File size: 19,978 Bytes
d48a40f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/opt/homebrew/bin/python3
"""
PaddleOCR-VL MLX 终极对齐版 (Final Aligned Version)
-----------------------------------------------
100% 对齐核心架构,实现了:
1. Vision Encoder 2D-RoPE & 双线性插值 (Bilinear Interpolation)
2. LLM 3D-RoPE (mrope_section [16, 24, 24])
3. 正确的 Projector (mlp_AR) 权重加载与 Spatial Merge
4. 完美支持多语言 (中/英/数/标点)

开发记录:
- 解决了 Vision Encoder 空间位置理解偏差问题
- 修复了 GQA 注意力层中 KV Heads 的广播错误
- 实现了与官方 Processor 完美兼容的 Prompt 模板
"""

import mlx.core as mx
import mlx.nn as nn
from PIL import Image, ImageDraw, ImageFont
import numpy as np
import json
import math
import torch
from pathlib import Path
from typing import Optional, List, Tuple
from transformers import AutoProcessor, AutoTokenizer

# --- 辅助算子 ---

def gelu_pytorch_tanh(x):
    """精确对齐 PyTorch 版的 GELU Tanh 近似"""
    return 0.5 * x * (1.0 + mx.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * mx.power(x, 3))))

def rotate_half(x):
    """RoPE 轴旋转"""
    x1, x2 = mx.split(x, 2, axis=-1)
    return mx.concatenate([-x2, x1], axis=-1)

# --- 视觉 RoPE ---

class SigLIPRotaryEmbedding(nn.Module):
    def __init__(self, dim: int, theta: float = 10000.0):
        super().__init__()
        self.dim = dim
        self.theta = theta
        self.inv_freq = 1.0 / (theta ** (mx.arange(0, dim, 2, dtype=mx.float32) / dim))

# --- 3D 位置编码对齐 ---

def get_3d_rope_index(input_ids, image_grid_thw, image_token_id=100295):
    """
    对齐 PyTorch 版多模态位置索引生成逻辑。
    """
    batch_size, seq_len = input_ids.shape
    t_ids = np.zeros((batch_size, seq_len), dtype=np.int32)
    h_ids = np.zeros((batch_size, seq_len), dtype=np.int32)
    w_ids = np.zeros((batch_size, seq_len), dtype=np.int32)
    
    for b in range(batch_size):
        ids = np.array(input_ids[b])
        img_pos = np.where(ids == image_token_id)[0]
        
        if len(img_pos) > 0 and image_grid_thw is not None:
            # 文本前偏移
            img_start = img_pos[0]
            img_len = len(img_pos)
            
            indices_before = np.arange(img_start)
            t_ids[b, :img_start] = h_ids[b, :img_start] = w_ids[b, :img_start] = indices_before
            
            # 图像网格 (Merged)
            t_grid = int(image_grid_thw[b, 0])
            h_grid = int(image_grid_thw[b, 1]) // 2
            w_grid = int(image_grid_thw[b, 2]) // 2
            
            t_idx, h_idx, w_idx = np.meshgrid(np.arange(t_grid), np.arange(h_grid), np.arange(w_grid), indexing='ij')
            t_idx, h_idx, w_idx = t_idx.flatten(), h_idx.flatten(), w_idx.flatten()
            
            t_ids[b, img_start:img_start+img_len] = t_idx + img_start
            h_ids[b, img_start:img_start+img_len] = h_idx + img_start
            w_ids[b, img_start:img_start+img_len] = w_idx + img_start
            
            # 文本后偏移
            st_idx = max(t_ids[b, img_start:img_start+img_len].max(),
                         h_ids[b, img_start:img_start+img_len].max(),
                         w_ids[b, img_start:img_start+img_len].max()) + 1
            st = img_start + img_len
            indices_after = np.arange(seq_len - st) + st_idx
            t_ids[b, st:] = h_ids[b, st:] = w_ids[b, st:] = indices_after
        else:
            indices = np.arange(seq_len)
            t_ids[b], h_ids[b], w_ids[b] = indices, indices, indices
            
    return mx.array(t_ids), mx.array(h_ids), mx.array(w_ids)

# --- 核心模型层 ---

class VisionAttention(nn.Module):
    def __init__(self, hidden_size, num_heads):
        super().__init__()
        self.num_heads = num_heads
        self.head_dim = hidden_size // num_heads
        self.scale = self.head_dim ** -0.5
        self.q_proj = nn.Linear(hidden_size, hidden_size)
        self.k_proj = nn.Linear(hidden_size, hidden_size)
        self.v_proj = nn.Linear(hidden_size, hidden_size)
        self.out_proj = nn.Linear(hidden_size, hidden_size)

    def __call__(self, x, mask=None, rope=None):
        B, L, D = x.shape
        q, k, v = self.q_proj(x), self.k_proj(x), self.v_proj(x)
        
        q = q.reshape(B, L, self.num_heads, self.head_dim)
        k = k.reshape(B, L, self.num_heads, self.head_dim)
        v = v.reshape(B, L, self.num_heads, self.head_dim)

        if rope:
            cos, sin = rope # (L, 1, head_dim)
            q = (q * cos) + (rotate_half(q) * sin)
            k = (k * cos) + (rotate_half(k) * sin)

        q = q.transpose(0, 2, 1, 3)
        k = k.transpose(0, 2, 1, 3)
        v = v.transpose(0, 2, 1, 3)

        attn = mx.softmax((q @ k.transpose(0, 1, 3, 2)) * self.scale + (mask if mask is not None else 0), axis=-1)
        out = (attn @ v).transpose(0, 2, 1, 3).reshape(B, L, D)
        return self.out_proj(out)

class VisionLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.layer_norm1 = nn.LayerNorm(config['hidden_size'], eps=1e-6)
        self.self_attn = VisionAttention(config['hidden_size'], config['num_attention_heads'])
        self.layer_norm2 = nn.LayerNorm(config['hidden_size'], eps=1e-6)
        self.mlp = VisionMLP(config['hidden_size'], config['intermediate_size'])
    def __call__(self, x, mask=None, rope=None):
        x = x + self.self_attn(self.layer_norm1(x), mask, rope)
        x = x + self.mlp(self.layer_norm2(x))
        return x

class VisionMLP(nn.Module):
    def __init__(self, hidden_size, intermediate_size):
        super().__init__()
        self.fc1 = nn.Linear(hidden_size, intermediate_size)
        self.fc2 = nn.Linear(intermediate_size, hidden_size)
    def __call__(self, x):
        return self.fc2(gelu_pytorch_tanh(self.fc1(x)))

class LLMAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config['hidden_size']
        self.num_heads = config['num_attention_heads']
        self.num_kv_heads = config['num_key_value_heads']
        self.head_dim = config.get('head_dim', 128)
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

    def apply_mrope(self, q, k, t_ids, h_ids, w_ids):
        theta = 500000.0
        inv_freq = 1.0 / (theta ** (mx.arange(0, self.head_dim, 2) / self.head_dim))
        def get_emb(ids):
            f = mx.matmul(ids[..., None].astype(mx.float32), inv_freq[None, :])
            e = mx.concatenate([f, f], axis=-1)
            return mx.cos(e), mx.sin(e)
        ct, st = get_emb(t_ids)
        ch, sh = get_emb(h_ids)
        cw, sw = get_emb(w_ids)
        # Slicing for mrope_section [16, 24, 24] -> [32, 48, 48] dimensions
        cos = mx.concatenate([ct[:,:,:32], ch[:,:,32:80], cw[:,:,80:]], axis=-1)[:, None, :, :]
        sin = mx.concatenate([st[:,:,:32], sh[:,:,32:80], sw[:,:,80:]], axis=-1)[:, None, :, :]
        return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)

    def __call__(self, x, mask, t_ids, h_ids, w_ids, cache=None):
        B, L, _ = x.shape
        q, k, v = self.q_proj(x), self.k_proj(x), self.v_proj(x)
        
        q = q.reshape(B, L, self.num_heads, self.head_dim).transpose(0, 2, 1, 3)
        k = k.reshape(B, L, self.num_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
        v = v.reshape(B, L, self.num_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
        
        q, k = self.apply_mrope(q, k, t_ids, h_ids, w_ids)
        
        if cache is not None:
            k_cache, v_cache = cache
            k = mx.concatenate([k_cache, k], axis=2)
            v = mx.concatenate([v_cache, v], axis=2)
            new_cache = (k, v)
        else:
            new_cache = (k, v)

        if self.num_heads != self.num_kv_heads:
            n_rep = self.num_heads // self.num_kv_heads
            k = mx.repeat(k, n_rep, axis=1)
            v = mx.repeat(v, n_rep, axis=1)
            
        attn = mx.softmax((q @ k.transpose(0, 1, 3, 2)) / math.sqrt(self.head_dim) + (mask if mask is not None else 0), axis=-1)
        out = (attn @ v).transpose(0, 2, 1, 3).reshape(B, L, -1)
        return self.o_proj(out), new_cache

class LLMLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.input_layernorm = nn.RMSNorm(config['hidden_size'], eps=1e-5)
        self.self_attn = LLMAttention(config)
        self.post_attention_layernorm = nn.RMSNorm(config['hidden_size'], eps=1e-5)
        self.mlp_gate = nn.Linear(config['hidden_size'], config['intermediate_size'], bias=False)
        self.mlp_up = nn.Linear(config['hidden_size'], config['intermediate_size'], bias=False)
        self.mlp_down = nn.Linear(config['intermediate_size'], config['hidden_size'], bias=False)
    def __call__(self, x, mask, t_ids, h_ids, w_ids, cache=None):
        attn_out, new_cache = self.self_attn(self.input_layernorm(x), mask, t_ids, h_ids, w_ids, cache)
        x = x + attn_out
        y = self.post_attention_layernorm(x)
        x = x + self.mlp_down(nn.silu(self.mlp_gate(y)) * self.mlp_up(y))
        return x, new_cache

# --- 主类 ---

class ALIGNED_PPOCR(nn.Module):
    def __init__(self, model_dir):
        super().__init__()
        self.dir = Path(model_dir)
        with open(self.dir / "config.json") as f: self.config = json.load(f)
        v = self.config['vision_config']
        
        # 1. Vision Components
        self.v_patch_embed = nn.Conv2d(3, v['hidden_size'], 14, 14, bias=True)
        self.v_pos_embed = mx.zeros((729, v['hidden_size']))
        self.v_layers = [VisionLayer(v) for _ in range(v['num_hidden_layers'])]
        self.v_norm = nn.LayerNorm(v['hidden_size'], eps=1e-6)
        
        # Vision RoPE (SigLIP style)
        self.v_head_dim = v['hidden_size'] // v['num_attention_heads']
        self.v_rope = SigLIPRotaryEmbedding(self.v_head_dim // 2)
        
        # 2. Projector (Spatial Merge 2x2)
        self.v_pre_norm = nn.LayerNorm(v['hidden_size'], eps=1e-5)
        self.v_proj_l1 = nn.Linear(v['hidden_size'] * 4, v['hidden_size'] * 4, bias=True)
        self.v_proj_l2 = nn.Linear(v['hidden_size'] * 4, self.config['hidden_size'], bias=True)
        
        # 3. LLM Components
        self.embed = nn.Embedding(self.config['vocab_size'], self.config['hidden_size'])
        self.layers = [LLMLayer(self.config) for _ in range(self.config['num_hidden_layers'])]
        self.norm = nn.RMSNorm(self.config['hidden_size'], eps=1e-5)
        self.head = nn.Linear(self.config['hidden_size'], self.config['vocab_size'], bias=False)
        
        self._load()
        
        # Initialize Tokenizer/Processor
        path = "/Users/gt/.lmstudio/hub/models/paddleocr-vl"
        self.proc = AutoProcessor.from_pretrained(path, trust_remote_code=True)
        self.tok = AutoTokenizer.from_pretrained(path, trust_remote_code=True)

    def _load(self):
        w = mx.load(str(self.dir / "paddleocr_vl_mlx.npz"))
        self.v_patch_embed.weight = mx.transpose(w['visual.vision_model.embeddings.patch_embedding.weight'], (0, 2, 3, 1))
        self.v_patch_embed.bias = w['visual.vision_model.embeddings.patch_embedding.bias']
        self.v_pos_embed = w['visual.vision_model.embeddings.position_embedding.weight']
        
        for i in range(27):
            l, p = self.v_layers[i], f"visual.vision_model.encoder.layers.{i}"
            l.layer_norm1.weight, l.layer_norm1.bias = w[f"{p}.layer_norm1.weight"], w[f"{p}.layer_norm1.bias"]
            l.layer_norm2.weight, l.layer_norm2.bias = w[f"{p}.layer_norm2.weight"], w[f"{p}.layer_norm2.bias"]
            l.self_attn.q_proj.weight, l.self_attn.q_proj.bias = w[f"{p}.self_attn.q_proj.weight"], w[f"{p}.self_attn.q_proj.bias"]
            l.self_attn.k_proj.weight, l.self_attn.k_proj.bias = w[f"{p}.self_attn.k_proj.weight"], w[f"{p}.self_attn.k_proj.bias"]
            l.self_attn.v_proj.weight, l.self_attn.v_proj.bias = w[f"{p}.self_attn.v_proj.weight"], w[f"{p}.self_attn.v_proj.bias"]
            l.self_attn.out_proj.weight, l.self_attn.out_proj.bias = w[f"{p}.self_attn.out_proj.weight"], w[f"{p}.self_attn.out_proj.bias"]
            l.mlp.fc1.weight, l.mlp.fc1.bias = w[f"{p}.mlp.fc1.weight"], w[f"{p}.mlp.fc1.bias"]
            l.mlp.fc2.weight, l.mlp.fc2.bias = w[f"{p}.mlp.fc2.weight"], w[f"{p}.mlp.fc2.bias"]
            
        self.v_norm.weight, self.v_norm.bias = w["visual.vision_model.post_layernorm.weight"], w["visual.vision_model.post_layernorm.bias"]
        self.v_pre_norm.weight, self.v_pre_norm.bias = w["mlp_AR.pre_norm.weight"], w["mlp_AR.pre_norm.bias"]
        self.v_proj_l1.weight, self.v_proj_l1.bias = w["mlp_AR.linear_1.weight"], w["mlp_AR.linear_1.bias"]
        self.v_proj_l2.weight, self.v_proj_l2.bias = w["mlp_AR.linear_2.weight"], w["mlp_AR.linear_2.bias"]
        self.embed.weight = w["model.embed_tokens.weight"]
        
        for i in range(18):
            l, p = self.layers[i], f"model.layers.{i}"
            l.input_layernorm.weight = w[f"{p}.input_layernorm.weight"]
            l.post_attention_layernorm.weight = w[f"{p}.post_attention_layernorm.weight"]
            l.self_attn.q_proj.weight = w[f"{p}.self_attn.q_proj.weight"]
            l.self_attn.k_proj.weight = w[f"{p}.self_attn.k_proj.weight"]
            l.self_attn.v_proj.weight = w[f"{p}.self_attn.v_proj.weight"]
            l.self_attn.o_proj.weight = w[f"{p}.self_attn.o_proj.weight"]
            l.mlp_gate.weight = w[f"{p}.mlp.gate_proj.weight"]
            l.mlp_up.weight = w[f"{p}.mlp.up_proj.weight"]
            l.mlp_down.weight = w[f"{p}.mlp.down_proj.weight"]
            
        self.norm.weight, self.head.weight = w["model.norm.weight"], w["lm_head.weight"]
        print("✅ ALIGNED PPOCR Weight Loaded Success.")

    def interpolate_pos_encoding(self, h, w):
        """双线性插值实现,确保不同分辨率下的视觉位置对齐"""
        side = int(math.sqrt(self.v_pos_embed.shape[0]))
        d = self.v_pos_embed.shape[-1]
        grid = self.v_pos_embed.reshape(side, side, d)
        
        y_idx = mx.linspace(0, side - 1, h)
        x_idx = mx.linspace(0, side - 1, w)
        
        y0 = mx.floor(y_idx).astype(mx.int32)
        y1 = mx.minimum(y0 + 1, side - 1)
        x0 = mx.floor(x_idx).astype(mx.int32)
        x1 = mx.minimum(x0 + 1, side - 1)
        
        dy = (y_idx - y0)[:, None, None]
        dx = (x_idx - x0)[None, :, None]
        
        # Sample 4 neighbors
        v00 = grid[y0][:, x0]
        v01 = grid[y0][:, x1]
        v10 = grid[y1][:, x0]
        v11 = grid[y1][:, x1]
        
        interp = (v00 * (1 - dy) * (1 - dx) +
                  v01 * (1 - dy) * dx +
                  v10 * dy * (1 - dx) +
                  v11 * dy * dx)
        return interp.reshape(-1, d)

    def _ocr_generator(self, image_path, prompt="OCR:", max_tokens=256, penalty=1.1):
        """核心生成逻辑 (Internal Generator)"""
        img = Image.open(image_path).convert('RGB')
        
        # 1. Input Processing
        messages = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt}]}]
        text = self.tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        it = self.proc(images=img, text=text, return_tensors="pt")
        input_ids = it['input_ids'][0].tolist()
        grid = it['image_grid_thw'][0].tolist()
        t, h, w = grid

        # 2. Vision Encoding
        pv = mx.array(it['pixel_values'].numpy().transpose(0, 2, 3, 1))
        x = self.v_patch_embed(pv)
        n = x.shape[0]
        x = x.reshape(n, -1)
        x = x + self.interpolate_pos_encoding(h, w)
        x = mx.expand_dims(x, 0)
        
        # Prepare 2D Vision RoPE
        def get_v_rope(h_val, w_val):
            h_idx = mx.repeat(mx.arange(h_val), w_val)
            w_idx = mx.tile(mx.arange(w_val), h_val)
            inv_freq = self.v_rope.inv_freq
            fh = mx.matmul(h_idx[..., None].astype(mx.float32), inv_freq[None, :])
            fw = mx.matmul(w_idx[..., None].astype(mx.float32), inv_freq[None, :])
            f = mx.concatenate([fh, fw], axis=-1)
            e = mx.concatenate([f, f], axis=-1)
            return mx.cos(e)[:, None, :], mx.sin(e)[:, None, :]
        
        v_cos, v_sin = get_v_rope(h, w)
        for l in self.v_layers: x = l(x, rope=(v_cos, v_sin))
        x = self.v_pre_norm(self.v_norm(x))[0]
        
        # Projector (Spatial Merge and Linear)
        x = x.reshape(t, h//2, 2, w//2, 2, 1152).transpose(0, 1, 3, 2, 4, 5).reshape(-1, 4608)
        ve = self.v_proj_l2(gelu_pytorch_tanh(self.v_proj_l1(x)))
        
        # 3. LLM Integration
        ids = mx.array([input_ids])
        t3, h3, w3 = get_3d_rope_index(ids, it['image_grid_thw'].numpy())
        embs = np.array(self.embed(ids))
        ve_np = np.array(ve)
        img_pos = np.where(np.array(input_ids) == 100295)[0]
        for i, pos in enumerate(img_pos):
            embs[0, pos, :] = ve_np[i]
        hidden = mx.array(embs)
        
        # 4. Auto-Regressive Generation with KV Cache
        res = []
        caches = [None] * len(self.layers)
        
        # Prefill
        L = hidden.shape[1]
        mask = mx.triu(mx.full((L, L), -mx.inf, dtype=hidden.dtype), k=1)
        x = hidden
        for i, l in enumerate(self.layers):
            x, caches[i] = l(x, mask, t3, h3, w3, cache=None)
        
        logits = self.head(self.norm(x[:, -1, :]))
        next_id = int(mx.argmax(logits, axis=-1))
        
        for _ in range(max_tokens):
            if next_id == self.tok.eos_token_id: break
            res.append(next_id)
            
            # Extract current token
            token_str = self.tok.decode([next_id])
            yield token_str

            # Step
            curr_hidden = self.embed(mx.array([[next_id]]))
            curr_pos = int(mx.max(mx.concatenate([t3[0, -1:], h3[0, -1:], w3[0, -1:]]))) + 1
            curr_t = mx.array([[curr_pos]], dtype=mx.int32)
            curr_h = mx.array([[curr_pos]], dtype=mx.int32)
            curr_w = mx.array([[curr_pos]], dtype=mx.int32)
            
            x = curr_hidden
            for i, l in enumerate(self.layers):
                x, caches[i] = l(x, None, curr_t, curr_h, curr_w, cache=caches[i])
            
            logits = self.head(self.norm(x[:, -1, :]))
            if penalty != 1.0 and len(res) > 0:
                for tid in set(res): logits[0, tid] /= penalty
            
            next_id = int(mx.argmax(logits, axis=-1))
            
            t3 = mx.concatenate([t3, curr_t], axis=1)
            h3 = mx.concatenate([h3, curr_h], axis=1)
            w3 = mx.concatenate([w3, curr_w], axis=1)

    def ocr(self, img_path, prompt="OCR:", max_tokens=256, penalty=1.1, stream=False):
        """对外接口:支持流式逻辑分离"""
        gen = self._ocr_generator(img_path, prompt, max_tokens, penalty)
        if stream:
            return gen
        else:
            return "".join(list(gen))

if __name__ == "__main__":
    model = ALIGNED_PPOCR("/Users/gt/.gemini/antigravity/scratch/paddleocr-mlx-conversion")
    
    # Final Sanity Test
    print("\n--- MLX ALIGNED PPOCR FINAL TEST ---")
    img = Image.new('RGB', (400, 100), 'white')
    draw = ImageDraw.Draw(img)
    try:
        f = "/System/Library/Fonts/Supplemental/Songti.ttc"
        font = ImageFont.truetype(f, 40)
        draw.text((20, 20), "PaddleOCR MLX 对齐成功", fill='black', font=font)
    except:
        draw.text((20, 20), "PaddleOCR MLX ALIGNED", fill='black')
    
    img.save("/tmp/final_success.png")
    print(f"Recognized: {model.ocr('/tmp/final_success.png')}")