2nd/10_Wiki/Topics/Other/Just-In-Time (JIT).md

---
id: wiki-2026-0508-just-in-time-jit
title: Just-In-Time (JIT)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [JIT Compilation, Dynamic Compilation, Tracing JIT]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [compiler, optimization, runtime, performance, llvm]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: jax
---

# Just-In-Time (JIT)

## 매 한 줄
> **"매 compile 매 first call, 매 reuse 매 hot path"**. JIT compilation 매 source / bytecode / IR 의 native code 의 runtime translation — 매 profile-guided 의 hot region 의 optimize. 2026 ML 시대 매 JAX `jit`, PyTorch 2.x `torch.compile`, Mojo, JuliaLang 매 mainstream.

## 매 핵심

### 매 JIT 의 mechanics
- **Trace**: 매 input shape / dtype 의 capture 매 computational graph.
- **Specialize**: 매 fixed shapes 의 specialized kernel 의 generate.
- **Cache**: 매 (function, signature) → compiled artifact.
- **Recompile**: 매 shape change → cache miss → recompile (avoid in hot loop).

### 매 vs AOT
- **AOT (ahead-of-time)**: rustc, gcc — startup 빠름, 매 dynamic dispatch 부족.
- **JIT**: 매 runtime info 의 use → better inlining, 매 startup 의 cost.
- **Hybrid**: PyPy, V8, .NET — interpret first, JIT after N invocations.

### 매 ML JIT 의 specifics
- **Static shape**: JAX `jit` 매 traced shape 의 specialize — dynamic shape 매 retrace.
- **XLA / Triton backend**: 매 fused kernels — memory bandwidth dominant.
- **Compilation cache**: persistent disk cache 매 cold-start 의 mitigate.

### 매 응용
1. ML training loop (JAX, torch.compile).
2. Numerical Python (Numba `@njit`).
3. JavaScript engines (V8, JSC).
4. Database query plans (Snowflake, DuckDB).

## 💻 패턴

### Pattern 1: JAX jit (2026 standard)
```python
import jax
import jax.numpy as jnp

@jax.jit
def attention(q, k, v):
    scores = jnp.einsum("bhqd,bhkd->bhqk", q, k) / jnp.sqrt(q.shape[-1])
    weights = jax.nn.softmax(scores, axis=-1)
    return jnp.einsum("bhqk,bhkd->bhqd", weights, v)

# First call: trace + compile (slow)
# Subsequent: cached (fast)
out = attention(q, k, v)
```

### Pattern 2: torch.compile (PyTorch 2.x)
```python
import torch

model = MyTransformer().cuda()
compiled = torch.compile(model, mode="reduce-overhead", fullgraph=True)

for batch in dataloader:
    out = compiled(batch)  # 매 first batch 매 slow, subsequent 매 fast
    out.backward()
```

### Pattern 3: Static argnums (avoid retrace)
```python
from functools import partial

@partial(jax.jit, static_argnums=(1,))
def topk(logits, k):
    return jax.lax.top_k(logits, k)

# 매 k=10 매 specialized — 매 k=20 매 separate compilation
topk(logits, 10)
topk(logits, 20)  # new compile
```

### Pattern 4: Numba JIT (Python → LLVM)
```python
from numba import njit
import numpy as np

@njit(cache=True, fastmath=True)
def mandelbrot(c, max_iter=100):
    z = 0.0 + 0.0j
    for i in range(max_iter):
        z = z * z + c
        if z.real * z.real + z.imag * z.imag > 4.0:
            return i
    return max_iter
```

### Pattern 5: AOT cache 의 prewarm
```python
import os
os.environ["JAX_COMPILATION_CACHE_DIR"] = "/var/cache/jax"

import jax
jax.config.update("jax_persistent_cache_min_entry_size_bytes", 0)
jax.config.update("jax_persistent_cache_min_compile_time_secs", 1.0)

# 매 first deployment 매 prewarm script 의 run — 매 next pods cold-start fast.
```

### Pattern 6: Recompilation detection
```python
import jax
from collections import Counter

class CompileCounter:
    def __init__(self):
        self.count = Counter()

    def trace(self, fn_name: str, sig: tuple):
        self.count[(fn_name, sig)] += 1
        if self.count[(fn_name, sig)] > 3:
            print(f"매 thrash: {fn_name} recompiled {self.count[(fn_name, sig)]} times")

# Usage: hook into jax.config or torch dynamo logger
```

### Pattern 7: Mojo JIT (2026)
```mojo
fn matmul[M: Int, N: Int, K: Int](a: Tensor, b: Tensor) -> Tensor:
    # 매 compile-time specialization 매 shapes — 매 SIMD auto-vectorize.
    var c = Tensor[DType.float32](M, N)
    for i in range(M):
        for j in range(N):
            var s: Float32 = 0
            for k in range(K):
                s += a[i, k] * b[k, j]
            c[i, j] = s
    return c
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Numerical Python tight loop | Numba `@njit`. |
| ML training | JAX `jit` 또는 `torch.compile`. |
| Variable shapes | Avoid JIT 또는 `dynamic=True`. |
| One-shot script | 매 JIT overhead 매 not worth. |
| Long-running server | JIT + persistent cache. |

**기본값**: ML 매 `torch.compile(mode="reduce-overhead")` 또는 `jax.jit`. Tight numerical loop 매 Numba.

## 🔗 Graph
- 부모: [[Performance-Optimization]]
- 변형: [[Tracing-JIT]]
- 응용: [[JAX]] · [[torch.compile]] · [[V8 Engine]]
- Adjacent: [[XLA]] · [[Triton]]

## 🤖 LLM 활용
**언제**: ML training/serving where compile cost amortizes (>100 calls), tight numerical loops, long-running services.
**언제 X**: One-shot scripts, code with constantly-changing shapes, debugging (use eager mode).

## ❌ 안티패턴
- **JIT in hot Python loop with varying shapes**: 매 retrace 매 every call — slower than eager.
- **No persistent cache**: 매 cold start 매 30s+ compile every deploy.
- **JIT debugging**: 매 stacktrace 매 useless — eager 의 disable JIT first.
- **Premature JIT**: profile first — 매 80% code 매 not bottleneck.

## 🧪 검증 / 중복
- Verified: JAX docs (2026), PyTorch 2.x docs, "Engineering a Compiler" (Cooper & Torczon), V8 design docs.
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — full content with JAX/torch.compile/Numba/Mojo 2026 patterns |