Upload 44 files

.gitattributes

@@ -372,3 +372,4 @@ shrek-large/step_6510 filter=lfs diff=lfs merge=lfs -text
 shrek-large/step_7812 filter=lfs diff=lfs merge=lfs -text
 shrek-large/step_9114 filter=lfs diff=lfs merge=lfs -text
 shrek-tiny/step_37758_all_preds.0 filter=lfs diff=lfs merge=lfs -text
+sudoku-extreme/shrek-tiny/step_37758_all_preds.0 filter=lfs diff=lfs merge=lfs -text

sudoku-extreme/shrek-tiny/all_config.yaml

@@ -0,0 +1,38 @@
+arch:
+  H_cycles: 2
+  H_layers: 2
+  L_cycles: 2
+  L_layers: 2
+  enable_error_injection: true
+  enable_stagnation_delta: true
+  expansion: 4
+  halt_exploration_prob: 0.1
+  halt_max_steps: 16
+  hidden_size: 512
+  loss:
+    loss_type: stablemax_cross_entropy
+    name: losses@ACTLossHead
+  name: hrm.hrm_act_v1@HierarchicalReasoningModel_ACTV1
+  num_heads: 8
+  pos_encodings: rope
+  puzzle_emb_ndim: 512
+beta1: 0.9
+beta2: 0.95
+checkpoint_every_eval: true
+checkpoint_path: checkpoints/SHREK_Ablation_Sudoku/SHREK_Tiny_Vanilla_Sudoku
+data_path: ../../dataset/data/sudoku-extreme-1k-aug-1000
+ema: true
+ema_rate: 0.999
+epochs: 40000
+eval_interval: 1000
+eval_save_outputs: []
+global_batch_size: 768
+lr: 0.0001
+lr_min_ratio: 1.0
+lr_warmup_steps: 2000
+project_name: SHREK_Ablation_Sudoku
+puzzle_emb_lr: 0.0001
+puzzle_emb_weight_decay: 1.0
+run_name: SHREK_Tiny_Vanilla_Sudoku
+seed: 0
+weight_decay: 1.0

sudoku-extreme/shrek-tiny/hrm_act_v1.py

@@ -0,0 +1,428 @@
+from typing import Tuple, List, Dict, Optional
+from dataclasses import dataclass
+import math
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from pydantic import BaseModel
+
+from models.common import trunc_normal_init_
+from models.layers import rms_norm, SwiGLU, Attention, RotaryEmbedding, CosSin, CastedEmbedding, CastedLinear
+from models.sparse_embedding import CastedSparseEmbedding
+from models.hrm.error_singals import get_error_signal  # SHREK: error signal module
+
+
+@dataclass
+class HierarchicalReasoningModel_ACTV1InnerCarry:
+    z_H: torch.Tensor
+    z_L: torch.Tensor
+    # SHREK: prev_pred stores last step's argmax predictions for flip rate computation.
+    # zeros = fresh start (first step after init or reset gives flip_rate ≈ 1.0)
+    prev_pred: torch.Tensor        # (B, seq_len) int32
+    # SHREK: Q-values cached in carry — no longer used for Q-targets (see Bug 4),
+    # but kept because removing them causes torch.compile regression.
+    prev_q_halt: torch.Tensor      # (B,) float32
+    prev_q_continue: torch.Tensor  # (B,) float32
+
+
+@dataclass
+class HierarchicalReasoningModel_ACTV1Carry:
+    inner_carry: HierarchicalReasoningModel_ACTV1InnerCarry
+    
+    steps: torch.Tensor
+    halted: torch.Tensor
+    
+    current_data: Dict[str, torch.Tensor]
+
+
+class HierarchicalReasoningModel_ACTV1Config(BaseModel):
+    batch_size: int
+    seq_len: int
+    puzzle_emb_ndim: int = 0
+    num_puzzle_identifiers: int
+    vocab_size: int
+
+    H_cycles: int
+    L_cycles: int
+
+    H_layers: int
+    L_layers: int
+
+    # Transformer config
+    hidden_size: int
+    expansion: float
+    num_heads: int
+    pos_encodings: str
+
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 10000.0
+    
+    # Halting Q-learning config
+    halt_max_steps: int
+    halt_exploration_prob: float
+
+    # SHREK: error injection warmup — ramps alpha from 0 to alpha_max over warmup steps.
+    # Prevents small models from collapsing before the error estimator is accurate.
+    alpha_max: float = 0.01
+    alpha_warmup_steps: int = 5000
+
+    # SHREK ablation flags — disable components from training scripts
+    enable_error_injection: bool = True    # Error-Conditioned Injection (flip rate + learned estimator)
+    enable_stagnation_delta: bool = True   # Stagnation-Aware Q-head (delta fed to Q-head)
+
+    forward_dtype: str = "bfloat16"
+
+
+class HierarchicalReasoningModel_ACTV1Block(nn.Module):
+    def __init__(self, config: HierarchicalReasoningModel_ACTV1Config) -> None:
+        super().__init__()
+
+        self.self_attn = Attention(
+            hidden_size=config.hidden_size,
+            head_dim=config.hidden_size // config.num_heads,
+            num_heads=config.num_heads,
+            num_key_value_heads=config.num_heads,
+            causal=False
+        )
+        self.mlp = SwiGLU(
+            hidden_size=config.hidden_size,
+            expansion=config.expansion,
+        )
+        self.norm_eps = config.rms_norm_eps
+
+    def forward(self, cos_sin: CosSin, hidden_states: torch.Tensor) -> torch.Tensor:
+        # Post Norm
+        # Self Attention
+        hidden_states = rms_norm(hidden_states + self.self_attn(cos_sin=cos_sin, hidden_states=hidden_states), variance_epsilon=self.norm_eps)
+        # Fully Connected
+        hidden_states = rms_norm(hidden_states + self.mlp(hidden_states), variance_epsilon=self.norm_eps)
+        return hidden_states
+
+
+class HierarchicalReasoningModel_ACTV1ReasoningModule(nn.Module):
+    def __init__(self, layers: List[HierarchicalReasoningModel_ACTV1Block]):
+        super().__init__()
+
+        self.layers = torch.nn.ModuleList(layers)
+
+    def forward(self, hidden_states: torch.Tensor, input_injection: torch.Tensor, **kwargs) -> torch.Tensor:
+        # Input injection (add)
+        hidden_states = hidden_states + input_injection
+        # Layers
+        for layer in self.layers:
+            hidden_states = layer(hidden_states=hidden_states, **kwargs)
+
+        return hidden_states
+
+
+class HierarchicalReasoningModel_ACTV1_Inner(nn.Module):
+    def __init__(self, config: HierarchicalReasoningModel_ACTV1Config) -> None:
+        super().__init__()
+        self.config = config
+        self.forward_dtype = getattr(torch, self.config.forward_dtype)
+
+        # I/O
+        self.embed_scale  = math.sqrt(self.config.hidden_size)
+        embed_init_std = 1.0 / self.embed_scale
+
+        self.embed_tokens = CastedEmbedding(self.config.vocab_size, self.config.hidden_size, init_std=embed_init_std, cast_to=self.forward_dtype)
+        self.lm_head      = CastedLinear(self.config.hidden_size, self.config.vocab_size, bias=False)
+        # SHREK: Q-head input size depends on stagnation delta flag
+        q_head_input_size = self.config.hidden_size + (1 if self.config.enable_stagnation_delta else 0)
+        self.q_head       = CastedLinear(q_head_input_size, 2, bias=True)
+
+        # SHREK: error_encoder maps the scalar error score -> hidden_size vector for injection into z_H
+        # alpha follows a linear warmup schedule (0 → alpha_max over warmup steps).
+        # This lets the error estimator train before its signal affects z_H.
+        self.error_encoder  = nn.Linear(1, self.config.hidden_size)
+        # SHREK: step counter for alpha warmup (not a learned parameter)
+        self.register_buffer('_alpha_step', torch.tensor(0, dtype=torch.long))
+        # SHREK: error_estimator reads z_H and predicts how wrong the model is.
+        # trained via auxiliary loss in pretrain.py using the real lm_loss as target.
+        # catches "stuck but wrong" — a model confidently on the wrong answer.
+        # flip rate catches oscillation; estimator catches confident-but-wrong.
+        self.error_estimator = nn.Linear(self.config.hidden_size, 1)
+
+        self.puzzle_emb_len = -(self.config.puzzle_emb_ndim // -self.config.hidden_size)  # ceil div
+        if self.config.puzzle_emb_ndim > 0:
+            # Zero init puzzle embeddings
+            self.puzzle_emb = CastedSparseEmbedding(self.config.num_puzzle_identifiers, self.config.puzzle_emb_ndim,
+                                                    batch_size=self.config.batch_size, init_std=0, cast_to=self.forward_dtype)
+
+        # LM Blocks
+        if self.config.pos_encodings == "rope":
+            self.rotary_emb = RotaryEmbedding(dim=self.config.hidden_size // self.config.num_heads,
+                                              max_position_embeddings=self.config.seq_len + self.puzzle_emb_len,
+                                              base=self.config.rope_theta)
+        elif self.config.pos_encodings == "learned":
+            self.embed_pos = CastedEmbedding(self.config.seq_len + self.puzzle_emb_len, self.config.hidden_size, init_std=embed_init_std, cast_to=self.forward_dtype)
+        else:
+            raise NotImplementedError()
+
+        # Reasoning Layers
+        self.H_level = HierarchicalReasoningModel_ACTV1ReasoningModule(layers=[HierarchicalReasoningModel_ACTV1Block(self.config) for _i in range(self.config.H_layers)])
+        self.L_level = HierarchicalReasoningModel_ACTV1ReasoningModule(layers=[HierarchicalReasoningModel_ACTV1Block(self.config) for _i in range(self.config.L_layers)])
+        
+        # Initial states
+        self.H_init = nn.Buffer(trunc_normal_init_(torch.empty(self.config.hidden_size, dtype=self.forward_dtype), std=1), persistent=True)
+        self.L_init = nn.Buffer(trunc_normal_init_(torch.empty(self.config.hidden_size, dtype=self.forward_dtype), std=1), persistent=True)
+
+        # Q head special init
+        # Init Q to (almost) zero for faster learning during bootstrapping
+        with torch.no_grad():
+            self.q_head.weight.zero_()
+            self.q_head.bias.fill_(-5)  # type: ignore
+
+    def _input_embeddings(self, input: torch.Tensor, puzzle_identifiers: torch.Tensor):
+        # Token embedding
+        embedding = self.embed_tokens(input.to(torch.int32))
+
+        # Puzzle embeddings
+        if self.config.puzzle_emb_ndim > 0:
+            puzzle_embedding = self.puzzle_emb(puzzle_identifiers)
+            
+            pad_count = self.puzzle_emb_len * self.config.hidden_size - puzzle_embedding.shape[-1]
+            if pad_count > 0:
+                puzzle_embedding = F.pad(puzzle_embedding, (0, pad_count))
+
+            embedding = torch.cat((puzzle_embedding.view(-1, self.puzzle_emb_len, self.config.hidden_size), embedding), dim=-2)
+
+        # Position embeddings
+        if self.config.pos_encodings == "learned":
+            # scale by 1/sqrt(2) to maintain forward variance
+            embedding = 0.707106781 * (embedding + self.embed_pos.embedding_weight.to(self.forward_dtype))
+
+        # Scale
+        return self.embed_scale * embedding
+
+    def empty_carry(self, batch_size: int):
+        return HierarchicalReasoningModel_ACTV1InnerCarry(
+            z_H=torch.empty(batch_size, self.config.seq_len + self.puzzle_emb_len, self.config.hidden_size, dtype=self.forward_dtype),
+            z_L=torch.empty(batch_size, self.config.seq_len + self.puzzle_emb_len, self.config.hidden_size, dtype=self.forward_dtype),
+            # SHREK: zeros = no previous prediction — first step gives flip_rate ≈ 1.0
+            # device=H_init.device ensures prev_pred is on CUDA, matching z_H and logits
+            prev_pred=torch.zeros(batch_size, self.config.seq_len, dtype=torch.int32, device=self.H_init.device),
+            prev_q_halt=torch.full((batch_size,), -5.0, device=self.H_init.device),
+            prev_q_continue=torch.full((batch_size,), -5.0, device=self.H_init.device),
+        )
+        
+    def reset_carry(self, reset_flag: torch.Tensor, carry: HierarchicalReasoningModel_ACTV1InnerCarry):
+        # SHREK: zero out prev_pred for reset sequences so they start fresh.
+        # a reset sequence is one that just halted — it will solve a new puzzle next.
+        # zeroing prev_pred means first step gives flip_rate ≈ 1.0 (maximum uncertainty).
+        new_prev_pred = torch.where(
+            reset_flag.view(-1, 1),
+            torch.zeros_like(carry.prev_pred),
+            carry.prev_pred
+        )
+        new_prev_q_halt     = torch.where(reset_flag, torch.full_like(carry.prev_q_halt,     -5.0), carry.prev_q_halt)
+        new_prev_q_continue = torch.where(reset_flag, torch.full_like(carry.prev_q_continue, -5.0), carry.prev_q_continue)
+
+        return HierarchicalReasoningModel_ACTV1InnerCarry(
+            z_H=torch.where(reset_flag.view(-1, 1, 1), self.H_init, carry.z_H),
+            z_L=torch.where(reset_flag.view(-1, 1, 1), self.L_init, carry.z_L),
+            prev_pred=new_prev_pred,
+            prev_q_halt=new_prev_q_halt,
+            prev_q_continue=new_prev_q_continue,
+        )
+
+
+    # SHREK: removed task_type parameter — error signal is now universal (no task rules needed)
+    def forward(self, carry: HierarchicalReasoningModel_ACTV1InnerCarry, batch: Dict[str, torch.Tensor], require_trace=False):
+    #  -> Tuple[HierarchicalReasoningModel_ACTV1InnerCarry, torch.Tensor, Tuple[torch.Tensor, torch.Tensor], torch.Tensor]:
+        seq_info = dict(
+            cos_sin=self.rotary_emb() if hasattr(self, "rotary_emb") else None,
+        )
+
+        # Input encoding
+        input_embeddings = self._input_embeddings(batch["inputs"], batch["puzzle_identifiers"])
+
+        z_H_trace = []
+
+        # Forward iterations
+        with torch.no_grad():
+            z_H, z_L = carry.z_H, carry.z_L
+
+            for _H_step in range(self.config.H_cycles):
+                for _L_step in range(self.config.L_cycles):
+                    if not ((_H_step == self.config.H_cycles - 1) and (_L_step == self.config.L_cycles - 1)):
+                        z_L = self.L_level(z_L, z_H + input_embeddings, **seq_info)
+
+                if not (_H_step == self.config.H_cycles - 1):
+                    z_H = self.H_level(z_H, z_L, **seq_info)
+                    if require_trace:
+                        z_H_trace.append(z_H.detach().cpu().clone())
+
+        assert not z_H.requires_grad and not z_L.requires_grad
+
+        # 1-step grad
+        z_L = self.L_level(z_L, z_H + input_embeddings, **seq_info)
+        z_H = self.H_level(z_H, z_L, **seq_info)
+
+        if require_trace:
+            z_H_trace.append(z_H.detach().cpu().clone())
+
+        # LM Outputs — decode z_H into token predictions
+        output = self.lm_head(z_H)[:, self.puzzle_emb_len:]   # (B, seq_len, vocab_size)
+
+        # SHREK Component 1: Combined Error Signal
+        # Signal A — flip rate: what fraction of tokens changed from last step?
+        #   catches oscillation (model changing its mind between wrong options)
+        flip_err, current_pred = get_error_signal(output, carry.prev_pred)     # (B,), (B, seq_len)
+
+        # Signal B — learned estimator: reads z_H and predicts how wrong the model is.
+        #   catches stuck-but-wrong (model confidently on wrong answer without oscillating)
+        #   average over content positions (skip puzzle embedding prefix positions)
+        #   CRITICAL: detach z_H_mean so aux_loss only trains error_estimator weights,
+        #   NOT the main reasoning layers. Without detach, aux_loss sends parasitic
+        #   gradients through z_H → H_level/L_level that conflict with lm_loss.
+        z_H_mean    = z_H[:, self.puzzle_emb_len:].mean(dim=1).detach()        # (B, hidden_size)
+        learned_err = torch.sigmoid(self.error_estimator(z_H_mean.float()))    # (B, 1)
+        learned_err = learned_err.squeeze(-1)                                  # (B,)
+
+        # SHREK: combined error = 50/50 blend of both signals
+        # flip_err works immediately from step 1 (no learning needed)
+        # learned_err becomes accurate over training and takes over as the stronger signal
+        error = 0.5 * flip_err + 0.5 * learned_err                            # (B,)
+
+        # SHREK Component 2: Stagnation-Aware Q-head
+        # Ablation: can be disabled via config.enable_stagnation_delta=False
+        if self.config.enable_stagnation_delta:
+            # Compute how much z_H changed during this reasoning step.
+            # CRITICAL: detach so Q-head gradients don't flow through all positions of z_H.
+            # Without detach, Q-head loss would route gradients through all 82 positions,
+            # drowning out lm_loss. With detach, Q-head gradients only flow through position 0.
+            z_H_f = z_H.detach().float()
+            z_H_start = carry.z_H.detach().float()
+            delta = torch.norm(z_H_f - z_H_start, dim=(1, 2)) / (torch.norm(z_H_start, dim=(1, 2)) + 1e-6)  # (B,)
+            # Q-head reads CLS token + stagnation delta
+            q_input = torch.cat([z_H[:, 0].to(torch.float32), delta.unsqueeze(-1)], dim=-1)  # (B, hidden_size+1)
+        else:
+            # Fallback: same as original HRM — Q-head reads CLS token only
+            q_input = z_H[:, 0].to(torch.float32)  # (B, hidden_size)
+        q_logits = self.q_head(q_input).to(torch.float32)  # (B, 2)
+
+        # SHREK: inject combined error into z_H (AFTER Q-head, only affects carry for next step)
+        # Ablation: can be disabled via config.enable_error_injection=False
+        if self.config.enable_error_injection:
+            # error_encoder maps scalar -> hidden_size vector
+            # alpha follows linear warmup: 0 → alpha_max over warmup steps
+            # scaled by 1/sqrt(hidden_size) so injection is proportional to model size
+            error_emb = self.error_encoder(error.unsqueeze(-1))                    # (B, hidden_size)
+            # SHREK: compute alpha from warmup schedule (not learned)
+            # During warmup, alpha ramps linearly from 0 to alpha_max.
+            # After warmup, alpha stays at alpha_max.
+            # Uses torch.clamp instead of Python min() to stay compatible with torch.compile.
+            with torch.no_grad():
+                if self.training:
+                    self._alpha_step += 1
+                alpha = self.config.alpha_max * torch.clamp(self._alpha_step / self.config.alpha_warmup_steps, max=1.0)
+            scale = math.sqrt(self.config.hidden_size)
+            z_H = z_H + (alpha * error_emb.unsqueeze(1) / scale).to(z_H.dtype)   # (B, seq_len, hidden_size)
+
+        # New carry: store error-injected z_H so next ACT step starts from it
+        new_carry = HierarchicalReasoningModel_ACTV1InnerCarry(
+            z_H=z_H.detach(),
+            z_L=z_L.detach(),
+            # SHREK: store current predictions — next step compares against these for flip rate
+            prev_pred=current_pred.detach(),
+            # SHREK: Q-values written to carry for torch.compile compatibility (not read for Q-targets)
+            prev_q_halt=q_logits[..., 0].detach(),
+            prev_q_continue=q_logits[..., 1].detach(),
+        )
+
+        # SHREK: also return learned_err so pretrain.py can compute auxiliary loss
+        # auxiliary loss trains the estimator: predicted error should match real lm_loss
+        if require_trace:
+            return z_H_trace, new_carry, output, (q_logits[..., 0], q_logits[..., 1]), learned_err
+        else:
+            return new_carry, output, (q_logits[..., 0], q_logits[..., 1]), learned_err
+
+
+class HierarchicalReasoningModel_ACTV1(nn.Module):
+    """ACT wrapper."""
+
+    def __init__(self, config_dict: dict):
+        super().__init__()
+        self.config = HierarchicalReasoningModel_ACTV1Config(**config_dict)
+        self.inner = HierarchicalReasoningModel_ACTV1_Inner(self.config)
+
+    @property
+    def puzzle_emb(self):
+        return self.inner.puzzle_emb
+
+    def initial_carry(self, batch: Dict[str, torch.Tensor]):
+        batch_size = batch["inputs"].shape[0]
+
+        return HierarchicalReasoningModel_ACTV1Carry(
+            inner_carry=self.inner.empty_carry(batch_size),  # Empty is expected, it will be reseted in first pass as all sequences are halted.
+            
+            steps=torch.zeros((batch_size, ), dtype=torch.int32),
+            halted=torch.ones((batch_size, ), dtype=torch.bool),  # Default to halted
+            
+            current_data={k: torch.empty_like(v) for k, v in batch.items()}
+        )
+        
+    # SHREK: removed task_type parameter — error signal is universal, no task rules needed
+    def forward(self, carry: HierarchicalReasoningModel_ACTV1Carry, batch: Dict[str, torch.Tensor], require_trace=False):
+        # -> Tuple[HierarchicalReasoningModel_ACTV1Carry, Dict[str, torch.Tensor], torch.Tensor]:
+        # Update data, carry (removing halted sequences)
+        new_inner_carry = self.inner.reset_carry(carry.halted, carry.inner_carry)
+
+        new_steps = torch.where(carry.halted, 0, carry.steps)
+
+        new_current_data = {k: torch.where(carry.halted.view((-1, ) + (1, ) * (batch[k].ndim - 1)), batch[k], v) for k, v in carry.current_data.items()}
+
+        # SHREK: run inner forward — unpack learned_err for auxiliary loss in pretrain.py
+        if require_trace:
+            z_H_trace, new_inner_carry, logits, (q_halt_logits, q_continue_logits), learned_err = self.inner(new_inner_carry, new_current_data, require_trace=require_trace)
+        else:
+            new_inner_carry, logits, (q_halt_logits, q_continue_logits), learned_err = self.inner(new_inner_carry, new_current_data)
+
+        outputs = {
+            "logits": logits,
+            "q_halt_logits": q_halt_logits,
+            "q_continue_logits": q_continue_logits,
+            "learned_err": learned_err,   # SHREK: for auxiliary loss in pretrain.py
+        }
+        
+        with torch.no_grad():
+            # Step
+            new_steps = new_steps + 1
+            is_last_step = new_steps >= self.config.halt_max_steps
+            
+            halted = is_last_step
+
+            # if training, and ACT is enabled
+            if self.training and (self.config.halt_max_steps > 1):
+                # Halt signal
+                # NOTE: During evaluation, always use max steps, this is to guarantee the same halting steps inside a batch for batching purposes
+                halted = halted | (q_halt_logits > q_continue_logits)
+
+                # Exploration
+                min_halt_steps = (torch.rand_like(q_halt_logits) < self.config.halt_exploration_prob) * torch.randint_like(new_steps, low=2, high=self.config.halt_max_steps + 1)
+
+                halted = halted & (new_steps >= min_halt_steps)
+
+                # SHREK: Q-target via double forward pass (same as original HRM).
+                # Run inner() a second time to get NEXT step's Q-values (step T+1).
+                # Save/restore _alpha_step so the second call doesn't double-count
+                # the warmup — without this, alpha reaches full strength at step ~2500
+                # instead of 5000, destabilizing early training.
+                # NOTE: prev_q fields in carry are not read here — they must stay
+                # in the dataclass for torch.compile compatibility.
+                saved_alpha_step = self.inner._alpha_step.clone()
+                next_q_halt_logits, next_q_continue_logits = self.inner(new_inner_carry, new_current_data)[-2]
+                self.inner._alpha_step.copy_(saved_alpha_step)
+
+                outputs["target_q_continue"] = torch.sigmoid(
+                    torch.where(is_last_step,
+                        next_q_halt_logits,
+                        torch.maximum(next_q_halt_logits, next_q_continue_logits))
+                )
+
+        if require_trace:
+            return HierarchicalReasoningModel_ACTV1Carry(new_inner_carry, new_steps, halted, new_current_data), outputs, new_steps, (q_halt_logits > q_continue_logits), z_H_trace
+        else:
+            return HierarchicalReasoningModel_ACTV1Carry(new_inner_carry, new_steps, halted, new_current_data), outputs, new_steps, (q_halt_logits > q_continue_logits)

sudoku-extreme/shrek-tiny/losses.py

@@ -0,0 +1,130 @@
+from typing import Any, Tuple, Dict, Sequence, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+IGNORE_LABEL_ID = -100
+
+
+def s(x, epsilon=1e-30):
+    return torch.where(
+        x<0,
+        1/(1-x+ epsilon),
+        x + 1
+    )
+
+
+def log_stablemax(x, dim=-1):
+    s_x = s(x)
+    return torch.log(s_x/torch.sum(s_x, dim=dim, keepdim=True))
+
+
+def stablemax_cross_entropy(logits, labels, ignore_index: int = -100):
+    logprobs = log_stablemax(logits.to(torch.float64), dim=-1)
+
+    valid_mask = labels != ignore_index
+    transformed_labels = torch.where(valid_mask, labels, 0)
+    prediction_logprobs = torch.gather(logprobs, index=transformed_labels.to(torch.long).unsqueeze(-1), dim=-1).squeeze(-1)
+
+    return -torch.where(valid_mask, prediction_logprobs, 0)
+
+
+def softmax_cross_entropy(logits, labels, ignore_index: int = -100):
+    # Cast logits to f32
+    # Flatten logits
+    return F.cross_entropy(logits.to(torch.float32).view(-1, logits.shape[-1]), labels.to(torch.long).view(-1), ignore_index=ignore_index, reduction="none").view(labels.shape)
+
+
+class ACTLossHead(nn.Module):
+    def __init__(self, model: nn.Module, loss_type: str):
+        super().__init__()
+        self.model = model
+        self.loss_fn = globals()[loss_type]
+        
+    def initial_carry(self, *args, **kwargs):
+        return self.model.initial_carry(*args, **kwargs)  # type: ignore
+
+    def forward(
+        self,
+        return_keys: Sequence[str],
+        require_trace=False,
+        # Model args
+        **model_kwargs,
+    ) -> Tuple[Any, torch.Tensor, Dict[str, torch.Tensor], Optional[Dict[str, torch.Tensor]], torch.Tensor]:
+        # Model logits
+        # B x SeqLen x D
+        if require_trace:
+            new_carry, outputs, steps, act_halt, z_H_trace = self.model(**model_kwargs, require_trace=require_trace)
+        else:
+            new_carry, outputs, steps, act_halt = self.model(**model_kwargs)
+        labels = new_carry.current_data["labels"]
+        alpha = 1.0
+        ds_mask = (alpha ** (16-steps.detach())).unsqueeze(dim=1)
+        # print(ds_mask.shape)
+
+        # Correctness
+        with torch.no_grad():
+            mask = labels != IGNORE_LABEL_ID
+            loss_counts = mask.sum(-1)
+            loss_divisor = loss_counts.clamp_min(1).unsqueeze(-1)  # Avoid NaNs in division
+
+            is_correct = mask & (torch.argmax(outputs["logits"], dim=-1) == labels)
+            seq_is_correct = is_correct.sum(-1) == loss_counts
+            
+            # # Metrics (halted)
+            # valid_metrics = new_carry.halted & (loss_counts > 0)
+            valid_metrics = (loss_counts > 0)
+
+            metrics = {
+                "count": valid_metrics.sum(),
+                
+                "accuracy":       torch.where(valid_metrics, (is_correct.to(torch.float32) / loss_divisor).sum(-1), 0).sum(),
+                "exact_accuracy": (valid_metrics & seq_is_correct).sum(),
+
+                "q_halt_accuracy": (valid_metrics & ((outputs["q_halt_logits"] >= 0) == seq_is_correct)).sum(),
+                "steps":          torch.where(valid_metrics, new_carry.steps, 0).sum(),
+            }
+
+        # Losses
+        # FIXME: Assuming the batch is always full
+        loss = self.loss_fn(outputs["logits"], labels, ignore_index=IGNORE_LABEL_ID)
+        # print(loss.shape)
+        lm_loss = (ds_mask * loss / loss_divisor).sum()
+        q_halt_loss = F.binary_cross_entropy_with_logits(outputs["q_halt_logits"], seq_is_correct.to(outputs["q_halt_logits"].dtype), reduction="sum")
+
+        metrics.update({
+            "lm_loss": lm_loss.detach(),
+            "q_halt_loss": q_halt_loss.detach(),
+        })
+
+        # Q continue (bootstrapping target loss)
+        q_continue_loss = 0
+        if "target_q_continue" in outputs:
+            q_continue_loss = F.binary_cross_entropy_with_logits(outputs["q_continue_logits"], outputs["target_q_continue"], reduction="sum")
+
+            metrics["q_continue_loss"] = q_continue_loss.detach()
+
+        # SHREK: auxiliary loss — trains error_estimator to predict per-sample lm_loss.
+        # Computed here (not in pretrain.py) so learned_err still has gradients.
+        # per_sample_lm_loss is detached so only error_estimator weights get updated by aux_loss.
+        aux_loss = 0
+        if "learned_err" in outputs:
+            per_sample_lm_loss = (ds_mask * loss / loss_divisor).sum(dim=1).detach()  # (B,)
+            # SHREK: min-max normalize target to 0-1 so it matches sigmoid output range.
+            # When all losses are equal (early training), target = 0 (no distinguishable error).
+            # When losses vary, target spreads across 0-1 giving the estimator a useful signal.
+            lm_min = per_sample_lm_loss.min()
+            lm_max = per_sample_lm_loss.max()
+            target = (per_sample_lm_loss - lm_min) / (lm_max - lm_min + 1e-6)       # (B,) in [0, 1]
+            aux_loss = F.mse_loss(outputs["learned_err"], target, reduction="sum")
+            metrics["aux_loss"] = aux_loss.detach()
+
+        # Filter outputs for return
+        detached_outputs = {k: outputs[k].detach() for k in return_keys if k in outputs}
+
+        if require_trace:
+            return z_H_trace, new_carry, lm_loss + 0.5 * (q_halt_loss + q_continue_loss) + 0.1 * aux_loss, metrics, detached_outputs, new_carry.halted.all(), new_carry.halted & act_halt
+        else:
+            return new_carry, lm_loss + 0.5 * (q_halt_loss + q_continue_loss) + 0.1 * aux_loss, metrics, detached_outputs, new_carry.halted.all(), new_carry.halted & act_halt

sudoku-extreme/shrek-tiny/step_10416

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7a9b87f3c1736f07844356e56f16a81f097f5b66d4dcaa24b0a465c7b55dc30
+size 54598517

sudoku-extreme/shrek-tiny/step_11718

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b79a29f9494f6cb0d96e889eda427029b53669c02e3c5e1966fdcb73dd35f85
+size 54598517

sudoku-extreme/shrek-tiny/step_1302

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d4c324d3cce0b23a822d4dcec4f92acf7efc837d868179842bff282b5fb021e
+size 54598483

sudoku-extreme/shrek-tiny/step_13020

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a500c15a465e806f8167f00548fe27a98116b40d1f8b6badab15e7bb330cd45
+size 54598517

sudoku-extreme/shrek-tiny/step_14322

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:674eba287ffc5b434065d82a9949595b56fa6d970128db62ce82456c8235012c
+size 54598517

sudoku-extreme/shrek-tiny/step_15624

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56717a02415f060c374bc6b944f3a277e7a794c4b20f7f012e3d2182bc97f9e8
+size 54598517

sudoku-extreme/shrek-tiny/step_16926

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4a66ee19c4b5cc832db9348557711086ea6e8388af8292f41cc947fc9b7f2b5
+size 54598517

sudoku-extreme/shrek-tiny/step_18228

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d13ade3a56642ffcff2d8ccf3e98cb4cf0bead1d02abcf186608d0af27670eb
+size 54598517

sudoku-extreme/shrek-tiny/step_19530

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:884efd1672ae1e942cad2e2afe0ea175babee5890869f08f2e9e01652349783f
+size 54598517

sudoku-extreme/shrek-tiny/step_20832

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37b32d3e9fef7ef49525019dc270e3e49ec2fad229304397e38067bfc5c7d2e5
+size 54598517

sudoku-extreme/shrek-tiny/step_22134

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0e74df690b60c7b83eddf3de2a8851d6faffd52ff73afd03f8dd89939b5eeb9
+size 54598517

sudoku-extreme/shrek-tiny/step_23436

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f77d1dd7a3fc14f6e7fe2ea3af70cb636cce5fe7063e392ada4ccd52802bbcb8
+size 54598517

sudoku-extreme/shrek-tiny/step_24738

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e81f5eade0b0ca6ea43399bc0a6d3106e6b8ed9f9f9cbe6b4d63aa6805e8b76c
+size 54598517

sudoku-extreme/shrek-tiny/step_2604

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bebb441ecdf5fcbcce3043451382c3351e2f2ff6babeae408bf65ddf34ffd1a8
+size 54598483

sudoku-extreme/shrek-tiny/step_26040

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c8c7f4337650caf5b9691c3cb1773ab39eec11d23a29a80fa094226512a80de
+size 54598517

sudoku-extreme/shrek-tiny/step_27342

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c9caa7822e6db807cbb90a022eb22eb9f177671c457a5e224ce0403cb99e2ed
+size 54598517

sudoku-extreme/shrek-tiny/step_28644

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b20b27a5c01420bd424788f4150109ffbd02e5b120192dc3a29a65ed66f4e5b
+size 54598517

sudoku-extreme/shrek-tiny/step_29946

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af8468106a2c5d22677e1eaf3c6c28106680efc19999169fd39ca185959a5a20
+size 54598517

sudoku-extreme/shrek-tiny/step_31248

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c132a5392ceae910c20c8183eeccf16d4ab6c4f2a0c183576558cd3aa49063b6
+size 54598517

sudoku-extreme/shrek-tiny/step_32550

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08e19ad87f3890577bbcc2d144a59f72d9d9d3170e941c658f6ffdacd8725c45
+size 54598517

sudoku-extreme/shrek-tiny/step_33852

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:141ae2f139b928feb4776acf1970694c795dfc8a7d32e48339eecf3b19888a2a
+size 54598517

sudoku-extreme/shrek-tiny/step_35154

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d0be9e4e7e011aeafb1bf8297d2069c474631192b7b643fdefb851aa1f42dc8
+size 54598517

sudoku-extreme/shrek-tiny/step_36456

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6962ea1a05e33cef44b4a1cba0f2a3a7265ec675265712a252e50702ed1de47d
+size 54598517

sudoku-extreme/shrek-tiny/step_37758

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0837a3d63e0b84b98b8eaddfc8a67bb96f3a8c26377fbd4d3f04f960ec654fc3
+size 54598517

sudoku-extreme/shrek-tiny/step_37758_all_preds.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93543b6f9821425ca72da9e5fcd966ae60528b8f423e3f4872cc1d4a8a38bb20
+size 1033379377

sudoku-extreme/shrek-tiny/step_3906

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:255ef492d8d77bb3f8d828f9adeafafb7d71efe29d623f1f7a70b599007207d7
+size 54598483

sudoku-extreme/shrek-tiny/step_39060

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8500994e181ac6e1347d0d1aa644f44110e75d88764b275496cae0bc3e5f7087
+size 54598517

sudoku-extreme/shrek-tiny/step_40362

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:583a24e7d74d0b2f2253e95c14661aad1a00ed597680842aeca760b6543b4b69
+size 54598517

sudoku-extreme/shrek-tiny/step_41664

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b09be0d8af7ca33b14709752f26318ea0ab0a59f0ef94c8e437804cdfd212b5e
+size 54598517

sudoku-extreme/shrek-tiny/step_42966

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c2fc33a697eb461c49d4627a80129d7c4b1af5077986b0f27d180b13252a893
+size 54598517

sudoku-extreme/shrek-tiny/step_44268

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1907fc6e2a3433cc87c82e874054485453f340bae67ec05945d43ffdcd6b868
+size 54598517

sudoku-extreme/shrek-tiny/step_45570

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86fd6adf26bb571d55cdc52f638d15866487eed7b09cec1d464b37cd8c25a7c5
+size 54598517

sudoku-extreme/shrek-tiny/step_46872

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0aaaa7d1c8040553db710d5027875380edc6675a90c1b074f635f69d23f89b9
+size 54598517

sudoku-extreme/shrek-tiny/step_48174

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48794e8438aa40d7053ca0204d02ada22e5e74098d80039e3f7cf3158b061a45
+size 54598517

sudoku-extreme/shrek-tiny/step_49476

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe7030727a222d3451f7b143f477b498b06f2ed2050b1f2b156d2100912f6ca6
+size 54598517

sudoku-extreme/shrek-tiny/step_50778

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cb76ae9942dc09f964508a64766189b5d7c5263a10c67bd960035ac9a68e26f
+size 54598517

sudoku-extreme/shrek-tiny/step_5208

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa53481158d956c5240552968e6ad19a31ba55758daa1e180079af9a5f8faf9d
+size 54598483

sudoku-extreme/shrek-tiny/step_52080

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae1e1035991097d9d09958d0b366a21b10d00cd4e02cefaac8cb47fdc663e3dd
+size 54598517

sudoku-extreme/shrek-tiny/step_6510

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb056827cf216f10bca7b5855c490a9702c164402e489abfe14b6ee82e393514
+size 54598483

sudoku-extreme/shrek-tiny/step_7812

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a803fd6913179026762f281155adb767e7486a03e5078f52e667835ebffd58db
+size 54598483

sudoku-extreme/shrek-tiny/step_9114

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a06f3a6b7a17e779d43ba63e7671c97d15ea2f6d9e251f136a03b7340266b1f
+size 54598483