From 704bb2f222acdc5b974587148c3a0fa0726f9ebb Mon Sep 17 00:00:00 2001
From: xrip <xrip@xrip.ru>
Date: Wed, 9 Sep 2026 12:04:40 +0300
Subject: [PATCH] build: add SM75 dense and MoE scripts

---
 build-dense.sh | 17 +++++++++++++++++
 build-moe.sh   | 17 +++++++++++++++++
 2 files changed, 34 insertions(+)
 create mode 100755 build-dense.sh
 create mode 100755 build-moe.sh

diff --git a/build-dense.sh b/build-dense.sh
new file mode 100755
index 000000000..83df4d74c
--- /dev/null
+++ b/build-dense.sh
@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+root=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)
+
+# Force MMQ for supported quantized operations; floating-point operations still use cuBLAS.
+cmake -S "$root" -B "$root/build-dense" \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_C_COMPILER=gcc-14 -DCMAKE_CXX_COMPILER=g++-14 \
+    -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
+    -DCMAKE_CUDA_HOST_COMPILER=g++-14 -DCMAKE_CUDA_ARCHITECTURES=75 \
+    -DGGML_NATIVE=ON -DGGML_OPENMP=ON \
+    -DGGML_CUDA=ON -DGGML_VULKAN=OFF -DGGML_BLAS=OFF \
+    -DGGML_CUDA_FORCE_MMQ=ON -DGGML_CUDA_FORCE_CUBLAS=OFF \
+    -DLLAMA_BUILD_TESTS=ON -DLLAMA_BUILD_TOOLS=ON
+cmake --build "$root/build-dense" --parallel "${CMAKE_BUILD_PARALLEL_LEVEL:-12}" \
+    --target llama-server llama-cli llama-bench test-backend-ops
diff --git a/build-moe.sh b/build-moe.sh
new file mode 100755
index 000000000..905c52d0a
--- /dev/null
+++ b/build-moe.sh
@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+root=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)
+
+# Keep the existing MoE build's forced-cuBLAS setting.
+cmake -S "$root" -B "$root/build-cublas" \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_C_COMPILER=gcc-14 -DCMAKE_CXX_COMPILER=g++-14 \
+    -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
+    -DCMAKE_CUDA_HOST_COMPILER=g++-14 -DCMAKE_CUDA_ARCHITECTURES=75 \
+    -DGGML_NATIVE=ON -DGGML_OPENMP=ON \
+    -DGGML_CUDA=ON -DGGML_VULKAN=OFF -DGGML_BLAS=OFF \
+    -DGGML_CUDA_FORCE_MMQ=OFF -DGGML_CUDA_FORCE_CUBLAS=ON \
+    -DLLAMA_BUILD_TESTS=ON -DLLAMA_BUILD_TOOLS=ON
+cmake --build "$root/build-cublas" --parallel "${CMAKE_BUILD_PARALLEL_LEVEL:-12}" \
+    --target llama-server llama-cli llama-bench test-backend-ops