chess-sim/code/sim/rl_check.py

"""Checks before an RL training run (sim/rl_policy.py, sim/rl_env.py, eval_policy.py --fast).

1. same_as_eval: with the noise off and one scene at a time, the RL environment and sampler reproduce
   eval_policy.py exactly on the same scenes (success, frames, closest approach, closing error):
   a) eval_policy --fast vs normal; b) the RL path with the action expert in bfloat16 (as loaded) vs
   eval_policy; c) the RL path with the expert in float32 (as trained here) vs eval_policy --expert-fp32.
2. noise: success with each candidate noise schedule (sigma at the first and last denoising step)
   against no noise, on the same scenes (what exploration costs).
3. likelihood: log-probabilities recomputed from stored chains and prefixes (other minibatch
   compositions) equal those at sampling time, so the PPO ratio starts at 1.
4. update: after one PPO step only the intended weights changed; the backbone is bit-identical.
5. reload: a saved checkpoint keeps the baseline's normalization, loads in eval_policy.load_policy
   (float32 expert, full precision) and gives the same chunk for the same x_0.
Writes <out>/rl_check.md and .json. Needs anti-aliasing off (SIM_OFFSAMPLES=0).

Run (pod):  MUJOCO_GL=egl PHASE2_PROFILE=baseline SIM_OFFSAMPLES=0 .venv/bin/python sim/rl_check.py \
                --policy models/baseline --out sim/reports/rl/check
"""
from __future__ import annotations

import argparse
import hashlib
import json
import os
import subprocess
import sys
import tempfile
from pathlib import Path

HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE))
FIELDS = ("success", "frames", "closest_mm", "close_lateral_mm")


def eval_run(policy, n, seed, out, fast, fp32=False):
    out = Path(out)
    if not (out / "eval_results.json").exists():
        cmd = [sys.executable, str(HERE / "eval_policy.py"), "--policy", policy, "--episodes", str(n), "--workers", "2",
               "--seed", str(seed), "--video-episodes", "0", "--out", str(out)]
        cmd += (["--fast"] if fast else []) + (["--expert-fp32"] if fp32 else [])
        subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, env=dict(os.environ))
    return {e["index"]: e for e in json.loads((out / "eval_results.json").read_text())["episodes"]}


def compare(a: dict, b: dict):
    diffs = [(i, f, a[i].get(f), b[i].get(f)) for i in sorted(a) for f in FIELDS if a[i].get(f) != b[i].get(f)]
    return dict(identical=not diffs, differences=diffs[:10], episodes=len(a),
                success=(sum(bool(x["success"]) for x in a.values()), sum(bool(x["success"]) for x in b.values())))


def digest(t):
    return hashlib.sha1(t.detach().float().cpu().numpy().tobytes()).hexdigest()


def set_schedule(pol, s0, s1):
    """Geometric noise schedule from s0 (first denoising step) to s1 (last), as FlowRL builds it."""
    import math

    import torch

    frac = torch.linspace(0.0, 1.0, pol.n_steps, device=pol.device)[:, None]
    pol.log_sigma.data = (math.log(s0) + frac * (math.log(s1) - math.log(s0))).expand(pol.n_steps, pol.adim).clone()
    pol.sigma_min, pol.sigma_max = 0.5 * min(s0, s1), 1.5 * max(s0, s1)


def main():
    ap = argparse.ArgumentParser(description=__doc__)
    ap.add_argument("--policy", required=True)
    ap.add_argument("--out", required=True)
    ap.add_argument("--same-episodes", type=int, default=8)
    ap.add_argument("--noise-episodes", type=int, default=60)
    ap.add_argument("--seed", type=int, default=9_000_003)
    ap.add_argument("--noise-seed", type=int, default=9_400_003)
    ap.add_argument("--schedules", default="0.1:0.002,0.2:0.002,0.3:0.003",
                    help="candidate noise schedules, first:last denoising step, comma-separated")
    args = ap.parse_args()
    out = Path(args.out)
    out.mkdir(parents=True, exist_ok=True)
    report = dict(args=vars(args))
    schedules = [tuple(float(x) for x in s.split(":")) for s in args.schedules.split(",")]

    # 1a / reference runs of eval_policy.py (before this process takes GPU memory).
    normal = eval_run(args.policy, args.same_episodes, args.seed, out / "eval_normal", fast=False)
    fast = eval_run(args.policy, args.same_episodes, args.seed, out / "eval_fast", fast=True)
    fp32 = eval_run(args.policy, args.same_episodes, args.seed, out / "eval_fast_fp32", fast=True, fp32=True)
    report["fast_vs_normal"] = compare(fast, normal)
    report["eval_fp32_vs_bf16"] = compare(fp32, normal)
    print("fast vs normal:", report["fast_vs_normal"], "| eval fp32 vs bf16 expert:", report["eval_fp32_vs_bf16"], flush=True)

    import numpy as np
    import torch

    from episode import load_config
    from rl_env import VecChessEnv
    from rl_policy import FlowRL
    from rl_train import ppo_update, run_episodes, summarize

    cfg = load_config()
    instr = cfg["dataset"]["instruction"]

    def same(pol):
        venv = VecChessEnv(1, cfg)
        eps = run_episodes(pol, venv, args.seed, list(range(args.same_episodes)), noise=False, early_stop=False,
                           check_expert=True, record=False, eval_noise_seeds=True)
        venv.close()
        return {e["info"]["index"]: dict(e["result"], index=e["info"]["index"]) for e in eps}

    # 1b. the RL path with the expert in bfloat16, as eval_policy loads it.
    pol = FlowRL(args.policy, instr, expert_fp32=False)
    report["rl_bf16_vs_eval"] = compare(same(pol), normal)
    print("rl bf16 vs eval:", report["rl_bf16_vs_eval"], flush=True)
    del pol
    torch.cuda.empty_cache()

    # 1c. the RL path with the expert in float32 vs eval_policy --expert-fp32.
    pol = FlowRL(args.policy, instr, *schedules[0], expert_fp32=True)
    base_digests = {n: digest(p) for n, p in pol.model.named_parameters()}
    report["rl_fp32_vs_eval_fp32"] = compare(same(pol), fp32)
    print("rl fp32 vs eval fp32:", report["rl_fp32_vs_eval_fp32"], flush=True)

    # 2. what exploration costs, per candidate schedule (batched, same scenes).
    venv = VecChessEnv(8, cfg)
    idx = list(range(args.noise_episodes))
    torch.manual_seed(0)
    report["noise"] = {"off": summarize(run_episodes(pol, venv, args.noise_seed, idx, noise=False, early_stop=False,
                                                     check_expert=False, record=False))}
    for s0, s1 in schedules:
        set_schedule(pol, s0, s1)
        report["noise"][f"{s0}:{s1}"] = summarize(run_episodes(pol, venv, args.noise_seed, idx, noise=True,
                                                               early_stop=False, check_expert=False, record=False))
        print("noise", f"{s0}:{s1}", report["noise"][f"{s0}:{s1}"], flush=True)
    print("noise off", report["noise"]["off"], flush=True)

    # 3. likelihood of stored chains with their stored prefixes, other minibatch compositions.
    s0, s1 = schedules[0]
    set_schedule(pol, s0, s1)
    eps = run_episodes(pol, venv, args.noise_seed + 1, list(range(8)), noise=True, early_stop=True, check_expert=False,
                       record=True)
    venv.close()
    dec = [d for e in eps for d in e["decisions"]][:24]
    with torch.no_grad():
        chain = torch.stack([d["chain"] for d in dec]).to(pol.device)
        old = torch.stack([d["logp"] for d in dec]).to(pol.device)
        new_all, _ = pol.evaluate(None, chain, prefix=pol.stack_prefix([d["prefix"] for d in dec]))
        new_one = torch.cat([pol.evaluate(None, d["chain"][None].to(pol.device), prefix=pol.stack_prefix([d["prefix"]]))[0]
                             for d in dec[:8]])
    report["likelihood"] = dict(
        schedule=f"{s0}:{s1}", decisions=len(dec), logp_scale=round(float(old.abs().mean()), 1),
        max_abs_diff_batch=float((new_all - old).abs().max()), max_ratio_minus_1_batch=float(((new_all - old).exp() - 1).abs().max()),
        max_abs_diff_single=float((new_one - old[:8]).abs().max()),
        max_ratio_minus_1_single=float(((new_one - old[:8]).exp() - 1).abs().max()))
    print("likelihood:", report["likelihood"], flush=True)

    # 4. one PPO step: which weights change.
    for d in dec:
        d["adv"], d["ret"] = float(np.random.randn()), float(np.random.rand())
    opt = torch.optim.AdamW([dict(params=pol.policy_params(), lr=3e-6), dict(params=list(pol.critic.parameters()), lr=3e-4)])
    ns = argparse.Namespace(epochs=1, minibatch=8, clip=0.2, vf_coef=0.5, max_grad_norm=1.0, target_kl=1e9)
    upd = ppo_update(pol, opt, dec, ns, policy_on=True)
    changed = [n for n, p in pol.model.named_parameters() if digest(p) != base_digests[n]]
    groups = {}
    for n in changed:
        g = n.split(".")[0] if not n.startswith("vlm_with_expert.") else ".".join(n.split(".")[:2])
        groups[g] = groups.get(g, 0) + 1
    trained = set(pol.trained_names)
    with torch.no_grad():
        after, _ = pol.evaluate(None, chain, prefix=pol.stack_prefix([d["prefix"] for d in dec]))
    report["update"] = dict(step=upd, changed_tensors=len(changed), trained_tensors=len(trained),
                            changed_outside_trained=[n for n in changed if n not in trained][:10],
                            trained_unchanged=len(trained - set(changed)), changed_groups=groups,
                            backbone_identical=not any(n.startswith("vlm_with_expert.vlm.") for n in changed),
                            state_proj_identical=not any(n.startswith("state_proj.") for n in changed),
                            logp_change_after_step=float((after - new_all).abs().mean()))
    print("update:", report["update"], flush=True)

    # 5. save, reload (as eval_policy --expert-fp32 loads it), same chunk.
    with tempfile.TemporaryDirectory() as tmp:
        pol.save(tmp)
        from safetensors.torch import load_file

        from eval_policy import load_policy
        norm_same = all(
            all(torch.equal(a[k], b[k]) for k in a)
            for f in Path(args.policy).glob("policy_*processor*.safetensors")
            for a, b in [(load_file(str(f)), load_file(str(Path(tmp) / f.name)))])
        loaded, _, _, _ = load_policy(tmp, expert_fp32=True)
        mine = dict(pol.model.named_parameters())
        weights_same = all(torch.equal(p.detach(), mine[n].detach()) for n, p in loaded.model.named_parameters()
                           if n.startswith("vlm_with_expert.lm_expert."))
        del loaded
        re = FlowRL(tmp, instr, *schedules[0], expert_fp32=True)
        with torch.no_grad():
            obs = [d for d in dec[:4]]
            x0 = pol.x0(4)
            p1 = pol.stack_prefix([d["prefix"] for d in obs])
            c1 = pol.evaluate(None, chain[:4], prefix=p1)[0]
            c2 = re.evaluate(None, chain[:4], prefix=re.stack_prefix([d["prefix"] for d in obs]))[0]
        report["reload"] = dict(normalization_identical=bool(norm_same), expert_weights_identical=bool(weights_same),
                                max_logp_diff=float((c1 - c2).abs().max()), files=sorted(p.name for p in Path(tmp).iterdir()))
    print("reload:", report["reload"], flush=True)

    (out / "rl_check.json").write_text(json.dumps(report, indent=1, default=str))
    lines = ["# RL checks", "", f"Policy `{args.policy}`.", "",
             f"1. Same as eval_policy ({args.same_episodes} scenes, seed {args.seed}, noise off, one scene at a time):",
             f"   - eval_policy --fast vs normal: {report['fast_vs_normal']['identical']}",
             f"   - RL path, expert bfloat16 vs eval_policy: {report['rl_bf16_vs_eval']['identical']} "
             f"{'' if report['rl_bf16_vs_eval']['identical'] else report['rl_bf16_vs_eval']['differences'][:3]}",
             f"   - RL path, expert float32 vs eval_policy --expert-fp32: {report['rl_fp32_vs_eval_fp32']['identical']} "
             f"{'' if report['rl_fp32_vs_eval_fp32']['identical'] else report['rl_fp32_vs_eval_fp32']['differences'][:3]}",
             f"   - eval_policy, expert float32 vs bfloat16: successes {report['eval_fp32_vs_bf16']['success']}",
             f"2. Noise schedules ({args.noise_episodes} scenes, seed {args.noise_seed}): success / clean lift / closing error median:",
             ] + [f"   - {k}: {v['success']:.1%} / {v['clean_lift']:.1%} / {v['close_mm_median']} mm" for k, v in report["noise"].items()] + [
             f"3. Likelihood ({report['likelihood']['schedule']}), {report['likelihood']['decisions']} stored chains: max |ratio - 1| "
             f"{report['likelihood']['max_ratio_minus_1_batch']:.2e} (one minibatch), "
             f"{report['likelihood']['max_ratio_minus_1_single']:.2e} (one at a time); log-prob scale {report['likelihood']['logp_scale']}.",
             f"4. One PPO step (lr 3e-6): {report['update']['changed_tensors']} tensors changed of {report['update']['trained_tensors']} "
             f"trained; outside the trained set: {report['update']['changed_outside_trained'] or 'none'}; backbone identical "
             f"{report['update']['backbone_identical']}; state_proj identical {report['update']['state_proj_identical']}; "
             f"mean log-prob change {report['update']['logp_change_after_step']:.3g}; KL {report['update']['step'].get('kl')}.",
             f"5. Save/reload: normalization identical {report['reload']['normalization_identical']}; expert weights identical "
             f"after eval_policy's float32 load {report['reload']['expert_weights_identical']}; max log-prob difference "
             f"{report['reload']['max_logp_diff']:.2e}."]
    (out / "rl_check.md").write_text("\n".join(lines) + "\n")
    print("\n".join(lines))


if __name__ == "__main__":
    main()