Upload SR2 ARC-AGI checkpoints (all evaluator steps and configs)

.gitattributes

@@ -0,0 +1 @@
+step_* filter=lfs diff=lfs merge=lfs -text

Arc-2-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV3 spectacular-dragon/all_config.yaml

@@ -0,0 +1,41 @@
+arch:
+  H_cycles: 1
+  H_layers: 16
+  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.sr2@HierarchicalReasoningModel_ACTV3
+  num_heads: 8
+  pos_encodings: rope
+  puzzle_emb_ndim: 512
+beta1: 0.9
+beta2: 0.95
+checkpoint_every_eval: true
+checkpoint_path: checkpoints/Arc-2-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV3
+  spectacular-dragon
+data_path: data/arc-2-aug-1000
+ema_decay: 0.999
+ema_device: cuda
+ema_enabled: true
+ema_use_buffers: true
+epochs: 300000
+eval_interval: 10000
+eval_save_outputs: []
+evaluators:
+- name: arc@ARC
+global_batch_size: 768
+load_checkpoint: null
+lr: 0.0001
+lr_min_ratio: 1.0
+lr_warmup_steps: 2000
+project_name: Arc-2-aug-1000 ACT-torch
+puzzle_emb_lr: 0.01
+puzzle_emb_weight_decay: 0.1
+run_name: HierarchicalReasoningModel_ACTV3 spectacular-dragon
+seed: 0
+target_q_update_every: 4
+weight_decay: 0.1

Arc-2-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV3 spectacular-dragon/losses.py

@@ -0,0 +1,105 @@
+from typing import Any, Tuple, Dict, Set, 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: Set[str],
+        # 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
+        new_carry, outputs = self.model(**model_kwargs)
+        labels = new_carry.current_data["labels"]
+
+        # Correctness
+        with torch.no_grad():
+            # Preds
+            outputs["preds"] = torch.argmax(outputs["logits"], dim=-1)
+
+            # Correctness
+            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 & (outputs["preds"] == labels)
+            seq_is_correct = is_correct.sum(-1) == loss_counts
+            
+            # Metrics (halted)
+            valid_metrics = new_carry.halted & (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
+        lm_loss = (self.loss_fn(outputs["logits"], labels, ignore_index=IGNORE_LABEL_ID) / 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()
+
+        # Filter outputs for return
+        detached_outputs = {k: outputs[k].detach() for k in return_keys if k in outputs}
+
+        return new_carry, lm_loss + 0.5 * (q_halt_loss + q_continue_loss), metrics, detached_outputs, new_carry.halted.all()

Arc-2-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV3 spectacular-dragon/sr2.py

@@ -0,0 +1,360 @@
+"""
+HRM ACT V2: Transformer Baseline for Architecture Ablation
+
+This is an architecture ablation of the Hierarchical Reasoning Model (HRM).
+Key changes from V1:
+1. REMOVED hierarchical split (no separate H and L levels)
+2. REMOVED inner cycles (no H_cycles/L_cycles loops within reasoning)
+3. KEPT ACT outer loop structure intact
+4. KEPT all data preprocessing, embeddings, and evaluation infrastructure
+
+Architecture: Single-level transformer that processes the full 30x30 grid as a
+900-token sequence, with the same positional encodings and sparse embeddings as V1.
+
+"""
+
+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
+
+
+@dataclass
+class HierarchicalReasoningModel_ACTV3InnerCarry:
+    z_H: torch.Tensor
+
+
+@dataclass
+class HierarchicalReasoningModel_ACTV3Carry:
+    inner_carry: HierarchicalReasoningModel_ACTV3InnerCarry
+
+    steps: torch.Tensor
+    halted: torch.Tensor
+
+    current_data: Dict[str, torch.Tensor]
+
+
+class HierarchicalReasoningModel_ACTV3Config(BaseModel):
+    batch_size: int
+    seq_len: int
+    puzzle_emb_ndim: int = 0
+    num_puzzle_identifiers: int
+    vocab_size: int
+
+    H_cycles: int
+
+    H_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
+    act_enabled: bool = True  # If False, always run halt_max_steps (no early stopping during training)
+    act_inference: bool = False  # If True, use adaptive computation during inference
+
+    forward_dtype: str = "bfloat16"
+
+
+class HierarchicalReasoningModel_ACTV3Block(nn.Module):
+    def __init__(self, config: HierarchicalReasoningModel_ACTV3Config) -> 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_ACTV3ReasoningModule(nn.Module):
+    """
+    说明（已修改）：
+    - 原来通过传入 List[Block] 构造若干“参数彼此独立”的层。
+    - 现在改为只持有一个共享的 block，并通过 `self.repeats` 在前向中重复调用同一个 block。
+    - 为了减少对外部代码的影响，仍然保留 `self.layers` 属性，但其只包含一个共享 block。
+    """
+    def __init__(self, block: HierarchicalReasoningModel_ACTV3Block, repeats: int):
+        super().__init__()
+        # 仅注册一个共享 block（保持属性名 layers 以避免外部依赖破坏）
+        self.layers = torch.nn.ModuleList([block])
+        self.repeats = int(repeats)
+
+    def forward(self, hidden_states: torch.Tensor, input_injection: torch.Tensor, **kwargs) -> torch.Tensor:
+        # Input injection (add)
+
+        # hidden_states = hidden_states + input_injection
+        # 使用同一个 block 重复 n 次（权重完全共享）
+        shared_block = self.layers[0]
+        for _ in range(self.repeats):
+            hidden_states = shared_block(hidden_states=hidden_states + input_injection, **kwargs)
+        return hidden_states
+
+
+class HierarchicalReasoningModel_ACTV3_Inner(nn.Module):
+    def __init__(self, config: HierarchicalReasoningModel_ACTV3Config) -> 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)
+        self.q_head = CastedLinear(self.config.hidden_size, 2, bias=True)
+
+        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_ACTV3ReasoningModule(
+        #     layers=[HierarchicalReasoningModel_ACTV3Block(self.config) for _i in range(self.config.H_layers)]
+        # )
+        
+        H_block = HierarchicalReasoningModel_ACTV3Block(self.config)
+        self.H_level = HierarchicalReasoningModel_ACTV3ReasoningModule(
+            block=H_block,
+            repeats=self.config.H_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,
+        )
+
+        # 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_ACTV3InnerCarry(
+            z_H=torch.empty(
+                batch_size,
+                self.config.seq_len + self.puzzle_emb_len,
+                self.config.hidden_size,
+                dtype=self.forward_dtype,
+            ),
+        )
+
+    def reset_carry(self, reset_flag: torch.Tensor, carry: HierarchicalReasoningModel_ACTV3InnerCarry):
+        return HierarchicalReasoningModel_ACTV3InnerCarry(
+            z_H=torch.where(reset_flag.view(-1, 1, 1), self.H_init, carry.z_H),
+        )
+
+    def forward(
+        self, carry: HierarchicalReasoningModel_ACTV3InnerCarry, batch: Dict[str, torch.Tensor], carry_steps: torch.Tensor
+    ) -> Tuple[HierarchicalReasoningModel_ACTV3InnerCarry, torch.Tensor, Tuple[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"])
+        
+        # 仅第一步注入 input
+        first_step_mask = (carry_steps == 0).view(-1, 1, 1).to(device=input_embeddings.device, dtype=input_embeddings.dtype)
+        gated_injection = input_embeddings * first_step_mask  # [B,S_full,D]
+
+        # 1-step grad
+        z_H = self.H_level(carry.z_H, gated_injection, **seq_info)
+
+        # LM Outputs
+        new_carry = HierarchicalReasoningModel_ACTV3InnerCarry(
+            z_H=z_H.detach(),
+        )  # New carry no grad
+        output = self.lm_head(z_H)[:, self.puzzle_emb_len :]
+
+        # Q head
+        q_logits = self.q_head(z_H[:, 0]).to(torch.float32)
+
+        return new_carry, output, (q_logits[..., 0], q_logits[..., 1])
+
+
+class HierarchicalReasoningModel_ACTV3(nn.Module):
+    """ACT wrapper."""
+
+    def __init__(self, config_dict: dict):
+        super().__init__()
+        self.config = HierarchicalReasoningModel_ACTV3Config(**config_dict)
+        self.inner = HierarchicalReasoningModel_ACTV3_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_ACTV3Carry(
+            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()},
+        )
+
+    def forward(
+        self,
+        carry: HierarchicalReasoningModel_ACTV3Carry,
+        batch: Dict[str, torch.Tensor],
+        compute_target_q: bool = False,
+    ) -> Tuple[HierarchicalReasoningModel_ACTV3Carry, Dict[str, 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()
+        }
+
+        # Forward inner model
+        new_inner_carry, logits, (q_halt_logits, q_continue_logits) = self.inner(
+            new_inner_carry, new_current_data, new_steps
+        )
+
+        outputs = {"logits": logits, "q_halt_logits": q_halt_logits, "q_continue_logits": q_continue_logits}
+
+        with torch.no_grad():
+            # Step
+            new_steps = new_steps + 1
+            is_last_step = new_steps >= self.config.halt_max_steps
+
+            halted = is_last_step
+
+            # Check if adaptive computation should be used
+            use_adaptive = (self.config.halt_max_steps > 1) and (
+                (self.training and self.config.act_enabled)
+                or (not self.training and self.config.act_inference)
+            )
+
+            if use_adaptive:
+                # Halt signal based on Q-values (but always halt at max steps)
+                q_halt_signal = q_halt_logits > q_continue_logits
+                halted = halted | q_halt_signal
+
+                # Store actual steps used for logging (only during inference)
+                if not self.training:
+                    outputs["actual_steps"] = new_steps.float()
+
+                # Exploration (only during training)
+                if self.training:
+                    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)
+
+                # Compute target Q (only during training)
+                # NOTE: No replay buffer and target networks for computing target Q-value.
+                # As batch_size is large, there're many parallel envs.
+                # Similar concept as PQN https://arxiv.org/abs/2407.04811
+                if self.training and compute_target_q:
+                    next_q_halt_logits, next_q_continue_logits = self.inner(
+                        new_inner_carry, new_current_data, new_steps
+                    )[-1]
+
+                    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),
+                        )
+                    )
+
+        return HierarchicalReasoningModel_ACTV3Carry(
+            new_inner_carry, new_steps, halted, new_current_data
+        ), outputs

Arc-2-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV3 spectacular-dragon/step_1818933

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8436b037d0f366ca02870c66dfd9a22590f17677aee752bf3ac3f43b2089a10
+size 4311137979

Arc-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV3 quizzical-labradoodle/all_config.yaml

@@ -0,0 +1,41 @@
+arch:
+  H_cycles: 1
+  H_layers: 16
+  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.sr2@HierarchicalReasoningModel_ACTV3
+  num_heads: 8
+  pos_encodings: rope
+  puzzle_emb_ndim: 512
+beta1: 0.9
+beta2: 0.95
+checkpoint_every_eval: true
+checkpoint_path: checkpoints/Arc-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV3
+  quizzical-labradoodle
+data_path: data/arc-aug-1000
+ema_decay: 0.999
+ema_device: cuda
+ema_enabled: true
+ema_use_buffers: true
+epochs: 300000
+eval_interval: 10000
+eval_save_outputs: []
+evaluators:
+- name: arc@ARC
+global_batch_size: 768
+load_checkpoint: null
+lr: 0.0001
+lr_min_ratio: 1.0
+lr_warmup_steps: 2000
+project_name: Arc-aug-1000 ACT-torch
+puzzle_emb_lr: 0.01
+puzzle_emb_weight_decay: 0.1
+run_name: HierarchicalReasoningModel_ACTV3 quizzical-labradoodle
+seed: 0
+target_q_update_every: 4
+weight_decay: 0.1