aisimulate-core 0.12.0

Engine-neutral inference simulation, deterministic replay, and performance modeling
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<title>AIC Perf Engine — Python + Rust Split (Design Doc)</title>
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<style>
  :root {
    --bg: #fdfdfd;
    --fg: #222;
    --muted: #6a6a6a;
    --accent: #2563eb;
    --accent-soft: #eff6ff;
    --warn: #c2410c;
    --warn-soft: #fff7ed;
    --good: #047857;
    --good-soft: #ecfdf5;
    --code-bg: #f5f5f4;
    --border: #e5e5e5;
    --sidebar-bg: #fafafa;
  }
  * { box-sizing: border-box; }
  body {
    margin: 0;
    font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto,
                 "Helvetica Neue", Arial, "PingFang SC", "Hiragino Sans GB",
                 "Microsoft YaHei", sans-serif;
    color: var(--fg);
    background: var(--bg);
    line-height: 1.65;
    font-size: 15.5px;
  }
  .layout { display: flex; max-width: 1200px; margin: 0 auto; }
  aside.toc {
    position: sticky;
    top: 0;
    align-self: flex-start;
    width: 260px;
    height: 100vh;
    overflow-y: auto;
    padding: 28px 18px;
    background: var(--sidebar-bg);
    border-right: 1px solid var(--border);
    font-size: 13.5px;
  }
  aside.toc h2 {
    margin-top: 0;
    font-size: 13px;
    text-transform: uppercase;
    letter-spacing: 0.06em;
    color: var(--muted);
  }
  aside.toc ol { padding-left: 22px; margin: 0; }
  aside.toc ol ol { padding-left: 16px; margin-top: 4px; }
  aside.toc li { margin: 4px 0; }
  aside.toc a { color: var(--fg); text-decoration: none; }
  aside.toc a:hover { color: var(--accent); }
  main { padding: 36px 48px 80px; max-width: 920px; }
  h1 { font-size: 2.1em; margin: 0 0 8px; line-height: 1.2; }
  h1 + .subtitle { color: var(--muted); margin-bottom: 28px; font-size: 1.05em; }
  h2 { font-size: 1.5em; margin-top: 52px; padding-bottom: 6px;
       border-bottom: 1px solid var(--border); }
  h3 { font-size: 1.18em; margin-top: 32px; color: #1a1a1a; }
  h4 { font-size: 1.0em; margin-top: 22px; color: var(--muted);
       text-transform: uppercase; letter-spacing: 0.04em; }
  p, ul, ol { margin: 12px 0; }
  ul, ol { padding-left: 28px; }
  li { margin: 4px 0; }
  code {
    font-family: "SF Mono", Menlo, Consolas, "Liberation Mono", monospace;
    font-size: 0.88em;
    background: var(--code-bg);
    padding: 1.5px 5px;
    border-radius: 3px;
  }
  pre {
    background: var(--code-bg);
    padding: 14px 18px;
    border-radius: 6px;
    overflow-x: auto;
    font-size: 13px;
    line-height: 1.5;
    border: 1px solid var(--border);
  }
  pre code { background: transparent; padding: 0; font-size: inherit; }
  table {
    border-collapse: collapse;
    width: 100%;
    margin: 16px 0;
    font-size: 14px;
  }
  th, td {
    border: 1px solid var(--border);
    padding: 8px 12px;
    text-align: left;
    vertical-align: top;
  }
  th { background: var(--code-bg); font-weight: 600; }
  blockquote {
    margin: 16px 0;
    padding: 12px 16px;
    border-left: 3px solid var(--accent);
    background: var(--accent-soft);
    color: #1d4ed8;
  }
  .callout {
    margin: 18px 0;
    padding: 12px 16px 12px 18px;
    border-radius: 6px;
    border-left: 3px solid;
  }
  .callout .label {
    display: block;
    font-size: 12px;
    text-transform: uppercase;
    letter-spacing: 0.06em;
    font-weight: 600;
    margin-bottom: 6px;
  }
  .callout.tldr { background: var(--accent-soft); border-color: var(--accent); }
  .callout.tldr .label { color: var(--accent); }
  .callout.warn { background: var(--warn-soft); border-color: var(--warn); }
  .callout.warn .label { color: var(--warn); }
  .callout.good { background: var(--good-soft); border-color: var(--good); }
  .callout.good .label { color: var(--good); }
  .ascii {
    font-family: "SF Mono", Menlo, Consolas, monospace;
    font-size: 12.5px;
    line-height: 1.35;
    background: #fbfaf6;
    border: 1px solid var(--border);
    border-radius: 6px;
    padding: 16px;
    margin: 16px 0;
    white-space: pre;
    overflow-x: auto;
  }
  hr {
    border: none;
    border-top: 1px solid var(--border);
    margin: 32px 0;
  }
  .pill {
    display: inline-block;
    padding: 1px 8px;
    border-radius: 10px;
    font-size: 12px;
    font-weight: 600;
    background: var(--code-bg);
    color: var(--muted);
    margin-left: 6px;
  }
  .pill.ok { background: var(--good-soft); color: var(--good); }
  .pill.tbd { background: var(--warn-soft); color: var(--warn); }
  footer { margin-top: 60px; padding-top: 16px;
    border-top: 1px solid var(--border);
    color: var(--muted); font-size: 13px; }
  @media (max-width: 900px) {
    .layout { flex-direction: column; }
    aside.toc { position: static; width: 100%; height: auto;
                border-right: none; border-bottom: 1px solid var(--border); }
    main { padding: 20px; }
  }
</style>
</head>
<body>
<div class="layout">

<aside class="toc">
  <h2>Contents</h2>
  <ol>
    <li><a href="#tldr">TL;DR</a></li>
    <li><a href="#motivation">Motivation</a>
      <ol>
        <li><a href="#current-state">Current state</a></li>
        <li><a href="#pain-points">Pain points</a></li>
      </ol>
    </li>
    <li><a href="#proposal">Proposal: Python + Rust Split</a>
      <ol>
        <li><a href="#layers">Three-layer architecture</a></li>
        <li><a href="#engine">The Engine abstraction</a></li>
        <li><a href="#two-callers">Two callers, one Rust core</a></li>
      </ol>
    </li>
    <li><a href="#poc">PoC Implementation</a>
      <ol>
        <li><a href="#poc-layout">Layout</a></li>
        <li><a href="#poc-flow">End-to-end flow</a></li>
        <li><a href="#poc-build">Build / run</a></li>
      </ol>
    </li>
    <li><a href="#validation">Validation Results</a></li>
    <li><a href="#decisions">Open Decisions</a></li>
    <li><a href="#roadmap">Roadmap</a></li>
    <li><a href="#appendix">Appendix: Naming Notes</a></li>
  </ol>
</aside>

<main>

<h1>AIC Perf Engine — Python + Rust Split</h1>
<div class="subtitle">Design doc &middot; greenfield PoC validated locally &middot; June 2026</div>

<div class="callout tldr">
  <span class="label">TL;DR</span>
  <p>Split AIC's perf engine into two layers: <strong>Python</strong> owns model
  building (HF parsing, op-list composition, quant resolution), <strong>Rust</strong>
  owns execution (compiled artifact + per-op latency math, hot path). Both
  Python sweep callers and external Rust callers (e.g. Dynamo Mocker) use
  the <strong>same</strong> Rust core function via two binding paths
  (PyO3 / direct crate use).</p>
  <p>The model-building knowledge stays in Python (one source of truth), but
  the hot path runs in Rust at GIL-free, rayon-friendly speed.</p>
  <p>A throwaway PoC under <code>poc/</code> validates the architecture
  end-to-end: 14/14 parity tests pass bit-identical against a pure-Python
  reference; an external Rust binary loads the same artifact and runs
  multi-threaded.</p>
</div>

<!-- ======================================================================= -->
<h2 id="motivation">1. Motivation</h2>

<h3 id="current-state">1.1 Current state</h3>

<p>AIConfigurator predicts LLM serving performance for many <em>(parallel
config, batch_size, ctx_tokens, quant_mode, ...)</em> combinations under SLA
targets. Today the entire pipeline is Python:</p>

<div class="ascii">
Task spec  →  sweep loop  →  Predictor strategy  →  backend.run_static
                                                       ↓
                                                   for op in model.context_ops:
                                                       op.query(database, ...)
                                                       ↓
                                                   sum, scale, return
</div>

<p>Two Rust hooks already exist:</p>
<ul>
  <li><code>aic-core/rust/aiconfigurator-core/</code> — a ctypes-backed sidecar for
    Dynamo Mocker. Receives a model identity via JSON; internally
    <em>re-derives</em> the op list from name + parallel sizes.</li>
  <li><code>sdk/rust_engine_step.py</code> — Python wrapper for the above.</li>
</ul>

<h3 id="pain-points">1.2 Pain points</h3>

<table>
  <thead><tr><th>Pain point</th><th>Cost today</th></tr></thead>
  <tbody>
    <tr><td>Python GIL serializes sweep evaluation</td>
        <td>Sweep with ~2,250 points takes ~0.6s; mostly DB look-ups + DataFrame ops on a single core.</td></tr>
    <tr><td>Model knowledge duplicated in Python and Rust</td>
        <td><code>aiconfigurator-core/model.rs</code> (~350 lines) re-derives the op list. Every new Python model needs a parallel Rust update.</td></tr>
    <tr><td>External Rust callers (Mocker) must round-trip through Python for build</td>
        <td>JSON ↔ struct conversion every per-point call; coupled to Python interpreter lifecycle.</td></tr>
    <tr><td>DataFrame-heavy hot loop</td>
        <td>pandas/concat ops dominate wallclock at low op counts; type drift bugs (object vs int64) caught only at parity.</td></tr>
  </tbody>
</table>

<!-- ======================================================================= -->
<h2 id="proposal">2. Proposal: Python + Rust Split</h2>

<h3 id="layers">2.1 Three-layer architecture</h3>

<p>The new design separates <strong>what to compute</strong> (Python's
job — model topology, quant semantics, layer counts) from
<strong>how fast / how multi-threaded to compute it</strong> (Rust's job).
Between them sits a <strong>compiled artifact</strong> — a plain-data
representation of the op list — that both languages share.</p>

<div class="ascii">
                Python                                          Rust
   ─────────────────────────────                  ─────────────────────────────
   parse HF config (model arch)
   ↓
   instantiate Operation objects
   (one per layer-level op)
   ↓
   produce list[OpSpec]

   ─────── compile/build ─────────►   build Engine: Vec&lt;OpSpec&gt; + metadata
                                                    │
                                                    ↓
   for (batch, seq, mode) in sweep:       run_static_internal(engine, db, batch, seq, mode)
       call into Engine ────────►           for op in engine.op_list:
                                                lookup db, scale, accumulate
                                            return HashMap&lt;op_name, latency&gt;
   ◄────── result ──────────────────
</div>

<h3 id="engine">2.2 The Engine abstraction</h3>

<p>The new central object is the <strong>Engine</strong>: a Rust
struct holding a compiled op list plus metadata. Its sole runtime
method is <code>run_static_internal(db, batch, seq, mode) → HashMap&lt;name, latency&gt;</code>.</p>

<table>
  <thead><tr><th>Property</th><th>Engine</th><th>V1 InferenceSession (deleted)</th><th>V2 stateless Predictor (current)</th></tr></thead>
  <tbody>
    <tr><td>State held</td><td>Compiled op list + metadata</td><td>model + db + backend</td><td>None (everything passed per call)</td></tr>
    <tr><td>Hot path location</td><td>Rust</td><td>Python</td><td>Python (delegates to backend)</td></tr>
    <tr><td>Bound to one model?</td><td>Yes (the compiled one)</td><td>Yes</td><td>No</td></tr>
    <tr><td>Sweep methods?</td><td>No (sweep lives in Python sweep.py)</td><td>Yes (<code>find_best_*</code>)</td><td>No</td></tr>
    <tr><td>Callable from Rust?</td><td>Yes (Mocker, future tools)</td><td>No</td><td>No</td></tr>
  </tbody>
</table>

<p>The Engine is conceptually the <em>execution unit</em> that V1's
InferenceSession <em>tried</em> to be. V2 went stateless as an overcorrection
against V1's three-things-in-one design; the PoC shows that
<strong>strategy</strong> (stateless Predictor) and <strong>execution unit</strong>
(stateful Engine) are different layers and both belong.</p>

<h3 id="two-callers">2.3 Two callers, one Rust core</h3>

<div class="ascii">
                    ┌──────────────────────────────────┐
                    │  Engine::run_static_internal     │
                    │  (rust/aic-step/src/engine.rs)   │   ◄── single source of truth
                    │  pure Rust, zero Python          │
                    └──────────────▲───────────────────┘
                                   │
                  ┌────────────────┴───────────────┐
                  │                                │
        ┌─────────┴───────────┐         ┌──────────┴──────────┐
        │ PyO3 binding        │         │ Rust crate direct   │
        │ Engine.run_static   │         │ engine.run_static_  │
        │   (PyEngine wrap)   │         │   internal(...)     │
        └─────────┬───────────┘         └──────────┬──────────┘
                  │                                │
        Python: Task / sweep              External Rust callers
        (releases GIL on call)            (Mocker, future tools)
</div>

<p>Both bindings hit the same <code>run_static_internal</code>. The PyO3
wrapper releases the GIL during execution, so Python-side multithreading
(or batch entry points) parallelizes without the GIL bottleneck.</p>

<!-- ======================================================================= -->
<h2 id="poc">3. PoC Implementation</h2>

<h3 id="poc-layout">3.1 Layout</h3>

<p>Greenfield, self-contained under <code>poc/</code>. Zero changes to
<code>src/aiconfigurator/</code>.</p>

<pre><code>poc/
├── Cargo.toml                # standalone workspace
├── pyproject.toml            # maturin config
├── pytest.ini                # pytest scope isolation
├── README.md
├── src/
│   ├── lib.rs                # crate entry; pyo3 gated behind `python` feature
│   ├── op.rs                 # OpSpec enum (Gemm only in PoC)
│   ├── db.rs                 # Database + arrow-rs parquet loader
│   ├── engine.rs             # Engine + run_static_internal (hot path)
│   ├── py.rs                 # PyO3 bindings (gated by feature)
│   └── bin/mocker_demo.rs    # external Rust caller demo
├── python/
│   ├── aic_step/             # facade re-exporting Rust symbols
│   └── aic_poc/              # MockLLMModel + pure-Python reference impl
├── data/
│   ├── build_gemm_parquet.py
│   └── gemm_perf.parquet     # (generated)
└── tests/
    └── test_e2e.py           # 14 parity + round-trip tests
</code></pre>

<h3 id="poc-flow">3.2 End-to-end flow</h3>

<h4>Python path</h4>
<pre><code>from aic_poc import MockLLMModel
from aic_step import build_engine, DbHandle

model = MockLLMModel(hidden=4096, n_layers=32)  # Python: build op list

engine = build_engine(model)                     # PyO3: Python → Rust
db = DbHandle.load("data/gemm_perf.parquet")     # arrow-rs: load DB

# Hot path: PyO3 wrapper releases GIL, executes pure-Rust core,
# returns Python dict[str, float].
result = engine.run_static(db, batch_size=8, seq_len=2048, mode="static")
</code></pre>

<h4>External Rust caller (mocker_demo)</h4>
<pre><code>use aic_step::{Database, Engine, StaticMode};

let engine = Arc::new(Engine::load_bin("compiled.bin")?);
let db = Arc::new(Database::load_gemm_parquet("data/gemm_perf.parquet")?);

// Same hot path as Python; no PyO3 in the picture.
let result = engine.run_static_internal(&amp;db, 8, 2048, StaticMode::Full)?;

// Trivial rayon fan-out across many points:
points.par_iter().for_each(|p| {
    let _ = engine.run_static_internal(&amp;db, p.batch, p.seq, StaticMode::Full);
});
</code></pre>

<h3 id="poc-build">3.3 Build / run</h3>

<pre><code># one-time (Python 3.13 validation environment)
brew install rust
python3.13 -m venv .venv &amp;&amp; source .venv/bin/activate
pip install maturin pyarrow pytest

# build PyO3 extension into the venv
cd poc/
maturin develop --release

# generate test data and run parity tests
python data/build_gemm_parquet.py
pytest                                          # 14/14 pass

# build the pure-Rust binary (no libpython)
cargo build --release --bin mocker_demo --no-default-features
./target/release/mocker_demo compiled.bin data/gemm_perf.parquet
</code></pre>

<div class="callout warn">
  <span class="label">Design Note</span>
  <p><code>pyo3</code> is an <em>optional</em> dependency behind a
  <code>python</code> Cargo feature. With it, the lib outputs a Python
  extension (<code>aic_step._native.so</code>). Without it (the binary
  build), the lib is a plain Rust crate and the binary links zero
  Python symbols. This is the cleanest way to dual-purpose one crate
  for both callers.</p>
</div>

<!-- ======================================================================= -->
<h2 id="validation">4. Validation Results</h2>

<table>
  <thead><tr><th>#</th><th>Claim</th><th>Result</th></tr></thead>
  <tbody>
    <tr>
      <td>1</td>
      <td>PyO3 binding builds, Python imports cleanly</td>
      <td><span class="pill ok">PASS</span> <code>import aic_step</code> works after <code>maturin develop</code></td>
    </tr>
    <tr>
      <td>2</td>
      <td>Python model → Rust Engine round-trip works</td>
      <td><span class="pill ok">PASS</span> <code>build_engine(MockLLMModel(...))</code> produces a usable Engine with 3 ops</td>
    </tr>
    <tr>
      <td>3</td>
      <td>Rust execution bit-identical to Python reference</td>
      <td><span class="pill ok">PASS</span> 14/14 tests; rel_tol = abs_tol = 1e-12 across {static_ctx, static_gen, static} × {1, 8, 32 batch} × {1024, 2048, 4096 seq}</td>
    </tr>
    <tr>
      <td>4</td>
      <td>Engine save/load .bin round-trip</td>
      <td><span class="pill ok">PASS</span> bincode artifact reloads to identical Engine</td>
    </tr>
    <tr>
      <td>5</td>
      <td>External Rust caller runs the same Engine</td>
      <td><span class="pill ok">PASS</span> <code>mocker_demo</code> loads <code>compiled.bin</code>, runs 8 points, output matches Python</td>
    </tr>
    <tr>
      <td>6</td>
      <td>Multi-threading via rayon works</td>
      <td><span class="pill ok">PASS</span> <code>rayon::par_iter</code> across points; outputs identical to sequential. Wallclock not yet meaningful at N=8 (thread-spawn overhead dominates); stress N at next iteration.</td>
    </tr>
  </tbody>
</table>

<div class="callout good">
  <span class="label">Bit-Identical</span>
  <p>Every parameterization of the parity test asserts
  <code>math.isclose(rust_val, py_val, rel_tol=1e-12, abs_tol=1e-12)</code>
  per op. All 27 (op × test) checks passed on first build. The
  arithmetic is simple enough that Rust f64 and Python float
  match exactly.</p>
</div>

<!-- ======================================================================= -->
<h2 id="decisions">5. Open Decisions</h2>

<h3>5.1 Naming</h3>

<p>Picked <strong>Engine</strong> (over <code>CompiledModel</code> /
<code>InferenceSession</code> / <code>PerfModel</code>) because:</p>
<ul>
  <li>Aligns with existing <code>rust_engine_step.py</code> in the repo.</li>
  <li>TensorRT/vLLM call compiled artifacts "engine" — pattern match.</li>
  <li>Avoids resurrecting <code>InferenceSession</code>, which V2 just
    deleted, and which falsely implies "running inference"
    (we predict perf, we don't run the model).</li>
  <li>Short. Clean. <code>Engine.run_static(...)</code> reads well.</li>
</ul>

<p>If "engine" ever conflicts with real LLM-runtime "engine" in shared
docs, escalate to <strong>PerfEngine</strong>; rename cost is trivial.</p>

<h3>5.2 Predictor vs Engine</h3>

<p>The current Phase 1 SDK has a <code>Predictor</code> Protocol
(stateless strategy). The Engine is a different layer — stateful
execution unit. Both belong:</p>

<table>
  <thead><tr><th>Layer</th><th>Statefulness</th><th>Role</th><th>Examples</th></tr></thead>
  <tbody>
    <tr><td>Predictor</td><td>stateless</td><td>How to evaluate one point</td>
        <td><code>AnalyticPredictor</code>, <code>MockerPredictor</code> (future)</td></tr>
    <tr><td>Engine</td><td>stateful (compiled artifact)</td><td>Concrete execution unit</td>
        <td><code>Engine</code> (Rust), wrapped by <code>EnginePredictor</code></td></tr>
  </tbody>
</table>

<p>The integration path: an <code>EnginePredictor</code> implements
the Predictor Protocol; internally it caches Engines per
<code>(model_id, parallel_config)</code>. Caller-side API is unchanged.</p>

<h3>5.3 Artifact format</h3>
<p>PoC uses <code>bincode</code> for the Engine serialization (fast,
compact, Rust-native). Considered Arrow IPC and JSON; bincode wins on
size/speed. JSON debugging is achievable via a dump tool if needed.</p>

<h3>5.4 DB layout</h3>
<p>PoC's DB is per-op (GEMM only). Production DB has ~10 op kinds
(attention, MoE, comm, MLA, etc.). The PoC's table abstraction
(<code>Database { gemm: GemmTable, ... }</code>) maps each op kind to
its own Rust struct, matching today's "<em>op owns its data</em>" Python
layout. Extending: one new struct per op, ~50–200 lines each.</p>

<!-- ======================================================================= -->
<h2 id="roadmap">6. Roadmap (if we proceed)</h2>

<ol>
  <li>
    <strong>Decide</strong>: Adopt the Engine architecture? Or keep PoC
    on the shelf? Inputs to the decision: the validation results above,
    plus a stress bench (next item).
  </li>
  <li>
    <strong>Stress bench</strong>: extend the PoC to N=10k points
    sequential vs rayon. Measure real wallclock speedup. (Expected:
    near-linear up to core count.)
  </li>
  <li>
    <strong>Op coverage</strong>: incrementally port the remaining ~9 op
    kinds (attention, MoE, comm, MLA, custom_allreduce, ...). One PR
    per op; parity test per op.
  </li>
  <li>
    <strong>DB completeness</strong>: extend <code>DbHandle</code> to
    cover non-GEMM tables; interpolation/extrapolation logic moves to
    Rust (current Python <code>perf_database.py</code> arithmetic
    ports cleanly).
  </li>
  <li>
    <strong>Integration into Task.run()</strong>: introduce an
    <code>EnginePredictor</code> that the existing Phase 1
    <code>Predictor</code> Protocol can return; gate behind a Task
    field for opt-in (parity tests both paths).
  </li>
  <li>
    <strong>Production wheels</strong>: maturin CI for
    macOS / Linux / Windows × Python 3.11–3.13. <code>uv add aic_step</code>-friendly.
  </li>
</ol>

<p>Estimated effort (single dev with LLM assist): items 2–3 are
~3–4 weeks; item 4 is ~2 weeks; item 5 is ~1 week; item 6 is ~1 week.</p>

<!-- ======================================================================= -->
<h2 id="appendix">7. Appendix: Naming Notes</h2>

<p>For internal sanity, the four words floating in this design:</p>

<table>
  <thead><tr><th>Word</th><th>Meaning here</th><th>What it is NOT</th></tr></thead>
  <tbody>
    <tr><td><code>Predictor</code></td>
        <td>Strategy (stateless interface) — Phase 1 V2 abstraction</td>
        <td>The thing that holds the compiled model</td></tr>
    <tr><td><code>Engine</code></td>
        <td>Execution unit — holds compiled op list, runs <code>run_static</code></td>
        <td>An LLM runtime; the predictor strategy itself</td></tr>
    <tr><td><code>compile_model()</code> / <code>build_engine()</code></td>
        <td>Marshal Python <code>Operation</code> objects into a Rust
            <code>Engine</code> (PyO3 type conversion + struct construction)</td>
        <td>"Compile" in the LLVM/JIT sense — there's no codegen</td></tr>
    <tr><td><code>OpSpec</code></td>
        <td>Plain-data Rust representation of one op (enum variant + params)</td>
        <td>An <code>Operation</code> Python class — that lives in Python</td></tr>
  </tbody>
</table>

<footer>
  PoC source: <code>poc/</code> on branch <code>rust_poc</code>.
  All 14 parity tests pass locally; pushed to <code>origin/rust_poc</code>
  for review.
</footer>

</main>
</div>
</body>
</html>