Datasets:

eduzrh
/

training-model-material

"""Sequence models for the HEA MD trajectory task.
All models share the signature ``forward(static, x) -> dict(pred=(B,T,F))``
where ``pred`` is in the *normalised* target space defined by data.Normalizer.
Implemented:
  * MLP        - point-wise baseline (no history)
  * GRU / LSTM - recurrent baselines
  * Transformer- causal encoder baseline
  * Mamba      - selective state-space (S6) backbone
  * Mamba2     - SSD-style multi-head gated SSM backbone
  * PIMamba    - Mamba + static conditioning + micro->macro coupling
"""
from __future__ import annotations

import math
from typing import Optional

import torch
import torch.nn as nn
import torch.nn.functional as F


# ---------------------------------------------------------------------------
# selective scan (chunked, numerically stable, pure PyTorch)
# ---------------------------------------------------------------------------
def selective_scan(u, delta, A, B, C, D=None, chunk: int = 8):
    """Linear time-invariant-in-A selective scan.
    u     : (b, T, d)
    delta : (b, T, d)
    A     : (d, n)   negative real
    B, C  : (b, T, n)
    returns (b, T, d)
    """
    b, T, d = u.shape
    if A.dim() == 2:
        A = A.unsqueeze(0).unsqueeze(0)      # (1,1,d,n)
    elif A.dim() == 3:
        A = A.unsqueeze(1)                   # (B,1,d,n)
    n = A.shape[-1]
    dt = delta.clamp(1e-4, 0.2)
    dev = u.device
    h = torch.zeros(b, d, n, device=dev, dtype=u.dtype)
    pad = (chunk - T % chunk) % chunk
    if pad:
        u = F.pad(u, (0, 0, 0, pad))
        dt = F.pad(dt, (0, 0, 0, pad))
        B = F.pad(B, (0, 0, 0, pad))
        C = F.pad(C, (0, 0, 0, pad))
    outs = []
    Tp = u.shape[1]
    for s in range(0, Tp, chunk):
        uc = u[:, s:s + chunk]
        dtc = dt[:, s:s + chunk]
        Bc = B[:, s:s + chunk]
        Cc = C[:, s:s + chunk]
        a = torch.exp(dtc.unsqueeze(-1) * A)                 # (b,c,d,n)
        bu = (dtc * uc).unsqueeze(-1) * Bc.unsqueeze(2)      # (b,c,d,n)
        P = torch.cumprod(a, dim=1)
        hloc = P * torch.cumsum(bu / P.clamp_min(1e-20), dim=1)
        hc = hloc + P * h.unsqueeze(1)
        h = hc[:, -1]
        y = (hc * Cc.unsqueeze(2)).sum(-1)
        outs.append(y)
    y = torch.cat(outs, dim=1)[:, :T]
    if D is not None:
        y = y + D.unsqueeze(0).unsqueeze(0) * u[:, :T]
    return y


class RMSNorm(nn.Module):
    def __init__(self, d, eps=1e-5):
        super().__init__()
        self.w = nn.Parameter(torch.ones(d))
        self.eps = eps

    def forward(self, x):
        return self.w * x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)


class MambaBlock(nn.Module):
    def __init__(self, d_model, d_state=16, expand=2, d_conv=4, chunk=32):
        super().__init__()
        self.d_inner = int(expand * d_model)
        self.d_state = d_state
        self.chunk = chunk
        self.norm = RMSNorm(d_model)
        self.in_proj = nn.Linear(d_model, 2 * self.d_inner, bias=False)
        self.conv = nn.Conv1d(self.d_inner, self.d_inner, d_conv,
                              groups=self.d_inner, padding=d_conv - 1)
        self.x_proj = nn.Linear(self.d_inner, 2 * d_state + 1, bias=False)
        self.dt_proj = nn.Linear(1, self.d_inner, bias=True)
        A = torch.empty(self.d_inner, d_state)
        nn.init.uniform_(A, 0.5, 8.0)          # positive magnitude
        self.A_log = nn.Parameter(torch.log(A))
        self.D = nn.Parameter(torch.ones(self.d_inner))
        self.out_proj = nn.Linear(self.d_inner, d_model, bias=False)
        with torch.no_grad():
            self.dt_proj.bias.uniform_(math.log(0.001), math.log(0.05))

    def forward(self, x):
        b, T, _ = x.shape
        h = self.norm(x)
        xz = self.in_proj(h)
        xin, z = xz.chunk(2, dim=-1)
        xin = self.conv(xin.transpose(1, 2))[:, :, :T].transpose(1, 2)
        xin = F.silu(xin)
        proj = self.x_proj(xin)
        dt, Bc, Cc = torch.split(proj, [1, self.d_state, self.d_state], dim=-1)
        dt = F.softplus(self.dt_proj(dt))
        A = -torch.exp(self.A_log)
        y = selective_scan(xin, dt, A, Bc, Cc, self.D, chunk=self.chunk)
        y = y * F.silu(z)
        return x + self.out_proj(y)


class Mamba2Block(nn.Module):
    """Simplified SSD-style block: multi-head scalar decay + shared B/C groups."""

    def __init__(self, d_model, nheads=4, d_state=64, expand=1, d_conv=4,
                 n_groups=1, chunk=32):
        super().__init__()
        assert d_model % nheads == 0
        self.d_model = d_model
        self.nheads = nheads
        self.headdim = d_model // nheads
        self.d_state = d_state
        self.n_groups = n_groups
        self.chunk = chunk
        self.norm = RMSNorm(d_model)
        self.in_proj = nn.Linear(d_model, 2 * d_model + 2 * d_state * n_groups + nheads,
                                 bias=False)
        self.conv = nn.Conv1d(d_model, d_model, d_conv, groups=d_model,
                              padding=d_conv - 1)
        self.A_log = nn.Parameter(torch.log(torch.empty(nheads).uniform_(0.5, 8.0)))
        self.D = nn.Parameter(torch.ones(nheads))
        self.dt_bias = nn.Parameter(torch.empty(nheads).uniform_(-4.0, -2.0))
        self.norm_g = RMSNorm(d_model)
        self.out_proj = nn.Linear(d_model, d_model, bias=False)

    def forward(self, x):
        b, T, _ = x.shape
        h = self.norm(x)
        nh, hd, ng, ds = self.nheads, self.headdim, self.n_groups, self.d_state
        split = [self.d_model, self.d_model, ng * ds, ng * ds, nh]
        xin, z, Bc, Cc, dt = torch.split(self.in_proj(h), split, dim=-1)
        xin = self.conv(xin.transpose(1, 2))[:, :, :T].transpose(1, 2)
        xin = F.silu(xin).view(b, T, nh, hd)
        dt = F.softplus(dt + self.dt_bias).view(b, T, nh)     # (b,T,nh)
        Bc = Bc.view(b, T, ng, ds)
        Cc = Cc.view(b, T, ng, ds)
        if ng != nh:
            rep = nh // ng
            Bc = Bc.repeat_interleave(rep, dim=2)
            Cc = Cc.repeat_interleave(rep, dim=2)
        # fold heads into batch, state dim = 1 -> reuse scalar scan
        xh = xin.permute(0, 2, 1, 3).reshape(b * nh, T, hd)
        dth = dt.permute(0, 2, 1).reshape(b * nh, T, 1).expand(-1, -1, hd)
        Bh = Bc.permute(0, 2, 1, 3).reshape(b * nh, T, ds)
        Ch = Cc.permute(0, 2, 1, 3).reshape(b * nh, T, ds)
        A = ((-torch.exp(self.A_log)).view(1, nh, 1, 1)
             .expand(b, nh, hd, 1).reshape(b * nh, 1, hd, 1))
        y = selective_scan(xh, dth, A, Bh, Ch, None, chunk=self.chunk)
        y = y + xh * self.D.view(1, nh, 1, 1).expand(b, nh, hd, 1).reshape(b * nh, 1, hd)
        y = y.view(b, nh, T, hd).permute(0, 2, 1, 3).reshape(b, T, self.d_model)
        y = self.norm_g(y) * F.silu(z)
        return x + self.out_proj(y)


class FiLM(nn.Module):
    def __init__(self, cond_dim, d_model):
        super().__init__()
        self.to_scale = nn.Linear(cond_dim, d_model)
        self.to_shift = nn.Linear(cond_dim, d_model)

    def forward(self, h, cond):
        s = 1.0 + self.to_scale(cond).unsqueeze(1)
        b = self.to_shift(cond).unsqueeze(1)
        return h * s + b


# ---------------------------------------------------------------------------
# full models
# ---------------------------------------------------------------------------
class SequenceModel(nn.Module):
    """Common wrapper: static encoder + backbone + multi-task heads."""

    def __init__(self, n_static, n_out, d_model=64, n_layers=2, dropout=0.1,
                 n_phase=4, n_rho=7, coupling=True, **kwargs):
        super().__init__()
        self.n_phase, self.n_rho = n_phase, n_rho
        self.n_micro = n_phase + n_rho + 1
        self.static_enc = nn.Sequential(
            nn.Linear(n_static, d_model), nn.GELU(),
            nn.Dropout(dropout), nn.Linear(d_model, d_model))
        self.cond_dim = d_model
        self.drop = nn.Dropout(dropout)
        self.coupling = coupling
        # ---- independent output head per physical variable (step 2 要求) ----
        self.phase_head = nn.Linear(d_model, n_phase)
        mk = lambda: nn.Sequential(nn.Linear(d_model, max(8, d_model // 2)),
                                   nn.GELU(),
                                   nn.Linear(max(8, d_model // 2), 1))
        self.rho_heads = nn.ModuleList([mk() for _ in range(n_rho)])
        self.twin_head = mk()
        stress_in = 2 * d_model if coupling else d_model
        self.stress_head = nn.Sequential(
            nn.Linear(stress_in, d_model), nn.GELU(), nn.Linear(d_model, 1))
        self.micro_proj = nn.Linear(self.n_micro, d_model) if coupling else None
        assert n_out == 1 + self.n_micro, (n_out, self.n_micro)

    def head(self, h, static):
        phase = torch.softmax(self.phase_head(h), dim=-1)
        rho = torch.cat([F.softplus(mi(h)) for mi in self.rho_heads], dim=-1)
        twin = F.softplus(self.twin_head(h))
        micro_p = torch.cat([phase, rho, twin], dim=-1)
        if self.coupling:
            hs = torch.cat([h, self.micro_proj(micro_p)], dim=-1)
        else:
            hs = h
        sigma = self.stress_head(hs)
        return torch.cat([sigma, micro_p], dim=-1), micro_p


class MLPModel(SequenceModel):
    def __init__(self, n_static, n_out, seq_dim=5, d_model=128, **kw):
        super().__init__(n_static, n_out, d_model=d_model, **kw)
        self.inp = nn.Sequential(nn.Linear(seq_dim + n_static, d_model), nn.GELU(),
                                 nn.Linear(d_model, d_model), nn.GELU())

    def forward(self, static, x):
        b, T, _ = x.shape
        z = torch.cat([x, static.unsqueeze(1).expand(-1, T, -1)], dim=-1)
        h = self.inp(z)
        pred, micro = self.head(h, static)
        return {"pred": pred, "micro": micro, "h": h}


class RNNModel(SequenceModel):
    def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
                 rnn="gru", **kw):
        super().__init__(n_static, n_out, d_model=d_model, **kw)
        self.inp = nn.Linear(seq_dim, d_model)
        cls = nn.GRU if rnn == "gru" else nn.LSTM
        self.rnn = cls(d_model, d_model, num_layers=n_layers, batch_first=True,
                       dropout=0.1 if n_layers > 1 else 0.0)
        self.film = FiLM(d_model, d_model)

    def forward(self, static, x):
        cond = self.static_enc(static)
        h = self.inp(x)
        h, _ = self.rnn(h)
        h = self.film(h, cond)
        pred, micro = self.head(h, static)
        return {"pred": pred, "micro": micro, "h": h}


class TransformerModel(SequenceModel):
    def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
                 n_heads=4, **kw):
        super().__init__(n_static, n_out, d_model=d_model, **kw)
        self.inp = nn.Linear(seq_dim, d_model)
        self.pos = nn.Parameter(torch.randn(1, 512, d_model) * 0.02)
        layer = nn.TransformerEncoderLayer(d_model, n_heads, 4 * d_model,
                                           dropout=0.1, batch_first=True,
                                           norm_first=True, activation="gelu")
        self.enc = nn.TransformerEncoder(layer, n_layers, enable_nested_tensor=False)
        self.film = FiLM(d_model, d_model)
        self.register_buffer("mask", torch.triu(torch.ones(512, 512), diagonal=1).bool())

    def forward(self, static, x):
        b, T, _ = x.shape
        cond = self.static_enc(static)
        h = self.inp(x) + self.pos[:, :T]
        h = self.enc(h, mask=self.mask[:T, :T])
        h = self.film(h, cond)
        pred, micro = self.head(h, static)
        return {"pred": pred, "micro": micro, "h": h}


class MambaModel(SequenceModel):
    def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
                 d_state=16, expand=2, **kw):
        super().__init__(n_static, n_out, d_model=d_model, **kw)
        self.inp = nn.Linear(seq_dim, d_model)
        self.blocks = nn.ModuleList([MambaBlock(d_model, d_state, expand)
                                     for _ in range(n_layers)])
        self.norm = RMSNorm(d_model)
        self.cond_proj = nn.Linear(d_model, d_model)

    def forward(self, static, x):
        cond = self.static_enc(static)
        h = self.inp(x) + cond.unsqueeze(1)
        for blk in self.blocks:
            h = blk(h)
        h = self.norm(h)
        pred, micro = self.head(h, static)
        return {"pred": pred, "micro": micro, "h": h}


class Mamba2Model(SequenceModel):
    def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
                 nheads=4, d_state=64, **kw):
        super().__init__(n_static, n_out, d_model=d_model, **kw)
        self.inp = nn.Linear(seq_dim, d_model)
        self.blocks = nn.ModuleList([Mamba2Block(d_model, nheads=nheads,
                                                 d_state=d_state)
                                     for _ in range(n_layers)])
        self.norm = RMSNorm(d_model)

    def forward(self, static, x):
        cond = self.static_enc(static)
        h = self.inp(x) + cond.unsqueeze(1)
        for blk in self.blocks:
            h = blk(h)
        h = self.norm(h)
        pred, micro = self.head(h, static)
        return {"pred": pred, "micro": micro, "h": h}


class PIMamba(SequenceModel):
    """Physics-informed Mamba: FiLM conditioning + micro -> macro coupling."""

    def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
                 d_state=16, expand=2, coupling=True, **kw):
        super().__init__(n_static, n_out, d_model=d_model, coupling=coupling, **kw)
        self.inp = nn.Linear(seq_dim, d_model)
        self.blocks = nn.ModuleList([MambaBlock(d_model, d_state, expand)
                                     for _ in range(n_layers)])
        self.films = nn.ModuleList([FiLM(d_model, d_model) for _ in range(n_layers)])
        self.norm = RMSNorm(d_model)

    def forward(self, static, x):
        cond = self.static_enc(static)
        h = self.inp(x) + cond.unsqueeze(1)
        for blk, film in zip(self.blocks, self.films):
            h = film(blk(h), cond)
        h = self.norm(h)
        pred, micro = self.head(h, static)
        return {"pred": pred, "micro": micro, "h": h}


class PIMambaV2(SequenceModel):
    """PI-Mamba v2: Mamba-2 backbone + *Taylor-form* micro->macro coupling.
    The stress head reads ``[h, W z_micro, W2 phi(z_micro)]`` where ``phi``
    contains ``sqrt(rho)`` and ``rho * twin`` terms.  In physical units the
    Taylor relation is ``sigma = sigma_0 + M alpha G b sqrt(rho)``; here the
    model works in normalised units, so the square-root and the pair product
    give it the correct *functional form* while the affine map absorbs units.
    """

    def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
                 d_state=16, nheads=4, coupling=True, **kw):
        # build the shared heads first (no coupling), then install the
        # Taylor-form coupling and its own stress head
        SequenceModel.__init__(self, n_static, n_out, d_model=d_model,
                               coupling=False, **kw)
        self.inp = nn.Linear(seq_dim, d_model)
        self.blocks = nn.ModuleList([Mamba2Block(d_model, nheads, d_state)
                                     for _ in range(n_layers)])
        self.films = nn.ModuleList([FiLM(d_model, d_model) for _ in range(n_layers)])
        self.norm = RMSNorm(d_model)
        if coupling:
            n_lin = 2 * d_model
            self.micro_proj = nn.Linear(self.n_micro, d_model)
            self.micro_nl = nn.Linear(self.n_rho + 2, d_model)  # sqrt(dens) + rho*twin
            self.stress_head = nn.Sequential(
                nn.Linear(n_lin + d_model, d_model), nn.GELU(), nn.Linear(d_model, 1))
            self.coupling = True
        else:
            self.micro_proj = None
            self.micro_nl = None
            self.stress_head = nn.Sequential(
                nn.Linear(d_model, d_model), nn.GELU(), nn.Linear(d_model, 1))
            self.coupling = False

    def head(self, h, static):
        phase = torch.softmax(self.phase_head(h), dim=-1)
        rho = torch.cat([F.softplus(mi(h)) for mi in self.rho_heads], dim=-1)
        twin = F.softplus(self.twin_head(h))
        micro_p = torch.cat([phase, rho, twin], dim=-1)
        if self.coupling:
            dens = torch.cat([rho, twin], dim=-1)          # (b,T,n_rho+1)
            nl = torch.cat([torch.sqrt(dens + 1e-6), dens[..., :1] * dens[..., -1:]], dim=-1)
            hs = torch.cat([h, self.micro_proj(micro_p), self.micro_nl(nl)], dim=-1)
        else:
            hs = h
        sigma = self.stress_head(hs)
        return torch.cat([sigma, micro_p], dim=-1), micro_p

    def forward(self, static, x):
        cond = self.static_enc(static)
        h = self.inp(x) + cond.unsqueeze(1)
        for blk, film in zip(self.blocks, self.films):
            h = film(blk(h), cond)
        h = self.norm(h)
        pred, micro = self.head(h, static)
        return {"pred": pred, "micro": micro, "h": h}


def build_model(name: str, n_static: int, n_out: int, **kw) -> SequenceModel:
    name = name.lower()
    table = {"mlp": MLPModel, "gru": RNNModel, "lstm": RNNModel,
             "transformer": TransformerModel, "mamba": MambaModel,
             "mamba2": Mamba2Model, "pimamba": PIMamba, "pimamba2": PIMambaV2}
    # next-generation backbones (2024-2026) live in models_next.py to keep the
    # import graph acyclic and this file readable.
    from . import models_next as NX
    table.update({"llama": NX.LLaMAModel, "xlstm": NX.xLSTMModel,
                  "gdn": NX.GatedDeltaModel, "hymba": NX.HybridModel,
                  "koopman": NX.KoopmanModel, "node": NX.NODEModel,
                  "rssm": NX.RSSMModel, "flow": NX.FlowModel})
    if name not in table:
        raise ValueError(f"unknown model {name}")
    if name == "lstm":
        kw["rnn"] = "lstm"
    return table[name](n_static, n_out, **kw)

	"""Sequence models for the HEA MD trajectory task.

	All models share the signature ``forward(static, x) -> dict(pred=(B,T,F))``
	where ``pred`` is in the normalised target space defined by data.Normalizer.

	Implemented:
	* MLP - point-wise baseline (no history)
	* GRU / LSTM - recurrent baselines
	* Transformer- causal encoder baseline
	* Mamba - selective state-space (S6) backbone
	* Mamba2 - SSD-style multi-head gated SSM backbone
	* PIMamba - Mamba + static conditioning + micro->macro coupling
	"""
	from __future__ import annotations

	import math
	from typing import Optional

	import torch
	import torch.nn as nn
	import torch.nn.functional as F


	# ---------------------------------------------------------------------------
	# selective scan (chunked, numerically stable, pure PyTorch)
	# ---------------------------------------------------------------------------
	def selective_scan(u, delta, A, B, C, D=None, chunk: int = 8):
	"""Linear time-invariant-in-A selective scan.

	u : (b, T, d)
	delta : (b, T, d)
	A : (d, n) negative real
	B, C : (b, T, n)
	returns (b, T, d)
	"""
	b, T, d = u.shape
	if A.dim() == 2:
	A = A.unsqueeze(0).unsqueeze(0) # (1,1,d,n)
	elif A.dim() == 3:
	A = A.unsqueeze(1) # (B,1,d,n)
	n = A.shape[-1]
	dt = delta.clamp(1e-4, 0.2)
	dev = u.device
	h = torch.zeros(b, d, n, device=dev, dtype=u.dtype)
	pad = (chunk - T % chunk) % chunk
	if pad:
	u = F.pad(u, (0, 0, 0, pad))
	dt = F.pad(dt, (0, 0, 0, pad))
	B = F.pad(B, (0, 0, 0, pad))
	C = F.pad(C, (0, 0, 0, pad))
	outs = []
	Tp = u.shape[1]
	for s in range(0, Tp, chunk):
	uc = u[:, s:s + chunk]
	dtc = dt[:, s:s + chunk]
	Bc = B[:, s:s + chunk]
	Cc = C[:, s:s + chunk]
	a = torch.exp(dtc.unsqueeze(-1) * A) # (b,c,d,n)
	bu = (dtc * uc).unsqueeze(-1) * Bc.unsqueeze(2) # (b,c,d,n)
	P = torch.cumprod(a, dim=1)
	hloc = P * torch.cumsum(bu / P.clamp_min(1e-20), dim=1)
	hc = hloc + P * h.unsqueeze(1)
	h = hc[:, -1]
	y = (hc * Cc.unsqueeze(2)).sum(-1)
	outs.append(y)
	y = torch.cat(outs, dim=1)[:, :T]
	if D is not None:
	y = y + D.unsqueeze(0).unsqueeze(0) * u[:, :T]
	return y


	class RMSNorm(nn.Module):
	def __init__(self, d, eps=1e-5):
	super().__init__()
	self.w = nn.Parameter(torch.ones(d))
	self.eps = eps

	def forward(self, x):
	return self.w * x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)


	class MambaBlock(nn.Module):
	def __init__(self, d_model, d_state=16, expand=2, d_conv=4, chunk=32):
	super().__init__()
	self.d_inner = int(expand * d_model)
	self.d_state = d_state
	self.chunk = chunk
	self.norm = RMSNorm(d_model)
	self.in_proj = nn.Linear(d_model, 2 * self.d_inner, bias=False)
	self.conv = nn.Conv1d(self.d_inner, self.d_inner, d_conv,
	groups=self.d_inner, padding=d_conv - 1)
	self.x_proj = nn.Linear(self.d_inner, 2 * d_state + 1, bias=False)
	self.dt_proj = nn.Linear(1, self.d_inner, bias=True)
	A = torch.empty(self.d_inner, d_state)
	nn.init.uniform_(A, 0.5, 8.0) # positive magnitude
	self.A_log = nn.Parameter(torch.log(A))
	self.D = nn.Parameter(torch.ones(self.d_inner))
	self.out_proj = nn.Linear(self.d_inner, d_model, bias=False)
	with torch.no_grad():
	self.dt_proj.bias.uniform_(math.log(0.001), math.log(0.05))

	def forward(self, x):
	b, T, _ = x.shape
	h = self.norm(x)
	xz = self.in_proj(h)
	xin, z = xz.chunk(2, dim=-1)
	xin = self.conv(xin.transpose(1, 2))[:, :, :T].transpose(1, 2)
	xin = F.silu(xin)
	proj = self.x_proj(xin)
	dt, Bc, Cc = torch.split(proj, [1, self.d_state, self.d_state], dim=-1)
	dt = F.softplus(self.dt_proj(dt))
	A = -torch.exp(self.A_log)
	y = selective_scan(xin, dt, A, Bc, Cc, self.D, chunk=self.chunk)
	y = y * F.silu(z)
	return x + self.out_proj(y)


	class Mamba2Block(nn.Module):
	"""Simplified SSD-style block: multi-head scalar decay + shared B/C groups."""

	def __init__(self, d_model, nheads=4, d_state=64, expand=1, d_conv=4,
	n_groups=1, chunk=32):
	super().__init__()
	assert d_model % nheads == 0
	self.d_model = d_model
	self.nheads = nheads
	self.headdim = d_model // nheads
	self.d_state = d_state
	self.n_groups = n_groups
	self.chunk = chunk
	self.norm = RMSNorm(d_model)
	self.in_proj = nn.Linear(d_model, 2 * d_model + 2 * d_state * n_groups + nheads,
	bias=False)
	self.conv = nn.Conv1d(d_model, d_model, d_conv, groups=d_model,
	padding=d_conv - 1)
	self.A_log = nn.Parameter(torch.log(torch.empty(nheads).uniform_(0.5, 8.0)))
	self.D = nn.Parameter(torch.ones(nheads))
	self.dt_bias = nn.Parameter(torch.empty(nheads).uniform_(-4.0, -2.0))
	self.norm_g = RMSNorm(d_model)
	self.out_proj = nn.Linear(d_model, d_model, bias=False)

	def forward(self, x):
	b, T, _ = x.shape
	h = self.norm(x)
	nh, hd, ng, ds = self.nheads, self.headdim, self.n_groups, self.d_state
	split = [self.d_model, self.d_model, ng * ds, ng * ds, nh]
	xin, z, Bc, Cc, dt = torch.split(self.in_proj(h), split, dim=-1)
	xin = self.conv(xin.transpose(1, 2))[:, :, :T].transpose(1, 2)
	xin = F.silu(xin).view(b, T, nh, hd)
	dt = F.softplus(dt + self.dt_bias).view(b, T, nh) # (b,T,nh)
	Bc = Bc.view(b, T, ng, ds)
	Cc = Cc.view(b, T, ng, ds)
	if ng != nh:
	rep = nh // ng
	Bc = Bc.repeat_interleave(rep, dim=2)
	Cc = Cc.repeat_interleave(rep, dim=2)
	# fold heads into batch, state dim = 1 -> reuse scalar scan
	xh = xin.permute(0, 2, 1, 3).reshape(b * nh, T, hd)
	dth = dt.permute(0, 2, 1).reshape(b * nh, T, 1).expand(-1, -1, hd)
	Bh = Bc.permute(0, 2, 1, 3).reshape(b * nh, T, ds)
	Ch = Cc.permute(0, 2, 1, 3).reshape(b * nh, T, ds)
	A = ((-torch.exp(self.A_log)).view(1, nh, 1, 1)
	.expand(b, nh, hd, 1).reshape(b * nh, 1, hd, 1))
	y = selective_scan(xh, dth, A, Bh, Ch, None, chunk=self.chunk)
	y = y + xh * self.D.view(1, nh, 1, 1).expand(b, nh, hd, 1).reshape(b * nh, 1, hd)
	y = y.view(b, nh, T, hd).permute(0, 2, 1, 3).reshape(b, T, self.d_model)
	y = self.norm_g(y) * F.silu(z)
	return x + self.out_proj(y)


	class FiLM(nn.Module):
	def __init__(self, cond_dim, d_model):
	super().__init__()
	self.to_scale = nn.Linear(cond_dim, d_model)
	self.to_shift = nn.Linear(cond_dim, d_model)

	def forward(self, h, cond):
	s = 1.0 + self.to_scale(cond).unsqueeze(1)
	b = self.to_shift(cond).unsqueeze(1)
	return h * s + b


	# ---------------------------------------------------------------------------
	# full models
	# ---------------------------------------------------------------------------
	class SequenceModel(nn.Module):
	"""Common wrapper: static encoder + backbone + multi-task heads."""

	def __init__(self, n_static, n_out, d_model=64, n_layers=2, dropout=0.1,
	n_phase=4, n_rho=7, coupling=True, **kwargs):
	super().__init__()
	self.n_phase, self.n_rho = n_phase, n_rho
	self.n_micro = n_phase + n_rho + 1
	self.static_enc = nn.Sequential(
	nn.Linear(n_static, d_model), nn.GELU(),
	nn.Dropout(dropout), nn.Linear(d_model, d_model))
	self.cond_dim = d_model
	self.drop = nn.Dropout(dropout)
	self.coupling = coupling
	# ---- independent output head per physical variable (step 2 要求) ----
	self.phase_head = nn.Linear(d_model, n_phase)
	mk = lambda: nn.Sequential(nn.Linear(d_model, max(8, d_model // 2)),
	nn.GELU(),
	nn.Linear(max(8, d_model // 2), 1))
	self.rho_heads = nn.ModuleList([mk() for _ in range(n_rho)])
	self.twin_head = mk()
	stress_in = 2 * d_model if coupling else d_model
	self.stress_head = nn.Sequential(
	nn.Linear(stress_in, d_model), nn.GELU(), nn.Linear(d_model, 1))
	self.micro_proj = nn.Linear(self.n_micro, d_model) if coupling else None
	assert n_out == 1 + self.n_micro, (n_out, self.n_micro)

	def head(self, h, static):
	phase = torch.softmax(self.phase_head(h), dim=-1)
	rho = torch.cat([F.softplus(mi(h)) for mi in self.rho_heads], dim=-1)
	twin = F.softplus(self.twin_head(h))
	micro_p = torch.cat([phase, rho, twin], dim=-1)
	if self.coupling:
	hs = torch.cat([h, self.micro_proj(micro_p)], dim=-1)
	else:
	hs = h
	sigma = self.stress_head(hs)
	return torch.cat([sigma, micro_p], dim=-1), micro_p


	class MLPModel(SequenceModel):
	def __init__(self, n_static, n_out, seq_dim=5, d_model=128, **kw):
	super().__init__(n_static, n_out, d_model=d_model, **kw)
	self.inp = nn.Sequential(nn.Linear(seq_dim + n_static, d_model), nn.GELU(),
	nn.Linear(d_model, d_model), nn.GELU())

	def forward(self, static, x):
	b, T, _ = x.shape
	z = torch.cat([x, static.unsqueeze(1).expand(-1, T, -1)], dim=-1)
	h = self.inp(z)
	pred, micro = self.head(h, static)
	return {"pred": pred, "micro": micro, "h": h}


	class RNNModel(SequenceModel):
	def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
	rnn="gru", **kw):
	super().__init__(n_static, n_out, d_model=d_model, **kw)
	self.inp = nn.Linear(seq_dim, d_model)
	cls = nn.GRU if rnn == "gru" else nn.LSTM
	self.rnn = cls(d_model, d_model, num_layers=n_layers, batch_first=True,
	dropout=0.1 if n_layers > 1 else 0.0)
	self.film = FiLM(d_model, d_model)

	def forward(self, static, x):
	cond = self.static_enc(static)
	h = self.inp(x)
	h, _ = self.rnn(h)
	h = self.film(h, cond)
	pred, micro = self.head(h, static)
	return {"pred": pred, "micro": micro, "h": h}


	class TransformerModel(SequenceModel):
	def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
	n_heads=4, **kw):
	super().__init__(n_static, n_out, d_model=d_model, **kw)
	self.inp = nn.Linear(seq_dim, d_model)
	self.pos = nn.Parameter(torch.randn(1, 512, d_model) * 0.02)
	layer = nn.TransformerEncoderLayer(d_model, n_heads, 4 * d_model,
	dropout=0.1, batch_first=True,
	norm_first=True, activation="gelu")
	self.enc = nn.TransformerEncoder(layer, n_layers, enable_nested_tensor=False)
	self.film = FiLM(d_model, d_model)
	self.register_buffer("mask", torch.triu(torch.ones(512, 512), diagonal=1).bool())

	def forward(self, static, x):
	b, T, _ = x.shape
	cond = self.static_enc(static)
	h = self.inp(x) + self.pos[:, :T]
	h = self.enc(h, mask=self.mask[:T, :T])
	h = self.film(h, cond)
	pred, micro = self.head(h, static)
	return {"pred": pred, "micro": micro, "h": h}


	class MambaModel(SequenceModel):
	def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
	d_state=16, expand=2, **kw):
	super().__init__(n_static, n_out, d_model=d_model, **kw)
	self.inp = nn.Linear(seq_dim, d_model)
	self.blocks = nn.ModuleList([MambaBlock(d_model, d_state, expand)
	for _ in range(n_layers)])
	self.norm = RMSNorm(d_model)
	self.cond_proj = nn.Linear(d_model, d_model)

	def forward(self, static, x):
	cond = self.static_enc(static)
	h = self.inp(x) + cond.unsqueeze(1)
	for blk in self.blocks:
	h = blk(h)
	h = self.norm(h)
	pred, micro = self.head(h, static)
	return {"pred": pred, "micro": micro, "h": h}


	class Mamba2Model(SequenceModel):
	def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
	nheads=4, d_state=64, **kw):
	super().__init__(n_static, n_out, d_model=d_model, **kw)
	self.inp = nn.Linear(seq_dim, d_model)
	self.blocks = nn.ModuleList([Mamba2Block(d_model, nheads=nheads,
	d_state=d_state)
	for _ in range(n_layers)])
	self.norm = RMSNorm(d_model)

	def forward(self, static, x):
	cond = self.static_enc(static)
	h = self.inp(x) + cond.unsqueeze(1)
	for blk in self.blocks:
	h = blk(h)
	h = self.norm(h)
	pred, micro = self.head(h, static)
	return {"pred": pred, "micro": micro, "h": h}


	class PIMamba(SequenceModel):
	"""Physics-informed Mamba: FiLM conditioning + micro -> macro coupling."""

	def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
	d_state=16, expand=2, coupling=True, **kw):
	super().__init__(n_static, n_out, d_model=d_model, coupling=coupling, **kw)
	self.inp = nn.Linear(seq_dim, d_model)
	self.blocks = nn.ModuleList([MambaBlock(d_model, d_state, expand)
	for _ in range(n_layers)])
	self.films = nn.ModuleList([FiLM(d_model, d_model) for _ in range(n_layers)])
	self.norm = RMSNorm(d_model)

	def forward(self, static, x):
	cond = self.static_enc(static)
	h = self.inp(x) + cond.unsqueeze(1)
	for blk, film in zip(self.blocks, self.films):
	h = film(blk(h), cond)
	h = self.norm(h)
	pred, micro = self.head(h, static)
	return {"pred": pred, "micro": micro, "h": h}


	class PIMambaV2(SequenceModel):
	"""PI-Mamba v2: Mamba-2 backbone + Taylor-form micro->macro coupling.

	The stress head reads ``[h, W z_micro, W2 phi(z_micro)]`` where ``phi``
	contains ``sqrt(rho)`` and ``rho * twin`` terms. In physical units the
	Taylor relation is ``sigma = sigma_0 + M alpha G b sqrt(rho)``; here the
	model works in normalised units, so the square-root and the pair product
	give it the correct functional form while the affine map absorbs units.
	"""

	def __init__(self, n_static, n_out, seq_dim=5, d_model=64, n_layers=2,
	d_state=16, nheads=4, coupling=True, **kw):
	# build the shared heads first (no coupling), then install the
	# Taylor-form coupling and its own stress head
	SequenceModel.__init__(self, n_static, n_out, d_model=d_model,
	coupling=False, **kw)
	self.inp = nn.Linear(seq_dim, d_model)
	self.blocks = nn.ModuleList([Mamba2Block(d_model, nheads, d_state)
	for _ in range(n_layers)])
	self.films = nn.ModuleList([FiLM(d_model, d_model) for _ in range(n_layers)])
	self.norm = RMSNorm(d_model)
	if coupling:
	n_lin = 2 * d_model
	self.micro_proj = nn.Linear(self.n_micro, d_model)
	self.micro_nl = nn.Linear(self.n_rho + 2, d_model) # sqrt(dens) + rho*twin
	self.stress_head = nn.Sequential(
	nn.Linear(n_lin + d_model, d_model), nn.GELU(), nn.Linear(d_model, 1))
	self.coupling = True
	else:
	self.micro_proj = None
	self.micro_nl = None
	self.stress_head = nn.Sequential(
	nn.Linear(d_model, d_model), nn.GELU(), nn.Linear(d_model, 1))
	self.coupling = False

	def head(self, h, static):
	phase = torch.softmax(self.phase_head(h), dim=-1)
	rho = torch.cat([F.softplus(mi(h)) for mi in self.rho_heads], dim=-1)
	twin = F.softplus(self.twin_head(h))
	micro_p = torch.cat([phase, rho, twin], dim=-1)
	if self.coupling:
	dens = torch.cat([rho, twin], dim=-1) # (b,T,n_rho+1)
	nl = torch.cat([torch.sqrt(dens + 1e-6), dens[..., :1] * dens[..., -1:]], dim=-1)
	hs = torch.cat([h, self.micro_proj(micro_p), self.micro_nl(nl)], dim=-1)
	else:
	hs = h
	sigma = self.stress_head(hs)
	return torch.cat([sigma, micro_p], dim=-1), micro_p

	def forward(self, static, x):
	cond = self.static_enc(static)
	h = self.inp(x) + cond.unsqueeze(1)
	for blk, film in zip(self.blocks, self.films):
	h = film(blk(h), cond)
	h = self.norm(h)
	pred, micro = self.head(h, static)
	return {"pred": pred, "micro": micro, "h": h}


	def build_model(name: str, n_static: int, n_out: int, **kw) -> SequenceModel:
	name = name.lower()
	table = {"mlp": MLPModel, "gru": RNNModel, "lstm": RNNModel,
	"transformer": TransformerModel, "mamba": MambaModel,
	"mamba2": Mamba2Model, "pimamba": PIMamba, "pimamba2": PIMambaV2}
	# next-generation backbones (2024-2026) live in models_next.py to keep the
	# import graph acyclic and this file readable.
	from . import models_next as NX
	table.update({"llama": NX.LLaMAModel, "xlstm": NX.xLSTMModel,
	"gdn": NX.GatedDeltaModel, "hymba": NX.HybridModel,
	"koopman": NX.KoopmanModel, "node": NX.NODEModel,
	"rssm": NX.RSSMModel, "flow": NX.FlowModel})
	if name not in table:
	raise ValueError(f"unknown model {name}")
	if name == "lstm":
	kw["rnn"] = "lstm"
	return table[name](n_static, n_out, **kw)