#!/usr/bin/env python3 """Mixed-pool DP evaluation for PaddleOCR-VL table recognition. * K vLLM servers per GPU x G GPUs (DP = number of GPUs, K workers each). * All cohorts are merged into one pool and interleaved, so every worker sees a mix of datasets instead of finishing one dataset before the next. * Samples are pre-assigned to workers with greedy LPT on image pixel count, so each worker's total image resolution is about the same. * Scoring overlaps inference: finished predictions are batched and scored by a ProcessPool (TEDS is ~n^3, so it needs many CPU workers) while inference continues. TEDS == 1 implies TEDS-S == 1. Usage: eval_mix.py --model DIR --tag TAG --cohorts gate_pub_icdar gate_val_eval10k ... """ import argparse import base64 import heapq import json import os import pyarrow.parquet as pq import queue import random import signal import subprocess import sys import threading import time import urllib.request from collections import defaultdict from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor V5 = "/autodl-fs/data/povl_v5" sys.path.insert(0, "/autodl-fs/data/povl_optimization/protocol") import canonical_teds as ct # noqa: E402 from evaluation_policy import filter_evaluation,row_is_excluded,policy as evaluation_policy from fullpage_evaluation_policy import filter_pack_rows,route_existing_curriculum # 默认用**本地盘**的 vllm(2026-10-25): /autodl-fs 是网络盘, # `import vllm` 实测 175 秒(其中 CPU 只有 19 秒 -> 89% 是 I/O 等待); # 拷到本地 xfs 后只要 44 秒, 12 个 engine 一起起 14分47秒 -> 2分34秒。 _LOCAL_VLLM = "/root/autodl-tmp/vllm_local/bin/vllm" VLLM = os.environ.get("POVL_VLLM", _LOCAL_VLLM if os.path.exists(_LOCAL_VLLM) else "/autodl-fs/data/vllm_env/bin/vllm") PROMPT = "Table Recognition:" def log(m): print(f"[{time.strftime('%m-%d %H:%M:%S')}] {m}", flush=True) def apply_run_parallel_config(args): """Reload this run's evaluation layout at each invocation. Allows the active training controller to keep running while the next evaluation adopts the user's updated multi-instance configuration. Runs without this opt-in file retain their existing CLI behaviour. """ root = os.environ.get('POVL_RUN_DIR') path = os.path.join(root, 'eval_parallel.json') if root else '' if not path or not os.path.isfile(path): return args with open(path, encoding='utf-8') as f: config = json.load(f) for key in ('per_gpu', 'conc_per_worker', 'score_workers'): value = config.get(key, getattr(args, key)) if not isinstance(value, int) or isinstance(value, bool) or value <= 0: raise ValueError(f'Invalid evaluation parallel setting {key}={value!r}') setattr(args, key, value) log(f'Evaluation layout: {len(args.gpus)} GPUs x {args.per_gpu} vLLM instances ' f'= {len(args.gpus) * args.per_gpu}; concurrency/instance={args.conc_per_worker}; ' f'score_workers={args.score_workers}; config={path}') return args # ---------------------------------------------------------------------------- # scoring # ---------------------------------------------------------------------------- def _score_one_raw_impl(job): """消融路径: 不抽表格, 直接把模型原始输出丢进 TEDS。""" sid, pred, gt, gt_fmt = job out = {"sample_id": sid} if not isinstance(pred, str) or not pred.strip(): out.update(teds=0.0, teds_s=0.0, status="empty") return out try: res = ct.score_pair(pred, gt, "otsl", gt_fmt) except Exception as e: out.update(teds=0.0, teds_s=0.0, status=f"score_exc:{type(e).__name__}") return out if res.get("status") == "metric_oversize": out.update(teds=0.0, teds_s=0.0, status="pred_oversize") return out if res.get("status") == "gt_error": out.update(teds=None, teds_s=None, status="gt_error") return out t, s = res["teds"], res["teds_structure"] if t is not None and t >= 1.0: s = 1.0 out.update(teds=t, teds_s=s, status=res["status"]) # 空 cell 表按用户定的口径以 TEDS 为准, 常规表以 TEDS-S 为准; # 汇报时按 primary_metric 分开, 不要把两种口径混在一张表里排序 for k in ("gt_fill_ratio", "gt_structure_only", "primary_metric", "primary"): if k in res: out[k] = res[k] return out def _score_one_impl(job): sid, pred, gt, gt_fmt = job out = {"sample_id": sid} if not isinstance(pred, str) or not pred.strip(): out.update(teds=0.0, teds_s=0.0, status="empty") return out # 抽表格本体再算分: 模型被喂整页图时可能把页眉/正文一起吐出来, 那些不该进 TEDS。 # prose_frac 同时是诊断量 —— 它直接说明模型是在读表还是在复述整页。 # # fmt=None 自动识别方言。这里**不能**像以前那样硬编码 "otsl": 模型有时会吐 # HTML, 那样 extract_table 会返回空 -> 直接判 0 分(实测自检 243/300 失败)。 # 识别之后再把预测与 GT 统一归一到 canonical HTML, 两边同方言再比, # 消除 OTSL/HTML 的口径错配。 try: pred_t, info = ct.extract_table(pred, None) except Exception as e: out.update(teds=0.0, teds_s=0.0, status=f"extract_exc:{type(e).__name__}") return out out.update(extracted_format=info["format"], prose_frac=info["prose_frac"]) if not pred_t: out.update(teds=0.0, teds_s=0.0, status=f"no_table:{info['format']}") return out try: pred_html = ct.cells_to_canonical_html(ct.parse_table(pred_t, info["format"])) gt_html = ct.cells_to_canonical_html(ct.parse_table(gt, gt_fmt)) res = ct.score_pair(pred_html, gt_html, "html", "html") except Exception as e: # never let one pathological sample kill a batch out.update(teds=0.0, teds_s=0.0, status=f"score_exc:{type(e).__name__}") return out st = res.get("status") if st == "metric_oversize": try: g = ct.count_tree_nodes(ct.parse_table(gt_html, "html")) except Exception: g = 10 ** 9 if g <= 4000: # GT is scoreable; the prediction blew up -> model failure out.update(teds=0.0, teds_s=0.0, status="pred_oversize") else: out.update(teds=None, teds_s=None, status="gt_oversize") return out if st == "gt_error": out.update(teds=None, teds_s=None, status="gt_error") return out t, s = res["teds"], res["teds_structure"] if t is not None and t >= 1.0: s = 1.0 out.update(teds=t, teds_s=s, status=st, merge_f1=res.get("merged_rect_f1")) return out # ---------------------------------------------------------------------------- # servers # ---------------------------------------------------------------------------- def gpu_free_bytes(gpu: int) -> int: """查某张卡的空闲显存; 查不到就返回一个保守估计。""" try: out = subprocess.check_output( ["nvidia-smi", "-i", str(gpu), "--query-gpu=memory.free", "--format=csv,noheader,nounits"], timeout=20) return int(out.decode().strip().splitlines()[0]) * 1024 ** 2 except Exception: return 40 * 10 ** 9 def plan_workers(gpus, per_gpu_hint: int, util: float, reserve: float = 0.10): """按每卡空闲显存决定每卡几个 vLLM worker。 规则(用户 2026-10-18): 每卡 2-4 个 worker, 按 gpu_memory_utilization 算, 且**worker 往大来** —— 显存够就顶满 4 个, 不够再往下退, 低于 2 就退到 1。 真正的 KV 预算在 start_servers 里按 util/per_gpu 再分一次。 """ need = 4 * 10 ** 9 # 每个 engine 的模型权重 + 激活 + 一点 KV plan = [] for g in gpus: free = gpu_free_bytes(g) budget = int(free * (1 - reserve)) n = max(1, min(per_gpu_hint, budget // need)) plan.append((g, n, free)) return plan # 方案2(2026-10-25): 下一个模型的 engine 在**当前模型还在打分**的时候预启动。 # 打分是纯 CPU 的(TEDS 编辑距离), 那 10 分钟 GPU 完全闲着; 而开机要 6~14 分钟。 # 预启动的 engine 用标记文件记录"端口 -> 模型路径 + pid", 下一个 eval_mix # 发现自己要的端口上已经有**同一个模型**的 engine 在跑, 就直接复用。 PREBOOT_MARK = "/root/autodl-tmp/preboot_marks" class AdoptedProc: """占位: 复用的 engine 不是本进程起的, 但仍要负责在最后把它杀掉。 预启动时用了 start_new_session=True, 所以它自己是进程组组长, os.killpg(pid) 能整组收掉。 """ def __init__(self, pid): self.pid = pid self._alive = True def poll(self): return None if self._alive else 0 def kill(self): self._alive = False def _port_alive(port, timeout=3): try: urllib.request.urlopen(f"http://127.0.0.1:{port}/v1/models", timeout=timeout) return True except Exception: return False def _read_mark(port): p = os.path.join(PREBOOT_MARK, f"{port}.json") try: with open(p) as f: return json.load(f) except Exception: return None def start_servers(model, gpus, per_gpu, base_port, gmu, max_len, logdir, served, adopt=True, wait=True): procs = [] ready = [] # --- 预热 trust_remote_code 的动态模块 ----------------------------------- # PaddleOCR-VL 的 config/processor 是 trust_remote_code 的自定义类, transformers # 会把它**复制**到 ~/.cache/huggingface/modules/transformers_modules/// # 再 import。N 个 vLLM 进程同时做这件事时会互相踩: 有一次 12 个 engine 里 # 最后一个报 "module ... has no attribute 'PaddleOCRVLImageProcessor'" # (复制到一半就被另一个进程读走)。所以先在父进程里 import 一次, # 让缓存目录被完整写好, 后面所有 engine 直接命中缓存。 try: from transformers import AutoConfig, AutoProcessor # noqa: F401 AutoConfig.from_pretrained(model, trust_remote_code=True) try: AutoProcessor.from_pretrained(model, trust_remote_code=True) except Exception as e: log(f" 预热 processor 失败(忽略): {type(e).__name__} {str(e)[:120]}") except Exception as e: log(f" 预热动态模块失败(忽略): {type(e).__name__} {str(e)[:120]}") def wait_ready(p, port, budget=1800): deadline = time.time() + budget while True: if p.poll() is not None: return False try: urllib.request.urlopen( f"http://127.0.0.1:{port}/v1/models", timeout=5) return True except Exception: if time.time() > deadline: return False time.sleep(3) def tail(fh, n=25): try: fh.flush() with open(fh.name) as f: return "".join(f.readlines()[-n:]) except Exception: return "" # 同卡上第 N 个 engine 要给剩下的显存, 不是给"整卡空闲显存"。 # 之前所有 worker 都用同一个 gmu, 第 4 个起来时卡上已被前 3 个占了 ~67GB, # 只剩 11.9GB < 它要的 21GB -> "Free memory ... less than desired"。 TOTAL_BYTES = gpu_free_bytes(gpus[0]) TOTAL_GIB = TOTAL_BYTES / (1024 ** 3) or 94.0 # (2026-10-24) **并行起 engine**。 # vllm 0.32 起一个 engine 要 2~4 分钟(import + 权重 + inductor 编译 + # CUDA graph 捕获), 串行起 8 个要 ~30 分钟, 5 个模型就是 2.5 小时纯开机。 # 并行起不会互相踩: 每个 engine 的显存预算 this_gmu 是**固定**的 # (TOTAL_BYTES*0.36/per_gpu, 不随"当前空闲"变), 同卡 2 个 engine 共占 # ~37G(KV 15G x2 + 权重 2.5G x2 + context) < 85G。 # 设 POVL_PARALLEL_ENGINE_BOOT=0 可退回原来的串行行为。 PARALLEL_BOOT = os.environ.get("POVL_PARALLEL_ENGINE_BOOT", "1") != "0" specs = [] for gi, g in enumerate(gpus): for w in range(per_gpu): specs.append((g, w, base_port + gi * per_gpu + w)) for (g, w, port) in specs: # --- 方案2: 端口上已经有**同一个模型**的 engine(预启动的) -> 直接复用 --- if adopt: m = _read_mark(port) if m and m.get("model") == model and m.get("pid") and _port_alive(port): log(f" 复用预启动的 engine :{port} ({model})") procs.append((AdoptedProc(m["pid"]), None, port)) ready.append(port) continue # 现在这张卡真正还剩多少(前面几个 engine 已经占掉了), 按比例分配。 # gmu 是 0~1 的比例, 不是字节。 # 固定预算, 不做动态探测。vllm 在启动瞬间检查显存, 而同卡上一个 # engine 那时还在分配中, 按"当前空闲"算必然偏大 -> 第二个就报 # "Free memory ... less than desired"。模型只有 1.9G, 每 engine # 分 total/per_gpu 就够, 剩下的全给 KV cache。 this_gmu = int(TOTAL_BYTES * 0.36 / per_gpu) # KV cache 字节 # 实测一个 engine 实际吃 ~17.5G(KV + 1.9G 权重 + CUDA context) env = dict(os.environ, CUDA_VISIBLE_DEVICES=str(g), # 不要设 VLLM_ATTENTION_BACKEND: 一旦显式指定, vllm 会 # 绕开 paddlex 的 PaddleOCR-VL 插件、退回 # TransformersForCausalLM 回退路径, 而那条路径不认识 # 顶层模块 mlp_AR -> "no module or parameter named # 'mlp_AR'". 让插件自己选 backend(它内部会选 # FLASH_ATTN, 用的是我们本地源码编译的 flash_attn)。 VLLM_USE_FLASHINFER_SAMPLER="0", OMP_NUM_THREADS="2", MKL_NUM_THREADS="2") # vLLM 子进程必须用 **vllm_env 自己的 transformers(5.x)**: # transformers 4.57 把 PreTrainedConfig 改名成 PretrainedConfig(少个 T), # vllm 0.32 里 `from transformers import PreTrainedConfig` 会 ImportError。 # 训练用的 venv_train 若排在 PYTHONPATH 前面就会踩到 -> 把那截摘掉。 _pp = [p for p in (env.get("PYTHONPATH") or "").split(os.pathsep) if p and "venv_train" not in p] env["PYTHONPATH"] = os.pathsep.join(_pp) cmd = [VLLM, "serve", model, "--port", str(port), "--served-model-name", served, "--max-model-len", str(max_len), # 两个都得给: 只给 kv-cache-memory 的话, vllm 仍用默认 # gpu_memory_utilization=0.9 做启动校验 -> 同卡第二三个 engine # 必然 "Free memory ... less than desired"。 "--kv-cache-memory", str(this_gmu), "--gpu-memory-utilization", str(round(0.86 / per_gpu, 3)), # 必须 >= max_model_len, 否则新版 vllm 直接拒绝启动 "--max-num-batched-tokens", str(max(16384, max_len * 2)), "--max-num-seqs", "128", "--no-enable-prefix-caching", "--mm-processor-cache-gb", "0"] if os.environ.get("POVL_TRC", "1") == "1": cmd.append("--trust-remote-code") if os.environ.get('POVL_ALLOW_TRUSTED_MM_KWARGS')=='1': # Official PaddleOCR clients send per-request pixel budgets. This # opt-in route only serves trusted local benchmark workers. cmd+=['--trust-request-mm-kwargs','--host','127.0.0.1'] fh = open(f"{logdir}/vllm_g{g}_w{w}.log", "w") p = subprocess.Popen(cmd, env=env, stdout=fh, stderr=subprocess.STDOUT, start_new_session=True) procs.append((p, fh, port)) if not PARALLEL_BOOT: # 串行模式: 每个起来就单独等它就绪, 这样挂了能立刻定位到是哪个, # 也不会连坐杀掉已就绪的实例。 if not wait_ready(p, port): log(f" vLLM :{port} 启动失败\n{tail(fh)}") stop_servers(procs) raise RuntimeError(f"vLLM :{port} failed to start; see {logdir}") for (g, w, port), (p, fh, _pp) in zip(specs, procs): if not wait_ready(p, port): log(f" vLLM :{port} 启动失败\n{tail(fh)}") stop_servers(procs) raise RuntimeError(f"vLLM :{port} failed to start; see {logdir}") ready.append(port) log(f" vLLM :{port} ready ({len(ready)}/{len(gpus)*per_gpu}, " f"卡{g} 剩 {gpu_free_bytes(g)/(1024**3):.0f}G)") time.sleep(1) log(f"{len(procs)} vLLM servers ready on {sorted(gpus)}") return procs def stop_servers(procs): for p, _, _ in procs: if p.poll() is None: try: os.killpg(p.pid, signal.SIGTERM) except ProcessLookupError: pass t = time.time() + 30 while time.time() < t and any(p.poll() is None for p, _, _ in procs): time.sleep(0.5) for p, fh, _ in procs: if p.poll() is None: try: os.killpg(p.pid, signal.SIGKILL) except ProcessLookupError: pass if fh is not None: # 复用的 engine 没有日志句柄 fh.close() def preboot_servers(model, gpus, per_gpu, base_port, gmu, max_len, logdir, served): """**后台**起一批 engine, 不等它们就绪, 直接返回。 用途(方案2): 当前模型推理结束、engine 已停掉、GPU 空着, 但打分还要跑 10 分钟(CPU 密集)。这段时间把**下一个模型**的 engine 起好, 下一个 eval_mix 直接复用, 就省掉了 6~14 分钟的串行开机。 标记文件 PREBOOT_MARK/.json 记着 "端口 -> 模型 + pid", start_servers 的复用逻辑靠它判断"这个端口上的 engine 是不是我要的模型"。 """ os.makedirs(PREBOOT_MARK, exist_ok=True) specs = [] for gi, g in enumerate(gpus): for w in range(per_gpu): specs.append((g, w, base_port + gi * per_gpu + w)) total_bytes = gpu_free_bytes(gpus[0]) this_gmu = int(total_bytes * 0.36 / per_gpu) env = dict(os.environ, VLLM_USE_FLASHINFER_SAMPLER="0", OMP_NUM_THREADS="2", MKL_NUM_THREADS="2") _pp = [p for p in (env.get("PYTHONPATH") or "").split(os.pathsep) if p and "venv_train" not in p] env["PYTHONPATH"] = os.pathsep.join(_pp) for (g, w, port) in specs: # 旧的标记先清掉, 免得被误当成"已经就绪" for f in (os.path.join(PREBOOT_MARK, f"{port}.json"),): try: os.remove(f) except OSError: pass # 等端口空出来(上一个模型的 engine 刚被 kill, socket 可能还在 TIME_WAIT; # 不等的话新 engine 会 "Address already in use" 直接死 —— 2026-10-26 踩过) import socket as _socket for _ in range(60): _s = _socket.socket() _s.setsockopt(_socket.SOL_SOCKET, _socket.SO_REUSEADDR, 1) try: _s.bind(("127.0.0.1", port)) _s.close() break except OSError: _s.close() time.sleep(1) cmd = [VLLM, "serve", model, "--port", str(port), "--served-model-name", served, "--max-model-len", str(max_len), "--kv-cache-memory", str(this_gmu), "--gpu-memory-utilization", str(round(0.86 / per_gpu, 3)), "--max-num-batched-tokens", str(max(16384, max_len * 2)), "--max-num-seqs", "128", "--no-enable-prefix-caching", "--mm-processor-cache-gb", "0"] if os.environ.get("POVL_TRC", "1") == "1": cmd.append("--trust-remote-code") os.makedirs(logdir, exist_ok=True) fh = open(os.path.join(logdir, f"preboot_g{g}_w{w}.log"), "w") e = dict(env, CUDA_VISIBLE_DEVICES=str(g)) p = subprocess.Popen(cmd, env=e, stdout=fh, stderr=subprocess.STDOUT, start_new_session=True) with open(os.path.join(PREBOOT_MARK, f"{port}.json"), "w") as mf: json.dump({"model": model, "pid": p.pid, "port": port}, mf) log(f" 预启动 engine :{port} (pid {p.pid}) <- {model}") # ---------------------------------------------------------------------------- # load balancing # ---------------------------------------------------------------------------- def _apply_text_metric_eligibility(job, result): # Keep TEDS=N/A for empty-text GT even on early prediction failures. try: tbl=ct.parse_table(job[2],job[3]) so=ct.is_structure_only_gt(tbl) result['gt_structure_only']=so result['gt_fill_ratio']=ct.gt_fill_ratio(tbl) result['primary_metric']='teds_s' if so:result['teds']=None result['primary']=result.get('teds_s') except Exception: result.update(teds=None,teds_s=None,status='gt_error') return result def score_one(job): return _apply_text_metric_eligibility(job,_score_one_impl(job)) def score_one_raw(job): return _apply_text_metric_eligibility(job,_score_one_raw_impl(job)) def apply_gt_override(rows, path): """用外部 GT 覆盖表替换 rows 里的 gt/gt_format。返回覆盖条数。 匹配键: 先试 pack 的上游 id(rows[i]["_orig_id"]), 再试 (cohort, sample_id)。 两种键都支持, 因为有些包带上游 id, 有些不带。 """ ov = {} for line in open(path, encoding="utf-8"): line = line.strip() if not line: continue d = json.loads(line) gt = d.get("gt") if not gt: continue ov[d["sample_id"]] = (gt, d.get("gt_format") or "otsl") n = 0 for r in rows: for key in (r.get("_orig_id"), r.get("sample_id")): if key and key in ov: gt, fmt = ov[key] r["gt"], r["gt_format"] = gt, fmt n += 1 break return n def load_from_packs(pack_root, cohorts=None): """从 parquet shards 读评测行。 与 cohort jsonl 路径产出同构的 row dict, 区别只有: image = (shard, row_group, idx) 定位符, 由 _infer 惰性解析 gt/gt_format 取自 gt_format_native 指向的那一份 (OTSL->HTML 是有损方向, 见 eval_pack.otsl_roundtrip_ok) """ import eval_pack_reader as epr epr.PACK_ROOT = pack_root rows = [] for shard in epr.shard_files(pack_root): for rg, idx, meta, sid in epr.index_rows(shard, pack_root): if not meta.get("has_image"): continue # 缺图行无法推理 if cohorts and meta.get("cohort") not in cohorts: continue r = dict(meta) r["_cohort"] = meta.get("cohort") r["_shard"] = shard # 永远用 (shard, rg, idx) 合成 id, 不要用上游的 sample_id: # 84% 的行上游 id 是空的, 且少数 id 有重复(同一页面被切成多块), # 跨 pack join / bad-case 回溯都要靠这个唯一定位。 # 保留上游 id 到 _orig_id 以便对照。 r["_orig_id"] = sid r["sample_id"] = meta.get("canonical_sample_id") or f"{os.path.basename(shard)}#{rg}#{idx}" r["image"] = (shard, rg, idx) native = meta.get("gt_format_native") or "otsl" r["gt"] = meta["gt_otsl"] if native == "otsl" else meta["gt_html"] r["gt_format"] = native r["_px"] = meta["img_w"] * meta["img_h"] # 宽高已是列, 无需解码 rows.append(r) return filter_pack_rows(filter_evaluation(rows)) def image_pixels(path, row=None): """像素量, 供 lpt_assign 做负载均衡。 parquet 行直接用 img_w/img_h 列, 不解码。 注意: 这里**不能**像以前那样 except: return 0 —— 静默的 0 会让 LPT 权重 全零、负载均衡无声失效, 而且不会报任何错。解不开就抛。 """ if isinstance(path, tuple): if row is not None and row.get("img_w"): return row["img_w"] * row["img_h"] return 1 # 缺图行(缺图的行本就该被过滤掉) from PIL import Image with Image.open(path) as im: w, h = im.size return w * h def lpt_assign(rows, n_workers): """Greedy LPT on pixel count; return per-worker lists (each mixed-shuffled).""" order = sorted(range(len(rows)), key=lambda i: -rows[i]["_px"]) heap = [(0, w) for w in range(n_workers)] buckets = [[] for _ in range(n_workers)] for i in order: load, w = heapq.heappop(heap) buckets[w].append(rows[i]) heapq.heappush(heap, (load + rows[i]["_px"], w)) rng = random.Random(0) for b in buckets: rng.shuffle(b) # mix datasets and sizes within each worker return buckets # ---------------------------------------------------------------------------- # inference # ---------------------------------------------------------------------------- def shard_work_units(rows, n_workers): """把 rows 按 shard 聚成"工作单元", 按像素量降序。 动态调度(而不是静态切桶)的前提: 每个单元必须能独立完成。 单元按成本降序进队列, worker 领完一个跑完再去领下一个 —— 领用本身是独占的(Queue.get), 所以不会出现两个 worker 抢同一片。 """ by = {} for r in rows: by.setdefault(r["_shard"], []).append(r) units = [(sum(x["_px"] for x in rs), sh, rs) for sh, rs in by.items()] units.sort(key=lambda t: -t[0]) # 贵的先领, 避免长尾 log(f" 工作单元: {len(units)} 个 shard (按像素量降序), workers={n_workers}") return units def infer(port, row, max_tokens, served, timeout, penalties=None): try: return _infer(port, row, max_tokens, served, timeout, penalties) except Exception as e: return {"prediction": "", "finish_reason": "error", "error": f"{type(e).__name__}: {e}"[:300]} def _infer(port, row, max_tokens, served, timeout, penalties=None): # row["image"] 可能是散图路径, 也可能是 parquet 定位符 (shard, rg, idx)。 # 定位符必须在这里才解析成字节 —— 见 eval_pack_reader 的模块说明。 loc = row["image"] if isinstance(loc, tuple): import eval_pack_reader as epr raw = epr.image_bytes(*loc) # 从字节流探测, 不靠 img_format 列 —— reshard 的 pack 里没有这一列, # 靠它判会把所有 PNG 误标成 jpeg, vllm 解码失败 -> 每条都 error mime = ("image/png" if raw[:8] == b"\x89PNG\r\n\x1a\n" else "image/jpeg") else: mime = "image/png" if loc.lower().endswith(".png") else "image/jpeg" raw = open(loc, "rb").read() b64 = base64.b64encode(raw).decode() body = {"model": served, "temperature": 0, "max_tokens": max_tokens, "skip_special_tokens": False, # 复读惩罚(2026-10-21): 贪心解码下模型有时会陷进复读循环, # 一路生成到 max_tokens(16384) —— 推理时间全耗在这上面, 而且 # 这些样本的分数基本是 0。vLLM 的 repetition_penalty>1 可以在 # 出现重复时压低对应 token 的 logits, 让它换一条路。 **(penalties or {}), "messages": [{"role": "user", "content": [ {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}}, {"type": "text", "text": PROMPT}]}]} req = urllib.request.Request(f"http://127.0.0.1:{port}/v1/chat/completions", data=json.dumps(body).encode(), headers={"Content-Type": "application/json"}) t0 = time.time() for attempt in range(3): try: r = json.loads(urllib.request.urlopen(req, timeout=timeout).read()) ch = r["choices"][0] return {"prediction": ch["message"]["content"] or "", "finish_reason": ch.get("finish_reason"), "completion_tokens": r.get("usage", {}).get("completion_tokens"), "latency": round(time.time() - t0, 2)} except Exception as e: err = f"{type(e).__name__}: {e}" time.sleep(2 * (attempt + 1)) return {"prediction": "", "finish_reason": "error", "error": err[:300]} def main(): ap = argparse.ArgumentParser() ap.add_argument("--model", required=True) ap.add_argument("--tag", required=True) ap.add_argument("--cohorts", nargs="*", default=[], help="只跑这些 cohort; 留空 = 全跑(packs 模式下 benchmark 即 cohort)") ap.add_argument("--gpus", nargs="+", type=int, default=list(range(8))) ap.add_argument("--per-gpu", type=int, default=4) ap.add_argument("--gmu", type=float, default=0.0, help="0 = auto (0.92/per_gpu)") ap.add_argument("--conc-per-worker", type=int, default=24) ap.add_argument("--score-workers", type=int, default=96) ap.add_argument("--score-batch", type=int, default=256) ap.add_argument("--max-tokens", type=int, default=16384) ap.add_argument("--repetition-penalty", type=float, default=0.0, help=">0 时启用 vLLM repetition_penalty(打断复读循环)") ap.add_argument("--frequency-penalty", type=float, default=0.0) ap.add_argument("--presence-penalty", type=float, default=0.0) ap.add_argument("--max-model-len", type=int, default=20000) ap.add_argument("--base-port", type=int, default=18000) ap.add_argument("--limit", type=int, default=0) ap.add_argument("--timeout", type=int, default=1800) ap.add_argument("--apply-blocklist", action="store_true", help="在脚本内再过滤一次表过长/GT不可用的样本" "(默认不过滤, 请直接用 eval_packs_filtered)") ap.add_argument("--span-subset", default=SPAN_SUBSET_DEFAULT, help="15k 跨行跨列子集目录; 只做身份标记, 不额外推理") ap.add_argument("--exclude-cohorts", default="", help="逗号分隔的 cohort 名, 这些来源**不参与**评测。" "2026-10-25: tabrecset_en 的 GT 质量不行(374 条全空、" "161 条本身是单列填空题表), 暂时移出评测。") ap.add_argument("--preboot-model", default="", help="在**本模型打分期间**后台预启动这个模型的 engine " "(方案2)。下一个 eval_mix 会自动复用。") ap.add_argument("--gt-override", default="", help="GT 覆盖表(jsonl: sample_id/gt/gt_format)。" "评测包里 TabRecSet 的 1566 条 GT 是坏的(OTSL 解析成 1 列), " "用修好的 GT 覆盖。key 用 pack 的 sample_id(上游 id)或 " "(cohort, sample_id)。") ap.add_argument("--packs", default="", help="parquet shards 根目录; 非空则从 shard 读, 忽略 cohort jsonl") a = ap.parse_args() if route_existing_curriculum(a): return apply_run_parallel_config(a) outdir = f"{V5}/eval/runs/{a.tag}" logdir = f"{outdir}/logs" os.makedirs(logdir, exist_ok=True) global SPAN, PENALTIES SPAN = load_span_subset(a.span_subset) if a.span_subset else {} PENALTIES = {} if a.repetition_penalty and a.repetition_penalty > 0: PENALTIES["repetition_penalty"] = a.repetition_penalty if a.frequency_penalty: PENALTIES["frequency_penalty"] = a.frequency_penalty if a.presence_penalty: PENALTIES["presence_penalty"] = a.presence_penalty if PENALTIES: log(f"复读惩罚: {PENALTIES}") if SPAN: log(f"跨度子集: {len(SPAN)} 个 sample_id 会在汇总时分档") pred_path = f"/dev/shm/povl_v5/preds/{a.tag}.jsonl" os.makedirs(os.path.dirname(pred_path), exist_ok=True) score_path = f"{outdir}/scores.jsonl" rows, cohort_of = [], {} if a.packs: rows = load_from_packs(a.packs, a.cohorts) else: for c in a.cohorts: for line in open(f"{V5}/eval/cohorts/{c}.jsonl"): r = json.loads(line) r["_cohort"] = c # cohort 模式(不带 --packs)原来没给 _shard, 而 shard_work_units # 按 _shard 分工作单元 -> 直接 KeyError。packs 模式(184k 全量评测 # 用的那条路)不受影响, 这里只补上散图模式缺的字段。 r["_shard"] = c rows.append(r) # GT 覆盖(见 --gt-override): 评测包里 TabRecSet 的坏 GT 会把分数压到 1/8 if a.gt_override: n_ov = apply_gt_override(rows, a.gt_override) log(f"GT 覆盖 {n_ov} 条 <- {a.gt_override}") # 默认不做名单过滤 —— 用 data/eval_packs_filtered 那份已经把表过长 / GT 不可用 # 的样本剔掉的 parquet 即可, 过滤发生在数据侧而不是代码侧, 可审计。 # 只有在直接喂未过滤的包/cohort 时才需要在这里兜一层。 # 排除指定来源(GT 质量不行 / 暂时不参与评测的) if a.exclude_cohorts: drop = {c.strip() for c in a.exclude_cohorts.split(",") if c.strip()} before = len(rows) rows = [r for r in rows if (r.get("_cohort") or r.get("cohort")) not in drop] log(f"排除来源 {sorted(drop)}: {before} -> {len(rows)} 条") before_policy=len(rows) rows=filter_evaluation(rows) log(f'Evaluation policy: {evaluation_policy()}; additionally removed {before_policy-len(rows)} rows') if a.apply_blocklist: import eval_blocklist bl = eval_blocklist.Blocklist.load() rows, dropped = bl.filter_rows(rows) log(f"排除名单 {bl.summary()}, 本次剔除 {dict(dropped)} -> 剩 {len(rows)}") if a.limit: random.Random(1).shuffle(rows) rows = rows[:a.limit] # resume: skip already scored samples done = set() if os.path.exists(score_path): for line in open(score_path): try: d = json.loads(line) done.add((d["cohort"], d["sample_id"])) except Exception: pass rows = [r for r in rows if (r["_cohort"], r["sample_id"]) not in done] log(f"{len(rows)} samples to run ({len(done)} already scored), cohorts={a.cohorts}") if not rows: if os.path.exists(score_path): summarize(score_path, outdir, a) else: log("没有待评测样本, 且无历史 scores.jsonl") return if a.packs: # 宽高已经是 parquet 列, 不用再全量 stat 一遍 for r in rows: r["_px"] = r["img_w"] * r["img_h"] else: with ThreadPoolExecutor(64) as ex: for r, px in zip(rows, ex.map(lambda r: image_pixels(r["image"]), rows)): r["_px"] = px served = "m" # OOM / startup failure -> fall back to fewer vLLM workers per GPU while True: plan = plan_workers(a.gpus, a.per_gpu, a.gmu or 0.9) log(" GPU 规划: " + ", ".join( f"gpu{g}->{n}worker(空闲{free/1e9:.0f}G)" for g, n, free in plan)) per_gpu = max(n for _, n, _ in plan) n_workers = sum(n for _, n, _ in plan) # KV 预算: 按每卡空闲显存给每个 engine, 避免并发探测互相干扰 gmu = a.gmu if a.gmu > 1 else int( min(free for _, _, free in plan) * 0.55 / per_gpu) log(f" 每卡 {per_gpu} worker, 共 {n_workers} 个, " f"每 engine KV 预算 {gmu/1e9:.1f} GB") try: procs = start_servers(a.model, a.gpus, per_gpu, a.base_port, gmu, a.max_model_len, logdir, served) break except RuntimeError as e: if per_gpu <= 1: raise log(f"server start failed: {e}") raise n_workers = len(a.gpus) * a.per_gpu # 动态调度: 成本降序的 shard 队列, worker 独占领取(见 worker_loop) shard_q = queue.Queue() for cost, shard, srows in shard_work_units(rows, n_workers): shard_q.put((cost, shard, srows)) done_shards = [] all_px = sorted(r["_px"] for r in rows) log(f"LPT: {n_workers} workers ({len(a.gpus)} GPU x {a.per_gpu}), " f"{shard_q.qsize()} 个 shard 待领, " f"总像素 {sum(all_px)/1e6:.1f}M (p50={all_px[len(all_px)//2]/1e6:.2f}M)") ports = [port for _, _, port in procs] done_q = queue.Queue() total = len(rows) t_start = time.time() def worker_loop(widx): """一个 worker 一个 vLLM 服务, 循环从共享队列领取未处理的 shard。 队列是独占领取的: 一个 shard 被某个 worker 取走后, 其它 worker 再取就拿不到同一片。worker 跑完手上的 shard 会立刻去看有没有 别人没领的, 有就接着跑 —— 长尾 worker 不会早退, 空转的 worker 也不会闲着。 """ port = ports[widx] while True: try: _, shard, srows = shard_q.get_nowait() except queue.Empty: break t0 = time.time() with ThreadPoolExecutor(a.conc_per_worker) as ex: futs = {ex.submit(infer, port, r, a.max_tokens, served, a.timeout, PENALTIES): r for r in srows} for f in futs: res = f.result() done_q.put((futs[f], res)) done_shards.append((shard, len(srows), time.time() - t0)) log(f" [w{widx}] shard {shard} 完成 " f"({len(srows)} 行, {time.time()-t0:.0f}s), " f"剩余 {shard_q.qsize()}") threads = [threading.Thread(target=worker_loop, args=(i,), daemon=True) for i in range(n_workers)] for t in threads: t.start() def _score_unthrottle(): """推理结束后把打分进程提权并放开全部核。 打分在推理阶段压优先级是对的(TEDS 是纯 CPU 编辑距离, 抢推理的核会 拖慢 GPU 喂数)。但推理一结束 GPU 就闲着、CPU 全空, 这时还压着 就是纯浪费 —— 实测推理收尾时还剩 4,810 个批次, 尾部光排空就花了 42 分钟(占整个 ckpt 的 41%)。 """ # 注意: 不能靠子进程自己往全局 set 里 add —— fork 出来的副本不会 # 回传父进程。直接从 pool 的进程表里取 pid。 pids = [] try: pids = [p.pid for p in list(pool._processes.values()) if p.pid] except Exception: pids = list(SCORE_PIDS) n = 0 for pid in pids: try: os.setpriority(os.PRIO_PROCESS, pid, 0) os.sched_setaffinity(pid, set(range(len(os.sched_getaffinity(0))))) n += 1 except Exception: pass if n: log(f" 推理结束 -> {n} 个打分进程提权, 放开全部 CPU") def _score_init(): """打分进程降优先级。 TEDS 是纯 CPU 密集的编辑距离, 推理侧是 GPU 喂数, 两边抢同一批核。 打分让位: 推理跑满的时候打分慢一点无所谓, 反过来会拖慢喂数、 把 GPU 饿着。 """ try: os.nice(19) except Exception: pass # 打分池最多占一半的核, 剩下的留给推理 try: c = len(os.sched_getaffinity(0)) os.sched_setaffinity(0, set(range(0, max(1, c // 2)))) except Exception: pass pool = ProcessPoolExecutor(a.score_workers, initializer=lambda: (_score_init(), _register_pid())) pending = [] n_pred = n_scored = 0 pf = open(pred_path, "a") sf = open(score_path, "a") batch = [] def flush_scores(block=False): nonlocal n_scored, pending keep = [] for fut, meta in pending: if block or fut.done(): for res, r in zip(fut.result(), meta): res.update(cohort=r["_cohort"], dataset=r.get("dataset"), origin=r.get("origin"), has_numeric=r.get("has_numeric"), res_bucket=r.get("res_bucket"), metrics=r.get("metrics"), n_merge=r.get("n_merge"), difficulty=SPAN.get(r.get("sample_id") or ""), n_spans=r.get("n_spans"), finish_reason=r["_fr"], ctoks=r["_ct"], **({"error": r["_err"]} if r.get("_err") else {})) sf.write(json.dumps(res, ensure_ascii=False) + "\n") n_scored += 1 sf.flush() else: keep.append((fut, meta)) pending = keep def submit(b): jobs = [(r["sample_id"], r["_pred"], r["gt"], r["gt_format"]) for r in b] # chunksize=1 so one huge table cannot pin a whole chunk fut = pool.submit(_score_batch, jobs) pending.append((fut, b)) alive = lambda: any(t.is_alive() for t in threads) last = 0 while alive() or not done_q.empty(): try: r, res = done_q.get(timeout=1) except queue.Empty: flush_scores() continue dead = [port for p, _, port in procs if p.poll() is not None] if dead: raise RuntimeError(f"vLLM died: {dead}") rec = {"sample_id": r["sample_id"], "cohort": r["_cohort"], **res} pf.write(json.dumps(rec, ensure_ascii=False) + "\n") n_pred += 1 r["_pred"], r["_fr"], r["_ct"] = res["prediction"], res.get("finish_reason"), res.get("completion_tokens") r["_err"] = res.get("error") batch.append(r) if len(batch) >= max(1, a.score_batch // 16): submit(batch) batch = [] flush_scores() if time.time() - last > 30: el = time.time() - t_start log(f"pred {n_pred}/{total} ({n_pred/el:.1f}/s) scored {n_scored} pending_batches {len(pending)}") last = time.time() pf.close() if batch: submit(batch) stop_servers(procs) # 方案2: GPU 现在空着, 而打分还要跑一会儿(纯 CPU)。这段时间把下一个 # 模型的 engine 起好, 下一个 eval_mix 直接复用, 省掉串行的开机时间。 if a.preboot_model: try: preboot_servers(a.preboot_model, a.gpus, per_gpu, a.base_port, gmu, a.max_model_len, logdir, served) except Exception as e: log(f" 预启动失败(忽略): {type(e).__name__}: {e}") _score_unthrottle() log(f"inference done in {time.time()-t_start:.0f}s; draining scoring") flush_scores(block=True) pool.shutdown() sf.close() log(f"all done in {time.time()-t_start:.0f}s") summarize(score_path, outdir, a) def _score_batch(jobs): return [score_one(j) for j in jobs] def _register_pid(): """子进程把自己登记到 SCORE_PIDS, 供主进程推理结束后提权。""" try: import multiprocessing as _mp SCORE_PIDS.add(os.getpid()) except Exception: pass # ---------------------------------------------------------------------------- # 跨行跨列子集 (15k): 不重复推理, 只在跑全量时把命中的行标出来再分层汇报 # ---------------------------------------------------------------------------- SPAN_SUBSET_DEFAULT = "/autodl-tmp/span15k" SPAN = {} # sample_id -> difficulty, 由 --span-subset 填充 PENALTIES = {} # 复读惩罚参数, 由 CLI 填充 SCORE_PIDS = set() # 打分子进程 pid def load_span_subset(path): """-> {sample_id: difficulty}。子集是 shard_* 的子集, 推理照常跑一遍, 这里只做身份标记, 汇报时按 difficulty 聚合。""" import glob as _glob out = {} for f in sorted(_glob.glob(os.path.join(path, "*.parquet"))): t = pq.read_table(f, columns=["sample_id", "difficulty"]) for sid, d in zip(t.column("sample_id").to_pylist(), t.column("difficulty").to_pylist()): out.setdefault(sid, d) return out def summarize(score_path, outdir, a): agg = defaultdict(lambda: defaultdict(list)) def add(k, d): # teds 和 teds_s 可能只有一个有值: 纯结构集(GT 全是空 td)时 # TEDS 不适用(None), 只有 TEDS-S 有意义。**两个指标要各自计分母** —— # 之前统一用 "teds 非 None 的条数" 当分母, 于是 synthtabnet 29,990 条 # 里只有 40 条 teds 非 None, 汇总出来是 n=40 / TEDS-S=0.0000, # 而真实是 29,990 条 / 0.8510(2026-10-25 在 184k 全量汇总里发现的)。 if d.get("teds") is None: agg[k]["na"].append(1) else: agg[k]["teds"].append(d["teds"]) if d.get("teds_s") is None: agg[k]["na_s"].append(1) else: agg[k]["teds_s"].append(d["teds_s"]) if d.get("finish_reason") == "length": agg[k]["trunc"].append(1) for line in open(score_path): d = json.loads(line) if row_is_excluded(d):continue c = d["cohort"] add(c, d) if c.startswith("cx_") and d.get("dataset"): add(f"{c}/{d['dataset']}", d) # parquet 带的新维度: 自建/公开, 有无数值, 分辨率档 # 跨行跨列三档: 15k 子集的行会被额外聚合一次 if d.get("difficulty"): agg[f"_span/{d['difficulty']}"]["teds"].append( d["teds"] if d["teds"] is not None else 0) agg[f"_span/{d['difficulty']}"]["teds_s"].append( d["teds_s"] if d["teds_s"] is not None else 0) if d["teds"] is None: agg[f"_span/{d['difficulty']}"]["na"].append(1) if d.get("benchmark"): agg[f"_span/{d['difficulty']}/{d['benchmark']}"]["teds"].append( d["teds"] if d["teds"] is not None else 0) agg[f"_span/{d['difficulty']}/{d['benchmark']}"]["teds_s"].append( d["teds_s"] if d["teds_s"] is not None else 0) if d["teds"] is None: agg[f"_span/{d['difficulty']}/{d['benchmark']}"]["na"].append(1) for dim in ("origin", "has_numeric", "res_bucket"): v = d.get(dim) if v is not None: agg[f"_by/{dim}/{v}"]["teds"].append(d["teds"] if d["teds"] is not None else 0) agg[f"_by/{dim}/{v}"]["teds_s"].append(d["teds_s"] if d["teds_s"] is not None else 0) if d["teds"] is None: agg[f"_by/{dim}/{v}"]["na"].append(1) out = {} for c, v in sorted(agg.items()): n_t = len(v["teds"]) n_s = len(v["teds_s"]) if not n_t and not n_s: continue out[c] = { "n": max(n_t, n_s), "n_teds": n_t, "n_teds_s": n_s, # 每个指标用它自己的分母 "teds": round(sum(v["teds"]) / n_t, 5) if n_t else None, "teds_s": round(sum(v["teds_s"]) / n_s, 5) if n_s else None, "truncated": len(v["trunc"]), "na": len(v["na"]), "na_s": len(v.get("na_s", [])), } json.dump(out, open(f"{outdir}/summary.json", "w"), indent=1) for c, v in out.items(): t = " n/a " if v["teds"] is None else f"{v['teds']:.4f}" s_ = " n/a " if v["teds_s"] is None else f"{v['teds_s']:.4f}" print(f"{c:40s} n={v['n']:6d} (t {v['n_teds']:6d} / ts {v['n_teds_s']:6d})" f" TEDS={t} TEDS-S={s_} trunc={v['truncated']}") # mustard / synthtabnet 等的 GT 是纯结构(全空 td), 只有 TEDS-S 有意义; # 它们的 TEDS 列天然是 0, 不要当成模型问题读。 for c, v in out.items(): # 只有"绝大多数行没有 TEDS"才是纯结构集(少数几行缺 TEDS 只是 # 个别 GT 解析失败, 不是纯结构) if (not c.startswith("_by/") and v["n_teds_s"] > 0 and v["n_teds"] < 0.5 * v["n_teds_s"]): print(f" 注: {c} 为纯结构集, TEDS 不适用({v['n_teds']} 条有值), " f"请只看 TEDS-S") if __name__ == "__main__": main()