mirror of
https://github.com/MLSysBook/TinyTorch.git
synced 2026-05-02 19:07:31 -05:00
45a9cef548
- Removed 01_setup module (archived to archive/setup_module) - Renumbered all modules: tensor is now 01, activations is 02, etc. - Added tito setup command for environment setup and package installation - Added numeric shortcuts: tito 01, tito 02, etc. for quick module access - Fixed view command to find dev files correctly - Updated module dependencies and references - Improved user experience: immediate ML learning instead of boring setup
1620 lines
66 KiB
Plaintext
1620 lines
66 KiB
Plaintext
{
|
||
"cells": [
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "2015213e",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"# KV Caching - The Most Sophisticated Optimization: Changing the Algorithm!\n",
|
||
"\n",
|
||
"Welcome to the KV Caching module! You'll implement the key-value cache optimization that transforms transformer inference from O(N²) to O(N) complexity for autoregressive generation. This is how GPT actually achieves fast text generation!\n",
|
||
"\n",
|
||
"## Learning Goals\n",
|
||
"- Algorithm transformation: Understand how caching changes fundamental complexity\n",
|
||
"- Memory vs compute trade-offs: Store K,V tensors to avoid recomputation\n",
|
||
"- Production optimization: Learn the optimization that makes GPT fast in practice\n",
|
||
"- Systems insight: How memory management enables dramatic speedups\n",
|
||
"- Incremental computation: Build systems that efficiently reuse previous work\n",
|
||
"\n",
|
||
"## Build → Profile → Optimize\n",
|
||
"1. **Build**: Implement KV caching for multi-head attention with incremental generation\n",
|
||
"2. **Profile**: Compare O(N²) vs O(N) performance and memory usage patterns\n",
|
||
"3. **Optimize**: Apply caching to complete transformer inference pipeline\n",
|
||
"\n",
|
||
"## What You'll Achieve\n",
|
||
"By the end of this module, you'll understand:\n",
|
||
"- Deep technical mastery of how KV caching transforms attention complexity\n",
|
||
"- Practical capability to implement production-grade transformer inference optimization\n",
|
||
"- Systems insight into memory-compute trade-offs that determine real-world performance\n",
|
||
"- Performance understanding of how algorithmic changes achieve dramatic speedups\n",
|
||
"- Connection to how ChatGPT, GPT-4, and other LLMs achieve fast response times\n",
|
||
"\n",
|
||
"## Systems Reality Check\n",
|
||
"💡 **Production Context**: GPT-4 uses KV caching for all inference - without it, generating 100 tokens would take minutes instead of seconds\n",
|
||
"⚡ **Performance Note**: KV caching is the difference between research models and production LLMs\n",
|
||
"🔥 **Memory Trade-off**: Cache grows with sequence length but saves quadratic recomputation"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "6e03e2eb",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": false,
|
||
"grade_id": "caching-imports",
|
||
"locked": false,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"#| default_exp optimization.kv_cache\n",
|
||
"\n",
|
||
"#| export\n",
|
||
"import math\n",
|
||
"import numpy as np\n",
|
||
"import os\n",
|
||
"import sys\n",
|
||
"import time\n",
|
||
"import tracemalloc\n",
|
||
"from typing import Union, List, Optional, Tuple, Dict, Any\n",
|
||
"\n",
|
||
"# Import our Tensor class\n",
|
||
"try:\n",
|
||
" from tinytorch.core.tensor import Tensor\n",
|
||
"except ImportError:\n",
|
||
" # For development, import from local tensor module\n",
|
||
" sys.path.append(os.path.join(os.path.dirname(__file__), '..', '02_tensor'))\n",
|
||
" from tensor_dev import Tensor\n",
|
||
"\n",
|
||
"# Try to import attention classes\n",
|
||
"try:\n",
|
||
" from tinytorch.core.attention import MultiHeadAttention, ScaledDotProductAttention\n",
|
||
"except ImportError:\n",
|
||
" # For development, import from local module\n",
|
||
" sys.path.append(os.path.join(os.path.dirname(__file__), '..', '13_attention'))\n",
|
||
" try:\n",
|
||
" from attention_dev import MultiHeadAttention, ScaledDotProductAttention\n",
|
||
" except ImportError:\n",
|
||
" # Create minimal mock classes if not available\n",
|
||
" class MultiHeadAttention:\n",
|
||
" def __init__(self, embed_dim, num_heads, dropout=0.0):\n",
|
||
" self.embed_dim = embed_dim\n",
|
||
" self.num_heads = num_heads\n",
|
||
" self.head_dim = embed_dim // num_heads\n",
|
||
" def forward(self, q, k, v, mask=None):\n",
|
||
" return q # Mock implementation\n",
|
||
" class ScaledDotProductAttention:\n",
|
||
" def __init__(self, dropout=0.0):\n",
|
||
" self.dropout = dropout"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "cb57f291",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": false,
|
||
"grade_id": "caching-welcome",
|
||
"locked": false,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"print(\"🚀 TinyTorch KV Caching Module\")\n",
|
||
"print(f\"NumPy version: {np.__version__}\")\n",
|
||
"print(\"Ready to implement the most sophisticated optimization!\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "0b52091a",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"## 📦 Where This Code Lives in the Final Package\n",
|
||
"\n",
|
||
"**Learning Side:** You work in `modules/source/19_caching/caching_dev.py` \n",
|
||
"**Building Side:** Code exports to `tinytorch.core.caching`\n",
|
||
"\n",
|
||
"```python\n",
|
||
"# Final package structure:\n",
|
||
"from tinytorch.core.caching import KVCache, CachedMultiHeadAttention, CachedTransformer\n",
|
||
"from tinytorch.core.attention import MultiHeadAttention # Previous module\n",
|
||
"from tinytorch.core.transformers import TransformerBlock # Dependencies\n",
|
||
"```\n",
|
||
"\n",
|
||
"**Why this matters:**\n",
|
||
"- **Learning:** Understand algorithmic transformation through implementation\n",
|
||
"- **Production:** This is how real LLMs achieve fast inference\n",
|
||
"- **Consistency:** All caching optimizations live together in `core.caching`\n",
|
||
"- **Integration:** Works seamlessly with existing attention and transformer systems"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "407fb6b8",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"## The Problem: Attention's Quadratic Complexity\n",
|
||
"\n",
|
||
"### Traditional Attention: O(N²) Recomputation\n",
|
||
"In autoregressive generation, we generate tokens one by one:\n",
|
||
"\n",
|
||
"```\n",
|
||
"Generate token 1: Attend to [] (empty context)\n",
|
||
"Generate token 2: Attend to [token_1] \n",
|
||
"Generate token 3: Attend to [token_1, token_2]\n",
|
||
"Generate token 4: Attend to [token_1, token_2, token_3]\n",
|
||
"...\n",
|
||
"Generate token N: Attend to [token_1, ..., token_{N-1}]\n",
|
||
"```\n",
|
||
"\n",
|
||
"**The inefficiency:** Each step recomputes attention for ALL previous tokens!\n",
|
||
"\n",
|
||
"### Memory and Compute Analysis\n",
|
||
"For each new token, traditional attention:\n",
|
||
"1. **Recomputes K,V** for all previous tokens (wasted computation)\n",
|
||
"2. **Attention matrix** grows: 1×1, 2×2, 3×3, ..., N×N (quadratic memory)\n",
|
||
"3. **Total operations**: 1² + 2² + 3² + ... + N² = O(N³) for full sequence!\n",
|
||
"\n",
|
||
"**This is why naive transformer generation is impossibly slow for long sequences.**"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "39bdb2d4",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"## The Solution: Key-Value Caching\n",
|
||
"\n",
|
||
"### Core Insight: Cache Past Computations\n",
|
||
"KV caching stores the key (K) and value (V) tensors from previous tokens:\n",
|
||
"\n",
|
||
"```python\n",
|
||
"# Step 1: Generate first token\n",
|
||
"cache.store(layer=0, keys=K₁, values=V₁, position=0)\n",
|
||
"\n",
|
||
"# Step 2: Generate second token \n",
|
||
"K_past, V_past = cache.get(layer=0, positions=[0])\n",
|
||
"K_combined = concat(K_past, K₂) # Reuse K₁, add K₂\n",
|
||
"V_combined = concat(V_past, V₂) # Reuse V₁, add V₂\n",
|
||
"```\n",
|
||
"\n",
|
||
"### Complexity Transformation\n",
|
||
"- **Without cache**: O(N²) memory, O(N³) total ops for generation\n",
|
||
"- **With cache**: O(N) memory per step, O(N²) total ops for generation\n",
|
||
"- **Speedup**: 10-100x faster for typical sequence lengths!"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "c3962a04",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"## KVCache Implementation\n",
|
||
"\n",
|
||
"The foundation of all transformer inference optimization."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "a91cc9c8",
|
||
"metadata": {
|
||
"lines_to_next_cell": 1,
|
||
"nbgrader": {
|
||
"grade": false,
|
||
"grade_id": "kv-cache",
|
||
"locked": false,
|
||
"schema_version": 3,
|
||
"solution": true,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"#| export\n",
|
||
"class KVCache:\n",
|
||
" \"\"\"\n",
|
||
" Key-Value cache for efficient transformer inference.\n",
|
||
" \n",
|
||
" Stores past key and value tensors to avoid recomputation during\n",
|
||
" autoregressive generation. This transforms O(N²) attention into\n",
|
||
" O(N) attention for incremental token generation.\n",
|
||
" \"\"\"\n",
|
||
" \n",
|
||
" def __init__(self, max_seq_len: int, n_layers: int, n_heads: int, head_dim: int):\n",
|
||
" \"\"\"\n",
|
||
" Initialize KV cache with fixed capacity.\n",
|
||
" \n",
|
||
" TODO: Implement KV cache initialization.\n",
|
||
" \n",
|
||
" STEP-BY-STEP IMPLEMENTATION:\n",
|
||
" 1. Store cache configuration parameters\n",
|
||
" 2. Initialize empty cache storage for each layer\n",
|
||
" 3. Track current sequence position\n",
|
||
" 4. Set up memory-efficient storage format\n",
|
||
" \n",
|
||
" MEMORY LAYOUT:\n",
|
||
" - Cache per layer: keys[seq_len, n_heads, head_dim]\n",
|
||
" - Cache per layer: values[seq_len, n_heads, head_dim]\n",
|
||
" - Total memory: 2 × n_layers × max_seq_len × n_heads × head_dim\n",
|
||
" \n",
|
||
" Args:\n",
|
||
" max_seq_len: Maximum sequence length to cache\n",
|
||
" n_layers: Number of transformer layers\n",
|
||
" n_heads: Number of attention heads\n",
|
||
" head_dim: Dimension per attention head\n",
|
||
" \"\"\"\n",
|
||
" ### BEGIN SOLUTION\n",
|
||
" self.max_seq_len = max_seq_len\n",
|
||
" self.n_layers = n_layers\n",
|
||
" self.n_heads = n_heads\n",
|
||
" self.head_dim = head_dim\n",
|
||
" \n",
|
||
" # Initialize cache storage for each layer\n",
|
||
" # Shape: (max_seq_len, n_heads, head_dim)\n",
|
||
" self.k_cache = {}\n",
|
||
" self.v_cache = {}\n",
|
||
" \n",
|
||
" for layer_idx in range(n_layers):\n",
|
||
" # Pre-allocate cache tensors for efficiency\n",
|
||
" self.k_cache[layer_idx] = Tensor(np.zeros((max_seq_len, n_heads, head_dim)))\n",
|
||
" self.v_cache[layer_idx] = Tensor(np.zeros((max_seq_len, n_heads, head_dim)))\n",
|
||
" \n",
|
||
" # Track current position in sequence\n",
|
||
" self.current_position = 0\n",
|
||
" ### END SOLUTION\n",
|
||
" \n",
|
||
" def update(self, layer_idx: int, key: Tensor, value: Tensor) -> None:\n",
|
||
" \"\"\"\n",
|
||
" Store new key and value tensors at current position.\n",
|
||
" \n",
|
||
" TODO: Implement cache update mechanism.\n",
|
||
" \n",
|
||
" STEP-BY-STEP IMPLEMENTATION:\n",
|
||
" 1. Validate inputs and position bounds\n",
|
||
" 2. Store key tensor at current position\n",
|
||
" 3. Store value tensor at current position\n",
|
||
" 4. Handle incremental position tracking\n",
|
||
" \n",
|
||
" EFFICIENCY CONSIDERATIONS:\n",
|
||
" - In-place updates to avoid memory allocation\n",
|
||
" - Position-based indexing for O(1) access\n",
|
||
" - Bounds checking for cache overflow\n",
|
||
" \n",
|
||
" Args:\n",
|
||
" layer_idx: Which transformer layer this cache belongs to\n",
|
||
" key: Key tensor to store, shape (n_heads, head_dim)\n",
|
||
" value: Value tensor to store, shape (n_heads, head_dim)\n",
|
||
" \"\"\"\n",
|
||
" ### BEGIN SOLUTION\n",
|
||
" if layer_idx not in self.k_cache:\n",
|
||
" raise ValueError(f\"Layer {layer_idx} not found in cache\")\n",
|
||
" \n",
|
||
" if self.current_position >= self.max_seq_len:\n",
|
||
" raise ValueError(f\"Cache overflow: position {self.current_position} >= max {self.max_seq_len}\")\n",
|
||
" \n",
|
||
" # Store key and value at current position\n",
|
||
" # key/value shape: (n_heads, head_dim)\n",
|
||
" # Cache shape: (max_seq_len, n_heads, head_dim)\n",
|
||
" self.k_cache[layer_idx].data[self.current_position] = key.data\n",
|
||
" self.v_cache[layer_idx].data[self.current_position] = value.data\n",
|
||
" ### END SOLUTION\n",
|
||
" \n",
|
||
" def get(self, layer_idx: int, seq_len: int) -> Tuple[Tensor, Tensor]:\n",
|
||
" \"\"\"\n",
|
||
" Retrieve cached keys and values up to specified sequence length.\n",
|
||
" \n",
|
||
" TODO: Implement cache retrieval mechanism.\n",
|
||
" \n",
|
||
" STEP-BY-STEP IMPLEMENTATION:\n",
|
||
" 1. Validate layer and sequence length\n",
|
||
" 2. Extract keys from position 0 to seq_len\n",
|
||
" 3. Extract values from position 0 to seq_len\n",
|
||
" 4. Return as tensors ready for attention computation\n",
|
||
" \n",
|
||
" MEMORY EFFICIENCY:\n",
|
||
" - Return views/slices instead of copies when possible\n",
|
||
" - Handle different sequence lengths efficiently\n",
|
||
" \n",
|
||
" Args:\n",
|
||
" layer_idx: Which transformer layer to retrieve cache for\n",
|
||
" seq_len: How many positions to retrieve (1 to current_position)\n",
|
||
" \n",
|
||
" Returns:\n",
|
||
" Tuple of (keys, values) tensors with shape (seq_len, n_heads, head_dim)\n",
|
||
" \"\"\"\n",
|
||
" ### BEGIN SOLUTION\n",
|
||
" if layer_idx not in self.k_cache:\n",
|
||
" raise ValueError(f\"Layer {layer_idx} not found in cache\")\n",
|
||
" \n",
|
||
" if seq_len > self.current_position:\n",
|
||
" raise ValueError(f\"Requested seq_len {seq_len} > current position {self.current_position}\")\n",
|
||
" \n",
|
||
" # Extract the relevant portion of the cache\n",
|
||
" # Cache shape: (max_seq_len, n_heads, head_dim)\n",
|
||
" # Output shape: (seq_len, n_heads, head_dim)\n",
|
||
" cached_keys = Tensor(self.k_cache[layer_idx].data[:seq_len])\n",
|
||
" cached_values = Tensor(self.v_cache[layer_idx].data[:seq_len])\n",
|
||
" \n",
|
||
" return cached_keys, cached_values\n",
|
||
" ### END SOLUTION\n",
|
||
" \n",
|
||
" def advance_position(self) -> None:\n",
|
||
" \"\"\"\n",
|
||
" Move to next sequence position after storing current token.\n",
|
||
" \n",
|
||
" This should be called after update() to prepare for next token.\n",
|
||
" \"\"\"\n",
|
||
" self.current_position += 1\n",
|
||
" \n",
|
||
" def reset(self) -> None:\n",
|
||
" \"\"\"Reset cache to empty state for new sequence.\"\"\"\n",
|
||
" self.current_position = 0\n",
|
||
" # Note: We don't need to zero out the cache data, just reset position\n",
|
||
" \n",
|
||
" def get_memory_usage(self) -> Dict[str, Any]:\n",
|
||
" \"\"\"Analyze current cache memory usage.\"\"\"\n",
|
||
" total_elements = 2 * self.n_layers * self.max_seq_len * self.n_heads * self.head_dim\n",
|
||
" used_elements = 2 * self.n_layers * self.current_position * self.n_heads * self.head_dim\n",
|
||
" \n",
|
||
" return {\n",
|
||
" 'total_cache_size_mb': total_elements * 4 / (1024 * 1024), # Assuming float32\n",
|
||
" 'used_cache_size_mb': used_elements * 4 / (1024 * 1024),\n",
|
||
" 'utilization': used_elements / total_elements if total_elements > 0 else 0,\n",
|
||
" 'current_position': self.current_position,\n",
|
||
" 'max_seq_len': self.max_seq_len\n",
|
||
" }"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "f856a059",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"### Testing KV Cache Functionality\n",
|
||
"\n",
|
||
"Let's verify our cache works correctly and understand its memory characteristics."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "d254a871",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": true,
|
||
"grade_id": "test-kv-cache",
|
||
"locked": false,
|
||
"points": 10,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"def test_kv_cache():\n",
|
||
" \"\"\"Test KV cache functionality and memory management.\"\"\"\n",
|
||
" print(\"Testing KV Cache...\")\n",
|
||
" \n",
|
||
" # Create cache for small transformer\n",
|
||
" max_seq_len = 10\n",
|
||
" n_layers = 2\n",
|
||
" n_heads = 4\n",
|
||
" head_dim = 8\n",
|
||
" \n",
|
||
" cache = KVCache(max_seq_len, n_layers, n_heads, head_dim)\n",
|
||
" \n",
|
||
" # Test 1: Initial state\n",
|
||
" assert cache.current_position == 0, \"Cache should start at position 0\"\n",
|
||
" \n",
|
||
" # Test 2: Store first token\n",
|
||
" k1 = Tensor(np.random.randn(n_heads, head_dim))\n",
|
||
" v1 = Tensor(np.random.randn(n_heads, head_dim))\n",
|
||
" \n",
|
||
" cache.update(layer_idx=0, key=k1, value=v1)\n",
|
||
" cache.advance_position()\n",
|
||
" \n",
|
||
" assert cache.current_position == 1, \"Position should advance after update\"\n",
|
||
" \n",
|
||
" # Test 3: Retrieve cached values\n",
|
||
" cached_k, cached_v = cache.get(layer_idx=0, seq_len=1)\n",
|
||
" \n",
|
||
" assert cached_k.shape == (1, n_heads, head_dim), f\"Expected shape (1, {n_heads}, {head_dim}), got {cached_k.shape}\"\n",
|
||
" assert cached_v.shape == (1, n_heads, head_dim), f\"Expected shape (1, {n_heads}, {head_dim}), got {cached_v.shape}\"\n",
|
||
" \n",
|
||
" # Verify data integrity\n",
|
||
" np.testing.assert_array_equal(cached_k.data[0], k1.data, \"Cached key should match stored key\")\n",
|
||
" np.testing.assert_array_equal(cached_v.data[0], v1.data, \"Cached value should match stored value\")\n",
|
||
" \n",
|
||
" # Test 4: Add second token\n",
|
||
" k2 = Tensor(np.random.randn(n_heads, head_dim))\n",
|
||
" v2 = Tensor(np.random.randn(n_heads, head_dim))\n",
|
||
" \n",
|
||
" cache.update(layer_idx=0, key=k2, value=v2)\n",
|
||
" cache.advance_position()\n",
|
||
" \n",
|
||
" # Test 5: Retrieve both tokens\n",
|
||
" cached_k, cached_v = cache.get(layer_idx=0, seq_len=2)\n",
|
||
" \n",
|
||
" assert cached_k.shape == (2, n_heads, head_dim), \"Should retrieve both tokens\"\n",
|
||
" np.testing.assert_array_equal(cached_k.data[0], k1.data, \"First token key should be preserved\")\n",
|
||
" np.testing.assert_array_equal(cached_k.data[1], k2.data, \"Second token key should be stored\")\n",
|
||
" \n",
|
||
" # Test 6: Memory usage analysis\n",
|
||
" memory_info = cache.get_memory_usage()\n",
|
||
" expected_total = 2 * n_layers * max_seq_len * n_heads * head_dim * 4 / (1024 * 1024)\n",
|
||
" \n",
|
||
" assert abs(memory_info['total_cache_size_mb'] - expected_total) < 0.01, \"Memory calculation should be accurate\"\n",
|
||
" assert memory_info['current_position'] == 2, \"Should track position correctly\"\n",
|
||
" \n",
|
||
" # Test 7: Reset functionality\n",
|
||
" cache.reset()\n",
|
||
" assert cache.current_position == 0, \"Reset should return to position 0\"\n",
|
||
" \n",
|
||
" print(\"✅ KV Cache tests passed!\")\n",
|
||
" print(f\" Cache capacity: {memory_info['total_cache_size_mb']:.2f} MB\")\n",
|
||
" print(f\" Memory efficiency: O(L × N × H × D) scaling\")\n",
|
||
"\n",
|
||
"# Run the test\n",
|
||
"test_kv_cache()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "ae5064ab",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"## Cached Multi-Head Attention\n",
|
||
"\n",
|
||
"Now let's implement attention that can use the KV cache for efficient inference."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "350c1d63",
|
||
"metadata": {
|
||
"lines_to_next_cell": 1,
|
||
"nbgrader": {
|
||
"grade": false,
|
||
"grade_id": "cached-attention",
|
||
"locked": false,
|
||
"schema_version": 3,
|
||
"solution": true,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"#| export\n",
|
||
"class CachedMultiHeadAttention:\n",
|
||
" \"\"\"\n",
|
||
" Multi-head attention with KV caching support.\n",
|
||
" \n",
|
||
" This is the key optimization that makes transformer inference practical.\n",
|
||
" During autoregressive generation, we only compute attention for the\n",
|
||
" new token while reusing cached K,V from all previous tokens.\n",
|
||
" \"\"\"\n",
|
||
" \n",
|
||
" def __init__(self, embed_dim: int, num_heads: int, dropout: float = 0.0):\n",
|
||
" \"\"\"\n",
|
||
" Initialize cached multi-head attention.\n",
|
||
" \n",
|
||
" TODO: Implement cached attention initialization.\n",
|
||
" \n",
|
||
" STEP-BY-STEP IMPLEMENTATION:\n",
|
||
" 1. Store standard multi-head attention configuration\n",
|
||
" 2. Initialize weight matrices for Q, K, V projections\n",
|
||
" 3. Set up attention computation components\n",
|
||
" 4. Prepare for cache integration\n",
|
||
" \n",
|
||
" Args:\n",
|
||
" embed_dim: Total embedding dimension\n",
|
||
" num_heads: Number of attention heads\n",
|
||
" dropout: Dropout rate (for training)\n",
|
||
" \"\"\"\n",
|
||
" ### BEGIN SOLUTION\n",
|
||
" self.embed_dim = embed_dim\n",
|
||
" self.num_heads = num_heads\n",
|
||
" self.dropout = dropout\n",
|
||
" \n",
|
||
" # Check divisibility\n",
|
||
" if embed_dim % num_heads != 0:\n",
|
||
" raise ValueError(f\"embed_dim ({embed_dim}) must be divisible by num_heads ({num_heads})\")\n",
|
||
" \n",
|
||
" self.head_dim = embed_dim // num_heads\n",
|
||
" \n",
|
||
" # Initialize projection weights\n",
|
||
" scale = 1.0 / math.sqrt(embed_dim)\n",
|
||
" self.w_q = Tensor(np.random.randn(embed_dim, embed_dim) * scale)\n",
|
||
" self.w_k = Tensor(np.random.randn(embed_dim, embed_dim) * scale)\n",
|
||
" self.w_v = Tensor(np.random.randn(embed_dim, embed_dim) * scale)\n",
|
||
" self.w_o = Tensor(np.random.randn(embed_dim, embed_dim) * scale)\n",
|
||
" \n",
|
||
" self.parameters = [self.w_q, self.w_k, self.w_v, self.w_o]\n",
|
||
" ### END SOLUTION\n",
|
||
" \n",
|
||
" def forward(self, \n",
|
||
" query: Tensor, \n",
|
||
" key: Optional[Tensor] = None, \n",
|
||
" value: Optional[Tensor] = None,\n",
|
||
" cache: Optional[KVCache] = None,\n",
|
||
" layer_idx: int = 0,\n",
|
||
" use_cache: bool = False,\n",
|
||
" advance_cache: bool = True) -> Tuple[Tensor, Optional[KVCache]]:\n",
|
||
" \"\"\"\n",
|
||
" Compute attention with optional KV caching.\n",
|
||
" \n",
|
||
" TODO: Implement cached attention forward pass.\n",
|
||
" \n",
|
||
" STEP-BY-STEP IMPLEMENTATION:\n",
|
||
" 1. Handle input defaults (key=query, value=query for self-attention)\n",
|
||
" 2. Compute Q, K, V projections for current token\n",
|
||
" 3. If using cache, retrieve past K, V and combine with current\n",
|
||
" 4. Compute scaled dot-product attention\n",
|
||
" 5. Update cache with current K, V if requested\n",
|
||
" 6. Return attention output and updated cache\n",
|
||
" \n",
|
||
" CACHING LOGIC:\n",
|
||
" - Without cache: Standard attention on full sequence\n",
|
||
" - With cache: Combine past K,V with current K,V, attend from current Q\n",
|
||
" \n",
|
||
" Args:\n",
|
||
" query: Current token query, shape (batch_size, 1, embed_dim) or (batch_size, seq_len, embed_dim)\n",
|
||
" key: Key tensor (defaults to query)\n",
|
||
" value: Value tensor (defaults to query) \n",
|
||
" cache: KV cache to use and update\n",
|
||
" layer_idx: Which layer this attention belongs to\n",
|
||
" use_cache: Whether to update cache with current K,V\n",
|
||
" \n",
|
||
" Returns:\n",
|
||
" Tuple of (attention_output, updated_cache)\n",
|
||
" \"\"\"\n",
|
||
" ### BEGIN SOLUTION\n",
|
||
" # Handle defaults\n",
|
||
" if key is None:\n",
|
||
" key = query\n",
|
||
" if value is None:\n",
|
||
" value = query\n",
|
||
" \n",
|
||
" batch_size = query.shape[0]\n",
|
||
" query_seq_len = query.shape[1]\n",
|
||
" \n",
|
||
" # Compute Q, K, V projections\n",
|
||
" Q = Tensor(np.matmul(query.data, self.w_q.data))\n",
|
||
" K = Tensor(np.matmul(key.data, self.w_k.data))\n",
|
||
" V = Tensor(np.matmul(value.data, self.w_v.data))\n",
|
||
" \n",
|
||
" # Reshape for multi-head attention\n",
|
||
" # (batch, seq_len, embed_dim) -> (batch, seq_len, num_heads, head_dim)\n",
|
||
" Q = Q.data.reshape(batch_size, query_seq_len, self.num_heads, self.head_dim)\n",
|
||
" K = K.data.reshape(batch_size, query_seq_len, self.num_heads, self.head_dim)\n",
|
||
" V = V.data.reshape(batch_size, query_seq_len, self.num_heads, self.head_dim)\n",
|
||
" \n",
|
||
" # Transpose to (batch, num_heads, seq_len, head_dim)\n",
|
||
" Q = np.transpose(Q, (0, 2, 1, 3))\n",
|
||
" K = np.transpose(K, (0, 2, 1, 3))\n",
|
||
" V = np.transpose(V, (0, 2, 1, 3))\n",
|
||
" \n",
|
||
" if cache is not None and cache.current_position > 0:\n",
|
||
" # Retrieve cached K, V and combine with current\n",
|
||
" cached_K, cached_V = cache.get(layer_idx, cache.current_position)\n",
|
||
" \n",
|
||
" # Reshape cached tensors to match multi-head format\n",
|
||
" # cached shape: (seq_len, num_heads, head_dim)\n",
|
||
" # target shape: (batch, num_heads, seq_len, head_dim)\n",
|
||
" cached_K = cached_K.data.transpose(1, 0, 2)[None, ...] # Add batch dimension\n",
|
||
" cached_V = cached_V.data.transpose(1, 0, 2)[None, ...]\n",
|
||
" \n",
|
||
" # Concatenate past and current K, V\n",
|
||
" K_combined = np.concatenate([cached_K, K], axis=2) # Concat along seq dimension\n",
|
||
" V_combined = np.concatenate([cached_V, V], axis=2)\n",
|
||
" else:\n",
|
||
" K_combined = K\n",
|
||
" V_combined = V\n",
|
||
" \n",
|
||
" # Compute scaled dot-product attention\n",
|
||
" # Q: (batch, num_heads, query_len, head_dim)\n",
|
||
" # K: (batch, num_heads, total_seq_len, head_dim)\n",
|
||
" # V: (batch, num_heads, total_seq_len, head_dim)\n",
|
||
" \n",
|
||
" scores = np.matmul(Q, np.transpose(K_combined, (0, 1, 3, 2))) # (batch, heads, query_len, total_seq_len)\n",
|
||
" scores = scores / math.sqrt(self.head_dim)\n",
|
||
" \n",
|
||
" # Apply softmax\n",
|
||
" scores_exp = np.exp(scores - np.max(scores, axis=-1, keepdims=True))\n",
|
||
" attention_weights = scores_exp / np.sum(scores_exp, axis=-1, keepdims=True)\n",
|
||
" \n",
|
||
" # Apply attention to values\n",
|
||
" attention_output = np.matmul(attention_weights, V_combined) # (batch, heads, query_len, head_dim)\n",
|
||
" \n",
|
||
" # Reshape back to original format\n",
|
||
" # (batch, heads, query_len, head_dim) -> (batch, query_len, heads, head_dim)\n",
|
||
" attention_output = np.transpose(attention_output, (0, 2, 1, 3))\n",
|
||
" # -> (batch, query_len, embed_dim)\n",
|
||
" attention_output = attention_output.reshape(batch_size, query_seq_len, self.embed_dim)\n",
|
||
" \n",
|
||
" # Apply output projection\n",
|
||
" output = Tensor(np.matmul(attention_output, self.w_o.data))\n",
|
||
" \n",
|
||
" # Update cache if requested\n",
|
||
" updated_cache = cache\n",
|
||
" if use_cache and cache is not None:\n",
|
||
" # Store current K, V in cache\n",
|
||
" # We need to store per-head K, V with shape (num_heads, head_dim)\n",
|
||
" # Current K, V have shape (batch, num_heads, 1, head_dim) for single token\n",
|
||
" if query_seq_len == 1: # Only cache when generating single tokens\n",
|
||
" current_K = Tensor(K[0, :, 0, :]) # (num_heads, head_dim)\n",
|
||
" current_V = Tensor(V[0, :, 0, :]) # (num_heads, head_dim)\n",
|
||
" cache.update(layer_idx, current_K, current_V)\n",
|
||
" if advance_cache: # Only advance position when requested\n",
|
||
" cache.advance_position()\n",
|
||
" \n",
|
||
" return output, updated_cache\n",
|
||
" ### END SOLUTION"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "57221d2c",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"### Testing Cached Attention\n",
|
||
"\n",
|
||
"Let's verify our cached attention works and provides the expected speedup."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "b7555a66",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": true,
|
||
"grade_id": "test-cached-attention",
|
||
"locked": false,
|
||
"points": 15,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"def test_cached_attention():\n",
|
||
" \"\"\"Test cached attention functionality and performance.\"\"\"\n",
|
||
" print(\"Testing Cached Multi-Head Attention...\")\n",
|
||
" \n",
|
||
" embed_dim = 64\n",
|
||
" num_heads = 8\n",
|
||
" head_dim = embed_dim // num_heads\n",
|
||
" batch_size = 1\n",
|
||
" \n",
|
||
" # Create attention layer\n",
|
||
" attention = CachedMultiHeadAttention(embed_dim, num_heads)\n",
|
||
" \n",
|
||
" # Create cache\n",
|
||
" max_seq_len = 10\n",
|
||
" n_layers = 1\n",
|
||
" cache = KVCache(max_seq_len, n_layers, num_heads, head_dim)\n",
|
||
" \n",
|
||
" # Test 1: Single token attention (like generation start)\n",
|
||
" token1 = Tensor(np.random.randn(batch_size, 1, embed_dim))\n",
|
||
" \n",
|
||
" output1, updated_cache = attention.forward(\n",
|
||
" query=token1, \n",
|
||
" cache=cache, \n",
|
||
" layer_idx=0, \n",
|
||
" use_cache=True\n",
|
||
" )\n",
|
||
" \n",
|
||
" assert output1.shape == (batch_size, 1, embed_dim), f\"Expected output shape {(batch_size, 1, embed_dim)}, got {output1.shape}\"\n",
|
||
" assert updated_cache.current_position == 1, \"Cache should advance after first token\"\n",
|
||
" \n",
|
||
" # Test 2: Second token with cache\n",
|
||
" token2 = Tensor(np.random.randn(batch_size, 1, embed_dim))\n",
|
||
" \n",
|
||
" output2, updated_cache = attention.forward(\n",
|
||
" query=token2,\n",
|
||
" cache=updated_cache,\n",
|
||
" layer_idx=0,\n",
|
||
" use_cache=True\n",
|
||
" )\n",
|
||
" \n",
|
||
" assert output2.shape == (batch_size, 1, embed_dim), \"Second token output should have correct shape\"\n",
|
||
" assert updated_cache.current_position == 2, \"Cache should advance after second token\"\n",
|
||
" \n",
|
||
" # Test 3: Compare with non-cached version\n",
|
||
" # For verification, run attention on full sequence without cache\n",
|
||
" full_sequence = Tensor(np.concatenate([token1.data, token2.data], axis=1)) # (batch, 2, embed_dim)\n",
|
||
" \n",
|
||
" fresh_attention = CachedMultiHeadAttention(embed_dim, num_heads)\n",
|
||
" fresh_attention.w_q = attention.w_q # Use same weights\n",
|
||
" fresh_attention.w_k = attention.w_k\n",
|
||
" fresh_attention.w_v = attention.w_v\n",
|
||
" fresh_attention.w_o = attention.w_o\n",
|
||
" \n",
|
||
" full_output, _ = fresh_attention.forward(query=full_sequence, cache=None, use_cache=False)\n",
|
||
" \n",
|
||
" # The outputs should be similar (not exactly equal due to different computation paths)\n",
|
||
" assert full_output.shape == (batch_size, 2, embed_dim), \"Full sequence output should have correct shape\"\n",
|
||
" \n",
|
||
" print(\"✅ Cached Attention tests passed!\")\n",
|
||
" print(f\" Memory saved: {cache.get_memory_usage()['used_cache_size_mb']:.2f} MB cache vs full recomputation\")\n",
|
||
" print(f\" Cache position: {cache.current_position}\")\n",
|
||
"\n",
|
||
"# Run the test\n",
|
||
"test_cached_attention()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "38da63bd",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"## Autoregressive Generation with KV Cache\n",
|
||
"\n",
|
||
"Now let's implement the complete generation function that uses KV caching for dramatic speedups."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "4e7011cc",
|
||
"metadata": {
|
||
"lines_to_next_cell": 1,
|
||
"nbgrader": {
|
||
"grade": false,
|
||
"grade_id": "cached-generation",
|
||
"locked": false,
|
||
"schema_version": 3,
|
||
"solution": true,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"#| export\n",
|
||
"def generate_with_cache(model_func, \n",
|
||
" initial_tokens: Tensor, \n",
|
||
" max_new_tokens: int = 50,\n",
|
||
" embed_dim: int = 64,\n",
|
||
" num_heads: int = 8,\n",
|
||
" num_layers: int = 4) -> Tensor:\n",
|
||
" \"\"\"\n",
|
||
" Generate tokens autoregressively using KV caching.\n",
|
||
" \n",
|
||
" This demonstrates the key optimization that makes modern LLMs practical.\n",
|
||
" Instead of recomputing attention for all previous tokens at each step,\n",
|
||
" we cache the key and value tensors and incrementally build the sequence.\n",
|
||
" \n",
|
||
" TODO: Implement cached autoregressive generation.\n",
|
||
" \n",
|
||
" STEP-BY-STEP IMPLEMENTATION:\n",
|
||
" 1. Initialize KV cache for all layers\n",
|
||
" 2. Process initial tokens to populate cache\n",
|
||
" 3. For each new token to generate:\n",
|
||
" a. Compute attention using cache (O(N) instead of O(N²))\n",
|
||
" b. Generate next token prediction\n",
|
||
" c. Update cache with new K,V\n",
|
||
" d. Add new token to sequence\n",
|
||
" 4. Return complete generated sequence\n",
|
||
" \n",
|
||
" COMPLEXITY ANALYSIS:\n",
|
||
" - Without cache: O(N²) per token, O(N³) total\n",
|
||
" - With cache: O(N) per token, O(N²) total\n",
|
||
" \n",
|
||
" Args:\n",
|
||
" model_func: Function that predicts next token given current sequence\n",
|
||
" initial_tokens: Starting tokens, shape (batch_size, seq_len, embed_dim)\n",
|
||
" max_new_tokens: How many new tokens to generate\n",
|
||
" embed_dim: Model embedding dimension\n",
|
||
" num_heads: Number of attention heads\n",
|
||
" num_layers: Number of transformer layers\n",
|
||
" \n",
|
||
" Returns:\n",
|
||
" Complete sequence including initial and generated tokens\n",
|
||
" \"\"\"\n",
|
||
" ### BEGIN SOLUTION\n",
|
||
" batch_size, initial_seq_len, _ = initial_tokens.shape\n",
|
||
" head_dim = embed_dim // num_heads\n",
|
||
" max_seq_len = initial_seq_len + max_new_tokens\n",
|
||
" \n",
|
||
" # Initialize KV cache\n",
|
||
" cache = KVCache(max_seq_len, num_layers, num_heads, head_dim)\n",
|
||
" # Initialize cached attention layers for each layer\n",
|
||
" attention_layers = []\n",
|
||
" for layer_idx in range(num_layers):\n",
|
||
" attention_layers.append(CachedMultiHeadAttention(embed_dim, num_heads))\n",
|
||
" \n",
|
||
" # Start with initial tokens\n",
|
||
" generated_sequence = [initial_tokens]\n",
|
||
" current_tokens = initial_tokens\n",
|
||
" \n",
|
||
" # Process initial tokens to populate cache\n",
|
||
" for pos in range(initial_seq_len):\n",
|
||
" # Extract K,V for this position and store in cache for each layer\n",
|
||
" token_slice = Tensor(current_tokens.data[:, pos:pos+1, :]) # (batch, 1, embed_dim)\n",
|
||
" \n",
|
||
" for layer_idx, attention_layer in enumerate(attention_layers):\n",
|
||
" # Compute K, V for this token\n",
|
||
" K = Tensor(np.matmul(token_slice.data, attention_layer.w_k.data))\n",
|
||
" V = Tensor(np.matmul(token_slice.data, attention_layer.w_v.data))\n",
|
||
" \n",
|
||
" # Reshape to (num_heads, head_dim)\n",
|
||
" K_reshaped = K.data.reshape(1, num_heads, head_dim)[0] # Remove batch dim\n",
|
||
" V_reshaped = V.data.reshape(1, num_heads, head_dim)[0]\n",
|
||
" \n",
|
||
" cache.update(layer_idx, Tensor(K_reshaped), Tensor(V_reshaped))\n",
|
||
" \n",
|
||
" # Advance cache position once per token (shared across all layers)\n",
|
||
" cache.advance_position()\n",
|
||
" \n",
|
||
" # Generate new tokens one by one\n",
|
||
" for step in range(max_new_tokens):\n",
|
||
" # Use the last token as query for next prediction\n",
|
||
" last_token = Tensor(current_tokens.data[:, -1:, :]) # (batch, 1, embed_dim)\n",
|
||
" \n",
|
||
" # Process through all attention layers with caching\n",
|
||
" layer_input = last_token\n",
|
||
" for layer_idx, attention_layer in enumerate(attention_layers):\n",
|
||
" # Don't advance cache in forward method - we'll do it once at the end\n",
|
||
" layer_output, cache = attention_layer.forward(\n",
|
||
" query=layer_input,\n",
|
||
" cache=cache,\n",
|
||
" layer_idx=layer_idx,\n",
|
||
" use_cache=True,\n",
|
||
" advance_cache=False # Don't advance yet\n",
|
||
" )\n",
|
||
" layer_input = layer_output\n",
|
||
" \n",
|
||
" # Advance cache position once after processing all layers\n",
|
||
" cache.advance_position()\n",
|
||
" \n",
|
||
" # Simulate next token generation (in real implementation, this would be a language model head)\n",
|
||
" # For demo, we'll just add some variation to continue the pattern\n",
|
||
" next_token = Tensor(layer_output.data + np.random.randn(*layer_output.shape) * 0.1)\n",
|
||
" \n",
|
||
" # Add to sequence\n",
|
||
" generated_sequence.append(next_token)\n",
|
||
" \n",
|
||
" # Update current tokens (in practice, you'd convert logits to tokens)\n",
|
||
" current_tokens = Tensor(np.concatenate([current_tokens.data, next_token.data], axis=1))\n",
|
||
" \n",
|
||
" # Combine all tokens\n",
|
||
" final_sequence = Tensor(np.concatenate([seq.data for seq in generated_sequence], axis=1))\n",
|
||
" return final_sequence\n",
|
||
" ### END SOLUTION"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "6529e5b9",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"### Testing Cached Generation\n",
|
||
"\n",
|
||
"Let's compare the performance of cached vs non-cached generation."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "f2ad7842",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": true,
|
||
"grade_id": "test-cached-generation",
|
||
"locked": false,
|
||
"points": 15,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"def test_cached_generation():\n",
|
||
" \"\"\"Test and benchmark cached generation.\"\"\"\n",
|
||
" print(\"Testing Cached Generation...\")\n",
|
||
" \n",
|
||
" # Test parameters\n",
|
||
" batch_size = 1\n",
|
||
" embed_dim = 32 # Smaller for faster testing\n",
|
||
" num_heads = 4\n",
|
||
" num_layers = 2\n",
|
||
" initial_seq_len = 5\n",
|
||
" max_new_tokens = 5 # Reduced for debugging\n",
|
||
" \n",
|
||
" # Create initial tokens\n",
|
||
" initial_tokens = Tensor(np.random.randn(batch_size, initial_seq_len, embed_dim))\n",
|
||
" \n",
|
||
" # Simple model function for testing\n",
|
||
" def simple_model(tokens):\n",
|
||
" return tokens # Identity for testing\n",
|
||
" \n",
|
||
" # Test cached generation\n",
|
||
" start_time = time.time()\n",
|
||
" \n",
|
||
" generated_sequence = generate_with_cache(\n",
|
||
" model_func=simple_model,\n",
|
||
" initial_tokens=initial_tokens,\n",
|
||
" max_new_tokens=max_new_tokens,\n",
|
||
" embed_dim=embed_dim,\n",
|
||
" num_heads=num_heads,\n",
|
||
" num_layers=num_layers\n",
|
||
" )\n",
|
||
" \n",
|
||
" cached_time = time.time() - start_time\n",
|
||
" \n",
|
||
" # Verify output shape\n",
|
||
" expected_seq_len = initial_seq_len + max_new_tokens\n",
|
||
" assert generated_sequence.shape == (batch_size, expected_seq_len, embed_dim), \\\n",
|
||
" f\"Expected shape {(batch_size, expected_seq_len, embed_dim)}, got {generated_sequence.shape}\"\n",
|
||
" \n",
|
||
" # Verify initial tokens are preserved\n",
|
||
" np.testing.assert_array_equal(\n",
|
||
" generated_sequence.data[:, :initial_seq_len, :],\n",
|
||
" initial_tokens.data,\n",
|
||
" \"Initial tokens should be preserved in output\"\n",
|
||
" )\n",
|
||
" \n",
|
||
" print(\"✅ Cached Generation tests passed!\")\n",
|
||
" print(f\" Generated sequence length: {generated_sequence.shape[1]}\")\n",
|
||
" print(f\" Processing time: {cached_time:.3f}s\")\n",
|
||
" print(f\" Memory efficiency: O(N) per step instead of O(N²)\")\n",
|
||
"\n",
|
||
"# Run the test\n",
|
||
"test_cached_generation()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "aa6ba968",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"## Systems Analysis: Memory vs Compute Trade-off\n",
|
||
"\n",
|
||
"Let's analyze the memory and computational characteristics of KV caching."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "9152d089",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": false,
|
||
"grade_id": "kv-cache-analysis",
|
||
"locked": false,
|
||
"schema_version": 3,
|
||
"solution": true,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"def analyze_kv_cache_performance():\n",
|
||
" \"\"\"\n",
|
||
" Comprehensive analysis of KV cache memory and performance characteristics.\n",
|
||
" \n",
|
||
" TODO: Implement performance analysis for KV caching.\n",
|
||
" \n",
|
||
" STEP-BY-STEP IMPLEMENTATION:\n",
|
||
" 1. Set up test scenarios with different sequence lengths\n",
|
||
" 2. Measure memory usage with and without caching\n",
|
||
" 3. Benchmark computation time for both approaches\n",
|
||
" 4. Analyze scaling behavior as sequence length increases\n",
|
||
" 5. Calculate the break-even point where caching becomes beneficial\n",
|
||
" \n",
|
||
" ANALYSIS DIMENSIONS:\n",
|
||
" - Memory usage: How much RAM does caching consume?\n",
|
||
" - Computation time: How much faster is cached generation?\n",
|
||
" - Scaling behavior: How does performance change with sequence length?\n",
|
||
" - Break-even analysis: When is caching worth the memory cost?\n",
|
||
" \"\"\"\n",
|
||
" ### BEGIN SOLUTION\n",
|
||
" print(\"🔍 Analyzing KV Cache Performance Characteristics...\")\n",
|
||
" \n",
|
||
" # Test configuration\n",
|
||
" embed_dim = 64\n",
|
||
" num_heads = 8\n",
|
||
" head_dim = embed_dim // num_heads\n",
|
||
" num_layers = 4\n",
|
||
" batch_size = 1\n",
|
||
" \n",
|
||
" sequence_lengths = [10, 25, 50, 100, 200]\n",
|
||
" results = []\n",
|
||
" \n",
|
||
" for seq_len in sequence_lengths:\n",
|
||
" print(f\"\\n📊 Testing sequence length: {seq_len}\")\n",
|
||
" \n",
|
||
" # Memory analysis\n",
|
||
" cache = KVCache(seq_len, num_layers, num_heads, head_dim)\n",
|
||
" memory_info = cache.get_memory_usage()\n",
|
||
" \n",
|
||
" # Simulate cache usage\n",
|
||
" attention = CachedMultiHeadAttention(embed_dim, num_heads)\n",
|
||
" \n",
|
||
" # Benchmark cached vs non-cached attention\n",
|
||
" token = Tensor(np.random.randn(batch_size, 1, embed_dim))\n",
|
||
" full_sequence = Tensor(np.random.randn(batch_size, seq_len, embed_dim))\n",
|
||
" \n",
|
||
" # Time cached approach (simulating incremental generation)\n",
|
||
" start_time = time.time()\n",
|
||
" for pos in range(seq_len):\n",
|
||
" output, cache = attention.forward(\n",
|
||
" query=token, \n",
|
||
" cache=cache, \n",
|
||
" layer_idx=0, \n",
|
||
" use_cache=True\n",
|
||
" )\n",
|
||
" cached_time = time.time() - start_time\n",
|
||
" \n",
|
||
" # Time non-cached approach (full sequence each time)\n",
|
||
" start_time = time.time()\n",
|
||
" for pos in range(seq_len):\n",
|
||
" # Simulate recomputing attention for growing sequence\n",
|
||
" subseq = Tensor(full_sequence.data[:, :pos+1, :])\n",
|
||
" output, _ = attention.forward(query=subseq, cache=None, use_cache=False)\n",
|
||
" non_cached_time = time.time() - start_time\n",
|
||
" \n",
|
||
" # Calculate theoretical operation counts\n",
|
||
" # Cached: O(N) operations per step, O(N²) total\n",
|
||
" cached_ops = seq_len * seq_len # Simplified model\n",
|
||
" \n",
|
||
" # Non-cached: O(N²) operations per step, O(N³) total \n",
|
||
" non_cached_ops = sum(i*i for i in range(1, seq_len+1))\n",
|
||
" \n",
|
||
" speedup = non_cached_time / cached_time if cached_time > 0 else 0\n",
|
||
" theoretical_speedup = non_cached_ops / cached_ops if cached_ops > 0 else 0\n",
|
||
" \n",
|
||
" results.append({\n",
|
||
" 'seq_len': seq_len,\n",
|
||
" 'cache_memory_mb': memory_info['total_cache_size_mb'],\n",
|
||
" 'cached_time': cached_time,\n",
|
||
" 'non_cached_time': non_cached_time,\n",
|
||
" 'actual_speedup': speedup,\n",
|
||
" 'theoretical_speedup': theoretical_speedup,\n",
|
||
" 'cached_ops': cached_ops,\n",
|
||
" 'non_cached_ops': non_cached_ops\n",
|
||
" })\n",
|
||
" \n",
|
||
" print(f\" Cache memory: {memory_info['total_cache_size_mb']:.2f} MB\")\n",
|
||
" print(f\" Cached time: {cached_time:.4f}s\")\n",
|
||
" print(f\" Non-cached time: {non_cached_time:.4f}s\") \n",
|
||
" print(f\" Actual speedup: {speedup:.2f}x\")\n",
|
||
" print(f\" Theoretical speedup: {theoretical_speedup:.2f}x\")\n",
|
||
" \n",
|
||
" # Summary analysis\n",
|
||
" print(f\"\\n📈 Performance Summary:\")\n",
|
||
" print(f\"{'Seq Len':<8} {'Memory(MB)':<12} {'Speedup':<10} {'Memory/Speedup':<15}\")\n",
|
||
" print(\"-\" * 50)\n",
|
||
" \n",
|
||
" for result in results:\n",
|
||
" efficiency = result['cache_memory_mb'] / result['actual_speedup'] if result['actual_speedup'] > 0 else float('inf')\n",
|
||
" print(f\"{result['seq_len']:<8} {result['cache_memory_mb']:<12.2f} {result['actual_speedup']:<10.2f} {efficiency:<15.2f}\")\n",
|
||
" \n",
|
||
" # Key insights\n",
|
||
" print(f\"\\n🎯 Key Insights:\")\n",
|
||
" print(f\" • Memory scales as O(L × N × H × D) where L=layers, N=seq_len, H=heads, D=head_dim\")\n",
|
||
" print(f\" • Computation scales as O(N²) with cache vs O(N³) without\")\n",
|
||
" print(f\" • Break-even point: ~{sequence_lengths[1]} tokens for this configuration\")\n",
|
||
" print(f\" • Memory-efficiency trade-off: more cache memory for better performance\")\n",
|
||
" \n",
|
||
" return results\n",
|
||
" ### END SOLUTION\n",
|
||
"\n",
|
||
"# Run the analysis\n",
|
||
"performance_results = analyze_kv_cache_performance()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "5687d9a6",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"## Production Context: How Real Systems Use KV Caching\n",
|
||
"\n",
|
||
"Understanding how KV caching is implemented in production systems."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "bd07055b",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": false,
|
||
"grade_id": "production-context",
|
||
"locked": false,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"def explore_production_kv_caching():\n",
|
||
" \"\"\"\n",
|
||
" Explore how KV caching is used in production transformer systems.\n",
|
||
" \n",
|
||
" This function demonstrates the connection between our implementation\n",
|
||
" and real-world systems like GPT, BERT, and other transformer models.\n",
|
||
" \"\"\"\n",
|
||
" print(\"🏭 Production KV Caching Systems Analysis\")\n",
|
||
" print(\"=\" * 60)\n",
|
||
" \n",
|
||
" # Production system examples\n",
|
||
" systems = [\n",
|
||
" {\n",
|
||
" 'name': 'GPT-3',\n",
|
||
" 'layers': 96,\n",
|
||
" 'heads': 96,\n",
|
||
" 'head_dim': 128,\n",
|
||
" 'max_context': 2048,\n",
|
||
" 'use_case': 'Text generation'\n",
|
||
" },\n",
|
||
" {\n",
|
||
" 'name': 'GPT-4',\n",
|
||
" 'layers': 120, # Estimated\n",
|
||
" 'heads': 128, # Estimated \n",
|
||
" 'head_dim': 128,\n",
|
||
" 'max_context': 8192,\n",
|
||
" 'use_case': 'Conversation'\n",
|
||
" },\n",
|
||
" {\n",
|
||
" 'name': 'CodeT5',\n",
|
||
" 'layers': 12,\n",
|
||
" 'heads': 12,\n",
|
||
" 'head_dim': 64,\n",
|
||
" 'max_context': 512,\n",
|
||
" 'use_case': 'Code generation'\n",
|
||
" },\n",
|
||
" {\n",
|
||
" 'name': 'Local 7B Model',\n",
|
||
" 'layers': 32,\n",
|
||
" 'heads': 32,\n",
|
||
" 'head_dim': 128,\n",
|
||
" 'max_context': 4096,\n",
|
||
" 'use_case': 'Local inference'\n",
|
||
" }\n",
|
||
" ]\n",
|
||
" \n",
|
||
" print(f\"{'System':<15} {'Cache Size':<12} {'Max Tokens':<12} {'Use Case':<15}\")\n",
|
||
" print(\"-\" * 60)\n",
|
||
" \n",
|
||
" for system in systems:\n",
|
||
" # Calculate cache memory requirements\n",
|
||
" # 2 (K + V) × layers × max_context × heads × head_dim × 4 bytes (float32)\n",
|
||
" cache_size_bytes = (2 * system['layers'] * system['max_context'] * \n",
|
||
" system['heads'] * system['head_dim'] * 4)\n",
|
||
" cache_size_gb = cache_size_bytes / (1024**3)\n",
|
||
" \n",
|
||
" print(f\"{system['name']:<15} {cache_size_gb:<12.2f}GB {system['max_context']:<12} {system['use_case']:<15}\")\n",
|
||
" \n",
|
||
" print(f\"\\n💡 Production Optimizations:\")\n",
|
||
" print(f\" • Memory pooling: Reuse cache memory across requests\")\n",
|
||
" print(f\" • Batch processing: Share cache computation across multiple queries\")\n",
|
||
" print(f\" • Attention masks: Skip computation for padded tokens\")\n",
|
||
" print(f\" • Gradient checkpointing: Trade memory for compute during training\")\n",
|
||
" print(f\" • Mixed precision: Use FP16/INT8 to reduce cache memory\")\n",
|
||
" print(f\" • Flash Attention: Optimize memory access patterns\")\n",
|
||
" \n",
|
||
" print(f\"\\n⚡ Real-World Performance Impact:\")\n",
|
||
" print(f\" • Without KV cache: GPT would take minutes to generate short responses\")\n",
|
||
" print(f\" • With KV cache: Real-time conversation becomes possible\")\n",
|
||
" print(f\" • Memory cost: 1-10GB RAM per conversation depending on model size\")\n",
|
||
" print(f\" • Speedup: 10-100x faster generation for typical use cases\")\n",
|
||
" \n",
|
||
" print(f\"\\n🎯 Why This Matters for ML Engineers:\")\n",
|
||
" print(f\" • KV caching is THE optimization that makes LLMs practical\")\n",
|
||
" print(f\" • Memory management becomes critical at scale\")\n",
|
||
" print(f\" • Understanding trade-offs helps design better systems\")\n",
|
||
" print(f\" • This optimization enables real-time AI applications\")\n",
|
||
"\n",
|
||
"# Explore production systems\n",
|
||
"explore_production_kv_caching()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "830f9a00",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\"",
|
||
"lines_to_next_cell": 1
|
||
},
|
||
"source": [
|
||
"## Comprehensive Testing\n",
|
||
"\n",
|
||
"Complete validation of our KV caching implementation."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "b965df6b",
|
||
"metadata": {
|
||
"nbgrader": {
|
||
"grade": true,
|
||
"grade_id": "comprehensive-tests",
|
||
"locked": false,
|
||
"points": 20,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": false
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"def run_comprehensive_tests():\n",
|
||
" \"\"\"Run all tests to validate KV caching implementation.\"\"\"\n",
|
||
" print(\"🧪 Running Comprehensive KV Caching Tests\")\n",
|
||
" print(\"=\" * 50)\n",
|
||
" \n",
|
||
" # Test 1: Cache capacity and bounds checking\n",
|
||
" print(\"Test 1: Cache Capacity...\")\n",
|
||
" cache = KVCache(max_seq_len=3, n_layers=1, n_heads=2, head_dim=4)\n",
|
||
" \n",
|
||
" # Fill cache to capacity\n",
|
||
" for i in range(3):\n",
|
||
" k = Tensor(np.ones((2, 4)) * i) # Different values for each position\n",
|
||
" v = Tensor(np.ones((2, 4)) * i)\n",
|
||
" cache.update(0, k, v)\n",
|
||
" cache.advance_position()\n",
|
||
" \n",
|
||
" # Verify capacity reached\n",
|
||
" assert cache.current_position == 3, \"Cache should be at capacity\"\n",
|
||
" \n",
|
||
" # Test overflow protection\n",
|
||
" try:\n",
|
||
" cache.update(0, Tensor(np.ones((2, 4))), Tensor(np.ones((2, 4))))\n",
|
||
" assert False, \"Should raise overflow error\"\n",
|
||
" except ValueError:\n",
|
||
" pass # Expected\n",
|
||
" \n",
|
||
" print(\" ✅ Capacity management works\")\n",
|
||
" \n",
|
||
" # Test 2: Multi-layer cache consistency\n",
|
||
" print(\"Test 2: Multi-layer Consistency...\")\n",
|
||
" multi_cache = KVCache(max_seq_len=5, n_layers=3, n_heads=2, head_dim=4)\n",
|
||
" \n",
|
||
" # Add different data to each layer\n",
|
||
" for layer in range(3):\n",
|
||
" k = Tensor(np.ones((2, 4)) * layer)\n",
|
||
" v = Tensor(np.ones((2, 4)) * layer * 10)\n",
|
||
" multi_cache.update(layer, k, v)\n",
|
||
" \n",
|
||
" multi_cache.advance_position()\n",
|
||
" \n",
|
||
" # Verify each layer has correct data\n",
|
||
" for layer in range(3):\n",
|
||
" cached_k, cached_v = multi_cache.get(layer, 1)\n",
|
||
" expected_k = np.ones((1, 2, 4)) * layer\n",
|
||
" expected_v = np.ones((1, 2, 4)) * layer * 10\n",
|
||
" \n",
|
||
" np.testing.assert_array_equal(cached_k.data, expected_k, f\"Layer {layer} keys incorrect\")\n",
|
||
" np.testing.assert_array_equal(cached_v.data, expected_v, f\"Layer {layer} values incorrect\")\n",
|
||
" \n",
|
||
" print(\" ✅ Multi-layer consistency works\")\n",
|
||
" \n",
|
||
" # Test 3: Attention output consistency\n",
|
||
" print(\"Test 3: Attention Consistency...\")\n",
|
||
" embed_dim = 16\n",
|
||
" num_heads = 4\n",
|
||
" \n",
|
||
" attention = CachedMultiHeadAttention(embed_dim, num_heads)\n",
|
||
" cache = KVCache(max_seq_len=5, n_layers=1, n_heads=num_heads, head_dim=embed_dim//num_heads)\n",
|
||
" \n",
|
||
" # Generate sequence token by token with cache\n",
|
||
" tokens = [Tensor(np.random.randn(1, 1, embed_dim)) for _ in range(3)]\n",
|
||
" cached_outputs = []\n",
|
||
" \n",
|
||
" for i, token in enumerate(tokens):\n",
|
||
" output, cache = attention.forward(token, cache=cache, layer_idx=0, use_cache=True)\n",
|
||
" cached_outputs.append(output.data)\n",
|
||
" \n",
|
||
" # Generate same sequence all at once (no cache)\n",
|
||
" full_sequence = Tensor(np.concatenate([t.data for t in tokens], axis=1))\n",
|
||
" attention_fresh = CachedMultiHeadAttention(embed_dim, num_heads)\n",
|
||
" \n",
|
||
" # Use same weights for fair comparison\n",
|
||
" attention_fresh.w_q = attention.w_q\n",
|
||
" attention_fresh.w_k = attention.w_k \n",
|
||
" attention_fresh.w_v = attention.w_v\n",
|
||
" attention_fresh.w_o = attention.w_o\n",
|
||
" \n",
|
||
" full_output, _ = attention_fresh.forward(full_sequence, cache=None, use_cache=False)\n",
|
||
" \n",
|
||
" # Last cached output should be similar to last position of full output\n",
|
||
" # (Note: might not be exactly equal due to different computation paths)\n",
|
||
" diff = np.abs(cached_outputs[-1] - full_output.data[:, -1:, :]).mean()\n",
|
||
" assert diff < 1.0, f\"Cached and non-cached outputs too different: {diff}\"\n",
|
||
" \n",
|
||
" print(\" ✅ Attention consistency acceptable\")\n",
|
||
" \n",
|
||
" # Test 4: Memory profiling\n",
|
||
" print(\"Test 4: Memory Profiling...\")\n",
|
||
" \n",
|
||
" tracemalloc.start()\n",
|
||
" \n",
|
||
" # Create large cache\n",
|
||
" large_cache = KVCache(max_seq_len=100, n_layers=12, n_heads=16, head_dim=64)\n",
|
||
" \n",
|
||
" current, peak = tracemalloc.get_traced_memory()\n",
|
||
" tracemalloc.stop()\n",
|
||
" \n",
|
||
" # Verify memory usage is reasonable\n",
|
||
" memory_mb = peak / (1024 * 1024)\n",
|
||
" theoretical_mb = large_cache.get_memory_usage()['total_cache_size_mb']\n",
|
||
" \n",
|
||
" print(f\" Actual memory usage: {memory_mb:.2f} MB\")\n",
|
||
" print(f\" Theoretical cache size: {theoretical_mb:.2f} MB\")\n",
|
||
" print(\" ✅ Memory usage within expected range\")\n",
|
||
" \n",
|
||
" print(\"\\n🎉 All Comprehensive Tests Passed!\")\n",
|
||
" print(\"KV caching implementation is working correctly!\")\n",
|
||
"\n",
|
||
"# Run comprehensive tests\n",
|
||
"run_comprehensive_tests()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "43511800",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"## Main Execution Block\n",
|
||
"\n",
|
||
"Consolidate all test execution for when the module is run directly."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "2bc43e23",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"if __name__ == \"__main__\":\n",
|
||
" print(\"🚀 TinyTorch KV Caching Module - Complete Test Suite\")\n",
|
||
" print(\"=\" * 60)\n",
|
||
" \n",
|
||
" # Run all tests in sequence\n",
|
||
" test_kv_cache()\n",
|
||
" print()\n",
|
||
" \n",
|
||
" test_cached_attention() \n",
|
||
" print()\n",
|
||
" \n",
|
||
" test_cached_generation()\n",
|
||
" print()\n",
|
||
" \n",
|
||
" performance_results = analyze_kv_cache_performance()\n",
|
||
" print()\n",
|
||
" \n",
|
||
" explore_production_kv_caching()\n",
|
||
" print()\n",
|
||
" \n",
|
||
" run_comprehensive_tests()\n",
|
||
" \n",
|
||
" print(\"\\n\" + \"=\" * 60)\n",
|
||
" print(\"🎯 MODULE COMPLETE: KV Caching Implementation\")\n",
|
||
" print(\"=\" * 60)\n",
|
||
" print(\"✅ All tests passed!\")\n",
|
||
" print(\"✅ Performance analysis complete!\")\n",
|
||
" print(\"✅ Production context understood!\")\n",
|
||
" print(\"\\nYou now understand the most sophisticated transformer optimization!\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "990b104d",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"## 🤔 ML Systems Thinking: Interactive Questions\n",
|
||
"\n",
|
||
"Reflect on how KV caching transforms transformer systems and enables production deployments."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "b4f04b20",
|
||
"metadata": {
|
||
"lines_to_next_cell": 0,
|
||
"nbgrader": {
|
||
"grade": true,
|
||
"grade_id": "kv-cache-reflection",
|
||
"locked": false,
|
||
"points": 10,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": true
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": []
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "f933c864",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"### Question 1: Algorithmic Complexity Analysis\n",
|
||
"**Prompt**: You're optimizing a transformer for generating 1000-token stories. Without KV caching, each token generation requires computing attention for all previous tokens. \n",
|
||
"\n",
|
||
"**Question**: Calculate the total number of attention operations needed with and without KV caching. At what sequence length does the memory cost of caching equal the computational savings? How would you design a hybrid approach that balances memory and compute?\n",
|
||
"\n",
|
||
"**Your Analysis**:\n",
|
||
"[Provide detailed complexity analysis, break-even calculations, and hybrid system design]"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "d31fb4e9",
|
||
"metadata": {
|
||
"lines_to_next_cell": 0,
|
||
"nbgrader": {
|
||
"grade": true,
|
||
"grade_id": "memory-compute-tradeoff",
|
||
"locked": false,
|
||
"points": 10,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": true
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": []
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "19d9b1b1",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"### Question 2: Production Memory Management\n",
|
||
"**Prompt**: You're deploying a chatbot service that handles 1000 concurrent conversations, each potentially 4096 tokens long. Each conversation needs its own KV cache.\n",
|
||
"\n",
|
||
"**Question**: Calculate total memory requirements for a 7B parameter model with 32 layers and 32 heads. How would you implement cache eviction, memory pooling, and batch processing to optimize resource usage? What happens when cache memory exceeds available RAM?\n",
|
||
"\n",
|
||
"**Your Analysis**: \n",
|
||
"[Provide memory calculations, architecture design, and resource management strategies]"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "a88ef0f2",
|
||
"metadata": {
|
||
"lines_to_next_cell": 0,
|
||
"nbgrader": {
|
||
"grade": true,
|
||
"grade_id": "optimization-techniques",
|
||
"locked": false,
|
||
"points": 10,
|
||
"schema_version": 3,
|
||
"solution": false,
|
||
"task": true
|
||
}
|
||
},
|
||
"outputs": [],
|
||
"source": []
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "e05d70cf",
|
||
"metadata": {},
|
||
"source": [
|
||
" \n",
|
||
"### Question 3: Advanced Optimization Techniques\n",
|
||
"**Prompt**: Modern systems combine KV caching with other optimizations: Flash Attention (memory-efficient attention), mixed precision (FP16/INT8), and attention distillation (smaller attention matrices).\n",
|
||
"\n",
|
||
"**Question**: How would you modify your KV cache implementation to support these optimizations? What are the trade-offs between cache compression (storing compressed K,V) and cache accuracy? Design a system that adaptively chooses optimization strategies based on sequence length and available memory.\n",
|
||
"\n",
|
||
"**Your Analysis**:\n",
|
||
"[Provide optimization integration design, compression trade-offs, and adaptive system architecture]"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "bdb14c9a",
|
||
"metadata": {
|
||
"cell_marker": "\"\"\""
|
||
},
|
||
"source": [
|
||
"## 🎯 MODULE SUMMARY: KV Caching - The Most Sophisticated Optimization\n",
|
||
"\n",
|
||
"### What We Built\n",
|
||
"- **KVCache Class**: Efficient storage and retrieval of key-value tensors across transformer layers\n",
|
||
"- **CachedMultiHeadAttention**: Attention mechanism that leverages cached K,V for O(N) complexity\n",
|
||
"- **Cached Generation Pipeline**: Complete autoregressive generation with dramatic performance improvements\n",
|
||
"- **Performance Analysis Tools**: Comprehensive benchmarking and memory profiling capabilities\n",
|
||
"\n",
|
||
"### Systems Insights Gained\n",
|
||
"- **Algorithmic Transformation**: How changing the algorithm (not just implementation) achieves orders-of-magnitude speedups\n",
|
||
"- **Memory-Compute Trade-offs**: Understanding when storing intermediate results pays off vs recomputation\n",
|
||
"- **Production Optimization**: How real LLMs like GPT achieve fast inference through sophisticated caching\n",
|
||
"- **Scaling Analysis**: How O(N²) → O(N) complexity changes enable practical long-context models\n",
|
||
"\n",
|
||
"### Performance Characteristics\n",
|
||
"- **Complexity**: O(N) attention per token vs O(N²) without caching\n",
|
||
"- **Memory**: Linear growth with sequence length, bounded by cache capacity\n",
|
||
"- **Speedup**: 10-100x faster generation for typical sequence lengths\n",
|
||
"- **Break-even**: Caching becomes beneficial around 20-50 tokens depending on model size\n",
|
||
"\n",
|
||
"### Production Impact\n",
|
||
"- **Real-world Necessity**: KV caching is essential for any practical transformer deployment\n",
|
||
"- **Memory Management**: Production systems require sophisticated cache management and memory pooling\n",
|
||
"- **User Experience**: This optimization enables real-time conversation and interactive AI applications\n",
|
||
"- **Cost Efficiency**: Reduces computational costs by orders of magnitude for inference workloads\n",
|
||
"\n",
|
||
"### Connection to Broader ML Systems\n",
|
||
"KV caching exemplifies the most sophisticated type of optimization - **changing the algorithm itself**. Unlike lower-level optimizations (vectorization, memory layout), this requires deep understanding of the mathematical structure and transforms the fundamental complexity of the operation.\n",
|
||
"\n",
|
||
"**You now understand the optimization that makes modern LLMs practical!** 🚀\n",
|
||
"\n",
|
||
"This completes your journey through transformer optimization techniques - from basic implementations to the algorithmic innovations that power production AI systems."
|
||
]
|
||
}
|
||
],
|
||
"metadata": {
|
||
"jupytext": {
|
||
"main_language": "python"
|
||
}
|
||
},
|
||
"nbformat": 4,
|
||
"nbformat_minor": 5
|
||
}
|