scripts/runners/run_embed.py

#!/usr/bin/env python3
"""
Scherm benchmark — Embeddings runner (BGE-M3 e afins).

Mede throughput (embeddings/s) e latência de um modelo de embedding servido
via vLLM (--task embed, endpoint OpenAI /v1/embeddings). Varre concorrência e
tamanho de batch, N repetições, agrega mediana + IQR + percentis.
Mesma filosofia ACM do run_serving.py (seed, env logging, JSON+CSV).

Por que importa pro pitch Scherm: embedding é o coração do RAG (Scherm Assistente).
"Quantos documentos/s essa GPU indexa" é tão decisivo quanto tokens/s de chat.

Uso:
  python run_embed.py --served-model BAAI/bge-m3 --base-url http://localhost:8000 \
      --concurrencies 1 8 32 --batch-sizes 1 32 --reps 10 --out /work/results --tag GPU
"""
import argparse, json, os, statistics, time, subprocess, platform, urllib.request
from datetime import datetime, timezone
from concurrent.futures import ThreadPoolExecutor

def sh(c): return subprocess.run(c, shell=True, capture_output=True, text=True)
def gpu_name():
    return (sh("nvidia-smi --query-gpu=name --format=csv,noheader").stdout.strip().splitlines() or ["?"])[0]
def gpu_driver():
    return (sh("nvidia-smi --query-gpu=driver_version --format=csv,noheader").stdout.strip().splitlines() or [""])[0]
def vram_used():
    v=[int(x) for x in sh("nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits").stdout.split() if x.strip().isdigit()]
    return max(v) if v else None

def quantiles(xs):
    xs=sorted(x for x in xs if x is not None)
    if not xs: return {}
    n=len(xs)
    def pct(p):
        if n==1: return xs[0]
        k=(n-1)*p; f=int(k); c=min(f+1,n-1); return xs[f]+(xs[c]-xs[f])*(k-f)
    mean=statistics.fmean(xs); std=statistics.pstdev(xs) if n>1 else 0.0
    return {"n":n,"mean":round(mean,3),"std":round(std,3),"median":round(pct(.5),3),
            "q1":round(pct(.25),3),"q3":round(pct(.75),3),"p95":round(pct(.95),3),
            "p99":round(pct(.99),3),"min":round(xs[0],3),"max":round(xs[-1],3),
            "ci95_halfwidth":round(1.96*std/(n**.5),3) if n>1 else 0.0}

# Texto jurídico fixo (determinístico) ~ um chunk de RAG típico.
CHUNK=("Trata-se de acao em que se discute a soberania de dados em sistemas de "
       "inteligencia artificial operados on-premise por orgaos publicos. O pedido "
       "fundamenta-se na necessidade de manter o processamento de informacoes sigilosas "
       "dentro do perimetro institucional, sem dependencia de provedores de nuvem "
       "estrangeiros, em conformidade com a LGPD e com requisitos de seguranca nacional. ")*3

def one_request(base_url, model, n_inputs):
    body=json.dumps({"model":model,"input":[CHUNK]*n_inputs}).encode()
    req=urllib.request.Request(f"{base_url}/v1/embeddings",data=body,
        headers={"Content-Type":"application/json"})
    t0=time.time()
    try:
        with urllib.request.urlopen(req,timeout=300) as r:
            d=json.load(r)
        dt=time.time()-t0
        n=len(d.get("data",[]))
        dim=len(d["data"][0]["embedding"]) if n else 0
        return {"latency_s":dt,"n_embed":n,"dim":dim,"ok":True}
    except Exception as e:
        return {"ok":False,"err":str(e)[:140]}

def main():
    ap=argparse.ArgumentParser()
    ap.add_argument("--served-model",required=True)
    ap.add_argument("--base-url",default="http://localhost:8000")
    ap.add_argument("--concurrencies",nargs="+",type=int,default=[1,8,32,64])
    ap.add_argument("--batch-sizes",nargs="+",type=int,default=[1,32])  # inputs por request
    ap.add_argument("--reps",type=int,default=10)
    ap.add_argument("--warmups",type=int,default=1)
    ap.add_argument("--out",required=True); ap.add_argument("--tag",default="")
    ap.add_argument("--run-id",default="")  # sufixo único p/ evitar colisão de CSV
    a=ap.parse_args()
    os.makedirs(a.out,exist_ok=True)
    env={"timestamp_utc":datetime.now(timezone.utc).isoformat(),"modality":"embedding",
         "gpu_name":gpu_name(),"nvidia_driver":gpu_driver(),"served_model":a.served_model,
         "seed":1234,"reps":a.reps,"chunk_chars":len(CHUNK),"python":platform.python_version()}
    safe=a.served_model.replace("/","_"); raw={"env":env,"points":[]}; rows=[]
    for bs in a.batch_sizes:
        for conc in a.concurrencies:
            for _ in range(a.warmups): one_request(a.base_url,a.served_model,bs)
            emb_s=[]; lat=[]; vr=[]; dim=0
            for _ in range(a.reps):
                t0=time.time()
                with ThreadPoolExecutor(max_workers=conc) as ex:
                    res=list(ex.map(lambda _:one_request(a.base_url,a.served_model,bs),range(conc)))
                wall=time.time()-t0
                ok=[r for r in res if r.get("ok")]
                if ok:
                    total=sum(r["n_embed"] for r in ok)
                    emb_s.append(total/wall)  # embeddings/s agregado
                    lat.append(statistics.fmean([r["latency_s"] for r in ok])*1000)  # ms
                    vr.append(vram_used()); dim=ok[0]["dim"]
            agg={"embeddings_per_s":quantiles(emb_s),"latency_ms":quantiles(lat),"vram_mib":quantiles(vr)}
            raw["points"].append({"batch_size":bs,"concurrency":conc,"dim":dim,"n_ok":len(emb_s),"agg":agg})
            g=lambda k,kk:agg.get(k,{}).get(kk,"")
            rows.append([gpu_name(),a.served_model,bs,conc,dim,len(emb_s),
                         g("embeddings_per_s","median"),g("embeddings_per_s","q3"),
                         g("latency_ms","median"),g("latency_ms","p95"),g("vram_mib","median")])
            print(f"  [EMBED {a.served_model}] bs={bs} conc={conc} dim={dim} n={len(emb_s)} "
                  f"emb/s(med)={g('embeddings_per_s','median')} lat(med)={g('latency_ms','median')}ms "
                  f"vram={g('vram_mib','median')}MiB",flush=True)
    sfx=f"{a.tag}_{a.run_id}" if a.run_id else a.tag
    with open(f"{a.out}/embed_{sfx}_{safe}.json","w") as f: json.dump(raw,f,indent=2)
    csv=f"{a.out}/embed_{sfx}_{safe}.csv"; new=not os.path.exists(csv)
    with open(csv,"a") as f:
        if new: f.write("gpu,model,batch_size,concurrency,dim,n_ok,embeddings_per_s_median,embeddings_per_s_q3,latency_ms_median,latency_ms_p95,vram_mib_median\n")
        for r in rows: f.write(",".join(str(x) for x in r)+"\n")
    print(f"[OK] embed {a.served_model} -> {a.out} ({csv})",flush=True)

if __name__=="__main__": main()