Upload mythos-fineweb-dense.py with huggingface_hub

mythos-fineweb-dense.py

@@ -0,0 +1,1237 @@
+#!/usr/bin/env python3
+"""
+SpiderPortal v5-Dense: English pretraining on FineWeb-Edu with AdamW.
+
+Architecture: RDT (2 prelude + 6 recurrent + 2 coda) with:
+- MLA (Multi-Latent Attention): 10.7x KV cache compression + sliding window
+- Engram conditional memory at recurrent layers 1 and 4
+- Dense FFN (all params active, MoE conversion in Phase 2)
+- LTI Injection + ACT Halting + LoRA Adapter
+- 32k context (extendable to 256k at inference via YaRN)
+
+Config: hidden_size=2048, 6 recurrent layers, 32 experts (Phase 2), top-2 routing
+
+Single GPU:
+    python mythos-fineweb-dense.py
+
+Multi-GPU:
+    torchrun --nproc_per_node=$(python -c "import torch; print(torch.cuda.device_count())") mythos-fineweb-dense.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
+import sys
+
+# Configure loguru to file + stderr
+LOG_FILE = "train_spiderportal.log"
+logger.remove()
+logger.add(sys.stderr, format="{time:YYYY-MM-DD HH:mm:ss} | {level} | {message}")
+logger.add(LOG_FILE, rotation="100 MB", retention="10 days", format="{time:YYYY-MM-DD HH:mm:ss} | {level} | {message}")
+
+# Speed up CUDA memory allocation
+import torch
+torch.cuda.empty_cache()
+
+# Numba CPU fallback
+from triton_kernels import (
+    numba_dispatch,
+    NUMBA_AVAILABLE as _NUMBA_OK,
+)
+from numba_cuda_kernels import (
+    cuda_engram_hash,
+    cuda_engram_gate,
+    cuda_act_halting,
+    cuda_engram_conv1d,
+    cuda_available as _CUDA_OK,
+)
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
+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, field
+from typing import Optional, Tuple, Dict, List
+from torch.nn import CrossEntropyLoss
+from datasets import load_dataset
+from transformers import AutoTokenizer
+
+
+# ---------------------------------------------------------------------------
+# SpiderPortal Model Architecture (Dense + MLA + Engram)
+# ---------------------------------------------------------------------------
+
+@dataclass
+class SpiderPortalConfig:
+    vocab_size: int = 50257
+    hidden_size: int = 2048
+    num_hidden_layers: int = 6
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 4
+    intermediate_size: int = 8192
+    hidden_act: str = "silu"
+    num_experts: int = 32
+    num_experts_per_tok: int = 2
+    num_shared_experts: int = 1
+    router_aux_loss_coef: float = 0.05
+    max_loop_iters: int = 2
+    act_threshold: float = 0.5
+    max_position_embeddings: int = 32768
+    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 = 128
+    loop_embed_dim: int = 128
+    vision_hidden_size: int = 2048
+    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"
+
+    # MLA parameters (DeepSeek-V2 style, scaled for hidden_size=2048)
+    kv_lora_rank: int = 128
+    q_lora_rank: int = 256
+    qk_rope_head_dim: int = 64
+    qk_nope_head_dim: int = 64
+    v_head_dim: int = 64
+
+    # Engram parameters (DeepSeek conditional memory)
+    engram_layers: List[int] = field(default_factory=lambda: [1, 4])
+    engram_ngram_orders: Tuple[int, ...] = (2, 3)
+    engram_hash_heads: int = 4
+    engram_table_size: int = 65537  # prime number for hash table
+    engram_conv_kernel: int = 4
+    engram_conv_dilation: int = 3
+    engram_dim: int = 128  # per-head embedding dimension
+
+
+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):
+        # bf16-only RMSNorm: no dtype conversions inside forward.
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states
+
+
+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
+
+
+# ---------------------------------------------------------------------------
+# MLA: Multi-Latent Attention (DeepSeek-V2 style) + Sliding Window
+# ---------------------------------------------------------------------------
+
+class SpiderPortalMLA(nn.Module):
+    """Multi-Latent Attention with compressed KV cache and sliding window.
+
+    For hidden_size=2048, num_heads=16:
+      - qk_nope_head_dim=64, qk_rope_head_dim=64 → total head_dim=128
+      - kv_lora_rank=128 → 10.7x compression vs full 2048-dim KV
+      - v_head_dim=64 → value projection
+      - sliding_window=4096 → local attention range
+    """
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.kv_lora_rank = config.kv_lora_rank
+        self.q_lora_rank = config.q_lora_rank
+        self.qk_rope_head_dim = config.qk_rope_head_dim
+        self.qk_nope_head_dim = config.qk_nope_head_dim
+        self.v_head_dim = config.v_head_dim
+        self.head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
+        self.sliding_window = getattr(config, 'sliding_window', None)
+
+        # Q projection: optional low-rank → full Q
+        if self.q_lora_rank > 0:
+            self.q_a_proj = nn.Linear(config.hidden_size, self.q_lora_rank, bias=False)
+            self.q_a_layernorm = SpiderPortalRMSNorm(self.q_lora_rank)
+            self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * self.head_dim, bias=False)
+        else:
+            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
+
+        # KV compression: hidden → kv_lora_rank (shared latent)
+        self.kv_a_proj_with_mqa = nn.Linear(config.hidden_size, self.kv_lora_rank + self.qk_rope_head_dim, bias=False)
+        self.kv_a_layernorm = SpiderPortalRMSNorm(self.kv_lora_rank)
+        # Decompress: kv_lora_rank → nope heads + v heads
+        self.kv_b_proj = nn.Linear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+        )
+        # Output projection
+        self.o_proj = nn.Linear(self.num_heads * self.v_head_dim, config.hidden_size, bias=False)
+
+        # RoPE frequencies
+        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.qk_rope_head_dim, config.rope_theta, factor, orig_max_pos)
+        else:
+            inv_freq = 1.0 / (config.rope_theta ** (torch.arange(0, self.qk_rope_head_dim, 2).float() / self.qk_rope_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 _make_sliding_window_mask(self, q_len, kv_len, device, dtype):
+        """Unused: sliding_window disabled."""
+        return None
+
+    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):
+        bsz, q_len, _ = hidden_states.size()
+        # Q projection
+        if self.q_lora_rank > 0:
+            q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+        else:
+            q = self.q_proj(hidden_states)
+        q = q.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        q_nope, q_rope = torch.split(q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        # KV: compress to latent, then decompress
+        kv_hidden = self.kv_a_proj_with_mqa(hidden_states)
+        kv_latent, k_rope = torch.split(kv_hidden, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        kv_latent_norm = self.kv_a_layernorm(kv_latent)
+        kv_b_out = self.kv_b_proj(kv_latent_norm)
+        k_nope, v = torch.split(kv_b_out, [self.num_heads * self.qk_nope_head_dim, self.num_heads * self.v_head_dim], dim=-1)
+
+        k_nope = k_nope.view(bsz, q_len, self.num_heads, self.qk_nope_head_dim).transpose(1, 2)
+        v = v.view(bsz, q_len, self.num_heads, self.v_head_dim).transpose(1, 2)
+        k_rope = k_rope.unsqueeze(1)
+
+        # RoPE on Q and K rope parts
+        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_full = cos[position_ids].unsqueeze(1)
+        sin_full = sin[position_ids].unsqueeze(1)
+
+        q_rope = self._apply_rotary(q_rope, cos_full, sin_full)
+        k_rope = self._apply_rotary(k_rope, cos_full, sin_full)
+
+        # Assemble full K
+        k_rope_expanded = k_rope.expand(-1, self.num_heads, -1, -1)
+        k_full = torch.cat([k_nope, k_rope_expanded], dim=-1)
+        q_full = torch.cat([q_nope, q_rope], dim=-1)
+
+        # KV cache
+        if past_key_value is not None:
+            k_full = torch.cat([past_key_value[0], k_full], dim=2)
+            v = torch.cat([past_key_value[1], v], dim=2)
+        past_kv = (k_full, v) if use_cache else None
+
+        # Attention with SDPA — is_causal=True for flash-attention fast path
+        # No 4D causal mask needed; sliding window disabled, so pure causal.
+        attn_output = F.scaled_dot_product_attention(
+            q_full, k_full, v,
+            attn_mask=None,
+            dropout_p=self.config.attention_dropout if self.training else 0.0,
+            is_causal=True,
+        )
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
+        return self.o_proj(attn_output), past_kv
+
+
+# ---------------------------------------------------------------------------
+# Engram: Conditional Memory via Scalable Lookup (DeepSeek style)
+# ---------------------------------------------------------------------------
+
+def _tokenizer_compress(token_ids, vocab_size=50257):
+    """Simulate NFKC + lowercase canonical ID projection."""
+    return token_ids % (vocab_size * 77 // 100)
+
+
+class SpiderPortalEngram(nn.Module):
+    """Conditional memory module via NN-gram lookup.
+
+    Applied only at specific recurrent layers (config.engram_layers).
+    """
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.ngram_orders = list(config.engram_ngram_orders)
+        self.num_heads_per_order = config.engram_hash_heads
+        self.table_size = config.engram_table_size
+        self.d_mem = config.engram_dim
+
+        self.total_mem_dim = len(self.ngram_orders) * self.num_heads_per_order * self.d_mem
+
+        # Stacked embedding table with offsets: [orders, heads, table_size, d_mem]
+        # This matches the deepseek MultiHeadEmbedding principle.
+        self.embed = nn.Parameter(
+            torch.randn(len(self.ngram_orders), self.num_heads_per_order, self.table_size, self.d_mem) * 0.02
+        )
+
+        # Seeds per (order, head) in a stable head_counter ordering.
+        seeds = []
+        for _order in self.ngram_orders:
+            for h in range(self.num_heads_per_order):
+                seeds.append((h + 1) * 2654435761)
+        self.register_buffer("hash_seeds", torch.tensor(seeds, dtype=torch.int64), persistent=False)
+
+        self.W_k = nn.Linear(self.total_mem_dim, config.hidden_size, bias=False)
+        self.W_v = nn.Linear(self.total_mem_dim, config.hidden_size, bias=False)
+
+        self.conv = nn.Conv1d(
+            config.hidden_size, config.hidden_size,
+            kernel_size=config.engram_conv_kernel,
+            padding=config.engram_conv_kernel - 1,
+            groups=config.hidden_size,
+        )
+        self.conv_dilation = config.engram_conv_dilation
+
+        with torch.no_grad():
+            self.conv.weight.zero_()
+            if self.conv.bias is not None:
+                self.conv.bias.zero_()
+
+        self.q_norm = SpiderPortalRMSNorm(config.hidden_size)
+        self.k_norm = SpiderPortalRMSNorm(config.hidden_size)
+
+        # No caching: required for gradient checkpoint recomputation stability.
+        self._fwd_cache = None
+
+    def _compute_indices(self, compressed_ids, n, head_idx):
+        """Vectorized NN-gram hash indices for a single (order, head)."""
+        # Kept for backward compatibility; not used in the stacked embedding path.
+        bsz, seq_len = compressed_ids.shape
+        pad = torch.zeros(bsz, n - 1, dtype=compressed_ids.dtype, device=compressed_ids.device)
+        padded = torch.cat([pad, compressed_ids], dim=1)
+
+        indices_list = []
+        for i in range(n):
+            indices_list.append(padded[:, i:i + seq_len])
+        ngrams = torch.stack(indices_list, dim=-1)
+
+        seed = int(self.hash_seeds[head_idx].item())
+        h_val = torch.zeros(bsz, seq_len, dtype=torch.int64, device=compressed_ids.device)
+        for i in range(n):
+            h_val = h_val * 31 + ngrams[:, :, i]
+            h_val = h_val % self.table_size
+        h_val = (h_val * seed) % self.table_size
+        return h_val
+
+    def _compute_hash(self, compressed, n, head_counter, bsz, seq_len):
+        """Compute n-gram hash indices, with Numba CPU fallback."""
+        if not compressed.is_cuda and NUMBA_AVAILABLE:
+            import numpy as np
+            h_val_np = numba_dispatch(
+                "hash_indices",
+                compressed.cpu().numpy().astype(np.int64),
+                n, self.table_size,
+                int(self.hash_seeds[head_counter].item()),
+            )
+            if h_val_np is not None:
+                return torch.from_numpy(h_val_np).to(compressed.device)
+
+        pad = torch.zeros(bsz, n - 1, dtype=compressed.dtype, device=compressed.device)
+        padded = torch.cat([pad, compressed], dim=1)
+        ngrams = torch.stack([padded[:, i : i + seq_len] for i in range(n)], dim=-1)
+        h_val = torch.zeros(bsz, seq_len, dtype=torch.int64, device=compressed.device)
+        for i in range(n):
+            h_val = h_val * 31 + ngrams[:, :, i].to(torch.int64)
+            h_val = h_val % self.table_size
+        return h_val
+
+    def _retrieve(self, token_ids):
+        """Retrieve memory vectors for a batch of token sequences."""
+        bsz, seq_len = token_ids.shape
+        compressed = _tokenizer_compress(token_ids)
+
+        # Use Numba CUDA hash if faster (PyTorch path is default, ~0.2ms per call)
+        indices = cuda_engram_hash(
+            compressed, self.hash_seeds,
+            self.ngram_orders, self.num_heads_per_order, self.table_size,
+        )
+        if indices is not None:
+            all_parts = []
+            head_counter = 0
+            for order_idx, n in enumerate(self.ngram_orders):
+                head_indices = indices[:, :, head_counter:head_counter + self.num_heads_per_order]
+                emb_table = self.embed[order_idx]
+                idx = head_indices.permute(0, 2, 1).unsqueeze(-1).expand(-1, -1, -1, self.d_mem)
+                mem = torch.gather(emb_table.unsqueeze(0).expand(bsz, -1, -1, -1), dim=2, index=idx)
+                mem = mem.permute(0, 2, 1, 3).reshape(bsz, seq_len, self.num_heads_per_order * self.d_mem)
+                all_parts.append(mem)
+                head_counter += self.num_heads_per_order
+            return torch.cat(all_parts, dim=-1)
+
+        # PyTorch fallback (CPU or if CUDA kernel unavailable)
+        all_parts = []
+        head_counter = 0
+        for order_idx, n in enumerate(self.ngram_orders):
+            h_val = self._compute_hash(compressed, n, head_counter, bsz, seq_len)
+            seeds_slice = self.hash_seeds[head_counter : head_counter + self.num_heads_per_order]
+            indices_pt = (h_val.unsqueeze(-1) * seeds_slice.view(1, 1, -1)) % self.table_size
+            emb_table = self.embed[order_idx]
+            idx = indices_pt.permute(0, 2, 1).unsqueeze(-1).expand(-1, -1, -1, self.d_mem)
+            mem = torch.gather(emb_table.unsqueeze(0).expand(bsz, -1, -1, -1), dim=2, index=idx)
+            mem = mem.permute(0, 2, 1, 3).reshape(bsz, seq_len, self.num_heads_per_order * self.d_mem)
+            all_parts.append(mem)
+            head_counter += self.num_heads_per_order
+        return torch.cat(all_parts, dim=-1)
+
+    def forward(self, hidden_states, token_ids, layer_id: int):
+        mem = self._retrieve(token_ids)
+
+        q = hidden_states
+        k = self.W_k(mem)
+        v = self.W_v(mem)
+        q_norm = self.q_norm(q)
+        k_norm = self.k_norm(k)
+        alpha = torch.sigmoid(
+            (q_norm * k_norm).sum(dim=-1, keepdim=True) / math.sqrt(q.shape[-1])
+        )
+        v_gated = alpha * v
+        v_gated_t = v_gated.transpose(1, 2)
+        conv_out = self.conv(v_gated_t)
+        conv_out = conv_out[:, :, :v_gated_t.shape[-1]]
+        conv_out = conv_out.transpose(1, 2)
+
+        y = F.silu(conv_out) + v_gated
+        return y
+
+
+# ---------------------------------------------------------------------------
+# FFN Expert (dense)
+# ---------------------------------------------------------------------------
+
+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))
+
+
+# ---------------------------------------------------------------------------
+# Prelude/Coda Dense Layer (uses MLA)
+# ---------------------------------------------------------------------------
+
+class SpiderPortalDenseLayer(nn.Module):
+    """Prelude/coda dense layer with MLA attention."""
+    def __init__(self, config):
+        super().__init__()
+        self.self_attn = SpiderPortalMLA(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
+
+
+# ---------------------------------------------------------------------------
+# Recurrent Dense Layer (uses MLA + optional Engram)
+# ---------------------------------------------------------------------------
+
+class SpiderPortalRecurrentDenseLayer(nn.Module):
+    """Recurrent layer with MLA attention and optional Engram memory."""
+    def __init__(self, config, layer_idx, has_engram=False):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.has_engram = has_engram
+        self.self_attn = SpiderPortalMLA(config)
+        if has_engram:
+            self.engram = SpiderPortalEngram(config)
+        self.ffn = SpiderPortalExpert(config, intermediate_size=config.intermediate_size)
+        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)
+        self.post_engram_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps) if has_engram else None
+    def forward(self, hidden_states, token_ids=None, 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
+
+        if self.has_engram and token_ids is not None:
+            engram_out = self.engram(hidden_states, token_ids, layer_id=self.layer_idx)
+            hidden_states = hidden_states + engram_out
+            if self.post_engram_layernorm is not None:
+                hidden_states = self.post_engram_layernorm(hidden_states)
+
+        ffn_input = self.post_attention_layernorm(hidden_states)
+        ffn_output = self.ffn(ffn_input)
+        hidden_states = hidden_states + ffn_output
+        return hidden_states, 0.0, past_kv
+
+
+# ---------------------------------------------------------------------------
+# LTI Injection, ACT Halting, LoRA Adapter
+# ---------------------------------------------------------------------------
+
+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
+
+
+def checkpoint(func, *args, **kwargs):
+    """Gradient checkpointing wrapper — saves VRAM at ~20% compute cost."""
+    if torch.is_grad_enabled():
+        return torch.utils.checkpoint.checkpoint(func, *args, use_reentrant=False, **kwargs)
+    return func(*args, **kwargs)
+
+
+# ---------------------------------------------------------------------------
+# Full Model
+# ---------------------------------------------------------------------------
+
+class SpiderPortalDenseModel(nn.Module):
+    """Full RDT model with MLA attention + Engram memory at layers 1,4.
+
+    Architecture:
+        2x Prelude (MLA + dense FFN)
+        6x Recurrent (MLA + Engram@L1,L4 + dense FFN) — with gradient checkpointing
+        2x Coda (MLA + dense FFN)
+        LTI Injection + ACT Halting + LoRA Adapter
+    """
+    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([
+            SpiderPortalRecurrentDenseLayer(config, i, has_engram=(i in config.engram_layers))
+            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, token_ids=None):
+        n_loops = n_loops or 1
+        input_embedding = input_embedding if input_embedding is not None else hidden_states
+        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 = checkpoint(
+                    layer, hidden_states,
+                    token_ids=token_ids,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values[i] if t == 0 else None,
+                    use_cache=use_cache
+                )
+                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, 0.0, 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 = SpiderPortalDenseModel(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()
+        hidden_states, aux_loss, past_kv = self.model(
+            hidden_states, input_embedding=input_embedding,
+            attention_mask=None, position_ids=position_ids,
+            use_cache=use_cache, n_loops=n_loops, token_ids=input_ids
+        )
+        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))
+        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):
+    def __init__(self, tokenizer, seq_len: int, subset: str, rank: int, world_size: int):
+        self.tokenizer = tokenizer
+        self.seq_len = seq_len
+        self.subset = subset
+        self.rank = rank
+        self.world_size = world_size
+        
+        # Local tokenized data - USE mmapped binary for speed
+        LOCAL_TOKEN_FILE = "/data/fineweb_tokenized/train_tokens.bin"
+        
+        if os.path.exists(LOCAL_TOKEN_FILE):
+            # Use memory-mapped file for fast I/O
+            import numpy as np
+            self.use_local = True
+            self.local_file = LOCAL_TOKEN_FILE
+            # Memory map for zero-copy reading
+            self.mm = np.memmap(LOCAL_TOKEN_FILE, dtype='<u4', mode='r')
+            self.num_tokens = len(self.mm)
+            self.num_samples = self.num_tokens // seq_len
+        else:
+            self.use_local = False
+            
+    def __iter__(self):
+        if self.use_local:
+            # Fast: use memory-mapped array
+            worker = get_worker_info()
+            num_workers = worker.num_workers if worker else 1
+            worker_id = worker.id if worker else 0
+            
+            samples_per_worker = self.num_samples // (self.world_size * num_workers)
+            start_sample = (self.rank * num_workers + worker_id) * samples_per_worker
+            end_sample = start_sample + samples_per_worker
+            
+            # Batch read tokens - convert to numpy array slice then tensor
+            import numpy as np
+            for i in range(start_sample, end_sample):
+                start_idx = i * self.seq_len
+                # Direct slice from memory-mapped array (avoid extra copies when possible)
+                tokens = self.mm[start_idx : start_idx + self.seq_len + 1]
+                x_np = tokens[:-1].astype("int64", copy=False)
+                y_np = tokens[1:].astype("int64", copy=False)
+                yield torch.from_numpy(x_np), torch.from_numpy(y_np)
+        else:
+            # Fallback to HuggingFace
+            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
+# ---------------------------------------------------------------------------
+
+def get_lr(step: int, warmup: int, total: int, max_lr: float, min_lr: float) -> float:
+    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
+# ---------------------------------------------------------------------------
+
+def save_weights_only(model, step, epoch, ckpt_dir, ddp):
+    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-dense-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"):
+    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-dense-{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 load_checkpoint(model, optimizer, path, ddp):
+    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():
+    # ------------------------------------------------------------------
+    # 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
+
+    # ------------------------------------------------------------------
+    # 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 = int(os.environ.get("SEQ_LEN", "2048"))
+    micro_batch = int(os.environ.get("MICRO_BATCH", "32"))
+    target_tokens = int(os.environ.get("TARGET_TOKENS", "50_000_000"))
+    grad_accum = int(os.environ.get("GRAD_ACCUM", "1"))
+    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 = int(os.environ.get("CKPT_EVERY", "500"))
+    ckpt_dir = "checkpoints-dense"
+    dataset_subset = "sample-10BT"
+
+    if master:
+        logger.info(
+            f"[DENSE MLA+Engram] hidden=2048 | layers=6 | seq_len={seq_len} | micro_batch={micro_batch} | grad_accum={grad_accum} | "
+            f"global_batch_tokens={global_batch_tok:,} | total_steps={total_steps:,}"
+        )
+        logger.info(
+            "Attention: MLA (sliding_window disabled) | "
+            "Engram: layers [1,4] | Context: 32k | "
+            "Gradient checkpointing: enabled"
+        )
+
+    # ------------------------------------------------------------------
+    # Model
+    # ------------------------------------------------------------------
+    cfg = SpiderPortalConfig(
+        hidden_size=2048, num_hidden_layers=6, num_attention_heads=16,
+        num_key_value_heads=4, intermediate_size=4096,
+        num_experts=32, num_experts_per_tok=2, num_shared_experts=1,
+        router_aux_loss_coef=0.05, max_loop_iters=2,
+        prelude_layers=2, coda_layers=2, lora_rank=128,
+        rope_theta=10000000.0,
+        rope_scaling=None,
+        max_position_embeddings=32768,
+        sliding_window=0,
+        tie_word_embeddings=True,
+        kv_lora_rank=128, q_lora_rank=256,
+        qk_rope_head_dim=64, qk_nope_head_dim=64, v_head_dim=64,
+        engram_layers=[1, 4],
+        engram_ngram_orders=(2, 3),
+        engram_hash_heads=4,
+        engram_table_size=65537,
+        engram_dim=128,
+    )
+    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).to(torch.bfloat16)
+
+    if ddp:
+        mp_policy = MixedPrecision(
+            param_dtype=amp_dtype,
+            reduce_dtype=amp_dtype,
+            buffer_dtype=amp_dtype,
+        )
+        wrap_policy = ModuleWrapPolicy({SpiderPortalDenseLayer, SpiderPortalRecurrentDenseLayer})
+        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)
+        # torch.compile disabled - causes issues with custom model
+        # model = torch.compile(model, mode="reduce-overhead")
+
+        if os.environ.get("SKIP_MXFP8", "0") == "1":
+            if master:
+                logger.info("MXFP8 disabled via SKIP_MXFP8=1 — using native bf16")
+        else:
+            # Apply MXFP8 (emulated) via _to_mxfp8_then_scaled_mm autograd Function.
+            # Uses single Function node (mem-efficient) with EMULATED kernels for compute 12.0 compat.
+            # Weights stay bf16; quantize/dequantize wraps mm for fp8 bandwidth savings.
+            try:
+                from torchao.prototype.mx_formats.mx_linear import _to_mxfp8_then_scaled_mm
+                from torchao.prototype.mx_formats.config import ScaleCalculationMode
+                from torchao.quantization.quantize_.common.kernel_preference import KernelPreference
+
+                # MXFP8: try AUTO (native cuBLAS) first, fall back to EMULATED
+                _mxfp8_kernel = KernelPreference.EMULATED
+                try:
+                    _t_a = torch.randn(32, 2048, device='cuda', dtype=torch.bfloat16)
+                    _t_b = torch.randn(32, 2048, device='cuda', dtype=torch.bfloat16)
+                    _t_a_fp8 = _t_a.to(torch.float8_e4m3fn)
+                    _t_b_fp8 = _t_b.to(torch.float8_e4m3fn)
+                    _sa = torch.ones(32, 64, dtype=torch.float8_e8m0fnu, device='cuda')
+                    _sb = torch.ones(32, 64, dtype=torch.float8_e8m0fnu, device='cuda')
+                    from torchao.prototype.mx_formats.utils import to_blocked
+                    _sa_b = to_blocked(_sa)
+                    _sb_b = to_blocked(_sb)
+                    _t_out = torch._scaled_mm(_t_a_fp8, _t_b_fp8.t(), scale_a=_sa_b, scale_b=_sb_b, out_dtype=torch.bfloat16)
+                    _t_ref = _t_a @ _t_b.t()
+                    if (_t_out.float() - _t_ref.float()).abs().max().item() < 10.0:
+                        _mxfp8_kernel = KernelPreference.AUTO
+                        if master:
+                            logger.info("Native MXFP8 _scaled_mm available!")
+                except Exception:
+                    pass
+
+                class MXFP8Linear(nn.Module):
+                    def __init__(self, linear: nn.Linear):
+                        super().__init__()
+                        self.weight = nn.Parameter(linear.weight.data.clone().to(torch.bfloat16))
+                        self.bias = nn.Parameter(linear.bias.data.clone().to(torch.bfloat16)) if linear.bias is not None else None
+                        self.in_features = linear.in_features
+                        self.out_features = linear.out_features
+
+                    def forward(self, x: torch.Tensor) -> torch.Tensor:
+                        x = x.to(torch.bfloat16)
+                        in_dim = x.shape[-1]
+                        pad = (32 - in_dim % 32) % 32
+                        if pad:
+                            x = F.pad(x, (0, pad))
+                            w = F.pad(self.weight, (0, pad))
+                        else:
+                            w = self.weight
+                        orig = x.shape
+                        x_2d = x.reshape(-1, x.shape[-1])
+                        out = _to_mxfp8_then_scaled_mm(
+                            x_2d, w,
+                            kernel_preference=_mxfp8_kernel,
+                            scale_calculation_mode=ScaleCalculationMode.RCEIL,
+                            wgrad_with_hp=False,
+                        )
+                        out = out.reshape(*orig[:-1], out.shape[-1])
+                        if self.bias is not None:
+                            out = out + self.bias
+                        return out
+
+                total = sum(1 for _ in model.modules() if isinstance(_, nn.Linear))
+                count = 0
+                for name, mod in list(model.named_modules()):
+                    if isinstance(mod, nn.Linear):
+                        if mod.in_features % 32 != 0 or mod.out_features % 32 != 0:
+                            continue
+                        parent_name, _, child_name = name.rpartition(".")
+                        parent = model.get_submodule(parent_name) if parent_name else model
+                        setattr(parent, child_name, MXFP8Linear(mod))
+                        count += 1
+                if master:
+                    _mode = "AUTO" if _mxfp8_kernel == KernelPreference.AUTO else "EMULATED"
+                    logger.info(f"MXFP8 ({_mode}) quantized {count}/{total} Linear layers")
+            except Exception as e:
+                    if master:
+                        logger.info(f"MXFP8 not available, bf16 fallback: {e}")
+                    import traceback; traceback.print_exc()
+
+    amp_ctx = (
+            torch.amp.autocast(device_type="cuda", dtype=amp_dtype)
+            if "cuda" in device
+            else nullcontext()
+        )
+
+    amp_ctx = nullcontext() if ddp else amp_ctx
+
+    # Enable SDPA best kernels when available.
+    try:
+        from torch.nn.attention import sdpa_kernel
+
+        sdpa_ctx = sdpa_kernel(enable_flash=True, enable_mem_efficient=True, enable_math=True)
+    except Exception:
+        sdpa_ctx = nullcontext()
+
+    if master:
+        n_params = sum(p.numel() for p in model.parameters())
+        engram_params = sum(p.numel() for n, p in model.named_parameters() if 'engram' in n)
+        mla_params = sum(p.numel() for n, p in model.named_parameters() if 'self_attn' in n)
+        embed_params = sum(p.numel() for n, p in model.named_parameters() if 'embed_tokens' in n or 'lm_head' in n)
+        ffn_params = sum(p.numel() for n, p in model.named_parameters() if 'ffn' in n or 'gate_proj' in n or 'up_proj' in n or 'down_proj' in n)
+        other_params = n_params - engram_params - mla_params - embed_params - ffn_params
+        logger.info(
+            f"Parameters: {n_params:,} (all active) | "
+            f"Embeddings: {embed_params:,} | MLA: {mla_params:,} | "
+            f"FFN: {ffn_params:,} | Engram: {engram_params:,} | "
+            f"Other: {other_params:,} | AMP dtype: {amp_dtype}"
+        )
+
+    # ------------------------------------------------------------------
+    # Optimizer — dual optimizer for Engram embeddings
+    # ------------------------------------------------------------------
+    engram_params_list = [p for n, p in model.named_parameters() if 'engram' in n and 'embed' in n and 'proj' not in n]
+    backbone_params = [p for n, p in model.named_parameters() if not ('engram' in n and 'embed' in n and 'proj' not in n)]
+
+    optimizer = torch.optim.AdamW(
+        backbone_params, lr=lr, weight_decay=wd, betas=(0.9, 0.95), fused=False, foreach=True, eps=1e-8
+    )
+    if engram_params_list:
+        engram_optimizer = torch.optim.Adam(
+            engram_params_list, lr=lr * 5, betas=(0.9, 0.95), eps=1e-8
+        )
+    else:
+        engram_optimizer = None
+
+    # ------------------------------------------------------------------
+    # Resume from latest checkpoint
+    # ------------------------------------------------------------------
+    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-dense-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=4, pin_memory=True, prefetch_factor=1)
+
+    # ------------------------------------------------------------------
+    # 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
+
+    # Allow env override for quick debugging.
+    max_steps_override = os.environ.get("MAX_STEPS", None)
+    while step < total_steps:
+        if max_steps_override is not None and step >= int(max_steps_override):
+            break
+        cur_lr = get_lr(step, warmup_steps, total_steps, lr, lr * 0.1)
+        for g in optimizer.param_groups:
+            g["lr"] = cur_lr
+        if engram_optimizer:
+            for g in engram_optimizer.param_groups:
+                g["lr"] = cur_lr * 5
+
+        optimizer.zero_grad()
+        if engram_optimizer:
+            engram_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, sdpa_ctx:
+                output = model(x)
+                if master and step == start_step and micro_step == 0:
+                    peak_vram = torch.cuda.max_memory_allocated() / 1024**3
+                    logger.info(f"Reached first model forward  |  Peak VRAM: {peak_vram:.1f}GB")
+                if isinstance(output, dict):
+                    logits = output["logits"]
+                else:
+                    logits = output
+                loss = nn.functional.cross_entropy(
+                    logits.view(-1, vocab_size), y.view(-1)
+                )
+                loss = loss / grad_accum
+
+            loss.backward()
+            if master and step == start_step and micro_step == 0:
+                logger.info("Reached first 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()
+        if engram_optimizer:
+            engram_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:
+        dist.barrier()
+        dist.destroy_process_group()
+
+    if master:
+        logger.success("Training complete.")
+
+
+if __name__ == "__main__":
+    main()