JGZYES
28535b1c74
- pmai unified CLI (generate/chat/interactive/http/inspect/output/config) - GPT+MoE transformer, .pap (f32/fp16/q8) + GGUF loader (order+version adaptive) - llama/Mixtral arch: RoPE+GQA+SwiGLU+MoE (C==torch verified) - C llama BPE tokenizer (validated vs llama-cpp-python) - training framework + 0.1B/0.22B MoE models; quantization fp16/q8 - build artifacts to output/; HTTP API; config.yaml; scripts; openapi
25 行
885 B
Bash
25 行
885 B
Bash
#!/usr/bin/env bash
|
|
# 基准：测 pmai 生成速度（tokens/s 与端到端延迟）
|
|
# 用法: scripts/bench.sh <model> [n_tokens]
|
|
set -euo pipefail
|
|
MODEL="${1:-../models/moe-0.1b-fp16.pap}"
|
|
N="${2:-100}"
|
|
CDIR="$(cd "$(dirname "$0")/.." && pwd)"
|
|
BIN="$CDIR/output/pmai"
|
|
|
|
echo "== pmai 基准: $MODEL n=$N =="
|
|
start=$(date +%s.%N)
|
|
"$BIN" "$MODEL" --prompt "The" --n_tokens "$N" --temperature 0 --top_k 0 > /tmp/bench_out.txt
|
|
end=$(date +%s.%N)
|
|
|
|
dt=$(echo "$end - $start" | bc)
|
|
tok=$(echo "$N" | bc)
|
|
tps=$(echo "scale=2; $tok / $dt" | bc)
|
|
echo "耗时: ${dt}s 约 ${tps} tokens/s（含模型加载）"
|
|
|
|
# 仅生成时间（预热一次）
|
|
start=$(date +%s.%N)
|
|
"$BIN" "$MODEL" --prompt "The" --n_tokens "$N" --temperature 0 --top_k 0 > /dev/null
|
|
end=$(date +%s.%N)
|
|
echo "第二次（已缓存）: $(echo "$end - $start" | bc)s ~$(echo "scale=2; $N / ($end - $start)" | bc) tokens/s"
|