File size: 13,239 Bytes
02acc80
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
E2E Validation v2: Fixed resolution pipeline.
Forces all images to 1120x1540 so token count matches fixed vision encoder output.
Compares HF model output vs fixed PyTorch modules output.
"""

import os, sys, time, torch, torch.nn.functional as F
from pathlib import Path
from PIL import Image

# ── Constants ──
MODEL_DIR = "./models/LightOnOCR-2-1B"
FIXED_H, FIXED_W = 1120, 1540
PATCH_SIZE = 14
SPATIAL_MERGE = 2
MERGED_H = FIXED_H // PATCH_SIZE // SPATIAL_MERGE  # 40
MERGED_W = FIXED_W // PATCH_SIZE // SPATIAL_MERGE  # 55
NUM_IMG_TOKENS = MERGED_H * MERGED_W               # 2200
NUM_VPAD_TOKENS = MERGED_H - 1                      # 39

# Token IDs
IMAGE_TOKEN_ID = 151655
VISION_PAD_ID = 151654
VISION_END_ID = 151653
IM_START_ID = 151644
IM_END_ID = 151645
EOS_TOKEN_ID = 151645

# Decoder constants
NUM_LAYERS = 28
NUM_KV_HEADS = 8
NUM_HEADS = 16
HEAD_DIM = 128
HIDDEN_SIZE = 1024
MAX_SEQ_LEN = 4096  # Large enough for 2256 input + generation


def download_test_images():
    """Get real test images."""
    import requests
    os.makedirs("test_images", exist_ok=True)
    images = {}

    sources = {
        "receipt": "https://huggingface.co/datasets/hf-internal-testing/fixtures_ocr/resolve/main/SROIE-receipt.jpeg",
    }
    for name, url in sources.items():
        path = f"test_images/{name}.png"
        if not os.path.exists(path):
            print(f"  Downloading {name}...")
            try:
                resp = requests.get(url, timeout=30)
                resp.raise_for_status()
                with open(path, "wb") as f:
                    f.write(resp.content)
            except Exception as e:
                print(f"    FAILED: {e}")
                continue
        img = Image.open(path).convert("RGB")
        images[name] = img
        print(f"  {name}: {img.size}")

    # Synthetic doc
    img = Image.new("RGB", (800, 600), "white")
    from PIL import ImageDraw
    draw = ImageDraw.Draw(img)
    draw.text((50, 50), "Invoice #12345", fill="black")
    draw.text((50, 100), "Date: 2024-01-15", fill="black")
    draw.text((50, 150), "Item 1: Widget x5 @ $10.00 = $50.00", fill="black")
    draw.text((50, 200), "Item 2: Gadget x2 @ $24.99 = $49.98", fill="black")
    draw.text((50, 250), "Total: $99.98", fill="black")
    img.save("test_images/synthetic.png")
    images["synthetic"] = img
    print(f"  synthetic: {img.size}")

    return images


def build_fixed_input_ids(processor, text_prompt="OCR this document. Extract all text."):
    """
    Build input_ids with exactly NUM_IMG_TOKENS image tokens,
    matching our fixed 1120x1540 vision encoder output.
    Uses the processor's chat template but with a fixed-size image.
    """
    # Create a dummy image at exact target size to get correct token count
    dummy_img = Image.new("RGB", (FIXED_W, FIXED_H), "white")
    messages = [{"role": "user", "content": [
        {"type": "image"}, {"type": "text", "text": text_prompt}
    ]}]
    text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
    inputs = processor(text=text, images=[dummy_img], return_tensors="pt")

    input_ids = inputs["input_ids"]
    # Verify token counts
    ids_list = input_ids[0].tolist()
    n_img = ids_list.count(IMAGE_TOKEN_ID)
    n_pad = ids_list.count(VISION_PAD_ID)
    assert n_img == NUM_IMG_TOKENS, f"Expected {NUM_IMG_TOKENS} IMG tokens, got {n_img}"
    print(f"    Input template: {input_ids.shape[1]} tokens ({n_img} IMG, {n_pad} VPAD)")
    return input_ids


def preprocess_image_fixed(img, processor):
    """Resize image to exactly FIXED_H x FIXED_W and get pixel_values."""
    # Resize maintaining content, pad to target
    img_resized = img.resize((FIXED_W, FIXED_H), Image.LANCZOS)
    # Use processor's image normalization
    dummy_messages = [{"role": "user", "content": [{"type": "image"}]}]
    text = processor.apply_chat_template(dummy_messages, add_generation_prompt=True, tokenize=False)
    inputs = processor(text=text, images=[img_resized], return_tensors="pt")
    return inputs["pixel_values"]


def run_hf_model(images, processor):
    """Run original HF model with FIXED resolution preprocessing."""
    from transformers import AutoModelForImageTextToText
    from safetensors.torch import load_file

    print("\n[HF Model] Loading...")
    model = AutoModelForImageTextToText.from_pretrained(
        MODEL_DIR, dtype=torch.bfloat16, attn_implementation="sdpa", device_map="cpu",
    )
    state_dict = load_file(os.path.join(MODEL_DIR, "model.safetensors"))
    remapped = {k.replace("model.vision_encoder.", "model.vision_tower.")
                 .replace("model.vision_projection.", "model.multi_modal_projector."): v
                for k, v in state_dict.items()}
    model.load_state_dict(remapped, strict=False)
    model = model.to("cuda").eval()

    results = {}
    for name, img in images.items():
        print(f"\n  [{name}] HF generate (fixed {FIXED_H}x{FIXED_W})...")
        pixel_values = preprocess_image_fixed(img, processor).to("cuda")
        input_ids = build_fixed_input_ids(processor).to("cuda")
        attention_mask = torch.ones_like(input_ids)

        input_len = input_ids.shape[1]
        t0 = time.time()
        with torch.no_grad():
            output_ids = model.generate(
                input_ids=input_ids,
                pixel_values=pixel_values,
                attention_mask=attention_mask,
                image_sizes=torch.tensor([[FIXED_H, FIXED_W]], device="cuda"),
                max_new_tokens=512, do_sample=False, temperature=None, top_p=None,
            )
        elapsed = time.time() - t0
        new_ids = output_ids[0, input_len:]
        text = processor.tokenizer.decode(new_ids, skip_special_tokens=True)
        n_tok = len(new_ids)
        print(f"    {n_tok} tokens, {elapsed:.1f}s ({n_tok/elapsed:.1f} tok/s)")
        print(f"    Output: {text[:200]}...")
        results[name] = {"text": text, "tokens": n_tok, "time": elapsed}

    del model
    torch.cuda.empty_cache()
    return results


def run_fixed_modules(images, processor):
    """Run fixed PyTorch modules E2E with proper token matching."""
    sys.path.insert(0, ".")
    from export_vision import build_vision_module, load_original_model
    from export_decoder import TextDecoderFixed, build_decoder_module

    print("\n[Fixed Modules] Loading...")
    orig = load_original_model()

    vision = build_vision_module(orig)
    decoder = build_decoder_module(orig)
    embed_tokens = orig.model.language_model.embed_tokens

    device = "cuda"
    dtype = torch.bfloat16
    vision = vision.to(device).to(dtype).eval()
    decoder = decoder.to(device).to(dtype).eval()
    embed_tokens = embed_tokens.to(device).to(dtype)

    del orig
    torch.cuda.empty_cache()

    results = {}
    for name, img in images.items():
        print(f"\n  [{name}] Fixed modules E2E...")
        try:
            # Step 1: Get pixel values at fixed resolution
            pixel_values = preprocess_image_fixed(img, processor).to(device).to(dtype)
            input_ids = build_fixed_input_ids(processor).to(device)

            # Step 2: Vision encoder
            with torch.no_grad():
                image_features = vision(pixel_values)  # [1, 2200, 1024]
            print(f"    Vision output: {image_features.shape}")

            # Step 3: Build combined embeddings
            with torch.no_grad():
                text_embeds = embed_tokens(input_ids)  # [1, seq_len, 1024]

            ids_list = input_ids[0].tolist()
            img_positions = [i for i, t in enumerate(ids_list) if t == IMAGE_TOKEN_ID]
            assert len(img_positions) == image_features.shape[1], \
                f"Token/feature mismatch: {len(img_positions)} slots vs {image_features.shape[1]} features"

            combined = text_embeds.clone()
            # Scatter vision features into IMG token positions
            indices = torch.tensor(img_positions, device=device)
            combined[0, indices] = image_features[0]

            seq_len = combined.shape[1]
            print(f"    Combined seq: {seq_len}, scattering {len(img_positions)} features")

            # Step 4: Prefill β€” feed combined embeddings through decoder
            kv_caches = []
            for _ in range(NUM_LAYERS):
                k = torch.zeros(1, NUM_KV_HEADS, MAX_SEQ_LEN, HEAD_DIM, dtype=dtype, device=device)
                v = torch.zeros(1, NUM_KV_HEADS, MAX_SEQ_LEN, HEAD_DIM, dtype=dtype, device=device)
                kv_caches.extend([k, v])

            position_ids = torch.arange(seq_len, device=device).unsqueeze(0)
            cache_position = torch.arange(seq_len, device=device)

            # Causal mask for prefill
            mask = torch.full((1, 1, seq_len, MAX_SEQ_LEN), float("-inf"), dtype=dtype, device=device)
            for i in range(seq_len):
                mask[0, 0, i, :i+1] = 0.0

            # Monkey-patch embed_tokens for prefill (we already have embeddings)
            orig_embed = decoder.embed_tokens
            class PrefillEmbed(torch.nn.Module):
                def __init__(self, embeds):
                    super().__init__()
                    self.e = embeds
                def forward(self, x):
                    return self.e
            decoder.embed_tokens = PrefillEmbed(combined)

            t0 = time.time()
            with torch.no_grad():
                result = decoder(input_ids[:, :seq_len], mask, position_ids, cache_position, *kv_caches)

            decoder.embed_tokens = orig_embed

            logits = result[0]
            kv_caches = list(result[1:])
            next_token = logits[0, -1].argmax().item()

            generated = [next_token]
            cur_pos = seq_len

            # Step 5: Autoregressive decode
            for step in range(511):
                if next_token == EOS_TOKEN_ID or cur_pos >= MAX_SEQ_LEN:
                    break

                token_input = torch.tensor([[next_token]], device=device)
                pos_ids = torch.tensor([[cur_pos]], device=device)
                cache_pos = torch.tensor([cur_pos], device=device)

                # Decode mask: attend to all positions up to cur_pos
                dmask = torch.zeros(1, 1, 1, MAX_SEQ_LEN, dtype=dtype, device=device)
                dmask[0, 0, 0, cur_pos+1:] = float("-inf")

                with torch.no_grad():
                    result = decoder(token_input, dmask, pos_ids, cache_pos, *kv_caches)

                logits = result[0]
                kv_caches = list(result[1:])
                next_token = logits[0, -1].argmax().item()
                generated.append(next_token)
                cur_pos += 1

            elapsed = time.time() - t0
            text = processor.tokenizer.decode(generated, skip_special_tokens=True)
            n_tok = len(generated)
            print(f"    {n_tok} tokens, {elapsed:.1f}s ({n_tok/elapsed:.1f} tok/s)")
            print(f"    Output: {text[:200]}...")
            results[name] = {"text": text, "tokens": n_tok, "time": elapsed}

        except Exception as e:
            import traceback
            traceback.print_exc()
            results[name] = {"text": f"ERROR: {e}", "tokens": 0, "time": 0}

    return results


def levenshtein(s1, s2):
    if len(s1) < len(s2): return levenshtein(s2, s1)
    if len(s2) == 0: return len(s1)
    prev = list(range(len(s2) + 1))
    for i, c1 in enumerate(s1):
        curr = [i + 1]
        for j, c2 in enumerate(s2):
            curr.append(min(prev[j+1]+1, curr[j]+1, prev[j]+(c1!=c2)))
        prev = curr
    return prev[-1]


def compare(hf_results, fx_results, images):
    print("\n" + "="*70)
    print("E2E COMPARISON (Fixed 1120x1540 resolution)")
    print("="*70)

    for name in images:
        hf = hf_results[name]
        fx = fx_results[name]
        hf_t, fx_t = hf["text"], fx["text"]

        exact = hf_t.strip() == fx_t.strip()
        ed = levenshtein(hf_t, fx_t)
        max_len = max(len(hf_t), len(fx_t), 1)
        char_acc = 1.0 - ed / max_len

        ref_words = set(hf_t.lower().split())
        hyp_words = set(fx_t.lower().split())
        union = ref_words | hyp_words
        word_acc = len(ref_words & hyp_words) / len(union) if union else 1.0

        print(f"\n{'─'*70}")
        print(f"  [{name}]")
        print(f"  HF  ({hf['tokens']} tok): {hf_t[:250]}")
        print(f"  FIX ({fx['tokens']} tok): {fx_t[:250]}")
        print(f"  Exact: {'βœ… YES' if exact else '❌ NO'}")
        print(f"  Edit dist: {ed}, Char acc: {char_acc:.4f}, Word acc: {word_acc:.4f}")

    print("\n" + "="*70)


def main():
    print("LightOnOCR-2-1B E2E Validation v2")
    print(f"Fixed resolution: {FIXED_H}x{FIXED_W} β†’ {NUM_IMG_TOKENS} vision features")
    print(f"Device: cuda, Max seq: {MAX_SEQ_LEN}")
    print("="*70)

    images = download_test_images()

    from transformers import AutoProcessor
    processor = AutoProcessor.from_pretrained(MODEL_DIR)

    hf_results = run_hf_model(images, processor)

    # Free GPU for fixed modules
    torch.cuda.empty_cache()

    fx_results = run_fixed_modules(images, processor)

    compare(hf_results, fx_results, images)


if __name__ == "__main__":
    main()