Upload mythos-fineweb.py

mythos-fineweb.py

@@ -0,0 +1,915 @@
+#!/usr/bin/env python3
+"""
+SpiderPortal v5 pretraining on FineWeb-Edu with AdamW.
+
+Single GPU:
+    python mythos-fineweb.py
+
+Multi-GPU:
+    torchrun --nproc_per_node=$(python -c "import torch; print(torch.cuda.device_count())") mythos-fineweb.py
+"""
+
+import os
+import math
+import time
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as dist
+from loguru import logger
+from torch.distributed.fsdp import (
+    FullyShardedDataParallel as FSDP,
+    ShardingStrategy,
+    MixedPrecision,
+    FullStateDictConfig,
+    StateDictType,
+)
+from torch.distributed.fsdp.wrap import ModuleWrapPolicy
+from torch.utils.data import IterableDataset, DataLoader, get_worker_info
+from contextlib import nullcontext
+from dataclasses import dataclass
+from typing import Optional, Tuple, Dict, List
+from torch.nn import CrossEntropyLoss
+
+from datasets import load_dataset
+from transformers import AutoTokenizer
+
+
+# ---------------------------------------------------------------------------
+# SpiderPortal Model Architecture
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class SpiderPortalConfig:
+    vocab_size: int = 50278
+    hidden_size: int = 384
+    num_hidden_layers: int = 8
+    num_attention_heads: int = 8
+    num_key_value_heads: int = 2
+    intermediate_size: int = 1024
+    hidden_act: str = "silu"
+    num_experts: int = 64
+    num_experts_per_tok: int = 1
+    num_shared_experts: int = 1
+    router_aux_loss_coef: float = 0.05
+    max_loop_iters: int = 1
+    act_threshold: float = 0.5
+    max_position_embeddings: int = 131072
+    rope_theta: float = 10000000.0
+    rope_scaling: dict = None
+    sliding_window: int = 4096
+    attention_dropout: float = 0.0
+    rms_norm_eps: float = 1e-6
+    initializer_range: float = 0.02
+    use_cache: bool = True
+    tie_word_embeddings: bool = True
+    prelude_layers: int = 2
+    coda_layers: int = 2
+    lora_rank: int = 32
+    loop_embed_dim: int = 48
+    vision_hidden_size: int = 384
+    audio_hidden_size: int = 512
+    vision_num_frames: int = 60
+    vision_tokens_per_frame: int = 256
+    vision_temporal_tokens: int = 64
+    vision_temporal_layers: int = 2
+    model_type: str = "spiderportal"
+    torch_dtype: str = "bfloat16"
+
+
+def loop_index_embedding(h, loop_t, loop_dim, theta=10000.0):
+    freqs = 1.0 / (theta ** (torch.arange(0, loop_dim, 2, device=h.device, dtype=h.dtype) / loop_dim))
+    angles = loop_t * freqs
+    emb = torch.cat([angles.sin(), angles.cos()], dim=-1)[:loop_dim]
+    emb_full = torch.zeros(h.shape[-1], device=h.device, dtype=h.dtype)
+    emb_full[:loop_dim] = emb
+    return h + emb_full.unsqueeze(0).unsqueeze(0)
+
+
+class SpiderPortalRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight.to(input_dtype) * hidden_states.to(input_dtype)
+
+
+def compute_yarn_inv_freq(head_dim, rope_theta, factor, orig_max, beta_fast=32.0, beta_slow=1.0):
+    dim = head_dim
+    orig_inv_freq = 1.0 / (rope_theta ** (torch.arange(0, dim, 2).float() / dim))
+    pos_freqs = torch.arange(0, dim, 2).float() / dim
+    beta = (pos_freqs * math.log(rope_theta) / math.log(orig_max))
+    scale = torch.where(beta < beta_slow, torch.ones_like(beta), torch.where(beta > beta_fast, torch.ones_like(beta) / factor, 1.0 - (beta - beta_slow) / (beta_fast - beta_slow) * (1.0 - 1.0 / factor)))
+    return orig_inv_freq * scale
+
+
+class SpiderPortalGQA(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_kv_heads = config.num_key_value_heads
+        self.head_dim = config.hidden_size // config.num_attention_heads
+        self.num_key_value_groups = self.num_heads // self.num_kv_heads
+        self.q_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+        self.k_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+        self.attention_dropout = config.attention_dropout
+        rope_scaling = getattr(config, 'rope_scaling', None)
+        if rope_scaling and rope_scaling.get("type") == "yarn":
+            factor = rope_scaling.get("factor", 1.0)
+            orig_max_pos = rope_scaling.get("original_max_position_embeddings", config.max_position_embeddings)
+            inv_freq = compute_yarn_inv_freq(self.head_dim, config.rope_theta, factor, orig_max_pos)
+        else:
+            inv_freq = 1.0 / (config.rope_theta ** (torch.arange(0, self.head_dim, 2).float() / self.head_dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+    def _rotate_half(self, x):
+        x1 = x[..., :x.shape[-1] // 2]
+        x2 = x[..., x.shape[-1] // 2:]
+        return torch.cat((-x2, x1), dim=-1)
+    def _apply_rotary(self, x, cos, sin):
+        return (x * cos) + (self._rotate_half(x) * sin)
+    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):
+        bsz, q_len, _ = hidden_states.size()
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        if position_ids is None:
+            position_ids = torch.arange(q_len, device=hidden_states.device).unsqueeze(0).expand(bsz, -1)
+        max_pos = position_ids.max().item() + 1
+        seq_len = max(max_pos, q_len)
+        t = torch.arange(seq_len, device=hidden_states.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(t, self.inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        cos, sin = emb.cos(), emb.sin()
+        cos = cos[position_ids].unsqueeze(1)
+        sin = sin[position_ids].unsqueeze(1)
+        query_states = self._apply_rotary(query_states, cos, sin)
+        key_states = self._apply_rotary(key_states, cos, sin)
+        if past_key_value is not None:
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        past_kv = (key_states, value_states) if use_cache else None
+        key_states = key_states.repeat_interleave(self.num_key_value_groups, dim=1)
+        value_states = value_states.repeat_interleave(self.num_key_value_groups, dim=1)
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+        if attention_mask is not None:
+            attn_weights = attn_weights + attention_mask
+        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
+        attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        return self.o_proj(attn_output), past_kv
+
+
+class SpiderPortalExpert(nn.Module):
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        inter_size = intermediate_size or config.intermediate_size
+        self.gate_proj = nn.Linear(config.hidden_size, inter_size, bias=False)
+        self.up_proj = nn.Linear(config.hidden_size, inter_size, bias=False)
+        self.down_proj = nn.Linear(inter_size, config.hidden_size, bias=False)
+        self.act_fn = nn.SiLU()
+    def forward(self, hidden_states):
+        return self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
+
+
+class SpiderPortalRouter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.weight = nn.Parameter(torch.randn(config.hidden_size, config.num_experts) * config.initializer_range)
+        self.register_buffer("router_bias", torch.zeros(config.num_experts))
+    def forward(self, hidden_states):
+        router_logits = hidden_states.view(-1, hidden_states.size(-1)) @ self.weight
+        routing_weights = F.softmax(router_logits, dim=-1, dtype=torch.float32)
+        biased_logits = router_logits + self.router_bias
+        biased_weights = F.softmax(biased_logits, dim=-1, dtype=torch.float32)
+        top_weights, top_indices = torch.topk(biased_weights, self.num_experts_per_tok, dim=-1)
+        top_weights = top_weights / top_weights.sum(dim=-1, keepdim=True)
+        top_weights = top_weights.to(hidden_states.dtype)
+        mean_probs = routing_weights.mean(dim=0)
+        aux_loss = self.num_experts * (mean_probs * mean_probs).sum()
+        return top_weights, top_indices, aux_loss
+
+
+class SpiderPortalMoE(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.num_experts = config.num_experts
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.experts = nn.ModuleList([SpiderPortalExpert(config) for _ in range(config.num_experts)])
+        self.shared_expert = SpiderPortalExpert(config)
+        self.router = SpiderPortalRouter(config)
+    def forward(self, hidden_states):
+        batch_size, seq_len, hidden_dim = hidden_states.shape
+        top_weights, top_indices, aux_loss = self.router(hidden_states)
+        flat_hidden = hidden_states.view(-1, hidden_dim)
+        final_output = torch.zeros_like(flat_hidden)
+        for expert_idx in range(self.num_experts_per_tok):
+            expert_ids = top_indices[:, expert_idx]
+            expert_weights = top_weights[:, expert_idx:expert_idx+1]
+            for e in range(self.num_experts):
+                mask = expert_ids == e
+                if mask.any():
+                    expert_output = self.experts[e](flat_hidden[mask])
+                    final_output[mask] += expert_output * expert_weights[mask]
+        shared_output = self.shared_expert(flat_hidden)
+        final_output = final_output + shared_output
+        return final_output.view(batch_size, seq_len, hidden_dim), aux_loss
+
+
+class SpiderPortalDenseLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self_attn = SpiderPortalGQA(config)
+        dense_intermediate = config.hidden_size * 4 // 3
+        self.ffn = SpiderPortalExpert(config, intermediate_size=dense_intermediate)
+        self.input_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):
+        attn_input = self.input_layernorm(hidden_states)
+        attn_output, past_kv = self.self_attn(attn_input, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, use_cache=use_cache)
+        hidden_states = hidden_states + attn_output
+        ffn_input = self.post_attention_layernorm(hidden_states)
+        ffn_output = self.ffn(ffn_input)
+        hidden_states = hidden_states + ffn_output
+        return hidden_states, past_kv
+
+
+class SpiderPortalMoELayer(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.self_attn = SpiderPortalGQA(config)
+        self.moe = SpiderPortalMoE(config)
+        self.input_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):
+        attn_input = self.input_layernorm(hidden_states)
+        attn_output, past_kv = self.self_attn(attn_input, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, use_cache=use_cache)
+        hidden_states = hidden_states + attn_output
+        moe_input = self.post_attention_layernorm(hidden_states)
+        moe_output, aux_loss = self.moe(moe_input)
+        hidden_states = hidden_states + moe_output
+        return hidden_states, aux_loss, past_kv
+
+
+class LTIInjection(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.log_A = nn.Parameter(torch.full((config.hidden_size,), -2.0))
+        self.delta_t = nn.Parameter(torch.tensor(1.0))
+        self.B = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
+        with torch.no_grad():
+            self.B.weight.data.normal_(mean=0.0, std=0.01)
+    def get_A(self):
+        return -torch.exp(self.log_A)
+    def forward(self, h_t, e):
+        A = self.get_A()
+        return A * h_t + self.B(e)
+
+
+class ACTHalting(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.halt_predictor = nn.Linear(config.hidden_size, 1)
+        self.threshold = config.act_threshold
+    def forward(self, hidden_states):
+        return torch.sigmoid(self.halt_predictor(hidden_states))
+
+
+class LoRAAdapter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        rank = config.lora_rank
+        self.down = nn.Linear(config.hidden_size, rank, bias=False)
+        self.B = nn.Parameter(torch.randn(rank, config.hidden_size) * 0.02)
+        self.scale = nn.Embedding(config.max_loop_iters, rank)
+        with torch.no_grad():
+            self.scale.weight.data.zero_()
+            self.down.weight.data.normal_(mean=0.0, std=0.001)
+    def forward(self, x, loop_t):
+        max_t = self.scale.num_embeddings - 1
+        t_idx = min(loop_t, max_t)
+        s = self.scale(torch.tensor(t_idx, device=x.device))
+        down = self.down(x) * s
+        return down @ self.B
+
+
+class SpiderPortalMoEModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.prelude_layers = nn.ModuleList([SpiderPortalDenseLayer(config) for _ in range(config.prelude_layers)])
+        self.recurrent_layers = nn.ModuleList([SpiderPortalMoELayer(config, i) for i in range(config.num_hidden_layers)])
+        self.coda_layers = nn.ModuleList([SpiderPortalDenseLayer(config) for _ in range(config.coda_layers)])
+        self.norm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.injection = LTIInjection(config)
+        self.act_halting = ACTHalting(config)
+        self.lora_adapter = LoRAAdapter(config)
+        self.loop_embed_dim = config.loop_embed_dim
+    def forward(self, hidden_states, input_embedding=None, attention_mask=None, position_ids=None, past_key_values=None, use_cache=False, n_loops=None):
+        n_loops = n_loops or self.config.max_loop_iters
+        input_embedding = input_embedding if input_embedding is not None else hidden_states
+        total_aux_loss = 0.0
+        for layer in self.prelude_layers:
+            hidden_states, _ = layer(hidden_states, attention_mask=attention_mask, position_ids=position_ids)
+        e = hidden_states.clone()
+        B, T_seq, D = hidden_states.shape
+        halted = torch.zeros(B, T_seq, device=hidden_states.device, dtype=torch.bool)
+        cumulative_p = torch.zeros(B, T_seq, device=hidden_states.device, dtype=hidden_states.dtype)
+        h_out = torch.zeros_like(hidden_states)
+        past_key_values = past_key_values if past_key_values is not None else [None] * len(self.recurrent_layers)
+        for t in range(n_loops):
+            h_loop = loop_index_embedding(hidden_states, t, self.loop_embed_dim)
+            if t > 0:
+                injection = self.injection(hidden_states, input_embedding)
+                hidden_states = hidden_states + injection
+            new_past_key_values = []
+            for i, layer in enumerate(self.recurrent_layers):
+                hidden_states, aux_loss, past_kv = layer(hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_values[i] if t == 0 else None, use_cache=use_cache)
+                total_aux_loss = total_aux_loss + aux_loss
+                new_past_key_values.append(past_kv)
+            lora_delta = self.lora_adapter(hidden_states, t)
+            hidden_states = hidden_states + lora_delta
+            halt_prob = self.act_halting(hidden_states).squeeze(-1)
+            still_running = ~halted
+            remainder = (1.0 - cumulative_p).clamp(min=0)
+            weight = torch.where(cumulative_p + halt_prob >= self.config.act_threshold, remainder, halt_prob)
+            weight = weight * still_running.to(hidden_states.dtype)
+            h_out = h_out + weight.unsqueeze(-1) * hidden_states
+            cumulative_p = cumulative_p + halt_prob * still_running.to(hidden_states.dtype)
+            halted = halted | (cumulative_p >= self.config.act_threshold)
+            if halted.all() and not self.training:
+                break
+        never_halted = (~halted).to(hidden_states.dtype).unsqueeze(-1)
+        hidden_states = h_out + never_halted * hidden_states
+        for layer in self.coda_layers:
+            hidden_states, _ = layer(hidden_states, attention_mask=attention_mask, position_ids=position_ids)
+        hidden_states = self.norm(hidden_states)
+        return hidden_states, total_aux_loss, new_past_key_values
+
+
+class SpiderPortalForConditionalGeneration(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.model = SpiderPortalMoEModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        if config.tie_word_embeddings:
+            self.lm_head.weight = self.embed_tokens.weight
+        self.apply(self._init_weights)
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            if hasattr(self, 'model') and module is self.model.injection.B:
+                return
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+    def forward(self, input_ids, attention_mask=None, position_ids=None, labels=None, n_loops=None, use_cache=False):
+        hidden_states = self.embed_tokens(input_ids)
+        model_dtype = next(self.model.parameters()).dtype
+        hidden_states = hidden_states.to(model_dtype)
+        input_embedding = hidden_states.clone()
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        causal_mask = torch.full((attention_mask.size(0), 1, attention_mask.size(1), attention_mask.size(1)), 0.0, dtype=hidden_states.dtype, device=hidden_states.device)
+        causal_mask = causal_mask.masked_fill(~attention_mask.unsqueeze(1).unsqueeze(2), torch.finfo(hidden_states.dtype).min)
+        causal_mask = causal_mask.triu(1)
+        hidden_states, aux_loss, past_kv = self.model(hidden_states, input_embedding=input_embedding, attention_mask=causal_mask, position_ids=position_ids, use_cache=use_cache, n_loops=n_loops)
+        logits = self.lm_head(hidden_states)
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+            loss = loss + self.config.router_aux_loss_coef * aux_loss
+        return {"loss": loss, "logits": logits, "aux_loss": aux_loss, "past_key_values": past_kv}
+    def get_num_params(self):
+        total = sum(p.numel() for p in self.parameters())
+        return {"total": total, "trainable": total}
+
+
+# ---------------------------------------------------------------------------
+# Dataset
+# ---------------------------------------------------------------------------
+
+
+class FineWebEduDataset(IterableDataset):
+    """
+    Streaming FineWeb-Edu loader yielding fixed-length (input, target) pairs.
+
+    FineWeb-Edu is trillions of tokens, so `streaming=True` pulls shards on
+    demand instead of materializing to disk. Sharding is two-dimensional —
+    `world_size` ranks × `num_workers` DataLoader workers per rank — and each
+    `(rank, worker_id)` deterministically owns one shard of the global stream.
+    That gives disjoint coverage without any cross-process coordination.
+
+    Streaming datasets are not seekable, so a resumed run re-enters its shard
+    from the beginning. Acceptable at pretraining scale: the chance of
+    re-playing the same tokens before the run ends is negligible versus the
+    cost of a true resumable loader.
+    """
+
+    def __init__(self, tokenizer, seq_len: int, subset: str, rank: int, world_size: int):
+        """
+        Args:
+            tokenizer  -- HuggingFace tokenizer with .encode(str) -> list[int]
+            seq_len    -- context length; every yielded pair has this many tokens
+            subset     -- FineWeb-Edu config name (e.g. "sample-1BT", "sample-10BT")
+            rank       -- global rank of this process within the distributed job
+            world_size -- total number of distributed processes
+        """
+        self.tokenizer = tokenizer
+        self.seq_len = seq_len
+        self.subset = subset
+        self.rank = rank
+        self.world_size = world_size
+
+    def __iter__(self):
+        """
+        Yield `(input_ids, target_ids)` tensors of length `seq_len` forever.
+
+        Inputs and targets are shifted by one for next-token prediction —
+        `target[i] == input[i + 1]`. Documents are concatenated into a rolling
+        buffer and sliced into fixed-length chunks, packing short docs together
+        and splitting long ones. This keeps every step at the same shape,
+        which under FSDP avoids recompute from variable-length inputs and
+        removes the need for a pad-aware attention mask.
+        """
+        worker = get_worker_info()
+        num_workers = worker.num_workers if worker else 1
+        worker_id = worker.id if worker else 0
+
+        total_shards = self.world_size * num_workers
+        shard_index = self.rank * num_workers + worker_id
+
+        ds = load_dataset(
+            "HuggingFaceFW/fineweb-edu",
+            name=self.subset,
+            split="train",
+            streaming=True,
+        ).shard(num_shards=total_shards, index=shard_index)
+
+        buf = []
+        for sample in ds:
+            buf.extend(self.tokenizer.encode(sample["text"]))
+            while len(buf) >= self.seq_len + 1:
+                chunk = buf[: self.seq_len + 1]
+                buf = buf[self.seq_len + 1 :]
+                yield (
+                    torch.tensor(chunk[:-1], dtype=torch.long),
+                    torch.tensor(chunk[1:], dtype=torch.long),
+                )
+
+
+# ---------------------------------------------------------------------------
+# LR schedule: linear warmup → cosine decay
+# ---------------------------------------------------------------------------
+
+
+def get_lr(step: int, warmup: int, total: int, max_lr: float, min_lr: float) -> float:
+    """
+    Linear warmup → half-cosine decay to `min_lr`.
+
+    Standard language-model pretraining schedule. The warmup phase prevents
+    Adam's second-moment estimate from collapsing to a huge LR in the first
+    few steps when gradients are noisy. The cosine tail lets the model make
+    small, increasingly conservative updates near the end of training rather
+    than crashing to `min_lr` at a fixed step.
+
+    Behavior by region:
+        step < warmup                 → linear ramp 0 → max_lr
+        warmup ≤ step < total         → cosine decay max_lr → min_lr
+        step ≥ total                  → clamped at min_lr (safety for
+                                        off-by-one step counters at the end
+                                        of training)
+
+    Args:
+        step    -- current global optimizer step (0-indexed)
+        warmup  -- number of warmup steps before cosine decay begins
+        total   -- step at which the cosine reaches `min_lr`
+        max_lr  -- peak learning rate reached at the end of warmup
+        min_lr  -- floor learning rate at and after `total` steps
+
+    Returns:
+        Scalar learning rate for this step.
+    """
+    if step < warmup:
+        return max_lr * step / warmup
+    if step >= total:
+        return min_lr
+    decay = (step - warmup) / (total - warmup)
+    return min_lr + 0.5 * (max_lr - min_lr) * (1.0 + math.cos(math.pi * decay))
+
+
+# ---------------------------------------------------------------------------
+# Checkpointing — weights-only every 500 steps, full at epoch end + best
+# ---------------------------------------------------------------------------
+
+
+def save_weights_only(model, step, epoch, ckpt_dir, ddp):
+    """Save model weights only (~1.3GB for 3B bf16). For testing/transfer."""
+    if ddp:
+        with FSDP.state_dict_type(
+            model,
+            StateDictType.FULL_STATE_DICT,
+            FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
+        ):
+            model_state = model.state_dict()
+    else:
+        model_state = model.state_dict()
+
+    ckpt_path = os.path.join(ckpt_dir, f"spiderportal-v5-ep{epoch}-step{step}.pt")
+    tmp_path = ckpt_path + ".tmp"
+    torch.save(model_state, tmp_path)
+    os.replace(tmp_path, ckpt_path)
+    size_mb = os.path.getsize(ckpt_path) / (1024 * 1024)
+    return ckpt_path, size_mb
+
+
+def save_full_checkpoint(model, optimizer, step, epoch, cfg, vocab_size, ckpt_dir, ddp, master, ckpt_name="full"):
+    """Save model + optimizer state (~18GB for 3B bf16). For resume training."""
+    if ddp:
+        with FSDP.state_dict_type(
+            model,
+            StateDictType.FULL_STATE_DICT,
+            FullStateDictConfig(offload_to_cpu=True, rank0_only=True),
+        ):
+            model_state = model.state_dict()
+            optim_state = FSDP.optim_state_dict(model, optimizer)
+    else:
+        model_state = model.state_dict()
+        optim_state = optimizer.state_dict()
+
+    if not master:
+        return None, 0
+
+    os.makedirs(ckpt_dir, exist_ok=True)
+    final_path = os.path.join(ckpt_dir, f"spiderportal-v5-{ckpt_name}.pt")
+    tmp_path = final_path + ".tmp"
+    torch.save(
+        {
+            "step": step,
+            "epoch": epoch,
+            "model_state_dict": model_state,
+            "optimizer_state_dict": optim_state,
+            "cfg": cfg,
+            "vocab_size": vocab_size,
+        },
+        tmp_path,
+    )
+    os.replace(tmp_path, final_path)
+    size_mb = os.path.getsize(final_path) / (1024 * 1024)
+    return final_path, size_mb
+
+
+def delete_step_checkpoints(ckpt_dir):
+    """Delete all weights-only step checkpoints to free disk space."""
+    deleted = 0
+    for f in os.listdir(ckpt_dir):
+        if f.startswith("spiderportal-v5-ep") and "-step" in f and f.endswith(".pt"):
+            path = os.path.join(ckpt_dir, f)
+            try:
+                os.remove(path)
+                deleted += 1
+            except OSError:
+                pass
+    return deleted
+
+
+def load_checkpoint(model, optimizer, path, ddp):
+    """Restore model + optimizer from full checkpoint."""
+    ckpt = torch.load(path, map_location="cpu", weights_only=False)
+
+    if ddp:
+        with FSDP.state_dict_type(
+            model,
+            StateDictType.FULL_STATE_DICT,
+            FullStateDictConfig(offload_to_cpu=True, rank0_only=False),
+        ):
+            model.load_state_dict(ckpt["model_state_dict"])
+            optim_state = FSDP.optim_state_dict_to_load(
+                model=model,
+                optim=optimizer,
+                optim_state_dict=ckpt["optimizer_state_dict"],
+            )
+            optimizer.load_state_dict(optim_state)
+    else:
+        model.load_state_dict(ckpt["model_state_dict"])
+        optimizer.load_state_dict(ckpt["optimizer_state_dict"])
+
+    return int(ckpt["step"]), int(ckpt.get("epoch", 0))
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+
+def main():
+    """
+    End-to-end pretraining entry point.
+
+    Order matters: distributed init must run before any CUDA allocation, the
+    tokenizer must exist before the model is built (vocab_size flows into
+    cfg), and FSDP must wrap the model before the optimizer is constructed
+    (FSDP re-flattens parameters, so an optimizer built on the unwrapped
+    model would track stale param objects). Resume then loads state into the
+    already-constructed optimizer in-place.
+
+    Lifecycle:
+        1. Initialize torch.distributed (NCCL) if launched under torchrun.
+        2. Build tokenizer → derive vocab_size.
+        3. Construct OpenMythos with the 3B variant config.
+        4. Wrap in FSDP with FULL_SHARD + bf16/fp16 mixed precision (multi-GPU)
+           or move to device + autocast (single-GPU).
+        5. Build fused AdamW on (possibly sharded) parameters.
+        6. Resume from the latest checkpoint in `ckpt_dir` if one exists.
+        7. Stream FineWeb-Edu through grad-accumulation microbatches with
+           cosine LR schedule, per-step logging, and periodic checkpoints.
+        8. Write a final checkpoint if the last save wasn't aligned to
+           `ckpt_every`, then barrier + tear down the process group.
+
+    All hyperparameters are literal constants in this function by design —
+    pretraining runs are long-lived and each run pins exact settings; a
+    CLI/config layer is deliberately avoided to keep the file self-auditable.
+    """
+    # ------------------------------------------------------------------
+    # Distributed init
+    # ------------------------------------------------------------------
+    ddp = int(os.environ.get("RANK", -1)) != -1
+    if ddp:
+        dist.init_process_group("nccl")
+        rank = int(os.environ["RANK"])
+        local_rank = int(os.environ["LOCAL_RANK"])
+        world_size = int(os.environ["WORLD_SIZE"])
+        device = f"cuda:{local_rank}"
+        torch.cuda.set_device(device)
+    else:
+        rank = local_rank = 0
+        world_size = 1
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    master = rank == 0
+
+    if master:
+        logger.info(
+            f"GPUs: {torch.cuda.device_count()}  |  World size: {world_size}  |  Device: {device}"
+        )
+
+    # ------------------------------------------------------------------
+    # Tokenizer
+    # ------------------------------------------------------------------
+    tokenizer = AutoTokenizer.from_pretrained("gpt2")
+    tokenizer.pad_token = tokenizer.eos_token
+    vocab_size = tokenizer.vocab_size
+
+    if master:
+        logger.info(f"Tokenizer: gpt2  |  Vocab size: {vocab_size:,}")
+
+    # ------------------------------------------------------------------
+    # Hyperparameters
+    # ------------------------------------------------------------------
+    seq_len = 2048
+    micro_batch = 32
+    target_tokens = 1_000_000_000
+    grad_accum = max(1, 256 // (world_size * micro_batch))
+    global_batch_tok = world_size * micro_batch * grad_accum * seq_len
+    total_steps = target_tokens // global_batch_tok
+    warmup_steps = 200
+    lr = 3e-4
+    wd = 0.1
+    log_every = 10
+    ckpt_every = 500
+    ckpt_dir = "checkpoints"
+    dataset_subset = "sample-1BT"
+
+    if master:
+        logger.info(
+            f"seq_len={seq_len} | micro_batch={micro_batch} | grad_accum={grad_accum} | "
+            f"global_batch_tokens={global_batch_tok:,} | total_steps={total_steps:,}"
+        )
+
+    # ------------------------------------------------------------------
+    # Model
+    # ------------------------------------------------------------------
+    cfg = SpiderPortalConfig(
+        hidden_size=384, num_hidden_layers=8, num_attention_heads=8,
+        num_key_value_heads=2, intermediate_size=1024,
+        num_experts=64, num_experts_per_tok=1, num_shared_experts=1,
+        router_aux_loss_coef=0.05, max_loop_iters=1,
+        prelude_layers=2, coda_layers=2, lora_rank=32,
+        rope_theta=10000000.0,
+        rope_scaling={"type": "yarn", "factor": 4.0, "original_max_position_embeddings": 32768},
+        max_position_embeddings=131072, sliding_window=4096,
+        tie_word_embeddings=True,
+    )
+    cfg.vocab_size = vocab_size
+
+    bf16_ok = torch.cuda.is_available() and torch.cuda.is_bf16_supported()
+    amp_dtype = torch.bfloat16 if bf16_ok else torch.float16
+
+    model = SpiderPortalForConditionalGeneration(cfg)
+
+    if ddp:
+        mp_policy = MixedPrecision(
+            param_dtype=amp_dtype,
+            reduce_dtype=amp_dtype,
+            buffer_dtype=amp_dtype,
+        )
+        wrap_policy = ModuleWrapPolicy({SpiderPortalDenseLayer, SpiderPortalMoELayer})
+        model = FSDP(
+            model,
+            sharding_strategy=ShardingStrategy.FULL_SHARD,
+            mixed_precision=mp_policy,
+            auto_wrap_policy=wrap_policy,
+            device_id=local_rank,
+        )
+    else:
+        model = model.to(device)
+        amp_ctx = (
+            torch.amp.autocast(device_type="cuda", dtype=amp_dtype)
+            if "cuda" in device
+            else nullcontext()
+        )
+
+    # FSDP handles its own mixed precision; only need autocast for single-GPU
+    amp_ctx = nullcontext() if ddp else amp_ctx  # type: ignore[possibly-undefined]
+
+    if master:
+        n_params = sum(p.numel() for p in model.parameters())
+        logger.info(f"Parameters: {n_params:,}  |  AMP dtype: {amp_dtype}")
+
+    # Compile for 20-30% speedup (requires PyTorch 2.0+)
+    try:
+        model = torch.compile(model, mode="reduce-overhead")
+        if master:
+            logger.info("torch.compile: enabled")
+    except Exception:
+        if master:
+            logger.info("torch.compile: not available, using eager mode")
+
+    # ------------------------------------------------------------------
+    # Optimizer
+    # ------------------------------------------------------------------
+    optimizer = torch.optim.AdamW(
+        model.parameters(), lr=lr, weight_decay=wd, betas=(0.9, 0.95), fused=True
+    )
+
+    # ------------------------------------------------------------------
+    # Resume from latest checkpoint (if any)
+    # ------------------------------------------------------------------
+    start_step = 0
+    start_epoch = 1
+    best_loss = float("inf")
+    existing_ckpts = [f for f in os.listdir(ckpt_dir) if f.startswith("spiderportal-v5-ep") and f.endswith(".pt") and "-step" not in f] if os.path.isdir(ckpt_dir) else []
+    if existing_ckpts:
+        latest = os.path.join(ckpt_dir, sorted(existing_ckpts)[-1])
+        if master:
+            logger.info(f"Resuming from checkpoint: {latest}")
+        start_step, start_epoch = load_checkpoint(model, optimizer, latest, ddp)
+        if master:
+            logger.success(f"Resumed at step {start_step}, epoch {start_epoch}")
+
+    # ------------------------------------------------------------------
+    # Dataset + DataLoader
+    # ------------------------------------------------------------------
+    dataset = FineWebEduDataset(tokenizer, seq_len, dataset_subset, rank, world_size)
+    loader = DataLoader(dataset, batch_size=micro_batch, num_workers=8, pin_memory=True, prefetch_factor=2)
+
+    # ------------------------------------------------------------------
+    # Training loop
+    # ------------------------------------------------------------------
+    if master:
+        os.makedirs(ckpt_dir, exist_ok=True)
+
+    model.train()
+    data_iter = iter(loader)
+    t0 = time.perf_counter()
+    step = start_step
+    epoch = start_epoch
+    step_ckpt_files = []
+    tokens_in_epoch = 0
+    tokens_per_epoch = target_tokens
+
+    while step < total_steps:
+        cur_lr = get_lr(step, warmup_steps, total_steps, lr, lr * 0.1)
+        for g in optimizer.param_groups:
+            g["lr"] = cur_lr
+
+        optimizer.zero_grad()
+        loss_accum = 0.0
+
+        for micro_step in range(grad_accum):
+            try:
+                x, y = next(data_iter)
+            except StopIteration:
+                data_iter = iter(loader)
+                x, y = next(data_iter)
+
+            x = x.to(device if not ddp else f"cuda:{local_rank}", non_blocking=True)
+            y = y.to(device if not ddp else f"cuda:{local_rank}", non_blocking=True)
+
+            sync = (
+                nullcontext()
+                if (not ddp or micro_step == grad_accum - 1)
+                else model.no_sync()
+            )
+            with sync, amp_ctx:
+                logits = model(x)
+                loss = nn.functional.cross_entropy(
+                    logits.view(-1, vocab_size), y.view(-1)
+                )
+                loss = loss / grad_accum
+
+            loss.backward()
+            loss_accum += loss.item()
+
+        if ddp:
+            grad_norm = model.clip_grad_norm_(1.0)
+        else:
+            grad_norm = nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
+        optimizer.step()
+        step += 1
+        tokens_in_epoch += global_batch_tok
+
+        if master and step % log_every == 0:
+            dt = time.perf_counter() - t0
+            tok_per_sec = global_batch_tok * log_every / dt
+            tokens_seen = step * global_batch_tok
+            logger.info(
+                f"Epoch {epoch} | step {step:6d}/{total_steps} | loss {loss_accum:.4f} "
+                f"| gnorm {float(grad_norm):.2f} | lr {cur_lr:.2e} "
+                f"| {tok_per_sec / 1e6:.2f}M tok/s "
+                f"| {tokens_seen / 1e9:.2f}B tokens seen"
+            )
+            t0 = time.perf_counter()
+
+        if step % ckpt_every == 0 and master:
+            ckpt_path, size_mb = save_weights_only(model, step, epoch, ckpt_dir, ddp)
+            step_ckpt_files.append(ckpt_path)
+            logger.info(f"Saved weights-only: {os.path.basename(ckpt_path)} ({size_mb:.0f}MB)")
+
+        if tokens_in_epoch >= tokens_per_epoch:
+            epoch_loss = loss_accum
+            if master:
+                epoch_time = (time.perf_counter() - t0) / 60
+                logger.info(f"Epoch {epoch} complete | loss={epoch_loss:.4f} | Time: {epoch_time:.1f}min")
+
+                for f in step_ckpt_files:
+                    if os.path.exists(f):
+                        os.remove(f)
+                        logger.info(f"  Deleted step checkpoint: {os.path.basename(f)}")
+                step_ckpt_files.clear()
+
+                ckpt_path, size_mb = save_full_checkpoint(model, optimizer, step, epoch, cfg, vocab_size, ckpt_dir, ddp, master, f"ep{epoch}")
+                if ckpt_path:
+                    logger.info(f"Saved epoch checkpoint: {os.path.basename(ckpt_path)} ({size_mb:.0f}MB)")
+
+                if epoch_loss < best_loss:
+                    best_loss = epoch_loss
+                    ckpt_path, size_mb = save_full_checkpoint(model, optimizer, step, epoch, cfg, vocab_size, ckpt_dir, ddp, master, "best")
+                    if ckpt_path:
+                        logger.info(f"Saved best checkpoint: {os.path.basename(ckpt_path)} ({size_mb:.0f}MB)")
+
+            epoch += 1
+            tokens_in_epoch = 0
+
+    if step > start_step and master:
+        ckpt_path, size_mb = save_full_checkpoint(model, optimizer, step, epoch, cfg, vocab_size, ckpt_dir, ddp, master, f"final-ep{epoch}")
+        if ckpt_path:
+            logger.info(f"Saved final checkpoint: {os.path.basename(ckpt_path)} ({size_mb:.0f}MB)")
+
+    if ddp:
+        # Barrier so no rank exits while another is still finishing its
+        # checkpoint gather — avoids NCCL "process group destroyed" noise.
+        dist.barrier()
+        dist.destroy_process_group()
+
+    if master:
+        logger.success("Training complete.")
+
+
+if __name__ == "__main__":
+    main()