chess-sim/code/sim/rl_policy.py

"""SmolVLA as a stochastic policy for RL (ReinFlow-style noise injection), plus a critic.

SmolVLA turns Gaussian noise x_0 into a 50-step action chunk with 10 Euler steps of a learned
velocity field (VLAFlowMatching.sample_actions). Here each step adds Gaussian noise of a learned
per-step, per-joint size to the six real action dimensions:

    x_{k+1} = x_k + dt * v(x_k, t_k) + sigma_k * eps_k

so the probability of the chain x_1 ... x_10 given x_0 is a product of Gaussians that can be
computed and differentiated (ReinFlow). With the noise off this is exactly SmolVLA's sampling.

Trained: the action expert (`vlm_with_expert.lm_expert`) and the action and time projections
(`action_in_proj`, `action_out_proj`, `action_time_mlp_in`, `action_time_mlp_out`), and a critic (an
MLP on the frozen backbone's pooled prefix output and the normalized state); optionally the noise
sizes. Frozen: the vision-language backbone and `state_proj`, which feeds it, so the prefix (images,
instruction, state) is computed without gradients. The noise and the critic are only for training;
`save()` writes a plain SmolVLA checkpoint with the baseline's own processors (normalization).

Numerics (sim/reports/rl/check): the backbone runs in bfloat16 and gives batch-size-dependent
results, and small noise sizes turn tiny differences in the mean into large log-probability
differences. So the prefix (KV cache, pad mask, pooled features, state) of every decision is kept
from sampling time and reused in the update (`split_prefix` / `stack_prefix`), and the action
expert runs in float32 (`expert_fp32`, also needed for small updates to register at all). The
noise size falls geometrically from `sigma_start` (first denoising step, where the flow turns it
into a coherent change of the whole chunk) to `sigma_end` (last step, which adds it directly to
the joint targets as frame-to-frame jitter).
"""
from __future__ import annotations

import math
import shutil
import sys
from pathlib import Path

import torch
from torch import nn

HERE = Path(__file__).resolve().parent
sys.path.insert(0, str(HERE))

TRAINED = ("vlm_with_expert.lm_expert.", "action_in_proj.", "action_out_proj.", "action_time_mlp_in.",
           "action_time_mlp_out.")


def gauss_logp(x: torch.Tensor, mean: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor:
    """Log density of each element of x under N(mean, sigma^2): (B, T, A) -> (B, T, A). Kept per element:
    PPO clips each element's ratio (as per-token ratios in language-model PPO); the sum over the 300
    values of a step is far too sharp with the small noise of the last denoising steps."""
    z = (x - mean) / sigma
    return -0.5 * z * z - torch.log(sigma) - 0.5 * math.log(2 * math.pi)


class Critic(nn.Module):
    def __init__(self, feat_dim: int, state_dim: int, hidden: int = 256):
        super().__init__()
        self.net = nn.Sequential(nn.Linear(feat_dim + state_dim, hidden), nn.Tanh(), nn.Linear(hidden, hidden), nn.Tanh(),
                                 nn.Linear(hidden, 1))

    def forward(self, feat, state):
        return self.net(torch.cat([feat, state], dim=-1)).squeeze(-1)


class FlowRL(nn.Module):
    def __init__(self, path: str, instruction: str, sigma_start=0.1, sigma_end=0.002, learn_sigma=False,
                 expert_fp32=True):
        super().__init__()
        from eval_policy import load_policy

        policy, pre, post, device = load_policy(path, expert_fp32=expert_fp32)
        self.policy, self.pre, self.post, self.device = policy, pre, post, device
        self.src_path = Path(path)
        self.model = policy.model
        cfg = policy.config
        self.n_steps, self.chunk, self.adim = cfg.num_steps, cfg.chunk_size, cfg.action_feature.shape[0]
        self.max_adim = cfg.max_action_dim
        self.instruction = instruction
        for p in policy.parameters():
            p.requires_grad_(False)
        self.trained_names = []
        for name, p in self.model.named_parameters():
            if name.startswith(TRAINED):
                p.requires_grad_(True)
                self.trained_names.append(name)
        frac = torch.linspace(0.0, 1.0, self.n_steps, device=device)[:, None]
        init = math.log(sigma_start) + frac * (math.log(sigma_end) - math.log(sigma_start))
        self.log_sigma = nn.Parameter(init.expand(self.n_steps, self.adim).clone(), requires_grad=learn_sigma)
        self.sigma_min, self.sigma_max = 0.5 * min(sigma_start, sigma_end), 1.5 * max(sigma_start, sigma_end)
        hidden = self.model.vlm_with_expert.config.text_config.hidden_size
        self.critic = Critic(hidden, self.adim).to(device)
        policy.eval()

    # ------------------------------------------------------------------ inputs
    def sigma(self) -> torch.Tensor:
        return self.log_sigma.exp().clamp(self.sigma_min, self.sigma_max)

    def policy_params(self):
        return [p for n, p in self.model.named_parameters() if p.requires_grad]

    def batch(self, obs_list: list[dict]) -> dict:
        """Observations (numpy, as eval_policy builds them) through the policy's own preprocessor, stacked."""
        from lerobot.policies.utils import prepare_observation_for_inference

        items = []
        with torch.no_grad():
            for o in obs_list:
                ob = prepare_observation_for_inference(dict(o), self.device, self.instruction, "so101_follower")
                items.append(self.pre(ob))
        return {k: torch.cat([it[k] for it in items]) for k, v in items[0].items() if torch.is_tensor(v)}

    @torch.no_grad()
    def prefix(self, batch: dict):
        """Images, instruction and state through the frozen backbone: (pad masks, KV cache, pooled
        features, normalized state)."""
        from lerobot.policies.smolvla.modeling_smolvla import make_att_2d_masks
        from lerobot.utils.constants import OBS_LANGUAGE_ATTENTION_MASK, OBS_LANGUAGE_TOKENS

        images, img_masks = self.policy.prepare_images(batch)
        state = self.policy.prepare_state(batch)
        embs, pad, att = self.model.embed_prefix(images, img_masks, batch[OBS_LANGUAGE_TOKENS],
                                                 batch[OBS_LANGUAGE_ATTENTION_MASK], state=state)
        att2d = make_att_2d_masks(pad, att)
        pos = torch.cumsum(pad, dim=1) - 1
        outs, kv = self.model.vlm_with_expert.forward(attention_mask=att2d, position_ids=pos, past_key_values=None,
                                                      inputs_embeds=[embs, None], use_cache=self.policy.config.use_cache,
                                                      fill_kv_cache=True)
        h, m = outs[0].float(), pad.unsqueeze(-1).float()
        feat = (h * m).sum(1) / m.sum(1).clamp(min=1.0)
        return pad, kv, feat, state[:, :self.adim].float()

    def _times(self, bsize: int):
        dt = -1.0 / self.n_steps
        return dt, [torch.tensor(1.0 + k * dt, dtype=torch.float32, device=self.device).expand(bsize)
                    for k in range(self.n_steps)]

    # ------------------------------------------------------------------ sampling and likelihood
    @torch.no_grad()
    def sample(self, batch: dict, x0: torch.Tensor, noise: bool = True):
        """The chain x_0..x_n (B, n+1, T, max_adim), its log-probabilities per element (B, n, T, A), the
        critic's value (B,) and the prefix (pad, kv, feat, state) for the update. With noise=False,
        x_n is SmolVLA's own action chunk for this x_0."""
        prefix = self.prefix(batch)
        pad, kv, feat, state = prefix
        bsize = x0.shape[0]
        dt, times = self._times(bsize)
        sig = self.sigma()
        x, chain = x0, [x0]
        logp = x0.new_zeros(bsize, self.n_steps, self.chunk, self.adim)
        for k in range(self.n_steps):
            v = self.model.denoise_step(pad, kv, x, times[k])
            mean = x + dt * v
            if noise:
                nxt = mean.clone()
                eps = torch.randn(bsize, self.chunk, self.adim, device=self.device)
                nxt[..., :self.adim] = mean[..., :self.adim] + sig[k] * eps
                logp[:, k] = gauss_logp(nxt[..., :self.adim], mean[..., :self.adim], sig[k])
            else:
                nxt = mean
            x = nxt
            chain.append(x)
        return torch.stack(chain, 1), logp, self.critic(feat, state), prefix

    @staticmethod
    def split_prefix(prefix, j: int):
        """Decision j's prefix, on the CPU."""
        pad, kv, feat, state = prefix
        return (pad[j:j + 1].cpu(), {l: {k: t[j:j + 1].cpu() for k, t in d.items()} for l, d in kv.items()},
                feat[j:j + 1].cpu(), state[j:j + 1].cpu())

    def stack_prefix(self, items):
        """Stored prefixes of several decisions as one batch on the device."""
        dev = self.device
        kv = {l: {k: torch.cat([it[1][l][k] for it in items]).to(dev) for k in items[0][1][l]} for l in items[0][1]}
        return (torch.cat([it[0] for it in items]).to(dev), kv, torch.cat([it[2] for it in items]).to(dev),
                torch.cat([it[3] for it in items]).to(dev))

    def evaluate(self, batch: dict | None, chain: torch.Tensor, prefix=None):
        """Per-element log-probabilities (B, n, T, A) of a stored chain under the current weights (with
        gradients through the action expert and the noise sizes) and the critic's value (B,).
        Pass the stored `prefix` (stack_prefix) to reuse sampling time's; else it is recomputed."""
        pad, kv, feat, state = prefix if prefix is not None else self.prefix(batch)
        bsize = chain.shape[0]
        dt, times = self._times(bsize)
        sig = self.sigma()
        logp = []
        for k in range(self.n_steps):
            v = self.model.denoise_step(pad, kv, chain[:, k], times[k])
            mean = chain[:, k] + dt * v
            logp.append(gauss_logp(chain[:, k + 1, ..., :self.adim], mean[..., :self.adim], sig[k]))
        return torch.stack(logp, 1), self.critic(feat, state)

    @torch.no_grad()
    def actions(self, chain: torch.Tensor):
        """The executed chunk in LeRobot units (B, T, 6): the last chain step through the policy's
        own postprocessor (unnormalization), as eval_policy applies it to every action."""
        x = chain[:, -1, :, :self.adim]
        bsize = x.shape[0]
        a = self.post(x.reshape(bsize * self.chunk, self.adim))
        return a.reshape(bsize, self.chunk, self.adim).float().cpu().numpy()

    def x0(self, bsize: int, generators=None) -> torch.Tensor:
        """Initial noise as SmolVLA draws it (sample_noise); with per-episode CUDA generators seeded
        like eval_policy's torch.manual_seed, the same draws as eval_policy.py."""
        shape = (self.chunk, self.max_adim)
        if generators is None:
            return torch.normal(mean=0.0, std=1.0, size=(bsize, *shape), dtype=torch.float32, device=self.device)
        return torch.cat([torch.normal(mean=0.0, std=1.0, size=(1, *shape), dtype=torch.float32, device=self.device,
                                       generator=g) for g in generators])

    # ------------------------------------------------------------------ checkpoint
    def save(self, out: str | Path, extra: dict | None = None):
        """A plain SmolVLA checkpoint (config + weights) with the baseline's processor files, and
        the training-only parts (noise sizes, critic) in rl_state.pt."""
        out = Path(out)
        out.mkdir(parents=True, exist_ok=True)
        self.policy.save_pretrained(out)
        for f in self.src_path.glob("policy_*processor*"):
            shutil.copy2(f, out / f.name)
        torch.save(dict(log_sigma=self.log_sigma.detach().cpu(), critic=self.critic.state_dict(), **(extra or {})),
                   out / "rl_state.pt")