gubernac's picture
download
raw
1.93 kB
#!/usr/bin/env bash
# GPU job entrypoint: unpack cached binary, mount models, start orchestrator.
# Args: benchmark plan (e.g. "A C D")
set -euo pipefail
PLAN="${1:-A C D}"
echo "=== GPU job entry: plan='$PLAN' ==="
nvidia-smi || echo "WARNING: nvidia-smi failed"
nvidia-smi --query-gpu=name,memory.total,driver_version,clocks.sm,power.limit --format=csv
echo "=== unpack llama.cpp build ==="
mkdir -p /llamacpp/bin
tar -xzf /artifacts/bin/llamacpp-pr27342-1deefcca-cu13-sm90-sm120.tar.gz -C /llamacpp/bin
chmod +x /llamacpp/bin/* 2>/dev/null || true
ls -la /llamacpp/bin/ | head
# llama-server links sibling libs (libllama-server-impl.so etc.) from its own dir
export LD_LIBRARY_PATH="/llamacpp/bin:${LD_LIBRARY_PATH:-}"
# ensure runtime shared libs present on slim runtime images
# (libgomp1: OpenMP for ggml-cpu; libcurl4: HF downloads; both cheap)
MISSING=$(ldd /llamacpp/bin/llama-server 2>/dev/null | grep "not found" || true)
if [ -n "$MISSING" ]; then
echo "missing libs, installing: $MISSING"
apt-get update -qq >/dev/null 2>&1 || true
apt-get install -y -qq libgomp1 libcurl4 >/dev/null 2>&1 || true
fi
MISSING=$(ldd /llamacpp/bin/llama-server 2>/dev/null | grep "not found" || true)
[ -z "$MISSING" ] || { echo "STILL MISSING: $MISSING"; exit 1; }
/llamacpp/bin/llama-server --version
echo "=== link models from bucket mount ==="
mkdir -p /models
ln -sf /artifacts/models/Qwen3.8-27B-OBLITERATED-Q5_K_M.gguf /models/
ln -sf /artifacts/models/Qwen3.8-27B-OBLITERATED-Q4_K_M.gguf /models/
ln -sf /artifacts/models/Qwen3.8-27B-DFlash2-Q8_0.gguf /models/
ln -sf /artifacts/models/mtp-Qwen3.8-27B-Q8_0.gguf /models/
ls -la /models/
echo "=== launch orchestrator (plan: $PLAN) as main process ==="
export BENCH_PLAN="$PLAN"
# exec: the orchestrator becomes the job's main process; when it exits (STOP
# or plan complete) the job finishes. Backgrounding here would kill the job.
exec bash /input/server_orchestrator.sh

Xet Storage Details

Size:
1.93 kB
·
Xet hash:
23928ff8c6fc6f6572855fcf03581c156a94b9c982585bab926e0466d808e0bf

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.