modeling_pillars.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from transformers import PreTrainedModel
try:
    from .configuration_pillars import PillarsConfig
except ImportError:
    from configuration_pillars import PillarsConfig

try:
    from x_transformers import Encoder
except ImportError:
    raise ImportError("To use PILLARS-DAT, you must run: pip install x-transformers")

# --- UTILS ---

class ComplexDropout(nn.Module):
    def __init__(self, p=0.5):
        super().__init__()
        self.p = p
    def forward(self, z):
        if not self.training or self.p == 0.0: return z
        mask = torch.ones_like(z.real)
        mask = F.dropout(mask, self.p, self.training, inplace=False)
        return z * mask

class RobustPhaseNorm(nn.Module):
    def __init__(self, d_model, eps=1e-5):
        super().__init__()
        self.scale = nn.Parameter(torch.ones(d_model))
        self.eps = eps
    def forward(self, x):
        mag = torch.abs(x)
        rms = torch.sqrt(torch.mean(mag**2, dim=-1, keepdim=True) + self.eps)
        return (x / rms) * self.scale

class ModReLU(nn.Module):
    def __init__(self, features):
        super().__init__()
        self.b = nn.Parameter(torch.zeros(features))

    def forward(self, z):
        # 1. FORCE FLOAT32 FOR GEOMETRY
        # We must calculate magnitude in high precision to prevent
        # square-law overflow (Re^2 + Im^2) from killing the gradients.
        z_32 = z.to(torch.complex64)

        # 2. Calculate Magnitude (Safe)
        mag = torch.abs(z_32)

        # 3. Activation Logic (Still FP32)
        new_mag = F.relu(mag + self.b.float())

        # 4. Reconstruct Phase (Safe Division)
        # (z / mag) is the unit vector (phase)
        phase = z_32 / (mag + 1e-6)

        # 5. Result
        out = new_mag * phase

        # 6. Cast back to network dtype (BF16/FP16)
        return out.to(z.dtype)

class ComplexToRealBridge(nn.Module):
    def __init__(self, d_model):
        super().__init__()
        self.proj = nn.Linear(d_model * 2, d_model)
        self.norm = nn.LayerNorm(d_model)
    def forward(self, x_complex):
        cat = torch.cat([x_complex.real, x_complex.imag], dim=-1)
        return self.norm(self.proj(cat))

# ==========================================
# 4. DYNAMIC RoSE (Mamba-3 Engine)
# ==========================================
class DynamicRoSE(nn.Module):
    def __init__(self, num_embeddings, embedding_dim, max_period=10000.0):
        super().__init__()
        self.embedding_dim = embedding_dim

        # 1. Master Real Embedding (The "Particle")
        self.raw_embedding = nn.Embedding(num_embeddings, embedding_dim)

        # 2. Complex Adapter (The "Wave" Magnitude/Initial Phase)
        self.adapter = nn.Linear(embedding_dim, embedding_dim * 2)

        # 3. Static Frequencies (Positional)
        freqs = torch.exp(torch.arange(0, embedding_dim, dtype=torch.float32) * -(math.log(max_period) / embedding_dim))
        self.register_buffer('freqs', freqs)

        self.rotation_predictor = nn.Linear(embedding_dim, embedding_dim * 2)

    def forward(self, input_ids):
        # A. Raw Particle
        real_base = self.raw_embedding(input_ids)
        B, L, D = real_base.shape

        # B. Complex Wave Content
        complex_params = self.adapter(real_base)
        z_t = torch.complex(complex_params[..., :D], complex_params[..., D:])

        rot_raw = self.rotation_predictor(real_base)
        rot_x, rot_y = rot_raw.chunk(2, dim=-1)

        rot_mag = torch.sqrt(rot_x**2 + rot_y**2 + 1e-6)
        dynamic_rot = torch.complex(rot_x / rot_mag, rot_y / rot_mag)

        # D. Static Positional Rotation
        pos = torch.arange(L, device=input_ids.device).float()
        static_angles = torch.outer(pos, self.freqs) # [L, D]
        static_rot = torch.polar(torch.ones_like(static_angles), static_angles) # [L, D]

        z_final = z_t * static_rot.unsqueeze(0) * dynamic_rot

        return z_final, real_base

# ==========================================
# 5. HYENA FILTER
# ==========================================
class HyenaNeuralFilter(nn.Module):
    def __init__(self, d_model, max_len=1024, hidden_dim=64):
        super().__init__()
        self.d_model = d_model
        freqs = torch.exp(torch.arange(0, hidden_dim, 2, dtype=torch.float32) * -(math.log(10000.0) / hidden_dim))
        self.register_buffer("freqs", freqs)
        self.mlp = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim), nn.SiLU(),
            nn.Linear(hidden_dim, hidden_dim), nn.SiLU(),
            nn.Linear(hidden_dim, d_model * 2)
        )
    def forward(self, L, device):
        t = torch.linspace(0, 1, steps=L, device=device).unsqueeze(-1)
        emb = torch.cat([torch.sin(t * self.freqs), torch.cos(t * self.freqs)], dim=-1)
        out = self.mlp(emb).view(L, self.d_model, 2)
        return torch.complex(out[..., 0], out[..., 1])

# ==========================================
# 6. GATED HARMONIC CONVOLUTION (Lean)
# ==========================================
# @title 🛠️ Fixed PRISM Layer (Precision-Gated)

# @title 🛠️ Fixed PRISM Layer (Type-Safe)

class GatedHarmonicConvolution(nn.Module):
    def __init__(self, d_model, max_len=1024, dropout=0.1):
        super().__init__()
        self.d_model = d_model
        self.filter_len = max_len
        self.neural_filter = HyenaNeuralFilter(d_model, max_len=max_len)
        self.gate_proj = nn.Linear(d_model * 2, d_model * 2)

        self.mix_real = nn.Linear(d_model, d_model)
        self.mix_imag = nn.Linear(d_model, d_model)
        self.out_real = nn.Linear(d_model, d_model)
        self.out_imag = nn.Linear(d_model, d_model)

        self.activation = ModReLU(d_model)
        self.norm = RobustPhaseNorm(d_model)
        self.dropout = ComplexDropout(dropout)

    def forward(self, x, src_mask=None):
        residual = x
        x_norm = self.norm(x)
        if src_mask is not None:
             x_norm = x_norm.masked_fill(src_mask.unsqueeze(-1), 0.0)

        # 🛑 PRECISION GATE 🛑
        # Force operations to Float32 Complex to preserve Phase Physics
        with torch.amp.autocast('cuda', enabled=False):

            # --- THE FIX IS HERE ---
            # Old: x_32 = x_norm.float()  <-- This stripped the imaginary part
            # New: Explicit cast to Complex64
            x_32 = x_norm.to(torch.complex64)
            # -----------------------

            B, L, D = x_32.shape
            eff_L = min(L, self.filter_len)

            # 1. FFT (Now safe because x_32 is definitely complex)
            x_freq = torch.fft.fft(x_32, n=eff_L, dim=1, norm='ortho')

            # 2. Filter (Ensure filter is also complex64)
            h = self.neural_filter(eff_L, x.device).unsqueeze(0).to(torch.complex64)
            x_filtered = x_freq * h

            # 3. IFFT
            x_time = torch.fft.ifft(x_filtered, n=eff_L, dim=1, norm='ortho')

            if L > eff_L: x_time = F.pad(x_time, (0,0,0,L-eff_L))
            else: x_time = x_time[:, :L, :]

            # 4. Gating (Sigmoid logic)
            # Safe concatenation because x_32 is complex64
            x_cat = torch.cat([x_32.real, x_32.imag], dim=-1)

            # Cast weights to Float32 for the calculation
            gate_w = self.gate_proj.weight.to(torch.float32)
            gate_b = self.gate_proj.bias.to(torch.float32)

            gate_out = F.linear(x_cat, gate_w, gate_b)
            gates = torch.sigmoid(gate_out)

            g_r, g_i = gates.chunk(2, dim=-1)
            x_gated_32 = torch.complex(x_time.real * g_r, x_time.imag * g_i)

            # 🏁 EXIT GATE: Cast back to original dtype (likely BFloat16 from autocast)
            # We cast real/imag separately to be safe
            target_dtype = x.dtype
            # If x was complex, target is complex. If x was real, we have an issue.
            # Assuming x comes from autocast, it might be complex16.

            x_gated = x_gated_32.to(target_dtype)

        # 5. Mixing (Back in mixed precision)
        mr, mi = self.mix_real, self.mix_imag
        x_mixed = torch.complex(mr(x_gated.real) - mi(x_gated.imag), mr(x_gated.imag) + mi(x_gated.real))

        x_act = self.activation(x_mixed)

        or_, oi = self.out_real, self.out_imag
        out = torch.complex(or_(x_act.real) - oi(x_act.imag), or_(x_act.imag) + oi(x_act.real))

        return self.dropout(out) + residual
# ==========================================
# 7. MODEL WRAPPERS
# ==========================================
class PRISMEncoder(nn.Module):
    def __init__(self, num_layers, d_model, max_len, dropout=0.1):
        super().__init__()
        self.layers = nn.ModuleList([
            GatedHarmonicConvolution(d_model, max_len, dropout)
            for _ in range(num_layers)
        ])
        self.final_norm = RobustPhaseNorm(d_model)
    def forward(self, x, src_mask=None):
        for layer in self.layers:
            if self.training: x = torch.utils.checkpoint.checkpoint(layer, x, src_mask, use_reentrant=False)
            else: x = layer(x, src_mask)
        return self.final_norm(x)

class PRISM_WikiText_Model(nn.Module):
    def __init__(self, vocab_size, d_model, max_len, prism_depth=5, trans_depth=1, dropout=0.1):
        super().__init__()
        self.d_model = d_model

        # 1. PRISM Core (The Optical/Passive Part)
        self.rose = DynamicRoSE(vocab_size, d_model)
        self.prism_encoder = PRISMEncoder(prism_depth, d_model, max_len=max_len, dropout=dropout)
        self.bridge = ComplexToRealBridge(d_model)
        self.periscope_proj = nn.Sequential(nn.Linear(d_model * 2, d_model), nn.LayerNorm(d_model), nn.GELU())

        # 2. Refiner (The Digital/Active Part)
        # 🔄 SWAPPED: Replaced Standard Transformer with RoPE-Enabled Encoder
        if trans_depth > 0:
            self.refiner = Encoder(
                dim=d_model,
                depth=trans_depth,
                heads=8,
                rotary_pos_emb=True,
                attn_flash=True,
                attn_dropout=dropout,
                ff_dropout=dropout,

            )
        else:
            self.refiner = None

        # 3. Output
        self.lm_head = nn.Linear(d_model, vocab_size)
        self.lm_head.weight = self.rose.raw_embedding.weight

    def forward(self, input_ids):
        # A. Wave Physics
        wave_src, particle_src = self.rose(input_ids)
        wave_out = self.prism_encoder(wave_src)
        wave_real = self.bridge(wave_out)

        # B. Interface
        mixed_memory = self.periscope_proj(torch.cat([wave_real, particle_src], dim=-1))

        # C. Digital Refinement (Now with RoPE)
        if self.refiner:
            out = self.refiner(mixed_memory)
        else:
            out = mixed_memory

        return self.lm_head(out)

# ==========================================
# 1. SENSORY STREAM (Transformer + RoPE)
# ==========================================
class SensoryStream(nn.Module):
    def __init__(self, depth, d_model, dropout=0.1):
        super().__init__()
        self.encoder = Encoder(
            dim=d_model,
            depth=depth,
            heads=4,                # 256 dim / 64 head_dim = 4 heads
            attn_flash=True,        # Flash Attention
            rotary_pos_emb=True,    # <--- CRITICAL: RoPE Enabled
            attn_dropout=dropout,
            ff_dropout=dropout,
            use_rmsnorm=True,       # RMSNorm (Llama style)
            ff_glu=True             # SwiGLU (Llama style)
        )

    def forward(self, x):
        return self.encoder(x)

class Pillars_DAT_Model(PreTrainedModel):
    config_class = PillarsConfig
    
    def __init__(self, config):
        super().__init__(config)
        self.config = config
        self.d_model = config.d_model
        self.d_branch = config.d_branch
        
        # 1. Root
        self.rose = DynamicRoSE(config.vocab_size, config.d_model)
        
        # 2. Downsample
        self.particle_down = nn.Linear(config.d_model, config.d_branch)
        self.wave_down = nn.Linear(config.d_model * 2, config.d_branch * 2)
        
        # 3. Stream A: Sensory (Rate)
        self.stream_sensory = SensoryStream(depth=config.depth, d_model=config.d_branch, dropout=config.dropout)
        
        # 4. Stream B: Relational (Phase)
        self.stream_relational = PRISMEncoder(num_layers=config.depth, d_model=config.d_branch, max_len=config.seq_len, dropout=config.dropout)
        self.relational_bridge = ComplexToRealBridge(config.d_branch)
        
        # 5. Fusion
        self.fusion_proj = nn.Linear(config.d_branch * 2, config.d_model)
        self.fusion_norm = nn.LayerNorm(config.d_model)
        
        # 6. Refiner
        self.refiner = Encoder(
            dim=config.d_model, depth=1, heads=8, attn_flash=True,
            rotary_pos_emb=True, attn_dropout=config.dropout, ff_dropout=config.dropout
        )
        
        # 7. Output (Converted to Layer for HF Compatibility)
        self.lm_head = nn.Linear(config.d_model, config.vocab_size)
        
        # Tie Weights Explicitly
        self.lm_head.weight = self.rose.raw_embedding.weight

    def forward(self, input_ids, labels=None):
        # 1. Physics
        wave_src, particle_src = self.rose(input_ids)
        p_small = self.particle_down(particle_src)
        
        w_flat = torch.cat([wave_src.real, wave_src.imag], dim=-1)
        w_small_flat = self.wave_down(w_flat)
        w_small = torch.complex(w_small_flat[..., :self.d_branch], w_small_flat[..., self.d_branch:])
        
        # 2. Parallel Streams
        sensory_out = self.stream_sensory(p_small)
        relational_out_complex = self.stream_relational(w_small)
        relational_out = self.relational_bridge(relational_out_complex)
        
        # 3. Fusion
        stacked = torch.cat([sensory_out, relational_out], dim=-1)
        context = self.fusion_norm(self.fusion_proj(stacked))
        
        # 4. Refinement
        refined = self.refiner(context)
        
        # 5. Output
        logits = self.lm_head(refined)
        
        loss = None
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
            return {"loss": loss, "logits": logits}
            
        return logits