File size: 26,131 Bytes
09c07bd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
# MLX

[**Quickstart**](#quickstart) | [**Installation**](#installation) |
[**Documentation**](https://ml-explore.github.io/mlx/build/html/index.html) |
[**Examples**](#examples)

[![CircleCI](https://circleci.com/gh/ml-explore/mlx.svg?style=svg)](https://circleci.com/gh/ml-explore/mlx)

MLX is an array framework for machine learning on Apple silicon,
brought to you by Apple machine learning research.

Some key features of MLX include:

- **Familiar APIs**: MLX has a Python API that closely follows NumPy. MLX
   also has fully featured C++, [C](https://github.com/ml-explore/mlx-c), and
   [Swift](https://github.com/ml-explore/mlx-swift/) APIs, which closely mirror
   the Python API. MLX has higher-level packages like `mlx.nn` and
   `mlx.optimizers` with APIs that closely follow PyTorch to simplify building
   more complex models.

- **Composable function transformations**: MLX supports composable function
  transformations for automatic differentiation, automatic vectorization,
  and computation graph optimization.

- **Lazy computation**: Computations in MLX are lazy. Arrays are only
  materialized when needed.

- **Dynamic graph construction**: Computation graphs in MLX are constructed
  dynamically. Changing the shapes of function arguments does not trigger
  slow compilations, and debugging is simple and intuitive.

- **Multi-device**: Operations can run on any of the supported devices
  (currently the CPU and the GPU).

- **Unified memory**: A notable difference from MLX and other frameworks
  is the *unified memory model*. Arrays in MLX live in shared memory.
  Operations on MLX arrays can be performed on any of the supported
  device types without transferring data.

MLX is designed by machine learning researchers for machine learning
researchers. The framework is intended to be user-friendly, but still efficient
to train and deploy models. The design of the framework itself is also
conceptually simple. We intend to make it easy for researchers to extend and
improve MLX with the goal of quickly exploring new ideas.

The design of MLX is inspired by frameworks like
[NumPy](https://numpy.org/doc/stable/index.html),
[PyTorch](https://pytorch.org/), [Jax](https://github.com/google/jax), and
[ArrayFire](https://arrayfire.org/).

## Examples

The [MLX examples repo](https://github.com/ml-explore/mlx-examples) has a
variety of examples, including:

- [Transformer language model](https://github.com/ml-explore/mlx-examples/tree/main/transformer_lm) training.
- Large-scale text generation with
  [LLaMA](https://github.com/ml-explore/mlx-examples/tree/main/llms/llama) and
  finetuning with [LoRA](https://github.com/ml-explore/mlx-examples/tree/main/lora).
- Generating images with [Stable Diffusion](https://github.com/ml-explore/mlx-examples/tree/main/stable_diffusion).
- Speech recognition with [OpenAI's Whisper](https://github.com/ml-explore/mlx-examples/tree/main/whisper).

## Quickstart

See the [quick start
guide](https://ml-explore.github.io/mlx/build/html/usage/quick_start.html)
in the documentation.

## Installation

MLX is available on [PyPI](https://pypi.org/project/mlx/). To install MLX on
macOS, run:

```bash
pip install mlx
```

To install the CUDA backend on Linux, run:

```bash
pip install mlx[cuda]
```

To install a CPU-only Linux package, run:

```bash
pip install mlx[cpu]
```

Checkout the
[documentation](https://ml-explore.github.io/mlx/build/html/install.html#)
for more information on building the C++ and Python APIs from source.

## ファインチューニング手法: LoRAとPRT

このリポジトリでは、Apple Silicon上で効率的にLLMをファインチューニングするための2つの手法を実装・提供しています。

### LoRA (Low-Rank Adaptation)

#### 概要

LoRAは、元のモデルの重みを変更せずに、低ランク行列を追加することでファインチューニングを行う手法です。メモリ効率が良く、複数のタスク用アダプターを切り替えて使用できます。

#### アーキテクチャ

```
元の計算: output = W × input
LoRA適用: output = (W + B × A) × input

W: 元の重み行列(固定、変更されない)
A, B: 低ランク行列(学習される)
rank: 低ランクの次元数(通常4-16)
```

**特徴**:

- **更新パラメータ**: 全パラメータの0.1-1%(低ランク行列のみ)
- **メモリ使用量**: 小さい(モデルサイズの1.2-1.5倍)
- **保存容量**: 非常に小さい(元のモデルの0.1-1%)
- **学習時間**: 短い
- **性能**: Full Fine-tuningに近い

#### 実装方法

**基本的な使い方**:

```bash
# 仮想環境を有効化
source .venv/bin/activate

# LoRAファインチューニング
python -m mlx_lm lora \
    --model ./models/tinyllama-1.1b \
    --train --data ./finetune_data \
    --num-layers 8 \
    --lora-r 8 \
    --learning-rate 1e-4 \
    --iters 30
```

**推奨パラメータ**:

- `--lora-r`: 8(低ランクの次元数)
- `--num-layers`: 8(適用する層数、TinyLlamaの場合は全22層中8層)
- `--learning-rate`: 1e-4
- `--iters`: 30

**学習済みアダプターでの推論**:

```bash
python -m mlx_lm chat \
    --model ./models/tinyllama-1.1b \
    --adapter-path ./adapters \
    --prompt "### Instruction: 日本の首都はどこですか?\n### Response:"
```

#### アーキテクチャの詳細

**TinyLlama-1.1Bでの実装例**:

- **モデル**: TinyLlama-1.1B-Chat-v1.0
- **LoRA適用層数**: 8層(全22層中)
- **LoRA rank**: 8
- **LoRA scale**: 20.0
- **更新パラメータ**: 約0.209%(2.294M/1100.048M)
- **アダプターサイズ**: 約8.8MB(元のモデルの0.44%)

**パフォーマンス**:

- トレーニング速度: 約5-6 it/sec
- トークン速度: 約250-300 tokens/sec
- ピークメモリ: 約2.4-2.5 GB

詳細は [`examples/prt_llm/README.md`](examples/prt_llm/README.md) と [`knowledge/SLM_選定とファインチューニング手法と性能変化.md`](knowledge/SLM_選定とファインチューニング手法と性能変化.md) を参照してください。

---

### PRT (Portable Reward Tuning)

#### 概要

PRTは、ファインチューニングをreward最大化問題として再定式化し、rewardモデルを明示的に学習することで、**異なるベースモデル間で再利用可能な**ファインチューニング手法です。論文「Portable Reward Tuning: Towards Reusable Fine-Tuning across Different Pretrained Models」(arXiv:2502.12776v1)に基づいています。

#### このリポジトリでの呼称(PRT1/2/3)

このREADMEでは、実装の違いを分かりやすくするために以下のように呼び分けます(**いずれも「学習対象はRewardモデル」**です)。

- **PRT1**: LinearのRewardヘッド(baseのhidden → vocabのreward logits): `examples/prt_llm/prt_llm.py`
- **PRT2**: Tiny Causal TransformerのRewardモデル(base hidden → 小型Transformer → vocab): `examples/prt_llm/prt_llm2.py`
- **PRT3**: Rewardモデルをベースモデルの**完全コピー(別インスタンス)**として学習: `examples/prt_llm/prt_llm_v3.py`

#### アーキテクチャ

**PRT1(Linear Reward Head)の基本構造**:

```
Base Model (固定)
├── Transformer Layers (22層)
└── LM Head (Linear)

Reward Model (学習可能)
└── Reward Head (Linear)
    ├── Input: Base Modelのhidden states (2048次元)
    └── Output: Reward logits (32000次元)

PRT合成:
v_θ = log softmax(base_logits) + reward_logits / λ
π_PRT = softmax(v_θ)
```

**PRT2 (拡張版) - Tiny Causal Transformer Reward Model**:

```
RewardTransformerRewardModel
├─ proj_in: Linear(2048 → 512)
├─ blocks: CausalTransformerBlock × L (L=2, default)
│   └─ block
│      ├─ norm1: RMSNorm(512)
│      ├─ attn: CausalSelfAttention(d=512, heads=8) + ALiBi
│      ├─ norm2: RMSNorm(512)
│      └─ mlp: SwiGLU FFN(512 → 2048 → 512)
├─ norm_out: RMSNorm(512)
└─ vocab_head: Linear(512 → 32000)
```

**特徴**:

- **Base Modelのパラメータ**: 完全に固定(変更されない)
- **Reward Model**: 独立したパラメータ(学習される)
- **Portable性**: 同じtokenizerを使用する異なるベースモデル間でrewardモデルを再利用可能
- **推論コスト(PRT1/2)**: Base Modelのforward passは1回のみ(効率的)
- **パラメータ数**: PRT1は約65M、PRT2は約26M(L=2の場合)

#### 実装方法

**PRT1(Linear Reward Head版)の使い方**:

```bash
# 仮想環境を有効化
source .venv/bin/activate

# Rewardモデルの学習
python3 examples/prt_llm/prt_llm.py train \
    --model ./models/tinyllama-1.1b \
    --data ./finetune_data \
    --output-dir ./prt_reward \
    --learning-rate 1e-4 \
    --iters 30 \
    --lambda 1.0 \
    --entropy-coeff 0.05

# PRTによる生成
python3 examples/prt_llm/prt_llm.py generate \
    --model ./models/tinyllama-1.1b \
    --reward ./prt_reward \
    --prompt "### Instruction: 日本の首都はどこですか?\n### Response:" \
    --max-tokens 100 \
    --temp 0.7
```

**PRT2(Tiny Transformer Reward版)の使い方**:

```bash
# Rewardモデルの学習
python3 examples/prt_llm/prt_llm2.py train \
    --model ./models/tinyllama-1.1b \
    --data ./finetune_data \
    --output-dir ./prt2_reward \
    --lambda 1.0 \
    --entropy-coeff 0.05 \
    --reward-d-model 512 \
    --reward-layers 2 \
    --reward-heads 8 \
    --reward-ffn-dim 2048

# PRT2による生成
python3 examples/prt_llm/prt_llm2.py generate \
    --model ./models/tinyllama-1.1b \
    --reward ./prt2_reward \
    --prompt "### Instruction: 日本の首都はどこですか?\n### Response:" \
    --max-tokens 100 \
    --temp 0.7
```

#### 損失関数

PRTの損失関数は以下の通りです:

```python
# Algorithm 1 (論文準拠)
v_θ = log softmax(base_logits) + reward_logits / λ
π_PRT = softmax(v_θ)
loss = CE(π_PRT, y) + (-entropy_coeff * H(ρ_θ))

# ρ_θ = softmax(reward_logits) (reward-only分布)
# H(ρ_θ): Entropy正則化(崩壊防止)
```

**パラメータ説明**:

- `--lambda`: KL正則化係数(デフォルト: 1.0)
- `--entropy-coeff`: Entropy正則化係数(デフォルト: 0.05)
- `--reward-d-model`: PRT2のTransformer次元(デフォルト: 512)
- `--reward-layers`: PRT2の層数(デフォルト: 2)

#### PRT vs LoRA の比較

| 観点 | LoRA | PRT1 | PRT2 | PRT3 |
|------|------|------|------|------|
| **更新パラメータ** | 0.1-1% | 約6% | 約2.4% | ≈100%(Reward=Full Copy) |
| **メモリ使用量** | 小さい | 中程度 | 中程度 | 大(モデル2本) |
| **学習時間** | 短い | 中程度 | やや長い | 長い |
| **保存容量** | 非常に小さい | 小さい | 小さい | 大 |
| **Portable性** | なし(モデル依存) | **あり**(tokenizer一致が前提) | **あり** | **あり** |
| **推論速度** | 高速 | 高速 | 中速 | 低〜中(forward 2回) |
| **表現力** | 中程度 | 低(Linear) | 高(Transformer) | 高(Full Model) |

詳細は [`examples/prt_llm/README.md`](examples/prt_llm/README.md) と [`knowledge/PRT_ネットワーク構造解説.md`](knowledge/PRT_ネットワーク構造解説.md) を参照してください。

---

### PRT Implementation History & Architectures

開発の進展に伴い、PRTはその適用範囲をLLMからVLMへ、そして実験環境からHPC/実アプリケーションへと拡大してきました。以下に各バージョンの実装箇所と特徴を示します。

**開発履歴一覧 (Version Summary):**

| Ver | 環境 | モデル | LoRA | タスク | 実装箇所 (Path) | 特記事項 |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **PRT1** | Mac (MLX) | TinyLlama-1.1B | × | LLM | `examples/prt_llm/prt_llm.py` | 軽量Reward Head |
| **PRT2** | Mac (MLX) | TinyLlama-1.1B | × | LLM | `examples/prt_llm/prt_llm2.py` | 高表現力Reward |
| **PRT3** | Mac (MLX) | TinyLlama-1.1B | ◯ | LLM | `examples/prt_llm/prt_llm_v3.py` | **LLM標準 (Twin)** |
| **PRT4** | Mac (MLX) | DeepSeek-VL2 | ? | VLM | `examples/prt_flickr30k` | Frozen (MoE負荷大) |
| **PRT5** | Mac (MLX) | Qwen2-VL-2B | ◯ | VLM | `examples/prt5_qwen2vl` | **VLM標準 (Efficient)** |
| **PRT6** | Linux (PT) | Qwen2-VL-2B | ◯ | VLM | `examples/prt6_pytorch` | PyTorch移植版 |
| **PRT7** | HPC (Sing) | Qwen2-VL-2B | ◯ | VLM | `examples/prt7_singularity` | コンテナ化 |
| **PRT8** | Mac (MLX) | Qwen2-VL-2B | ◯ | VLM | `examples/prt8_vlm` | 汎用FW化 |
| **PRT9** | Mac (MLX) | Qwen2-VL-2B | ◯ | DriveLM | `examples/prt9_drivelm` | 自動運転QA |
| **PRT10**| Linux (PT) | Eagle2-1B | ◯ | DriveLM | `examples/prt10_pytorch_drivelm` | (Conceptual/WIP) |
| **PRT11**| **Mac** | **Eagle2-1B** | **◯** | **DriveLM** | `examples/prt11_drivelm` | **Mac実用版 (Active)** |
| **PRT12**| **Any** | **Qwen2.5-VL** | **◯** | **General** | `examples/prt12_qwen25vl` | **Transformers最小構成 (Wrapper)** |
| **PRT13**| **Any** | **Qwen2.5-VL** | **◯** | **General** | `examples/prt13_qwen25vl` | **Transformers標準設計 (Design A)** |

以下に各フェーズの実装詳細を示します。

#### Phase 1: LLM Foundations (PRT1 - PRT3)

初期のPRTは、LLM(TinyLlama等)を対象に、Rewardモデルの構造を探索しました。

**PRT1: Linear Reward Head**

- **構造**: ベースモデルの隠れ層から直接Linear層でRewardを計算。
- **特徴**: 最も軽量だが、表現力が限定的。
- **実装**: `examples/prt_llm/prt_llm.py`

**PRT2: Transformer Reward Model**

- **構造**: ベースモデルの隠れ層の上に、小さなTransformerブロックを追加。
- **特徴**: 表現力と計算コストのバランスが良い。
- **実装**: `examples/prt_llm/prt_llm2.py`

**PRT3: Twin Model Composition (Current Standard)**

- **構造**: Rewardモデルとして、ベースモデルと同じアーキテクチャのコピーを使用(別インスタンス)。
- **特徴**: 最も表現力が高く、LoRA等の既存技術をそのまま適用可能。現在のPRTの標準形。
- **実装**: `examples/prt_llm/prt_llm_v3.py`

```mermaid
graph TD
    subgraph PRT1 [PRT1: Linear]
        B1[Base Model] --> H1[Hidden]
        H1 --> L1[Linear Head]
        L1 --> R1[Reward Logits]
    end
    
    subgraph PRT2 [PRT2: Tiny Transformer]
        B2[Base Model] --> H2[Hidden]
        H2 --> T2[Tiny Transformer]
        T2 --> L2[Linear Head]
        L2 --> R2[Reward Logits]
    end
    
    subgraph PRT3 [PRT3: Twin Model]
        B3[Base Model] --> Log3[Base Logits]
        RM3[Reward Model (Copy)] --> R3[Reward Logits]
        
        Log3 -- "+" --> Sum
        R3 -- "/ lambda" --> Sum
    end
```

#### Phase 2: VLM Expansion on MLX (PRT4 - PRT5)

視覚情報を含むVLM(Vision-Language Model)への適用フェーズです。

**PRT4: MoE VLM Experiment (Archived)**

- **対象**: DeepSeek-VL2 (Mixture-of-Experts)
- **結果**: Mac M1でのMoE計算負荷とMLXの制約により開発凍結。
- **ディレクトリ**: `examples/prt_flickr30k`

**PRT5: Efficient VLM Tuning (Qwen2-VL)**

- **対象**: Qwen2-VL-2B-Instruct
- **構造**: Vision Encoderからの特徴量を、BaseとRewardの両方のLLMに入力します。メモリ効率のためBaseは4bit量子化されています。
- **実装**: `examples/prt5_qwen2vl`

```mermaid
graph LR
    Img[Image] --> V[Vision Encoder]
    Txt[Text Prompt] --> Emb[Embedding]
    
    V --> Concat
    Emb --> Concat
    
    Concat --> Base[Base LLM (4bit Frozen)]
    Concat --> Reward[Reward LLM (LoRA Trainable)]
    
    Base --> BL[Base Logits]
    Reward --> RL[Reward Logits]
    
    BL & RL --> PRT_Loss
```

#### Phase 3: Scaling to HPC (PRT6 - PRT7)

大学や研究所のHPC(スパコン)環境での学習を見据えたスケーリングフェーズです。

**PRT6: PyTorch & Accelerate Port**

- **内容**: MLXの実装をPyTorch + Hugging Face Accelerateに移植。
- **目的**: マルチGPU分散学習の実現。
- **実装**: `examples/prt6_pytorch`

**PRT7: Singularity Containerization**

- **内容**: PRT6をSingularity (Apptainer) コンテナにパッケージング。
- **目的**: 依存関係(CUDA, Libraries)の完全な再現性と、管理者権限のない共用スパコンでの実行。
- **実装**: `examples/prt7_singularity`

```mermaid
graph LR
    subgraph HPC_Node [HPC Compute Node]
        subgraph Singularity [Singularity Container (PRT7)]
            Script[run_singularity.sh]
            Env[Python venv / Dependencies]
            Code[PRT6 PyTorch Code]
            
            Script --> Env
            Env --> Code
        end
        GPU[NVIDIA A100/H100 GPUs]
        Code -.-> GPU
    end
```

#### Phase 4: Generalization & Application (PRT8 - PRT9)

実用化に向けたリファクタリングと、複雑なタスクへの応用です。

**PRT8: VLM Generalization Framework**

- **内容**: PRT5-7の知見を統合し、汎用的なVLM学習ライブラリとしてリファクタリング。
- **構造**: `PRTForCausalLM` クラスによるモデルの抽象化と、共通化されたデータハンドラ。
- **実装**: `examples/prt8_vlm`

**PRT9: Autonomous Driving (DriveLM)**

- **内容**: 自動運転QAデータセット「DriveLM」への適用。
- **特徴**: Preprocessによる特殊なJSON形式の変換と、高解像度画像の取り扱い。
- **実装**: `examples/prt9_drivelm`

```mermaid
graph TD
    Raw[DriveLM Raw Data] --> Prep[preprocess_drivelm.py]
    Prep --> JSONL[Formatted QA JSONL]
    
    JSONL --> Load[DataLoader]
    Img[High-Res Images] --> Load
    
    subgraph PRT9_Training
        Load --> PRT8[PRT8 Codebase]
        Mod[Qwen2-VL-2B] -.-> PRT8
    end
    
    PRT8 --> ModelFile[Saved Model]
```

#### Phase 5: Real-World Adaptation & Challenges (PRT10 - PRT11)

実世界での運用(特にMacローカル環境やGPUリソースの制約)に向けた適応フェーズです。実装にあたり以下の「壁」がありました。

**実装への壁 (Barriers):**

1. **PyTorch環境**: GPU環境の準備と依存関係の複雑さ。
2. **メモリ制約**: フルファインチューニングは不可能。LoRAであってもVLM(Vision Model + LLM)はメモリ的に厳しい。
3. **タスクの変化**: LLMからVLM(DriveLM)への移行。論文実装にあるViT(Vision Transformer)の実装が存在しないため、独自に構築する必要があった。

これらの課題を乗り越えるため、以下の2つのアプローチを実装しました。

| バージョン | 環境 | モデル | LoRA | タスク | ステータス |
| --- | --- | --- | --- | --- | --- |
| **論文実装** | PyTorch | LLM/(ViT) | QLoRA | Llama | (Reference) |
| **PRT10** | PyTorch (GPU) | VLM | LoRA | DriveLM | **実装済み** |
| **PRT11** | **Mac (Metal)** | **VLM** | **LoRA** | **DriveLM** | **稼働中 (Code Exists)** |
| **(Future)** | ? | VLM | Full FT | DriveLM | 未生成 |

**PRT10: PyTorch VLM (LoRA)**

- **環境**: Linux / PyTorch / CUDA
- **特徴**: PRT9のコンセプトをPyTorchネイティブ環境に移植。標準的なGPUサーバーでの学習を想定。
- **実装**: `examples/prt10_pytorch_drivelm` (Conceptual/WIP)

**PRT11: Mac Metal VLM (LoRA)**

- **環境**: Mac M1/M2/M3 (Metal/MPS)
- **特徴**:
  - `nvidia/Eagle2-1B` モデルを採用し、MacBook Pro等のローカル環境で学習可能に最適化。
  - **Dynamic Patching**: `transformers` ライブラリやモデルのリモートコードに対し、実行時に動的にパッチを適用することで、FlashAttention非対応環境(Mac)での動作を強制。
  - **Memory Efficient**: 1BクラスのモデルとLoRAを組み合わせることで、ユニファイドメモリ環境での動作を実現。
- **実装**: `examples/prt11_drivelm/train_prt11.py`

**PRT12: Transformers Minimal Implementation**

- **環境**: Any (PyTorch + Transformers)
- **特徴**:
  - `mlx` や複雑な依存関係を排除し、`transformers``peft` のみで構成された最小実装。
  - Qwen2.5-VL を対象に、Algorithm 1 (学習) と Algorithm 2 (推論) のロジックをブラックボックスなしで実装。
  - **教育的価値**: PRTの動作原理(Logits合成)を理解するための参照実装として最適です。
- **実装**: `examples/prt12_qwen25vl`

**PRT13: Transformers Standard Implementation (Design A)**

- **環境**: Any (PyTorch + Transformers)
- **特徴**:
  - PRTの「王道」とされる設計パターン(Design A)。
  - **構成**:
    - **Trainer継承**: `PRTTrainer` で学習ロジックをカプセル化し、標準エコシステム(SFT Trainer等)との親和性を確保。
    - **Generator分離**: `PRTGenerator` で推論ロジックを独立させ、**「モデルスワッピング実験(Baseモデルを入れ替えてAdapterのポータビリティを検証)」** を容易に。
  - **成果物**: 純粋なLoRAアダプタのみ(モデルコード不要)。
- **実装**: `examples/prt13_qwen25vl`

---

### PRT Architecture Details (Mechanism of Portability)

PRTの最大の利点である「ポータビリティ(学習時と推論時で異なるモデルサイズを扱える)」を実現するアーキテクチャの詳細です。

#### 1. Core Concept: Logits Composition

PRTはモデル内部の隠れ層ではなく、最終出力である **「語彙空間(Logits)」** のレベルで結合を行います。これにより、モデルサイズ(7B vs 2B)が異なっていても、使用している「辞書(Tokenizer)」さえ同じであれば計算が可能になります。

$$ v_\theta(y|x) = \log P_{\text{Base}}(y|x) + \frac{R_\phi(y|x)}{\lambda} $$

#### 2. Architecture: Training vs Inference

**学習フェーズ (Training Phase)**:
計算コストを抑えるため、**「小さなモデル同士」**(2B + 2B)で学習を行います。

- **Base Model (Reference)**: Qwen2-VL-**2B** (Frozen)
- **Reward Model**: Qwen2-VL-**2B** (Frozen) + **LoRA Adapter** (Trainable)

```mermaid
graph LR
    Dataset --> Input
    Input --> Base["Reference (2B)"]
    Input --> Reward["Reward (2B)"]
    
    subgraph RewardModel [Reward Model Composition]
        Reward --> LoRA["LoRA Adapter (Trainable)"]
        LoRA -.-> RewardOut[Logits]
    end
    
    Base -.-> Loss
    RewardOut --> Loss
    Loss --> Update["Update LoRA Params"]
    
    style LoRA fill:#f96,stroke:#333,stroke-width:2px
    style Base fill:#eee,stroke:#333,stroke-dasharray: 5 5
```

**推論フェーズ (Inference Phase)**:
獲得したRewardモデル(2B + LoRA)を、**「大きなモデル(7B)」** の推論プロセスに組み込みます。
重要な点として、**Rewardモデルを7Bモデルに移植するわけではありません**。Rewardモデルは2Bのまま独立して並走し、その出力を7Bモデルの出力に「加算」することで生成をガイドします。

- **Base Model**: Qwen2-VL-**7B** (Frozen)  <-- **ここだけ入れ替える**
- **Reward Model**: Qwen2-VL-**2B** (Frozen) + **LoRA Adapter** (Frozen) <-- **ずっと2Bのまま並走**
- **操作**: **Logits Level Addition (出力の単純合成)**

```mermaid
graph LR
    Prompt --> Input
    Input --> Base["Base (swapped to 7B)"]
    Input --> Reward["Reward (stays 2B)"]
    
    subgraph RewardComponents [Reward + LoRA]
        Reward --> LoRA_Inf["LoRA Adapter (Frozen)"]
        LoRA_Inf --> RewardLogits[Reward Logits]
    end

    Base --> BaseLogits[Base Logits]
    
    BaseLogits -- "Log Probability" --> Add
    RewardLogits -- "Reward (R/λ)" --> Add
    Add["Compose (Logits Addition) v = log P + R/λ"] --> Output
    
    style LoRA_Inf fill:#f96,stroke:#333,stroke-width:2px
```

このアーキテクチャにより、**「学習コストは2B相当、推論能力は7B相当」** という効率的な運用が可能になります。

#### 3. Full Fine-Tuning (Original PRT) vs LoRA

現在の実装 (`train_prt5.py`) はメモリ効率のため **LoRA** を使用していますが、論文通りの **Full Fine-Tuning**(全パラメータ学習)を行うことも可能です。
その場合はコードの以下の部分を変更します(ロジックやLoss関数は変更不要です)。

```python
# train_prt5.py の変更イメージ

# Current (LoRA)
reward_model.freeze()          # 1. 全体を凍結
apply_lora(reward_model)       # 2. LoRA層のみ追加・学習

# Target (Full FT)
# reward_model.freeze()        # 1. 凍結しない(または Unfreeze)
# apply_lora(reward_model)     # 2. LoRAは適用しない(スキップ)
```

これにより、Optimizerはモデル全体のパラメータを学習対象として更新するようになります。

---

## Contributing

Check out the [contribution guidelines](https://github.com/ml-explore/mlx/tree/main/CONTRIBUTING.md) for more information
on contributing to MLX. See the
[docs](https://ml-explore.github.io/mlx/build/html/install.html) for more
information on building from source, and running tests.

We are grateful for all of [our
contributors](https://github.com/ml-explore/mlx/tree/main/ACKNOWLEDGMENTS.md#Individual-Contributors). If you contribute
to MLX and wish to be acknowledged, please add your name to the list in your
pull request.

## Citing MLX

The MLX software suite was initially developed with equal contribution by Awni
Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. If you find
MLX useful in your research and wish to cite it, please use the following
BibTex entry:

```text
@software{mlx2023,
  author = {Awni Hannun and Jagrit Digani and Angelos Katharopoulos and Ronan Collobert},
  title = {{MLX}: Efficient and flexible machine learning on Apple silicon},
  url = {https://github.com/ml-explore},
  version = {0.0},
  year = {2023},
}
```