happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Fix Issue !19219 

Hangs were reported on Jetson Orin AGX if we set CUDA_SCALE_LAUNCH_QUEUES=4x. Reverting the previous PR (#19042) and updating the document to consider setting CUDA_SCALE_LAUNCH_QUEUES=4x for faster throughput on multi-GPU systems.
This commit is contained in:
Gaurav Garg 2026-01-31 12:29:04 +05:30
parent 89f10baad5
commit 027ef82d5c
2 changed files with 1 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
docs/build.md
View File

@ -252,9 +252,7 @@ CUDA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.ggu
The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU. The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU.
**Default behavior:** llama.cpp automatically sets `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs. Consider setting `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
See PR [#19042](https://github.com/ggml-org/llama.cpp/pull/19042) for performance benchmarks and technical details.
### Unified Memory ### Unified Memory

10
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -5049,16 +5049,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
static std::mutex mutex; static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex); std::lock_guard<std::mutex> lock(mutex);
if (!initialized) { if (!initialized) {
// Set CUDA_SCALE_LAUNCH_QUEUES before any CUDA API call to improve multi-GPU pipeline parallelism performance
// PR: https://github.com/ggml-org/llama.cpp/pull/19042
if (getenv("CUDA_SCALE_LAUNCH_QUEUES") == nullptr) {
#ifdef _WIN32
_putenv_s("CUDA_SCALE_LAUNCH_QUEUES", "4x");
#else
setenv("CUDA_SCALE_LAUNCH_QUEUES", "4x", 0); // don't overwrite if already set
#endif // _WIN32
}
ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context; ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32; const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;