tune and run local with turboquant

This commit is contained in:
2026-06-22 10:37:43 +07:00
parent a0713b260b
commit 2b5977b605
3 changed files with 39 additions and 6 deletions
+5 -2
View File
@@ -30,7 +30,7 @@ fi
: "${CACHE_TYPE_K:=q8_0}"
: "${CACHE_TYPE_V:=turbo2}"
: "${THREADS:=8}"
: "${IMAGE:=ghcr.io/ggml-org/llama.cpp:server-cuda}"
: "${IMAGE:=ghcr.io/ggml-org/llama.cpp:server-cuda13}"
: "${BIND_HOST:=0.0.0.0}"
# Разбить по пробелам для podman (пусто = только CDI-устройство).
@@ -42,6 +42,7 @@ podman run -d \
--replace \
--restart no \
--shm-size 1g \
--cpuset-cpus="0-11" \
--ulimit memlock=-1:-1 \
-p "${BIND_HOST}:${PORT}:8080" \
-v "${ROOT}/models:/models:ro" \
@@ -52,7 +53,9 @@ podman run -d \
-m "/models/${MODEL_FILE}" \
--host 0.0.0.0 \
--port 8080 \
-ngl "${N_GPU_LAYERS}" \
--n-cpu-moe "${N_CPU_MOE}" \
--spec-type draft-mtp --spec-draft-n-max 3 \
-c "${CTX_SIZE}" \
-np 1 \
-fa on \
@@ -61,7 +64,7 @@ podman run -d \
--no-mmap \
--mlock \
--ctx-checkpoints 1 \
--cache-ram 0 \
--cache-ram -1 \
--jinja \
--reasoning on \
--reasoning-budget -1 \