Gahow Wang
d5dcf1a5ab
runner/servers: add --pp for both engines (xserv --pp N; llama.cpp -sm layer over N GPUs). New drivers: pp_final.sh (sequential latency + per-GPU VRAM + byte-exact correctness), pp_diag.sh (single x2 vs pp4 x2 determinism control), pp_quality_full.sh / pp_llama_47.sh (AIME+GSM8K matrix, xserv on 0-3 || llama on 4-7), summarize_pp/summarize_fullq, pp_time.py latency probe. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
26 lines
1.3 KiB
Bash
26 lines
1.3 KiB
Bash
#!/usr/bin/env bash
|
|
# llama.cpp PP=1/2/4 quality (aime2025+gsm8k, 30 each) on physical GPUs 4-7,
|
|
# parallel with the xserv matrix on 0-3. Pass --llama-devices so the runner pins
|
|
# CUDA_VISIBLE_DEVICES to 4.. (it otherwise forces 0..N-1). Distinct port + dirs.
|
|
set -u
|
|
cd /opt/wjh/projects/xserv
|
|
export PATH=$HOME/.cargo/bin:/usr/local/cuda-12.9/bin:$PATH
|
|
export CUDA_HOME=/usr/local/cuda-12.9
|
|
GGUF=/opt/wjh/models/qwen3-8b/qwen3-8b-bf16.gguf
|
|
LBIN=third_party/llama.cpp/build/bin/llama-server
|
|
PROG=bench-out/LLAMA47_PROGRESS.md
|
|
: > "$PROG"; echo "# llama on GPU 4-7 — $(date)" >> "$PROG"
|
|
for pp in 1 2 4; do
|
|
dev=$(seq -s, 4 $((3+pp)))
|
|
out=bench-out/fullq-llama-pp$pp; rm -rf "$out"
|
|
echo "=== START llama pp=$pp dev=$dev $(date +%H:%M:%S) ===" >> "$PROG"
|
|
pkill -9 -f "llama-server.*18181" 2>/dev/null; sleep 2
|
|
python3 -u -m tools.bench.runner --systems llama.cpp --pp "$pp" --llama-devices "$dev" \
|
|
--llama-bin "$LBIN" --llama-gguf "$GGUF" --llama-port 18181 \
|
|
--suite quality --quality-tasks aime2025,gsm8k --quality-limit 30 \
|
|
--max-batch 1 --max-seq-len 4096 --out-dir "$out" >/tmp/fql-$pp.log 2>&1
|
|
echo "=== END llama pp=$pp rc=$? $(date +%H:%M:%S) $(ls $out/comparison-*.json 2>/dev/null|wc -l) json ===" >> "$PROG"
|
|
done
|
|
pkill -9 -f "llama-server.*18181" 2>/dev/null
|
|
echo "LLAMA47_DONE" >> "$PROG"
|