[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6QF2Q789EBPPN6PZPM0Z997":3,"skill-same:01M6QF2Q789EBPPN6PZPM0Z997":91},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":50,"commitSha":51,"skillDirectory":52,"forksCount":14,"repositoryHtmlUrl":53,"repositoryHomepageUrl":54,"ownerHtmlUrl":55,"ownerBlogUrl":56,"ownerDescription":57,"updatedAt":58,"install":59,"resources":72},"01M6QF2Q789EBPPN6PZPM0Z997","long-context",null,"Extend context windows of transformer models using RoPE, YaRN, ALiBi, and position interpolation techniques. Use when processing long documents (32k-128k+ tokens), extending pre-trained models beyond original context limits, or implementing efficient positional encodings. Covers rotary embeddings, attention biases, interpolation methods, and extrapolation strategies for LLMs.","Tashima-Tarsh\u002FDisha6.6","Tashima-Tarsh","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F269906498?v=4","01M6PBH98X03XERH0HFT4PFB1P","NITISH KUMAR","01M6PBH9DA0M6KM3Z12P848F0K",0,"MIT","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fkal-long-context\u002FSKILL.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"context","long","long context","bias","limits","interpolation","for","implementing","RoPE","Emerging Techniques","tokens","windows","attention","strategies","methods","use","multimodal","embeddings","yarn","documents","2026-11-10 10:12:30",false,"---\nname: long-context\ndescription: Extend context windows of transformer models using RoPE, YaRN, ALiBi, and position interpolation techniques. Use when processing long documents (32k-128k+ tokens), extending pre-trained models beyond original context limits, or implementing efficient positional encodings. Covers rotary embeddings, attention biases, interpolation methods, and extrapolation strategies for LLMs.\nlicense: MIT\nmetadata:\n  version: 1.0.0\n  category: multimodal-and-emerging\n  maintainer: Kalaris Labs\n  tags: Emerging Techniques, Long Context, RoPE, YaRN, ALiBi, Position Interpolation, Extended Context, Rotary Embeddings, Attention Bias, Context Extension, Positional Encoding\n  dependencies: transformers, torch, flash-attn\n---\n\n# Long Context: Extending Transformer Context Windows\n\n## When to Use This Skill\n\nUse Long Context techniques when you need to:\n- **Process long documents** (32k, 64k, 128k+ tokens) with transformer models\n- **Extend context windows** of pre-trained models (LLaMA, Mistral, etc.)\n- **Implement efficient positional encodings** (RoPE, ALiBi)\n- **Train models** with length extrapolation capabilities\n- **Deploy models** that handle variable-length inputs efficiently\n- **Fine-tune** existing models for longer contexts with minimal compute\n\n**Key Techniques**: RoPE (Rotary Position Embeddings), YaRN, ALiBi (Attention with Linear Biases), Position Interpolation\n\n**Papers**: RoFormer (arXiv 2104.09864), YaRN (arXiv 2309.00071), ALiBi (arXiv 2108.12409), Position Interpolation (arXiv 2306.15595)\n\n## Installation\n\n```bash\n# HuggingFace Transformers (includes RoPE, YaRN support)\npip install transformers torch\n\n# For custom implementations\npip install einops  # Tensor operations\npip install rotary-embedding-torch  # Standalone RoPE\n\n# Optional: FlashAttention for efficiency\npip install flash-attn --no-build-isolation\n```\n\n## Quick Start\n\n### RoPE (Rotary Position Embeddings)\n\n```python\nimport torch\nimport torch.nn as nn\n\nclass RotaryEmbedding(nn.Module):\n    \"\"\"Rotary Position Embeddings (RoPE).\"\"\"\n\n    def __init__(self, dim, max_seq_len=8192, base=10000):\n        super().__init__()\n        # Compute inverse frequencies\n        inv_freq = 1.0 \u002F (base ** (torch.arange(0, dim, 2).float() \u002F dim))\n        self.register_buffer(\"inv_freq\", inv_freq)\n        self.max_seq_len = max_seq_len\n\n    def forward(self, seq_len, device):\n        # Position indices\n        t = torch.arange(seq_len, device=device).type_as(self.inv_freq)\n\n        # Compute frequencies\n        freqs = torch.outer(t, self.inv_freq)  # (seq_len, dim\u002F2)\n\n        # Compute sin and cos\n        emb = torch.cat((freqs, freqs), dim=-1)  # (seq_len, dim)\n        return emb.cos(), emb.sin()\n\ndef rotate_half(x):\n    \"\"\"Rotate half the hidden dimensions.\"\"\"\n    x1, x2 = x.chunk(2, dim=-1)\n    return torch.cat((-x2, x1), dim=-1)\n\ndef apply_rotary_pos_emb(q, k, cos, sin):\n    \"\"\"Apply rotary embeddings to queries and keys.\"\"\"\n    # q, k shape: (batch, heads, seq_len, dim)\n    q_embed = (q * cos) + (rotate_half(q) * sin)\n    k_embed = (k * cos) + (rotate_half(k) * sin)\n    return q_embed, k_embed\n\n# Usage\nrope = RotaryEmbedding(dim=64, max_seq_len=8192)\ncos, sin = rope(seq_len=2048, device='cuda')\n\n# In attention layer\nq_rotated, k_rotated = apply_rotary_pos_emb(query, key, cos, sin)\n```\n\n### ALiBi (Attention with Linear Biases)\n\n```python\ndef get_alibi_slopes(num_heads):\n    \"\"\"Get ALiBi slope values for each attention head.\"\"\"\n    def get_slopes_power_of_2(n):\n        start = 2 ** (-(2 ** -(math.log2(n) - 3)))\n        ratio = start\n        return [start * (ratio ** i) for i in range(n)]\n\n    if math.log2(num_heads).is_integer():\n        return get_slopes_power_of_2(num_heads)\n    else:\n        # Closest power of 2\n        closest_power = 2 ** math.floor(math.log2(num_heads))\n        slopes = get_slopes_power_of_2(closest_power)\n        # Add extra slopes\n        extra = get_slopes_power_of_2(2 * closest_power)\n        slopes.extend(extra[0::2][:num_heads - closest_power])\n        return slopes\n\ndef create_alibi_bias(seq_len, num_heads):\n    \"\"\"Create ALiBi attention bias.\"\"\"\n    # Distance matrix\n    context_position = torch.arange(seq_len)\n    memory_position = torch.arange(seq_len)\n    relative_position = memory_position[None, :] - context_position[:, None]\n\n    # Get slopes\n    slopes = torch.tensor(get_alibi_slopes(num_heads))\n\n    # Apply slopes to distances\n    alibi = slopes[:, None, None] * relative_position[None, :, :]\n    return alibi  # (num_heads, seq_len, seq_len)\n\n# Usage in attention\nnum_heads = 8\nseq_len = 2048\nalibi_bias = create_alibi_bias(seq_len, num_heads).to('cuda')\n\n# Add bias to attention scores\n# attn_scores shape: (batch, num_heads, seq_len, seq_len)\nattn_scores = attn_scores + alibi_bias\nattn_weights = torch.softmax(attn_scores, dim=-1)\n```\n\n### Position Interpolation for LLaMA\n\n```python\nfrom transformers import LlamaForCausalLM, LlamaTokenizer\n\n# Original context: 2048 tokens\nmodel = LlamaForCausalLM.from_pretrained(\"meta-llama\u002FLlama-2-7b-hf\")\n\n# Extend to 32k with position interpolation\n# Modify RoPE base frequency\nmodel.config.rope_scaling = {\n    \"type\": \"linear\",\n    \"factor\": 16.0  # 2048 * 16 = 32768\n}\n\n# Or use dynamic scaling\nmodel.config.rope_scaling = {\n    \"type\": \"dynamic\",\n    \"factor\": 16.0\n}\n\n# Fine-tune with long documents (minimal steps needed)\n# Position interpolation works out-of-the-box after this config change\n```\n\n## Core Concepts\n\n### 1. RoPE (Rotary Position Embeddings)\n\n**How it works:**\n- Encodes absolute position via rotation matrix\n- Provides relative position dependency in attention\n- Enables length extrapolation\n\n**Mathematical formulation:**\n```\nq_m = (W_q * x_m) * e^(imθ)\nk_n = (W_k * x_n) * e^(inθ)\n\nwhere θ_j = base^(-2j\u002Fd) for j ∈ [0, d\u002F2)\n```\n\n**Advantages:**\n- Decaying inter-token dependency with distance\n- Compatible with linear attention\n- Better extrapolation than absolute position encodings\n\n### 2. YaRN (Yet another RoPE extensioN)\n\n**Key innovation:**\n- NTK-aware interpolation (Neural Tangent Kernel)\n- Attention temperature scaling\n- Efficient context extension (10× less tokens vs baselines)\n\n**Parameters:**\n```python\n# YaRN configuration\nyarn_config = {\n    \"scale\": 16,                    # Extension factor\n    \"original_max_position\": 2048,  # Base context\n    \"extrapolation_factor\": 1.0,    # NTK parameter\n    \"attn_factor\": 1.0,             # Attention scaling\n    \"beta_fast\": 32,                # High-frequency scale\n    \"beta_slow\": 1,                 # Low-frequency scale\n}\n```\n\n**Performance:**\n- Extends LLaMA to 128k tokens\n- 2.5× less training steps than baselines\n- State-of-the-art context window extension\n\n### 3. ALiBi (Attention with Linear Biases)\n\n**Core idea:**\n- No positional embeddings added to tokens\n- Apply distance penalty directly to attention scores\n- Bias proportional to key-query distance\n\n**Formula:**\n```\nattention_bias[i, j] = -m * |i - j|\n\nwhere m = slope for each attention head\n```\n\n**Advantages:**\n- 11% faster training vs sinusoidal embeddings\n- 11% less memory usage\n- Strong length extrapolation (train 1k, test 2k+)\n- Inductive bias towards recency\n\n### 4. Position Interpolation\n\n**Technique:**\n- Linearly down-scale position indices\n- Interpolate within trained range (vs extrapolate beyond)\n- Minimal fine-tuning required\n\n**Formula:**\n```\n# Original: position indices [0, 1, 2, ..., L]\n# Extended: position indices [0, 0.5, 1.0, ..., L\u002F2]\n# (for 2× extension)\n\nscaled_position[i] = i \u002F extension_factor\n```\n\n**Results:**\n- LLaMA 7B-65B extended to 32k tokens\n- 1000 fine-tuning steps sufficient\n- 600× better stability than extrapolation\n\n## Method Comparison\n\n| Method | Max Context | Training Needed | Memory | Extrapolation | Best For |\n|--------|-------------|-----------------|--------|---------------|----------|\n| **RoPE** | 8k-32k | Full pre-training | Moderate | Good | New models |\n| **YaRN** | 32k-128k | Minimal (10× efficient) | Moderate | Excellent | Extending existing models |\n| **ALiBi** | Unlimited | Full pre-training | Low (-11%) | Excellent | Training from scratch |\n| **Position Interpolation** | 32k+ | Minimal (1k steps) | Moderate | Poor (by design) | Quick extension |\n\n## Implementation Patterns\n\nDetails, code examples and parameter tables: [references\u002Fimplementation-patterns.md](references\u002Fimplementation-patterns.md). Read it when this step applies.\n\n## Fine-tuning for Long Context\n\n### Minimal Fine-tuning (Position Interpolation)\n\n```python\nfrom transformers import Trainer, TrainingArguments\n\n# Extend model config\nmodel.config.max_position_embeddings = 32768\nmodel.config.rope_scaling = {\"type\": \"linear\", \"factor\": 16.0}\n\n# Training args (minimal steps needed)\ntraining_args = TrainingArguments(\n    output_dir=\".\u002Fllama-32k\",\n    num_train_epochs=1,\n    max_steps=1000,           # Only 1000 steps!\n    per_device_train_batch_size=1,\n    gradient_accumulation_steps=16,\n    learning_rate=2e-5,\n    warmup_steps=100,\n    logging_steps=10,\n    save_steps=500,\n)\n\n# Train on long documents\ntrainer = Trainer(\n    model=model,\n    args=training_args,\n    train_dataset=long_document_dataset,  # 32k token sequences\n)\n\ntrainer.train()\n```\n\n### YaRN Fine-tuning\n\n```bash\n# Clone YaRN implementation\ngit clone https:\u002F\u002Fgithub.com\u002Fjquesnelle\u002Fyarn\ncd yarn\n\n# Fine-tune LLaMA with YaRN\npython scripts\u002Ftrain.py \\\n    --model meta-llama\u002FLlama-2-7b-hf \\\n    --scale 16 \\\n    --rope_theta 10000 \\\n    --max_length 32768 \\\n    --batch_size 1 \\\n    --gradient_accumulation 16 \\\n    --steps 400 \\\n    --learning_rate 2e-5\n```\n\n## Best Practices\n\n### 1. Choose the Right Method\n\n```python\n# For NEW models (training from scratch)\nuse_method = \"ALiBi\"  # Best extrapolation, lowest memory\n\n# For EXTENDING existing RoPE models\nuse_method = \"YaRN\"  # Most efficient extension (10× less data)\n\n# For QUICK extension with minimal compute\nuse_method = \"Position Interpolation\"  # 1000 steps\n\n# For MODERATE extension with good efficiency\nuse_method = \"Linear RoPE Scaling\"  # Built-in, simple\n```\n\n### 2. Scaling Factor Selection\n\n```python\n# Conservative (safer, better quality)\nscaling_factor = 2.0  # 8k → 16k\n\n# Moderate (good balance)\nscaling_factor = 4.0  # 8k → 32k\n\n# Aggressive (requires more fine-tuning)\nscaling_factor = 8.0  # 8k → 64k\nscaling_factor = 16.0  # 8k → 128k\n\n# Rule: Larger factors need more fine-tuning steps\nsteps_needed = 100 * scaling_factor  # Rough estimate\n```\n\n### 3. Fine-tuning Data\n\n```python\n# ✅ Good: Long documents matching target length\ntrain_data = [\n    {\"text\": long_doc_32k_tokens},  # Full 32k\n    {\"text\": long_doc_24k_tokens},  # Varied lengths\n    {\"text\": long_doc_16k_tokens},\n]\n\n# ❌ Bad: Short documents (won't learn long context)\ntrain_data = [\n    {\"text\": short_doc_2k_tokens},\n]\n\n# Use datasets like:\n# - PG-19 (books, long texts)\n# - arXiv papers\n# - Long-form conversations\n# - GitHub repositories (concatenated files)\n```\n\n### 4. Avoid Common Pitfalls\n\n```python\n# ❌ Bad: Applying position interpolation without fine-tuning\nmodel.config.rope_scaling = {\"type\": \"linear\", \"factor\": 16.0}\n# Model will perform poorly without fine-tuning!\n\n# ✅ Good: Fine-tune after scaling\nmodel.config.rope_scaling = {\"type\": \"linear\", \"factor\": 16.0}\nfine_tune(model, long_documents, steps=1000)\n\n# ❌ Bad: Too aggressive scaling without data\nscale_to_1M_tokens()  # Won't work without massive fine-tuning\n\n# ✅ Good: Incremental scaling\n# 8k → 16k → 32k → 64k (fine-tune at each step)\n```\n\n## Production Deployment\n\n### Inference with Long Context\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\n# Load long-context model\nmodel = AutoModelForCausalLM.from_pretrained(\n    \"togethercomputer\u002FLLaMA-2-7B-32K\",  # 32k context\n    torch_dtype=torch.float16,\n    device_map=\"auto\"\n)\ntokenizer = AutoTokenizer.from_pretrained(\"togethercomputer\u002FLLaMA-2-7B-32K\")\n\n# Process long document\nlong_text = \"...\" * 30000  # 30k tokens\ninputs = tokenizer(long_text, return_tensors=\"pt\", truncation=False).to('cuda')\n\n# Generate\noutputs = model.generate(\n    **inputs,\n    max_new_tokens=512,\n    temperature=0.7,\n)\n\nresponse = tokenizer.decode(outputs[0], skip_special_tokens=True)\n```\n\n### Memory Optimization\n\n```python\n# Use gradient checkpointing for fine-tuning\nmodel.gradient_checkpointing_enable()\n\n# Use Flash Attention 2\nmodel = AutoModelForCausalLM.from_pretrained(\n    \"meta-llama\u002FLlama-2-7b-hf\",\n    attn_implementation=\"flash_attention_2\",  # 2-3× faster\n    torch_dtype=torch.float16\n)\n\n# Use paged attention (vLLM)\nfrom vllm import LLM\n\nllm = LLM(\n    model=\"togethercomputer\u002FLLaMA-2-7B-32K\",\n    max_model_len=32768,  # 32k context\n    gpu_memory_utilization=0.9\n)\n```\n\n## Resources\n\n- **RoPE Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.09864 (RoFormer)\n- **YaRN Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.00071\n- **ALiBi Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.12409 (Train Short, Test Long)\n- **Position Interpolation**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.15595\n- **HuggingFace RoPE Utils**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftransformers\u002Fblob\u002Fmain\u002Fsrc\u002Ftransformers\u002Fmodeling_rope_utils.py\n- **YaRN Implementation**: https:\u002F\u002Fgithub.com\u002Fjquesnelle\u002Fyarn\n- **Together AI Blog**: https:\u002F\u002Fwww.together.ai\u002Fblog\u002Fllama-2-7b-32k\n\n## See Also\n\n- `references\u002Frope.md` - Detailed RoPE implementation and theory\n- `references\u002Fextension_methods.md` - YaRN, ALiBi, Position Interpolation comparisons\n- `references\u002Ffine_tuning.md` - Complete fine-tuning guide for context extension\n\n## Agent operating procedure\n\n1. **Check the environment.** Confirm GPU memory, model checkpoint and licence, framework versions and input formats.\n2. **Pin down the inputs.** Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.\n3. **Run a small version first.** Run inference on a single sample at low resolution or short length first.\n4. **Execute the full task** using the instructions and references above.\n5. **Validate the result.** Inspect outputs qualitatively and with task metrics; check licence restrictions for generated content.\n6. **Report.** State what was run (versions, commands, parameters), what was checked, and what is still uncertain.\n\n| If this happens | Do this |\n|---|---|\n| Out-of-memory on large inputs | Reduce resolution or sequence length, use half precision, or tile the input. |\n| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (`help()`, `--help`, official docs), adapt, and tell the user. Never invent an API. |\n| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |\n\n**Integrity rules**\n\n- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.\n- Check model and dataset licences before commercial or published use.\n- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.\n- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.\n","# Long Context: Extending Transformer Context Windows\n\n## When to Use This Skill\n\nUse Long Context techniques when you need to:\n- **Process long documents** (32k, 64k, 128k+ tokens) with transformer models\n- **Extend context windows** of pre-trained models (LLaMA, Mistral, etc.)\n- **Implement efficient positional encodings** (RoPE, ALiBi)\n- **Train models** with length extrapolation capabilities\n- **Deploy models** that handle variable-length inputs efficiently\n- **Fine-tune** existing models for longer contexts with minimal compute\n\n**Key Techniques**: RoPE (Rotary Position Embeddings), YaRN, ALiBi (Attention with Linear Biases), Position Interpolation\n\n**Papers**: RoFormer (arXiv 2104.09864), YaRN (arXiv 2309.00071), ALiBi (arXiv 2108.12409), Position Interpolation (arXiv 2306.15595)\n\n## Installation\n\n```bash\n# HuggingFace Transformers (includes RoPE, YaRN support)\npip install transformers torch\n\n# For custom implementations\npip install einops  # Tensor operations\npip install rotary-embedding-torch  # Standalone RoPE\n\n# Optional: FlashAttention for efficiency\npip install flash-attn --no-build-isolation\n```\n\n## Quick Start\n\n### RoPE (Rotary Position Embeddings)\n\n```python\nimport torch\nimport torch.nn as nn\n\nclass RotaryEmbedding(nn.Module):\n    \"\"\"Rotary Position Embeddings (RoPE).\"\"\"\n\n    def __init__(self, dim, max_seq_len=8192, base=10000):\n        super().__init__()\n        # Compute inverse frequencies\n        inv_freq = 1.0 \u002F (base ** (torch.arange(0, dim, 2).float() \u002F dim))\n        self.register_buffer(\"inv_freq\", inv_freq)\n        self.max_seq_len = max_seq_len\n\n    def forward(self, seq_len, device):\n        # Position indices\n        t = torch.arange(seq_len, device=device).type_as(self.inv_freq)\n\n        # Compute frequencies\n        freqs = torch.outer(t, self.inv_freq)  # (seq_len, dim\u002F2)\n\n        # Compute sin and cos\n        emb = torch.cat((freqs, freqs), dim=-1)  # (seq_len, dim)\n        return emb.cos(), emb.sin()\n\ndef rotate_half(x):\n    \"\"\"Rotate half the hidden dimensions.\"\"\"\n    x1, x2 = x.chunk(2, dim=-1)\n    return torch.cat((-x2, x1), dim=-1)\n\ndef apply_rotary_pos_emb(q, k, cos, sin):\n    \"\"\"Apply rotary embeddings to queries and keys.\"\"\"\n    # q, k shape: (batch, heads, seq_len, dim)\n    q_embed = (q * cos) + (rotate_half(q) * sin)\n    k_embed = (k * cos) + (rotate_half(k) * sin)\n    return q_embed, k_embed\n\n# Usage\nrope = RotaryEmbedding(dim=64, max_seq_len=8192)\ncos, sin = rope(seq_len=2048, device='cuda')\n\n# In attention layer\nq_rotated, k_rotated = apply_rotary_pos_emb(query, key, cos, sin)\n```\n\n### ALiBi (Attention with Linear Biases)\n\n```python\ndef get_alibi_slopes(num_heads):\n    \"\"\"Get ALiBi slope values for each attention head.\"\"\"\n    def get_slopes_power_of_2(n):\n        start = 2 ** (-(2 ** -(math.log2(n) - 3)))\n        ratio = start\n        return [start * (ratio ** i) for i in range(n)]\n\n    if math.log2(num_heads).is_integer():\n        return get_slopes_power_of_2(num_heads)\n    else:\n        # Closest power of 2\n        closest_power = 2 ** math.floor(math.log2(num_heads))\n        slopes = get_slopes_power_of_2(closest_power)\n        # Add extra slopes\n        extra = get_slopes_power_of_2(2 * closest_power)\n        slopes.extend(extra[0::2][:num_heads - closest_power])\n        return slopes\n\ndef create_alibi_bias(seq_len, num_heads):\n    \"\"\"Create ALiBi attention bias.\"\"\"\n    # Distance matrix\n    context_position = torch.arange(seq_len)\n    memory_position = torch.arange(seq_len)\n    relative_position = memory_position[None, :] - context_position[:, None]\n\n    # Get slopes\n    slopes = torch.tensor(get_alibi_slopes(num_heads))\n\n    # Apply slopes to distances\n    alibi = slopes[:, None, None] * relative_position[None, :, :]\n    return alibi  # (num_heads, seq_len, seq_len)\n\n# Usage in attention\nnum_heads = 8\nseq_len = 2048\nalibi_bias = create_alibi_bias(seq_len, num_heads).to('cuda')\n\n# Add bias to attention scores\n# attn_scores shape: (batch, num_heads, seq_len, seq_len)\nattn_scores = attn_scores + alibi_bias\nattn_weights = torch.softmax(attn_scores, dim=-1)\n```\n\n### Position Interpolation for LLaMA\n\n```python\nfrom transformers import LlamaForCausalLM, LlamaTokenizer\n\n# Original context: 2048 tokens\nmodel = LlamaForCausalLM.from_pretrained(\"meta-llama\u002FLlama-2-7b-hf\")\n\n# Extend to 32k with position interpolation\n# Modify RoPE base frequency\nmodel.config.rope_scaling = {\n    \"type\": \"linear\",\n    \"factor\": 16.0  # 2048 * 16 = 32768\n}\n\n# Or use dynamic scaling\nmodel.config.rope_scaling = {\n    \"type\": \"dynamic\",\n    \"factor\": 16.0\n}\n\n# Fine-tune with long documents (minimal steps needed)\n# Position interpolation works out-of-the-box after this config change\n```\n\n## Core Concepts\n\n### 1. RoPE (Rotary Position Embeddings)\n\n**How it works:**\n- Encodes absolute position via rotation matrix\n- Provides relative position dependency in attention\n- Enables length extrapolation\n\n**Mathematical formulation:**\n```\nq_m = (W_q * x_m) * e^(imθ)\nk_n = (W_k * x_n) * e^(inθ)\n\nwhere θ_j = base^(-2j\u002Fd) for j ∈ [0, d\u002F2)\n```\n\n**Advantages:**\n- Decaying inter-token dependency with distance\n- Compatible with linear attention\n- Better extrapolation than absolute position encodings\n\n### 2. YaRN (Yet another RoPE extensioN)\n\n**Key innovation:**\n- NTK-aware interpolation (Neural Tangent Kernel)\n- Attention temperature scaling\n- Efficient context extension (10× less tokens vs baselines)\n\n**Parameters:**\n```python\n# YaRN configuration\nyarn_config = {\n    \"scale\": 16,                    # Extension factor\n    \"original_max_position\": 2048,  # Base context\n    \"extrapolation_factor\": 1.0,    # NTK parameter\n    \"attn_factor\": 1.0,             # Attention scaling\n    \"beta_fast\": 32,                # High-frequency scale\n    \"beta_slow\": 1,                 # Low-frequency scale\n}\n```\n\n**Performance:**\n- Extends LLaMA to 128k tokens\n- 2.5× less training steps than baselines\n- State-of-the-art context window extension\n\n### 3. ALiBi (Attention with Linear Biases)\n\n**Core idea:**\n- No positional embeddings added to tokens\n- Apply distance penalty directly to attention scores\n- Bias proportional to key-query distance\n\n**Formula:**\n```\nattention_bias[i, j] = -m * |i - j|\n\nwhere m = slope for each attention head\n```\n\n**Advantages:**\n- 11% faster training vs sinusoidal embeddings\n- 11% less memory usage\n- Strong length extrapolation (train 1k, test 2k+)\n- Inductive bias towards recency\n\n### 4. Position Interpolation\n\n**Technique:**\n- Linearly down-scale position indices\n- Interpolate within trained range (vs extrapolate beyond)\n- Minimal fine-tuning required\n\n**Formula:**\n```\n# Original: position indices [0, 1, 2, ..., L]\n# Extended: position indices [0, 0.5, 1.0, ..., L\u002F2]\n# (for 2× extension)\n\nscaled_position[i] = i \u002F extension_factor\n```\n\n**Results:**\n- LLaMA 7B-65B extended to 32k tokens\n- 1000 fine-tuning steps sufficient\n- 600× better stability than extrapolation\n\n## Method Comparison\n\n| Method | Max Context | Training Needed | Memory | Extrapolation | Best For |\n|--------|-------------|-----------------|--------|---------------|----------|\n| **RoPE** | 8k-32k | Full pre-training | Moderate | Good | New models |\n| **YaRN** | 32k-128k | Minimal (10× efficient) | Moderate | Excellent | Extending existing models |\n| **ALiBi** | Unlimited | Full pre-training | Low (-11%) | Excellent | Training from scratch |\n| **Position Interpolation** | 32k+ | Minimal (1k steps) | Moderate | Poor (by design) | Quick extension |\n\n## Implementation Patterns\n\nDetails, code examples and parameter tables: [references\u002Fimplementation-patterns.md](references\u002Fimplementation-patterns.md). Read it when this step applies.\n\n## Fine-tuning for Long Context\n\n### Minimal Fine-tuning (Position Interpolation)\n\n```python\nfrom transformers import Trainer, TrainingArguments\n\n# Extend model config\nmodel.config.max_position_embeddings = 32768\nmodel.config.rope_scaling = {\"type\": \"linear\", \"factor\": 16.0}\n\n# Training args (minimal steps needed)\ntraining_args = TrainingArguments(\n    output_dir=\".\u002Fllama-32k\",\n    num_train_epochs=1,\n    max_steps=1000,           # Only 1000 steps!\n    per_device_train_batch_size=1,\n    gradient_accumulation_steps=16,\n    learning_rate=2e-5,\n    warmup_steps=100,\n    logging_steps=10,\n    save_steps=500,\n)\n\n# Train on long documents\ntrainer = Trainer(\n    model=model,\n    args=training_args,\n    train_dataset=long_document_dataset,  # 32k token sequences\n)\n\ntrainer.train()\n```\n\n### YaRN Fine-tuning\n\n```bash\n# Clone YaRN implementation\ngit clone https:\u002F\u002Fgithub.com\u002Fjquesnelle\u002Fyarn\ncd yarn\n\n# Fine-tune LLaMA with YaRN\npython scripts\u002Ftrain.py \\\n    --model meta-llama\u002FLlama-2-7b-hf \\\n    --scale 16 \\\n    --rope_theta 10000 \\\n    --max_length 32768 \\\n    --batch_size 1 \\\n    --gradient_accumulation 16 \\\n    --steps 400 \\\n    --learning_rate 2e-5\n```\n\n## Best Practices\n\n### 1. Choose the Right Method\n\n```python\n# For NEW models (training from scratch)\nuse_method = \"ALiBi\"  # Best extrapolation, lowest memory\n\n# For EXTENDING existing RoPE models\nuse_method = \"YaRN\"  # Most efficient extension (10× less data)\n\n# For QUICK extension with minimal compute\nuse_method = \"Position Interpolation\"  # 1000 steps\n\n# For MODERATE extension with good efficiency\nuse_method = \"Linear RoPE Scaling\"  # Built-in, simple\n```\n\n### 2. Scaling Factor Selection\n\n```python\n# Conservative (safer, better quality)\nscaling_factor = 2.0  # 8k → 16k\n\n# Moderate (good balance)\nscaling_factor = 4.0  # 8k → 32k\n\n# Aggressive (requires more fine-tuning)\nscaling_factor = 8.0  # 8k → 64k\nscaling_factor = 16.0  # 8k → 128k\n\n# Rule: Larger factors need more fine-tuning steps\nsteps_needed = 100 * scaling_factor  # Rough estimate\n```\n\n### 3. Fine-tuning Data\n\n```python\n# ✅ Good: Long documents matching target length\ntrain_data = [\n    {\"text\": long_doc_32k_tokens},  # Full 32k\n    {\"text\": long_doc_24k_tokens},  # Varied lengths\n    {\"text\": long_doc_16k_tokens},\n]\n\n# ❌ Bad: Short documents (won't learn long context)\ntrain_data = [\n    {\"text\": short_doc_2k_tokens},\n]\n\n# Use datasets like:\n# - PG-19 (books, long texts)\n# - arXiv papers\n# - Long-form conversations\n# - GitHub repositories (concatenated files)\n```\n\n### 4. Avoid Common Pitfalls\n\n```python\n# ❌ Bad: Applying position interpolation without fine-tuning\nmodel.config.rope_scaling = {\"type\": \"linear\", \"factor\": 16.0}\n# Model will perform poorly without fine-tuning!\n\n# ✅ Good: Fine-tune after scaling\nmodel.config.rope_scaling = {\"type\": \"linear\", \"factor\": 16.0}\nfine_tune(model, long_documents, steps=1000)\n\n# ❌ Bad: Too aggressive scaling without data\nscale_to_1M_tokens()  # Won't work without massive fine-tuning\n\n# ✅ Good: Incremental scaling\n# 8k → 16k → 32k → 64k (fine-tune at each step)\n```\n\n## Production Deployment\n\n### Inference with Long Context\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\n# Load long-context model\nmodel = AutoModelForCausalLM.from_pretrained(\n    \"togethercomputer\u002FLLaMA-2-7B-32K\",  # 32k context\n    torch_dtype=torch.float16,\n    device_map=\"auto\"\n)\ntokenizer = AutoTokenizer.from_pretrained(\"togethercomputer\u002FLLaMA-2-7B-32K\")\n\n# Process long document\nlong_text = \"...\" * 30000  # 30k tokens\ninputs = tokenizer(long_text, return_tensors=\"pt\", truncation=False).to('cuda')\n\n# Generate\noutputs = model.generate(\n    **inputs,\n    max_new_tokens=512,\n    temperature=0.7,\n)\n\nresponse = tokenizer.decode(outputs[0], skip_special_tokens=True)\n```\n\n### Memory Optimization\n\n```python\n# Use gradient checkpointing for fine-tuning\nmodel.gradient_checkpointing_enable()\n\n# Use Flash Attention 2\nmodel = AutoModelForCausalLM.from_pretrained(\n    \"meta-llama\u002FLlama-2-7b-hf\",\n    attn_implementation=\"flash_attention_2\",  # 2-3× faster\n    torch_dtype=torch.float16\n)\n\n# Use paged attention (vLLM)\nfrom vllm import LLM\n\nllm = LLM(\n    model=\"togethercomputer\u002FLLaMA-2-7B-32K\",\n    max_model_len=32768,  # 32k context\n    gpu_memory_utilization=0.9\n)\n```\n\n## Resources\n\n- **RoPE Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2104.09864 (RoFormer)\n- **YaRN Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2309.00071\n- **ALiBi Paper**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2108.12409 (Train Short, Test Long)\n- **Position Interpolation**: https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.15595\n- **HuggingFace RoPE Utils**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftransformers\u002Fblob\u002Fmain\u002Fsrc\u002Ftransformers\u002Fmodeling_rope_utils.py\n- **YaRN Implementation**: https:\u002F\u002Fgithub.com\u002Fjquesnelle\u002Fyarn\n- **Together AI Blog**: https:\u002F\u002Fwww.together.ai\u002Fblog\u002Fllama-2-7b-32k\n\n## See Also\n\n- `references\u002Frope.md` - Detailed RoPE implementation and theory\n- `references\u002Fextension_methods.md` - YaRN, ALiBi, Position Interpolation comparisons\n- `references\u002Ffine_tuning.md` - Complete fine-tuning guide for context extension\n\n## Agent operating procedure\n\n1. **Check the environment.** Confirm GPU memory, model checkpoint and licence, framework versions and input formats.\n2. **Pin down the inputs.** Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.\n3. **Run a small version first.** Run inference on a single sample at low resolution or short length first.\n4. **Execute the full task** using the instructions and references above.\n5. **Validate the result.** Inspect outputs qualitatively and with task metrics; check licence restrictions for generated content.\n6. **Report.** State what was run (versions, commands, parameters), what was checked, and what is still uncertain.\n\n| If this happens | Do this |\n|---|---|\n| Out-of-memory on large inputs | Reduce resolution or sequence length, use half precision, or tile the input. |\n| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (`help()`, `--help`, official docs), adapt, and tell the user. Never invent an API. |\n| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |\n\n**Integrity rules**\n\n- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.\n- Check model and dataset licences before commercial or published use.\n- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.\n- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.\n",{"name":5,"license":15,"metadata":44,"description":7},{"tags":45,"version":46,"category":47,"maintainer":48,"dependencies":49},"Emerging Techniques, Long Context, RoPE, YaRN, ALiBi, Position Interpolation, Extended Context, Rotary Embeddings, Attention Bias, Context Extension, Positional Encoding","1.0.0","multimodal-and-emerging","Kalaris Labs","transformers, torch, flash-attn","main","0da4485785ee599388900be50db1635489d53056",".agents\u002Fskills\u002Fkal-long-context","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6","https:\u002F\u002Fdisha6.6.thenitishkr.in\u002F","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh","www.thenitishkr","thenitishkr | AI • Cyber-security • Research\r\n","2026-11-08 03:33:25",{"source":8,"skillName":5,"commands":60,"prompt":70,"zipAvailable":71},[61,64,67],{"tool":62,"command":63},"npx","npx skills add Tashima-Tarsh\u002FDisha6.6 --skill long-context",{"tool":65,"command":66},"bunx","bunx skills add Tashima-Tarsh\u002FDisha6.6 --skill long-context",{"tool":68,"command":69},"pnpm","pnpm dlx skills add Tashima-Tarsh\u002FDisha6.6 --skill long-context","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fkal-long-context\u002FSKILL.md",true,[73,79,83,87],{"path":74,"resourceType":75,"resourceTypeDesc":76,"sizeBytes":77,"sourceHtmlUrl":78},"references\u002Fextension_methods.md","REFERENCE","参考文档",14537,"https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fkal-long-context\u002Freferences\u002Fextension_methods.md",{"path":80,"resourceType":75,"resourceTypeDesc":76,"sizeBytes":81,"sourceHtmlUrl":82},"references\u002Ffine_tuning.md",16431,"https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fkal-long-context\u002Freferences\u002Ffine_tuning.md",{"path":84,"resourceType":75,"resourceTypeDesc":76,"sizeBytes":85,"sourceHtmlUrl":86},"references\u002Fimplementation-patterns.md",2343,"https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fkal-long-context\u002Freferences\u002Fimplementation-patterns.md",{"path":88,"resourceType":75,"resourceTypeDesc":76,"sizeBytes":89,"sourceHtmlUrl":90},"references\u002Frope.md",12121,"https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fkal-long-context\u002Freferences\u002Frope.md",[92,123,147,174,199,222],{"skillNo":93,"name":94,"nameZh":6,"description":95,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":96,"categories":97,"keywords":103,"publishedAt":39},"01M6QRD07D4TGNDK6VJ4EHF8BV","youtube-full","Use when the user needs YouTube transcripts, video search, channel browsing, playlist extraction, or content monitoring. Trigger phrases: 'get the transcript for', 'search YouTube for', 'what are the latest videos on', 'list this playlist', 'monitor this channel', or any request involving a YouTube URL, video ID, or @handle. Do NOT use for downloading video or audio files, YouTube engagement data (likes, comments), or private\u002Fage-restricted videos.","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fyoutube-full\u002FSKILL.md",[98],{"categoryNo":99,"categoryName":100,"taxonomyType":101,"taxonomyTypeDesc":102},"01JBSEEDCAT000000000000016","图像与多媒体","SCENARIO","使用场景",[104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,25,120,121,122],"youtube full","youtube","full","transcript","search","monitoring","data","Video","Audio","list","files","do","get","age","trigger","URL","browsing","request","youtube transcripts",{"skillNo":124,"name":125,"nameZh":6,"description":126,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":127,"categories":128,"keywords":129,"publishedAt":39},"01M6QRCJ3112VYPA97YG7KGMC2","x-twitter-growth","X\u002FTwitter growth engine for building audience, crafting viral content, and analyzing engagement. Use when the user wants to grow on X\u002FTwitter, write tweets or threads, analyze their X profile, research competitors on X, plan a posting strategy, or optimize engagement. Complements social-content (generic multi-platform) with X-specific depth: algorithm mechanics, thread engineering, reply strategy, profile optimization, and competitive intelligence via web search.","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fx-twitter-growth\u002FSKILL.md",[],[130,131,132,133,108,134,135,136,137,138,139,140,141,142,34,143,144,145,146,25],"twitter","x twitter growth","growth","research","content","engineering","plan","web search","competitors","analyze","engine","write","strategy","platform","grow","posting","competitive",{"skillNo":148,"name":149,"nameZh":6,"description":150,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":151,"categories":152,"keywords":156,"publishedAt":39},"01M6QRBN3EHQNJY42RERP7A149","write-a-skill","Create new agent skills with proper structure, progressive disclosure, and bundled resources. Use when user wants to create, write, build, or author a new skill.","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fwrite-a-skill\u002FSKILL.md",[153],{"categoryNo":154,"categoryName":155,"taxonomyType":101,"taxonomyTypeDesc":102},"01JBSEEDCAT000000000000023","Skill与智能体开发",[157,141,158,159,160,161,162,163,34,164,165,50,166,167,168,169,170,171,172,173],"write a skill","github","create","https","productivity","matt pocock","direct","example","build","new skill","structure","new agent","resources","with","tree","progressive disclosure","author",{"skillNo":175,"name":176,"nameZh":6,"description":177,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":178,"categories":179,"keywords":180,"publishedAt":39},"01M6QRB6G8P33PNGAF9Y69EJ9T","wrap-up","Close out a launched Claude Managed Agent — recap every primitive the founder now owns, regenerate the single-file overview page, and suggest the next 1-2 upgrades. Use when the user says \"wrap up\", \"close this out\", \"what do I own now\", \"give me the summary\", \"recap the agent\", or when the orchestrator routes phase=wrap-up. primitives_inventory.py tables everything owned (agent, environment, session, memory, outcome, deployment); overview_page.py regenerates a self-contained .\u002Fmy-agent\u002Fagent-overview.html; upgrade_suggester.py ranks the next moves from recorded deferrals plus standing hardening steps. Companion to run-without-you; the last stop before phase=done.","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fwrap-up\u002FSKILL.md",[],[181,182,183,184,185,186,187,188,189,34,190,115,191,192,193,194,195,196,197,198],"wrap up","up","wrap","gemini","deployment","run","file","hardening","closeout","stop","upgrade","founder","close","companion","you","tables","directions","phase",{"skillNo":200,"name":201,"nameZh":6,"description":202,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":203,"categories":204,"keywords":206,"publishedAt":39},"01M6QRA5AYC1RBFZ39GTWGS3N8","workflow-builder","Design and write deterministic multi-agent workflow scripts (.js files in .claude\u002Fworkflows\u002F) for Claude Code's Workflow tool. Use when a user wants to build, create, author, scaffold, or run a custom Claude Code workflow, orchestrate sub-agents (fan-out, pipeline, loop, judge-panel), or automate a repeatable multi-step task across fresh-context agents.","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fworkflow-builder\u002FSKILL.md",[205],{"categoryNo":154,"categoryName":155,"taxonomyType":101,"taxonomyTypeDesc":102},[207,208,209,210,211,159,186,212,213,214,215,216,141,19,114,217,218,219,220,221],"workflow","workflow builder","builder","design","workflows","code","orchestrate","pipeline","loop","multi-agent","across","fresh","scripts","judge","claude code workflow",{"skillNo":223,"name":224,"nameZh":6,"description":225,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":226,"categories":227,"keywords":228,"publishedAt":39},"01M6QR7QMCZESKQA0ERW1E657M","weekly-review","Use when someone wants to run a weekly review, close open loops, audit stalled projects and commitments, get their system back to trusted, restart a lapsed review habit, or says \"\u002Fcs:weekly-review\". Walks David Allen's three-phase loop — GET CLEAR, GET CURRENT, GET CREATIVE — with deterministic scripts that inventory open loops, gate the checklist with named gaps, and score commitment health 0-100.","https:\u002F\u002Fgithub.com\u002FTashima-Tarsh\u002FDisha6.6\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fweekly-review\u002FSKILL.md",[],[229,230,231,232,233,234,235,236,237,215,238,239,240,241,242,243,116,244,245,198],"review","weekly review","weekly","deterministic scripts","reflect","feeds","inventory","open","team","projects","project-management","scanner","conversation","back","project","sprint","cadence"]