Delete AI-Transfer

AI-Transfer/README.md

@@ -1,81 +0,0 @@
-# SpiderPortal v5 — GPU Training Transfer Package
-
-## Contents
-- `scripts/train_single_gpu.py` — Optimized single-GPU training script
-- `scripts/train_ddp.py` — Original DDP script (for reference)
-- `data/spiderportal_combined.pkl` — Training dataset (491K samples, 537MB)
-- `notebooks/spiderportal_gpu.ipynb` — Original Kaggle notebook (for reference)
-
-## Quick Start
-
-### 1. Setup
-```bash
-pip install torch transformers pandas
-```
-
-### 2. Run Training
-```bash
-python scripts/train_single_gpu.py
-```
-
-## Configuration (edit in `train_single_gpu.py`)
-
-| Parameter | Default | RTX 6000 (96GB) |
-|-----------|---------|-----------------|
-| `BATCH_SIZE` | 64 | 64-128 |
-| `MAX_LEN` | 256 | 256-512 |
-| `EPOCHS` | 3 | 3 |
-| `N_LOOPS` | 2 | 2 |
-| `BASE_LR` | 2e-5 | 2e-5 |
-| `WARMUP_STEPS` | 1000 | 1000 |
-
-## Expected Performance (RTX PRO 6000, 96GB)
-
-| Metric | Value |
-|--------|-------|
-| Model size | 659M params |
-| Active params (MoE) | ~59M |
-| VRAM usage | ~15-25GB |
-| Batch size | 64 |
-| Steps per epoch | ~7,672 |
-| Time per epoch | 40-100 min |
-| **Total (3 epochs)** | **2-5 hours** |
-
-## Checkpoints
-
-Saved to `checkpoints/` directory:
-
-| Checkpoint type | What's saved | Size | Purpose |
-|----------------|--------------|------|---------|
-| Every 500 steps | Model weights only | ~1.3GB | Testing, transfer, inference |
-| End of epoch | Model + optimizer | ~6.6GB | Resume training |
-| Best loss | Model + optimizer | ~6.6GB | Resume training |
-
-Step checkpoints are auto-deleted at the start of each new epoch to free disk space.
-
-### Loading a weights-only checkpoint (inference)
-```python
-state_dict = torch.load("checkpoints/spiderportal-v5-ep1-step500.pt", map_location="cpu")
-model.load_state_dict(state_dict)
-```
-
-### Resuming training from an epoch checkpoint
-```python
-ckpt = torch.load("checkpoints/spiderportal-v5-ep1.pt", map_location="cpu")
-model.load_state_dict(ckpt["model_state_dict"])
-optimizer.load_state_dict(ckpt["optimizer_state_dict"])
-start_epoch = ckpt["epoch"]
-```
-
-### Peak disk usage during training: ~20GB
-
-## Model Architecture
-
-- **Type**: MoE-RDT (Mixture of Experts + Recurrent Depth Transformer)
-- **Total params**: 659M
-- **Active params**: ~59M (1 routed expert + 1 shared expert per token)
-- **Experts**: 64 routed + 1 shared
-- **Layers**: 2 prelude → 8 recurrent (MoE) → 2 coda
-- **Attention**: GQA (8 heads, 2 KV heads)
-- **Context**: 131K tokens (YaRN RoPE scaling)
-- **Loop**: ACT halting + LTI injection + LoRA adapters

AI-Transfer/data/spiderportal_combined.pkl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:706bb0870f785241df36c3dcc1925deb8a47ce6003dfe506b6c0751624c34e63
-size 563067610

AI-Transfer/notebooks/spiderportal_gpu.ipynb

@@ -1,466 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# SpiderPortal MoE-RDT v5 \u2014 Multi-GPU Training (DDP)\n",
-    "\n",
-    "Optimized for 2\u00d7 T4 GPUs (32GB total VRAM).\n",
-    "- **Total params**: ~659M (64 experts)\n",
-    "- **Active params**: ~59M per token\n",
-    "- **Training**: bf16, DDP, gradient accumulation\n",
-    "- **Expected time**: ~1-1.5 hours for 3 epochs"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!pip install -q transformers pandas safetensors\n",
-    "\n",
-    "import torch\n",
-    "import torch.nn as nn\n",
-    "import torch.nn.functional as F\n",
-    "import math, os, json, gc, random, time\n",
-    "from pathlib import Path\n",
-    "from dataclasses import dataclass\n",
-    "from typing import Optional, Tuple, Dict, List\n",
-    "from torch.nn import CrossEntropyLoss\n",
-    "\n",
-    "print(f\"PyTorch: {torch.__version__}\")\n",
-    "print(f\"CUDA available: {torch.cuda.is_available()}\")\n",
-    "print(f\"GPU count: {torch.cuda.device_count()}\")\n",
-    "for i in range(torch.cuda.device_count()):\n",
-    "    p = torch.cuda.get_device_properties(i)\n",
-    "    print(f\"  GPU {i}: {torch.cuda.get_device_name(i)} ({p.total_memory / 1e9:.1f}GB)\")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "@dataclass\n",
-    "class SpiderPortalConfig:\n",
-    "    vocab_size: int = 50278\n",
-    "    hidden_size: int = 384\n",
-    "    num_hidden_layers: int = 8\n",
-    "    num_attention_heads: int = 8\n",
-    "    num_key_value_heads: int = 2\n",
-    "    intermediate_size: int = 1024\n",
-    "    hidden_act: str = \"silu\"\n",
-    "    num_experts: int = 64\n",
-    "    num_experts_per_tok: int = 1\n",
-    "    num_shared_experts: int = 1\n",
-    "    router_aux_loss_coef: float = 0.05\n",
-    "    max_loop_iters: int = 4\n",
-    "    act_threshold: float = 0.5\n",
-    "    max_position_embeddings: int = 131072\n",
-    "    rope_theta: float = 10000000.0\n",
-    "    rope_scaling: dict = None\n",
-    "    sliding_window: int = 4096\n",
-    "    attention_dropout: float = 0.0\n",
-    "    rms_norm_eps: float = 1e-6\n",
-    "    initializer_range: float = 0.02\n",
-    "    use_cache: bool = True\n",
-    "    tie_word_embeddings: bool = True\n",
-    "    prelude_layers: int = 2\n",
-    "    coda_layers: int = 2\n",
-    "    lora_rank: int = 32\n",
-    "    loop_embed_dim: int = 48\n",
-    "    vision_hidden_size: int = 384\n",
-    "    audio_hidden_size: int = 512\n",
-    "    vision_num_frames: int = 60\n",
-    "    vision_tokens_per_frame: int = 256\n",
-    "    vision_temporal_tokens: int = 64\n",
-    "    vision_temporal_layers: int = 2\n",
-    "    model_type: str = \"spiderportal\"\n",
-    "    torch_dtype: str = \"bfloat16\"\n",
-    "\n",
-    "def loop_index_embedding(h, loop_t, loop_dim, theta=10000.0):\n",
-    "    freqs = 1.0 / (theta ** (torch.arange(0, loop_dim, 2, device=h.device, dtype=h.dtype) / loop_dim))\n",
-    "    angles = loop_t * freqs\n",
-    "    emb = torch.cat([angles.sin(), angles.cos()], dim=-1)[:loop_dim]\n",
-    "    emb_full = torch.zeros(h.shape[-1], device=h.device, dtype=h.dtype)\n",
-    "    emb_full[:loop_dim] = emb\n",
-    "    return h + emb_full.unsqueeze(0).unsqueeze(0)\n",
-    "\n",
-    "class SpiderPortalRMSNorm(nn.Module):\n",
-    "    def __init__(self, hidden_size, eps=1e-6):\n",
-    "        super().__init__()\n",
-    "        self.weight = nn.Parameter(torch.ones(hidden_size))\n",
-    "        self.variance_epsilon = eps\n",
-    "    def forward(self, hidden_states):\n",
-    "        input_dtype = hidden_states.dtype\n",
-    "        hidden_states = hidden_states.to(torch.float32)\n",
-    "        variance = hidden_states.pow(2).mean(-1, keepdim=True)\n",
-    "        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)\n",
-    "        return self.weight.to(input_dtype) * hidden_states.to(input_dtype)\n",
-    "\n",
-    "def compute_yarn_inv_freq(head_dim, rope_theta, factor, orig_max, beta_fast=32.0, beta_slow=1.0):\n",
-    "    dim = head_dim\n",
-    "    orig_inv_freq = 1.0 / (rope_theta ** (torch.arange(0, dim, 2).float() / dim))\n",
-    "    pos_freqs = torch.arange(0, dim, 2).float() / dim\n",
-    "    beta = (pos_freqs * math.log(rope_theta) / math.log(orig_max))\n",
-    "    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)))\n",
-    "    return orig_inv_freq * scale\n",
-    "\n",
-    "class SpiderPortalGQA(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.config = config\n",
-    "        self.hidden_size = config.hidden_size\n",
-    "        self.num_heads = config.num_attention_heads\n",
-    "        self.num_kv_heads = config.num_key_value_heads\n",
-    "        self.head_dim = config.hidden_size // config.num_attention_heads\n",
-    "        self.num_key_value_groups = self.num_heads // self.num_kv_heads\n",
-    "        self.q_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)\n",
-    "        self.k_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)\n",
-    "        self.v_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)\n",
-    "        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)\n",
-    "        self.attention_dropout = config.attention_dropout\n",
-    "        rope_scaling = getattr(config, 'rope_scaling', None)\n",
-    "        if rope_scaling and rope_scaling.get(\"type\") == \"yarn\":\n",
-    "            factor = rope_scaling.get(\"factor\", 1.0)\n",
-    "            orig_max_pos = rope_scaling.get(\"original_max_position_embeddings\", config.max_position_embeddings)\n",
-    "            inv_freq = compute_yarn_inv_freq(self.head_dim, config.rope_theta, factor, orig_max_pos)\n",
-    "        else:\n",
-    "            inv_freq = 1.0 / (config.rope_theta ** (torch.arange(0, self.head_dim, 2).float() / self.head_dim))\n",
-    "        self.register_buffer(\"inv_freq\", inv_freq, persistent=False)\n",
-    "    def _rotate_half(self, x):\n",
-    "        x1 = x[..., :x.shape[-1] // 2]\n",
-    "        x2 = x[..., x.shape[-1] // 2:]\n",
-    "        return torch.cat((-x2, x1), dim=-1)\n",
-    "    def _apply_rotary(self, x, cos, sin):\n",
-    "        return (x * cos) + (self._rotate_half(x) * sin)\n",
-    "    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):\n",
-    "        bsz, q_len, _ = hidden_states.size()\n",
-    "        query_states = self.q_proj(hidden_states)\n",
-    "        key_states = self.k_proj(hidden_states)\n",
-    "        value_states = self.v_proj(hidden_states)\n",
-    "        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)\n",
-    "        key_states = key_states.view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)\n",
-    "        value_states = value_states.view(bsz, q_len, self.num_kv_heads, self.head_dim).transpose(1, 2)\n",
-    "        if position_ids is None:\n",
-    "            position_ids = torch.arange(q_len, device=hidden_states.device).unsqueeze(0).expand(bsz, -1)\n",
-    "        max_pos = position_ids.max().item() + 1\n",
-    "        seq_len = max(max_pos, q_len)\n",
-    "        t = torch.arange(seq_len, device=hidden_states.device, dtype=self.inv_freq.dtype)\n",
-    "        freqs = torch.outer(t, self.inv_freq)\n",
-    "        emb = torch.cat((freqs, freqs), dim=-1)\n",
-    "        cos, sin = emb.cos(), emb.sin()\n",
-    "        cos = cos[position_ids].unsqueeze(1)\n",
-    "        sin = sin[position_ids].unsqueeze(1)\n",
-    "        query_states = self._apply_rotary(query_states, cos, sin)\n",
-    "        key_states = self._apply_rotary(key_states, cos, sin)\n",
-    "        if past_key_value is not None:\n",
-    "            key_states = torch.cat([past_key_value[0], key_states], dim=2)\n",
-    "            value_states = torch.cat([past_key_value[1], value_states], dim=2)\n",
-    "        past_kv = (key_states, value_states) if use_cache else None\n",
-    "        key_states = key_states.repeat_interleave(self.num_key_value_groups, dim=1)\n",
-    "        value_states = value_states.repeat_interleave(self.num_key_value_groups, dim=1)\n",
-    "        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)\n",
-    "        if attention_mask is not None:\n",
-    "            attn_weights = attn_weights + attention_mask\n",
-    "        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)\n",
-    "        attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)\n",
-    "        attn_output = torch.matmul(attn_weights, value_states)\n",
-    "        attn_output = attn_output.transpose(1, 2).contiguous()\n",
-    "        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)\n",
-    "        return self.o_proj(attn_output), past_kv\n",
-    "\n",
-    "class SpiderPortalExpert(nn.Module):\n",
-    "    def __init__(self, config, intermediate_size=None):\n",
-    "        super().__init__()\n",
-    "        inter_size = intermediate_size or config.intermediate_size\n",
-    "        self.gate_proj = nn.Linear(config.hidden_size, inter_size, bias=False)\n",
-    "        self.up_proj = nn.Linear(config.hidden_size, inter_size, bias=False)\n",
-    "        self.down_proj = nn.Linear(inter_size, config.hidden_size, bias=False)\n",
-    "        self.act_fn = nn.SiLU()\n",
-    "    def forward(self, hidden_states):\n",
-    "        return self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))\n",
-    "\n",
-    "class SpiderPortalRouter(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.num_experts = config.num_experts\n",
-    "        self.num_experts_per_tok = config.num_experts_per_tok\n",
-    "        self.weight = nn.Parameter(torch.randn(config.hidden_size, config.num_experts) * config.initializer_range)\n",
-    "        self.register_buffer(\"router_bias\", torch.zeros(config.num_experts))\n",
-    "    def forward(self, hidden_states):\n",
-    "        router_logits = hidden_states.view(-1, hidden_states.size(-1)) @ self.weight\n",
-    "        routing_weights = F.softmax(router_logits, dim=-1, dtype=torch.float32)\n",
-    "        biased_logits = router_logits + self.router_bias\n",
-    "        biased_weights = F.softmax(biased_logits, dim=-1, dtype=torch.float32)\n",
-    "        top_weights, top_indices = torch.topk(biased_weights, self.num_experts_per_tok, dim=-1)\n",
-    "        top_weights = top_weights / top_weights.sum(dim=-1, keepdim=True)\n",
-    "        top_weights = top_weights.to(hidden_states.dtype)\n",
-    "        mean_probs = routing_weights.mean(dim=0)\n",
-    "        aux_loss = self.num_experts * (mean_probs * mean_probs).sum()\n",
-    "        return top_weights, top_indices, aux_loss\n",
-    "\n",
-    "class SpiderPortalMoE(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.config = config\n",
-    "        self.num_experts = config.num_experts\n",
-    "        self.num_experts_per_tok = config.num_experts_per_tok\n",
-    "        self.experts = nn.ModuleList([SpiderPortalExpert(config) for _ in range(config.num_experts)])\n",
-    "        self.shared_expert = SpiderPortalExpert(config)\n",
-    "        self.router = SpiderPortalRouter(config)\n",
-    "    def forward(self, hidden_states):\n",
-    "        batch_size, seq_len, hidden_dim = hidden_states.shape\n",
-    "        top_weights, top_indices, aux_loss = self.router(hidden_states)\n",
-    "        flat_hidden = hidden_states.view(-1, hidden_dim)\n",
-    "        final_output = torch.zeros_like(flat_hidden)\n",
-    "        for expert_idx in range(self.num_experts_per_tok):\n",
-    "            expert_ids = top_indices[:, expert_idx]\n",
-    "            expert_weights = top_weights[:, expert_idx:expert_idx+1]\n",
-    "            for e in range(self.num_experts):\n",
-    "                mask = expert_ids == e\n",
-    "                if mask.any():\n",
-    "                    expert_output = self.experts[e](flat_hidden[mask])\n",
-    "                    final_output[mask] += expert_output * expert_weights[mask]\n",
-    "        shared_output = self.shared_expert(flat_hidden)\n",
-    "        final_output = final_output + shared_output\n",
-    "        return final_output.view(batch_size, seq_len, hidden_dim), aux_loss\n",
-    "\n",
-    "class SpiderPortalDenseLayer(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.self_attn = SpiderPortalGQA(config)\n",
-    "        dense_intermediate = config.hidden_size * 4 // 3\n",
-    "        self.ffn = SpiderPortalExpert(config, intermediate_size=dense_intermediate)\n",
-    "        self.input_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)\n",
-    "        self.post_attention_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)\n",
-    "    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):\n",
-    "        attn_input = self.input_layernorm(hidden_states)\n",
-    "        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)\n",
-    "        hidden_states = hidden_states + attn_output\n",
-    "        ffn_input = self.post_attention_layernorm(hidden_states)\n",
-    "        ffn_output = self.ffn(ffn_input)\n",
-    "        hidden_states = hidden_states + ffn_output\n",
-    "        return hidden_states, past_kv\n",
-    "\n",
-    "class SpiderPortalMoELayer(nn.Module):\n",
-    "    def __init__(self, config, layer_idx):\n",
-    "        super().__init__()\n",
-    "        self.layer_idx = layer_idx\n",
-    "        self.self_attn = SpiderPortalGQA(config)\n",
-    "        self.moe = SpiderPortalMoE(config)\n",
-    "        self.input_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)\n",
-    "        self.post_attention_layernorm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)\n",
-    "    def forward(self, hidden_states, attention_mask=None, position_ids=None, past_key_value=None, use_cache=False):\n",
-    "        attn_input = self.input_layernorm(hidden_states)\n",
-    "        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)\n",
-    "        hidden_states = hidden_states + attn_output\n",
-    "        moe_input = self.post_attention_layernorm(hidden_states)\n",
-    "        moe_output, aux_loss = self.moe(moe_input)\n",
-    "        hidden_states = hidden_states + moe_output\n",
-    "        return hidden_states, aux_loss, past_kv\n",
-    "\n",
-    "class LTIInjection(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.hidden_size = config.hidden_size\n",
-    "        self.log_A = nn.Parameter(torch.full((config.hidden_size,), -2.0))\n",
-    "        self.delta_t = nn.Parameter(torch.tensor(1.0))\n",
-    "        self.B = nn.Linear(config.hidden_size, config.hidden_size, bias=False)\n",
-    "        with torch.no_grad():\n",
-    "            self.B.weight.data.normal_(mean=0.0, std=0.01)\n",
-    "    def get_A(self):\n",
-    "        return -torch.exp(self.log_A)\n",
-    "    def forward(self, h_t, e):\n",
-    "        A = self.get_A()\n",
-    "        return A * h_t + self.B(e)\n",
-    "\n",
-    "class ACTHalting(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.halt_predictor = nn.Linear(config.hidden_size, 1)\n",
-    "        self.threshold = config.act_threshold\n",
-    "    def forward(self, hidden_states):\n",
-    "        return torch.sigmoid(self.halt_predictor(hidden_states))\n",
-    "\n",
-    "class LoRAAdapter(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        rank = config.lora_rank\n",
-    "        self.down = nn.Linear(config.hidden_size, rank, bias=False)\n",
-    "        self.B = nn.Parameter(torch.randn(rank, config.hidden_size) * 0.02)\n",
-    "        self.scale = nn.Embedding(config.max_loop_iters, rank)\n",
-    "        with torch.no_grad():\n",
-    "            self.scale.weight.data.zero_()\n",
-    "            self.down.weight.data.normal_(mean=0.0, std=0.001)\n",
-    "    def forward(self, x, loop_t):\n",
-    "        max_t = self.scale.num_embeddings - 1\n",
-    "        t_idx = min(loop_t, max_t)\n",
-    "        s = self.scale(torch.tensor(t_idx, device=x.device))\n",
-    "        down = self.down(x) * s\n",
-    "        return down @ self.B\n",
-    "\n",
-    "class SpiderPortalMoEModel(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.config = config\n",
-    "        self.prelude_layers = nn.ModuleList([SpiderPortalDenseLayer(config) for _ in range(config.prelude_layers)])\n",
-    "        self.recurrent_layers = nn.ModuleList([SpiderPortalMoELayer(config, i) for i in range(config.num_hidden_layers)])\n",
-    "        self.coda_layers = nn.ModuleList([SpiderPortalDenseLayer(config) for _ in range(config.coda_layers)])\n",
-    "        self.norm = SpiderPortalRMSNorm(config.hidden_size, eps=config.rms_norm_eps)\n",
-    "        self.injection = LTIInjection(config)\n",
-    "        self.act_halting = ACTHalting(config)\n",
-    "        self.lora_adapter = LoRAAdapter(config)\n",
-    "        self.loop_embed_dim = config.loop_embed_dim\n",
-    "    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",
-    "        n_loops = n_loops or self.config.max_loop_iters\n",
-    "        input_embedding = input_embedding if input_embedding is not None else hidden_states\n",
-    "        total_aux_loss = 0.0\n",
-    "        for layer in self.prelude_layers:\n",
-    "            hidden_states, _ = layer(hidden_states, attention_mask=attention_mask, position_ids=position_ids)\n",
-    "        e = hidden_states.clone()\n",
-    "        B, T_seq, D = hidden_states.shape\n",
-    "        halted = torch.zeros(B, T_seq, device=hidden_states.device, dtype=torch.bool)\n",
-    "        cumulative_p = torch.zeros(B, T_seq, device=hidden_states.device, dtype=hidden_states.dtype)\n",
-    "        h_out = torch.zeros_like(hidden_states)\n",
-    "        past_key_values = past_key_values if past_key_values is not None else [None] * len(self.recurrent_layers)\n",
-    "        for t in range(n_loops):\n",
-    "            h_loop = loop_index_embedding(hidden_states, t, self.loop_embed_dim)\n",
-    "            if t > 0:\n",
-    "                injection = self.injection(hidden_states, input_embedding)\n",
-    "                hidden_states = hidden_states + injection\n",
-    "            new_past_key_values = []\n",
-    "            for i, layer in enumerate(self.recurrent_layers):\n",
-    "                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)\n",
-    "                total_aux_loss = total_aux_loss + aux_loss\n",
-    "                new_past_key_values.append(past_kv)\n",
-    "            lora_delta = self.lora_adapter(hidden_states, t)\n",
-    "            hidden_states = hidden_states + lora_delta\n",
-    "            halt_prob = self.act_halting(hidden_states).squeeze(-1)\n",
-    "            still_running = ~halted\n",
-    "            remainder = (1.0 - cumulative_p).clamp(min=0)\n",
-    "            weight = torch.where(cumulative_p + halt_prob >= self.config.act_threshold, remainder, halt_prob)\n",
-    "            weight = weight * still_running.to(hidden_states.dtype)\n",
-    "            h_out = h_out + weight.unsqueeze(-1) * hidden_states\n",
-    "            cumulative_p = cumulative_p + halt_prob * still_running.to(hidden_states.dtype)\n",
-    "            halted = halted | (cumulative_p >= self.config.act_threshold)\n",
-    "            if halted.all() and not self.training:\n",
-    "                break\n",
-    "        never_halted = (~halted).to(hidden_states.dtype).unsqueeze(-1)\n",
-    "        hidden_states = h_out + never_halted * hidden_states\n",
-    "        for layer in self.coda_layers:\n",
-    "            hidden_states, _ = layer(hidden_states, attention_mask=attention_mask, position_ids=position_ids)\n",
-    "        hidden_states = self.norm(hidden_states)\n",
-    "        return hidden_states, total_aux_loss, new_past_key_values\n",
-    "\n",
-    "class SpiderPortalForConditionalGeneration(nn.Module):\n",
-    "    def __init__(self, config):\n",
-    "        super().__init__()\n",
-    "        self.config = config\n",
-    "        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)\n",
-    "        self.model = SpiderPortalMoEModel(config)\n",
-    "        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n",
-    "        if config.tie_word_embeddings:\n",
-    "            self.lm_head.weight = self.embed_tokens.weight\n",
-    "        self.apply(self._init_weights)\n",
-    "    def _init_weights(self, module):\n",
-    "        if isinstance(module, nn.Linear):\n",
-    "            if hasattr(self, 'model') and module is self.model.injection.B:\n",
-    "                return\n",
-    "            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n",
-    "            if module.bias is not None:\n",
-    "                module.bias.data.zero_()\n",
-    "        elif isinstance(module, nn.Embedding):\n",
-    "            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n",
-    "    def forward(self, input_ids, attention_mask=None, position_ids=None, labels=None, n_loops=None, use_cache=False):\n",
-    "        hidden_states = self.embed_tokens(input_ids)\n",
-    "        model_dtype = next(self.model.parameters()).dtype\n",
-    "        hidden_states = hidden_states.to(model_dtype)\n",
-    "        input_embedding = hidden_states.clone()\n",
-    "        if attention_mask is None:\n",
-    "            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)\n",
-    "        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)\n",
-    "        causal_mask = causal_mask.masked_fill(~attention_mask.unsqueeze(1).unsqueeze(2), torch.finfo(hidden_states.dtype).min)\n",
-    "        causal_mask = causal_mask.triu(1)\n",
-    "        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)\n",
-    "        logits = self.lm_head(hidden_states)\n",
-    "        loss = None\n",
-    "        if labels is not None:\n",
-    "            shift_logits = logits[..., :-1, :].contiguous()\n",
-    "            shift_labels = labels[..., 1:].contiguous()\n",
-    "            loss_fct = CrossEntropyLoss()\n",
-    "            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))\n",
-    "            loss = loss + self.config.router_aux_loss_coef * aux_loss\n",
-    "        return {\"loss\": loss, \"logits\": logits, \"aux_loss\": aux_loss, \"past_key_values\": past_kv}\n",
-    "    def get_num_params(self):\n",
-    "        total = sum(p.numel() for p in self.parameters())\n",
-    "        return {\"total\": total, \"trainable\": total}"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Multi-GPU DDP Training\n",
-    "# DDP requires a standalone script (mp.spawn doesn't work in notebooks)\n",
-    "# The script is at scripts/train_ddp.py\n",
-    "!python scripts/train_ddp.py"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Test generation (after training)\n",
-    "device = torch.device(\"cuda:0\")\n",
-    "config = SpiderPortalConfig(\n",
-    "    prelude_layers=2, coda_layers=2, lora_rank=32,\n",
-    "    rope_theta=10000000.0, tie_word_embeddings=True,\n",
-    ")\n",
-    "model = SpiderPortalForConditionalGeneration(config)\n",
-    "\n",
-    "checkpoint = torch.load(\"/kaggle/working/spiderportal-v5-ep1/model.pt\", map_location=\"cpu\")\n",
-    "model.load_state_dict(checkpoint)\n",
-    "model = model.to(torch.bfloat16).to(device)\n",
-    "model.eval()\n",
-    "\n",
-    "from transformers import AutoTokenizer\n",
-    "tokenizer = AutoTokenizer.from_pretrained(\"gpt2\")\n",
-    "tokenizer.pad_token = tokenizer.eos_token\n",
-    "\n",
-    "with torch.no_grad():\n",
-    "    prompt = \"Question: Instruction: What is the capital of France?\\nAnswer:\"\n",
-    "    input_ids = tokenizer(prompt, return_tensors=\"pt\")[\"input_ids\"].to(device)\n",
-    "    generated = input_ids.clone()\n",
-    "    for _ in range(32):\n",
-    "        outputs = model(generated, n_loops=1)\n",
-    "        next_token = torch.argmax(outputs[\"logits\"][:, -1, :], dim=-1, keepdim=True)\n",
-    "        generated = torch.cat([generated, next_token], dim=1)\n",
-    "        if next_token.item() == tokenizer.eos_token_id:\n",
-    "            break\n",
-    "    print(f\"Generated: {tokenizer.decode(generated[0])}\")"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 3",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "name": "python",
-   "version": "3.10.0"
-  },
-  "accelerator": "GPU"
- },
- "nbformat": 4,
- "nbformat_minor": 4
-}

AI-Transfer/scripts/train_ddp.py

@@ -1,601 +0,0 @@
-#!/usr/bin/env python3
-"""SpiderPortal v5 — Multi-GPU DDP Training.
-
-Run from Kaggle notebook cell:
-    !python train_ddp.py
-
-Or directly:
-    python -m torch.distributed.run --nproc_per_node=2 train_ddp.py
-"""
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import torch.distributed as dist
-import math
-import os
-import json
-import gc
-import random
-import time
-import subprocess
-from pathlib import Path
-from dataclasses import dataclass
-from typing import Optional, Tuple, Dict, List
-from torch.nn import CrossEntropyLoss
-from torch.nn.parallel import DistributedDataParallel as DDP
-
-@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 = 4
-    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}
-
-def train_ddp(local_rank, world_size):
-    dist.init_process_group("nccl", rank=local_rank, world_size=world_size)
-    torch.cuda.set_device(local_rank)
-    device = torch.device(f"cuda:{local_rank}")
-    is_master = local_rank == 0
-
-    if is_master:
-        print(f"Training on {world_size} GPUs (DDP)")
-        for i in range(world_size):
-            p = torch.cuda.get_device_properties(i)
-            print(f"  GPU {i}: {torch.cuda.get_device_name(i)} ({p.total_memory / 1e9:.1f}GB)")
-
-    config = 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=2,
-        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,
-    )
-
-    model = SpiderPortalForConditionalGeneration(config)
-    model = model.to(torch.bfloat16).to(device)
-    model = DDP(model, device_ids=[local_rank], find_unused_parameters=True)
-
-    if is_master:
-        params = model.module.get_num_params()
-        print(f"Model: {params['total']/1e6:.1f}M params")
-        print(f"Experts: {config.num_experts} routed + {config.num_shared_experts} shared")
-
-    BASE_LR = 2e-5
-    WARMUP_STEPS = 1000
-    optimizer = torch.optim.AdamW(model.parameters(), lr=BASE_LR, weight_decay=0.01)
-
-    import pandas as pd
-    data_dir = Path("/kaggle/input/datasets/cliworks/spiderp-custom-1")
-    all_records = []
-    if data_dir.exists():
-        pkl_file = data_dir / "spiderportal_combined.pkl"
-        if pkl_file.exists():
-            df = pd.read_pickle(pkl_file)
-            all_records = df.to_dict("records")
-        else:
-            for f in data_dir.glob("*.parquet"):
-                try:
-                    df_pq = pd.read_parquet(f)
-                    if all(c in df_pq.columns for c in ["instruction", "input", "output"]):
-                        all_records.extend(df_pq[["instruction", "input", "output"]].to_dict("records"))
-                except:
-                    pass
-
-    if not all_records:
-        if is_master:
-            print("No data found, creating synthetic data...")
-        all_records = [{"instruction": f"Question {i}: What is {i} + {i}?", "input": "", "output": f"The answer is {i+i}."} for i in range(10000)]
-
-    if is_master:
-        print(f"Loaded {len(all_records)} samples")
-
-    from transformers import AutoTokenizer
-    tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    tokenizer.pad_token = tokenizer.eos_token
-
-    BATCH_SIZE = 8
-    GRAD_ACCUM = 2
-    MAX_LEN = 256
-    EPOCHS = 3
-    N_LOOPS = 1
-
-    if is_master:
-        print(f"Batch size: {BATCH_SIZE}, Grad accum: {GRAD_ACCUM}")
-        print(f"Effective batch: {BATCH_SIZE * GRAD_ACCUM * world_size}")
-        print(f"LR: {BASE_LR} with {WARMUP_STEPS}-step warmup")
-
-    def build_prompt(sample):
-        instruction = str(sample.get("instruction", "")).strip()
-        inp = str(sample.get("input", "")).strip()
-        output = str(sample.get("output", "")).strip()
-        if inp:
-            return f"Question: Instruction: {instruction}\nInput: {inp}\nAnswer: {output}\n"
-        return f"Question: Instruction: {instruction}\nAnswer: {output}\n"
-
-    def format_sample(sample, tokenizer, max_len):
-        text = build_prompt(sample) + tokenizer.eos_token
-        enc = tokenizer(text, truncation=True, max_length=max_len, padding="max_length", return_tensors="pt")
-        input_ids = enc["input_ids"].squeeze(0)
-        labels = input_ids.clone()
-        prefix_ids = tokenizer("Question:", add_special_tokens=False)["input_ids"]
-        mask_len = min(len(prefix_ids), labels.shape[0])
-        labels[:mask_len] = -100
-        return {"input_ids": input_ids, "labels": labels}
-
-    shard_size = len(all_records) // world_size
-    start_idx = local_rank * shard_size
-    end_idx = start_idx + shard_size if local_rank < world_size - 1 else len(all_records)
-    local_samples = all_records[start_idx:end_idx]
-
-    if is_master:
-        print(f"Data sharding: {len(all_records)} total -> {len(local_samples)} per GPU")
-
-    global_step = 0
-    best_loss = float('inf')
-    start_time = time.time()
-
-    for epoch in range(1, EPOCHS + 1):
-        random.shuffle(local_samples)
-        total_loss = 0
-        num_batches = 0
-        optimizer.zero_grad()
-
-        for i in range(0, len(local_samples), BATCH_SIZE):
-            batch_samples = local_samples[i:i+BATCH_SIZE]
-            if len(batch_samples) < BATCH_SIZE:
-                continue
-
-            if global_step < WARMUP_STEPS:
-                lr = BASE_LR * (global_step + 1) / WARMUP_STEPS
-                for param_group in optimizer.param_groups:
-                    param_group['lr'] = lr
-
-            batch = [format_sample(s, tokenizer, MAX_LEN) for s in batch_samples]
-            input_ids = torch.stack([b["input_ids"] for b in batch]).to(device)
-            labels = torch.stack([b["labels"] for b in batch]).to(device)
-
-            if global_step == 0 and is_master:
-                print("  [First forward pass - CUDA graph building...]")
-
-            outputs = model(input_ids=input_ids, labels=labels, n_loops=N_LOOPS)
-            loss = outputs["loss"] / GRAD_ACCUM
-            loss.backward()
-
-            if (i // BATCH_SIZE + 1) % GRAD_ACCUM == 0:
-                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
-                optimizer.step()
-                optimizer.zero_grad()
-                global_step += 1
-
-            total_loss += loss.item() * GRAD_ACCUM
-            num_batches += 1
-
-            if is_master and ((i // BATCH_SIZE) == 0 or (global_step < 20) or (global_step % 50 == 0)):
-                avg_loss = total_loss / max(num_batches, 1)
-                elapsed = time.time() - start_time
-                steps_per_hour = (global_step + 1) / elapsed * 3600 if elapsed > 0 else 0
-                current_lr = optimizer.param_groups[0]['lr']
-                print(f"Epoch {epoch}/{EPOCHS} | Step {global_step} | avg_loss={avg_loss:.4f} | LR={current_lr:.2e} | {steps_per_hour:.1f} steps/hr")
-
-            if is_master and global_step > 0 and global_step % 500 == 0:
-                output_dir = Path(f"/kaggle/working/spiderportal-v5-ep{epoch}-step{global_step}")
-                output_dir.mkdir(parents=True, exist_ok=True)
-                state_dict = {k: v.cpu() for k, v in model.module.state_dict().items()}
-                torch.save(state_dict, output_dir / "model.pt")
-                print(f"Saved checkpoint at step {global_step}")
-
-        if is_master:
-            avg_loss = total_loss / max(num_batches, 1)
-            print(f"Epoch {epoch}/{EPOCHS} | avg_loss={avg_loss:.4f} | Time: {(time.time()-start_time)/60:.1f}min")
-            output_dir = Path(f"/kaggle/working/spiderportal-v5-ep{epoch}")
-            output_dir.mkdir(parents=True, exist_ok=True)
-            state_dict = {k: v.cpu() for k, v in model.module.state_dict().items()}
-            torch.save(state_dict, output_dir / "model.pt")
-            if avg_loss < best_loss:
-                best_loss = avg_loss
-
-    if is_master:
-        print(f"Training complete! Best loss: {best_loss:.4f}")
-        print(f"Total time: {(time.time() - start_time)/60:.1f} minutes")
-
-    dist.destroy_process_group()
-
-if __name__ == "__main__":
-    import sys
-    print(f"Script started, CUDA devices: {torch.cuda.device_count()}")
-    sys.stdout.flush()
-    
-    num_gpus = torch.cuda.device_count()
-    
-    if num_gpus <= 1:
-        dist.init_process_group = lambda *a, **k: None
-        dist.destroy_process_group = lambda: None
-        train_ddp(0, 1)
-    else:
-        print(f"Launching DDP with {num_gpus} GPUs...")
-        sys.stdout.flush()
-        import os
-        import signal
-        
-        # Set NCCL environment before forking
-        os.environ["MASTER_ADDR"] = "127.0.0.1"
-        os.environ["MASTER_PORT"] = "29500"
-        os.environ["WORLD_SIZE"] = str(num_gpus)
-        os.environ["LOCAL_WORLD_SIZE"] = str(num_gpus)
-        
-        procs = []
-        for rank in range(num_gpus):
-            print(f"Forking rank {rank}...")
-            sys.stdout.flush()
-            pid = os.fork()
-            if pid == 0:
-                # Child process
-                os.environ["RANK"] = str(rank)
-                os.environ["LOCAL_RANK"] = str(rank)
-                print(f"Child {rank} starting training...")
-                sys.stdout.flush()
-                train_ddp(rank, num_gpus)
-                os._exit(0)
-            else:
-                procs.append(pid)
-                print(f"Parent: forked child {rank} with PID {pid}")
-                sys.stdout.flush()
-        
-        # Parent process: wait for children
-        print("Parent waiting for children...")
-        sys.stdout.flush()
-        for pid in procs:
-            os.waitpid(pid, 0)

AI-Transfer/scripts/train_single_gpu.py

@@ -1,587 +0,0 @@
-#!/usr/bin/env python3
-"""SpiderPortal v5 — Single-GPU Optimized Training.
-
-For RTX PRO 6000 (96GB) or similar large-VRAM GPU.
-No DDP, maximal batch size, torch.compile, pre-tokenized data.
-
-Usage:
-    python train_single_gpu.py
-"""
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import math
-import os
-import json
-import gc
-import random
-import time
-from pathlib import Path
-from dataclasses import dataclass
-from typing import Optional, Tuple, Dict, List
-from torch.nn import CrossEntropyLoss
-
-@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 = 4
-    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}
-
-def train_single_gpu():
-    device = torch.device("cuda")
-    gpu_name = torch.cuda.get_device_name(0)
-    gpu_mem = torch.cuda.get_device_properties(0).total_memory / 1e9
-    print(f"GPU: {gpu_name} ({gpu_mem:.1f}GB)")
-
-    config = 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=2,
-        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,
-    )
-
-    print("Building model...")
-    model = SpiderPortalForConditionalGeneration(config)
-    model = model.to(torch.bfloat16).to(device)
-
-    params = model.get_num_params()
-    print(f"Model: {params['total']/1e6:.1f}M params")
-    print(f"Experts: {config.num_experts} routed + {config.num_shared_experts} shared")
-
-    try:
-        model = torch.compile(model, mode="reduce-overhead")
-        print("torch.compile: enabled")
-    except Exception:
-        print("torch.compile: not available, using eager mode")
-
-    BASE_LR = 2e-5
-    WARMUP_STEPS = 1000
-    optimizer = torch.optim.AdamW(model.parameters(), lr=BASE_LR, weight_decay=0.01)
-
-    import pandas as pd
-    data_dir = Path(__file__).parent / "data"
-    all_records = []
-    pkl_file = data_dir / "spiderportal_combined.pkl"
-    if pkl_file.exists():
-        print(f"Loading dataset from {pkl_file}...")
-        df = pd.read_pickle(pkl_file)
-        all_records = df.to_dict("records")
-    else:
-        print(f"No dataset found at {pkl_file}, creating synthetic data...")
-        all_records = [{"instruction": f"Question {i}: What is {i} + {i}?", "input": "", "output": f"The answer is {i+i}."} for i in range(10000)]
-
-    print(f"Loaded {len(all_records):,} samples")
-
-    from transformers import AutoTokenizer
-    tokenizer = AutoTokenizer.from_pretrained("gpt2")
-    tokenizer.pad_token = tokenizer.eos_token
-
-    BATCH_SIZE = 128
-    MAX_LEN = 256
-    EPOCHS = 3
-    N_LOOPS = 2
-
-    print(f"Batch size: {BATCH_SIZE} (no grad accum)")
-    print(f"Effective batch: {BATCH_SIZE}")
-    print(f"LR: {BASE_LR} with {WARMUP_STEPS}-step warmup")
-    print(f"Max seq len: {MAX_LEN}, N_LOOPS: {N_LOOPS}")
-
-    def build_prompt(sample):
-        instruction = str(sample.get("instruction", "")).strip()
-        inp = str(sample.get("input", "")).strip()
-        output = str(sample.get("output", "")).strip()
-        if inp:
-            return f"Question: Instruction: {instruction}\nInput: {inp}\nAnswer: {output}\n"
-        return f"Question: Instruction: {instruction}\nAnswer: {output}\n"
-
-    print("Pre-tokenizing dataset...")
-    prefix_ids = tokenizer("Question:", add_special_tokens=False)["input_ids"]
-    mask_len = len(prefix_ids)
-
-    pre_tokenized = []
-    for i, sample in enumerate(all_records):
-        instruction = str(sample.get("instruction", "")).strip()
-        inp = str(sample.get("input", "")).strip()
-        output = str(sample.get("output", "")).strip()
-        if inp:
-            text = f"Question: Instruction: {instruction}\nInput: {inp}\nAnswer: {output}\n" + tokenizer.eos_token
-        else:
-            text = f"Question: Instruction: {instruction}\nAnswer: {output}\n" + tokenizer.eos_token
-        enc = tokenizer(text, truncation=True, max_length=MAX_LEN, padding="max_length")
-        input_ids = enc["input_ids"]
-        labels = input_ids[:]
-        for j in range(min(mask_len, len(labels))):
-            labels[j] = -100
-        pre_tokenized.append((input_ids, labels))
-        if (i + 1) % 50000 == 0:
-            print(f"  Tokenized {i+1:,}/{len(all_records):,}")
-
-    print(f"Pre-tokenization complete: {len(pre_tokenized):,} samples")
-    del all_records
-    gc.collect()
-
-    global_step = 0
-    best_loss = float('inf')
-    start_time = time.time()
-    checkpoint_dir = Path("checkpoints")
-    checkpoint_dir.mkdir(exist_ok=True)
-    step_ckpt_files = []
-
-    for epoch in range(1, EPOCHS + 1):
-        if epoch > 1:
-            for f in step_ckpt_files:
-                if f.exists():
-                    f.unlink()
-                    print(f"  Deleted old step checkpoint: {f.name}")
-            step_ckpt_files.clear()
-            gc.collect()
-
-        indices = list(range(len(pre_tokenized)))
-        random.shuffle(indices)
-        total_loss = 0
-        num_batches = 0
-        optimizer.zero_grad()
-
-        for batch_start in range(0, len(indices), BATCH_SIZE):
-            batch_indices = indices[batch_start:batch_start + BATCH_SIZE]
-            if len(batch_indices) < BATCH_SIZE:
-                continue
-
-            if global_step < WARMUP_STEPS:
-                lr = BASE_LR * (global_step + 1) / WARMUP_STEPS
-                for param_group in optimizer.param_groups:
-                    param_group['lr'] = lr
-
-            batch_input_ids = []
-            batch_labels = []
-            for idx in batch_indices:
-                input_ids, labels = pre_tokenized[idx]
-                batch_input_ids.append(input_ids)
-                batch_labels.append(labels)
-
-            input_ids = torch.tensor(batch_input_ids, dtype=torch.long, device=device)
-            labels = torch.tensor(batch_labels, dtype=torch.long, device=device)
-
-            if global_step == 0:
-                print("  [First forward pass - compiling...]")
-
-            outputs = model(input_ids=input_ids, labels=labels, n_loops=N_LOOPS)
-            loss = outputs["loss"]
-            loss.backward()
-
-            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
-            optimizer.step()
-            optimizer.zero_grad()
-            global_step += 1
-
-            total_loss += loss.item()
-            num_batches += 1
-
-            if (batch_start // BATCH_SIZE) == 0 or global_step < 20 or global_step % 100 == 0:
-                avg_loss = total_loss / max(num_batches, 1)
-                elapsed = time.time() - start_time
-                steps_per_hour = (global_step + 1) / elapsed * 3600 if elapsed > 0 else 0
-                current_lr = optimizer.param_groups[0]['lr']
-                samples_per_sec = (global_step * BATCH_SIZE) / elapsed if elapsed > 0 else 0
-                print(f"Epoch {epoch}/{EPOCHS} | Step {global_step} | loss={avg_loss:.4f} | LR={current_lr:.2e} | {steps_per_hour:.0f} steps/hr | {samples_per_sec:.0f} samples/sec")
-
-            if global_step > 0 and global_step % 500 == 0:
-                ckpt_path = checkpoint_dir / f"spiderportal-v5-ep{epoch}-step{global_step}.pt"
-                state_dict = {k: v.cpu() for k, v in model.state_dict().items()}
-                torch.save(state_dict, ckpt_path)
-                step_ckpt_files.append(ckpt_path)
-                size_mb = ckpt_path.stat().st_size / (1024 * 1024)
-                print(f"Saved weights-only checkpoint: {ckpt_path.name} ({size_mb:.0f}MB)")
-
-        avg_loss = total_loss / max(num_batches, 1)
-        epoch_time = (time.time() - start_time) / 60
-        print(f"Epoch {epoch}/{EPOCHS} complete | avg_loss={avg_loss:.4f} | Time: {epoch_time:.1f}min")
-
-        ckpt_path = checkpoint_dir / f"spiderportal-v5-ep{epoch}.pt"
-        torch.save({
-            "step": global_step,
-            "epoch": epoch,
-            "model_state_dict": {k: v.cpu() for k, v in model.state_dict().items()},
-            "optimizer_state_dict": optimizer.state_dict(),
-            "config": config.__dict__,
-        }, ckpt_path)
-        size_mb = ckpt_path.stat().st_size / (1024 * 1024)
-        print(f"Saved epoch checkpoint: {ckpt_path.name} ({size_mb:.0f}MB)")
-
-        if avg_loss < best_loss:
-            best_loss = avg_loss
-            best_path = checkpoint_dir / "spiderportal-v5-best.pt"
-            torch.save({
-                "step": global_step,
-                "epoch": epoch,
-                "model_state_dict": {k: v.cpu() for k, v in model.state_dict().items()},
-                "optimizer_state_dict": optimizer.state_dict(),
-                "config": config.__dict__,
-            }, best_path)
-            size_mb = best_path.stat().st_size / (1024 * 1024)
-            print(f"Saved best checkpoint: {best_path.name} ({size_mb:.0f}MB)")
-
-    total_time = (time.time() - start_time) / 3600
-    print(f"\nTraining complete!")
-    print(f"Best loss: {best_loss:.4f}")
-    print(f"Total time: {total_time:.2f} hours")
-    print(f"Total steps: {global_step}")
-    print(f"Checkpoints saved to: {checkpoint_dir}")
-
-if __name__ == "__main__":
-    train_single_gpu()