: tensor[%zu] is null\n", __func__, i);
return GGML_CUDA_COMM_ALLREDUCE_FAILED;
}
if (ggml_nelements(tensors[i]) != ne || tensors[i]->type != type) {
GGML_LOG_ERROR("%s: internal failed: tensor[%zu] ne=%" PRId64 " type=%d expected ne=%" PRId64 " type=%d\n",
__func__, i, ggml_nelements(tensors[i]), (int) tensors[i]->type, ne, (int) type);
return GGML_CUDA_COMM_ALLREDUCE_FAILED;
}
if (!ggml_is_contiguously_allocated(tensors[i])) {
GGML_LOG_WARN("%s: internal tensor[%zu] is not contiguously allocated: ne=%" PRId64 " nbytes=%zu packed=%zu type=%d\n",
__func__, i, ne, ggml_nbytes(tensors[i]),
(size_t) ne * ggml_type_size(type) / ggml_blck_size(type), (int) type);
}
if (((uintptr_t) tensors[i]->data & 0xF) != 0) {
GGML_LOG_WARN("%s: internal tensor[%zu] data pointer is not 16-byte aligned: %p type=%d ne=%" PRId64 "\n",
__func__, i, tensors[i]->data, (int) type, ne);
}
}
return ggml_backend_cuda_comm_allreduce_internal(comm_ctx, tensors)