File size: 32,692 Bytes
5585d19
 
7b88814
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5585d19
 
7b88814
5585d19
7b88814
5585d19
 
7b88814
 
 
 
 
 
 
5585d19
 
7b88814
5585d19
 
 
 
 
 
 
 
 
 
 
 
 
 
7b88814
 
3e06f75
7b88814
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3f0ff37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7b88814
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e35bf09
 
 
5585d19
 
7b88814
e35bf09
 
 
 
 
 
 
 
5585d19
e35bf09
7b88814
 
 
 
 
 
 
 
 
e35bf09
5585d19
 
 
 
 
7b88814
5585d19
 
 
7b88814
5585d19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7b88814
5585d19
 
e35bf09
5585d19
 
 
 
 
 
 
 
7b88814
5585d19
7b88814
5585d19
 
7b88814
5585d19
 
 
 
 
 
 
 
 
 
 
 
 
7b88814
5585d19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e35bf09
5585d19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e35bf09
5585d19
 
 
 
 
 
 
 
 
 
 
 
e35bf09
5585d19
 
 
 
 
 
7b88814
5585d19
 
 
 
 
 
 
7b88814
5585d19
 
 
e35bf09
5585d19
 
 
 
 
 
 
 
 
 
 
 
e35bf09
5585d19
7b88814
 
 
 
 
 
 
 
 
 
 
 
 
 
5585d19
 
 
e35bf09
5585d19
 
 
 
7b88814
 
e35bf09
7b88814
 
 
 
 
 
 
e35bf09
 
5585d19
 
 
 
 
 
 
e35bf09
5585d19
 
 
 
 
 
e35bf09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7b88814
 
 
 
 
5585d19
7b88814
5585d19
7b88814
 
 
 
 
 
 
 
 
 
5585d19
 
 
7b88814
 
5585d19
7b88814
 
5585d19
 
 
 
7b88814
5585d19
 
 
7b88814
 
 
5585d19
 
7b88814
5585d19
7b88814
 
 
5585d19
 
7b88814
5585d19
7b88814
5585d19
7b88814
5585d19
 
7b88814
5585d19
7b88814
5585d19
 
7b88814
5585d19
7b88814
5585d19
7b88814
 
 
 
 
 
 
 
 
 
 
 
 
 
5585d19
 
7b88814
5585d19
7b88814
5585d19
 
 
e35bf09
5585d19
 
7b88814
 
e35bf09
7b88814
 
 
 
e35bf09
 
 
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
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
#!/usr/bin/env python3
"""
QuantPPI — Hugging Face ZeroGPU Gradio Space + Cloud FastAPI Backend Entrypoint.

Supports all 10 Protein Language Models from QuantPPI.ipynb:
  - ESM-2: 8M (D=320), 35M (D=480), 150M (D=640), 650M (D=1280), 3B (D=2560), 15B (D=5120)
  - ESM-C: 300M (D=960), 600M (D=1152), 6B (D=2560)
  - ESM-3: 1.4B (D=1536)

Architecture:
1. `spaces` is imported prior to any `torch` import as required by Hugging Face ZeroGPU.
2. Precomputed corpus embeddings (.h5) for all 10 PLMs stream on-demand from `HF_DATASET_REPO`
   (`trizzhal9000/quantppi-embeddings`).
3. Custom raw FASTA queries execute via EvolutionaryScale Forge API (`ESM_FORGE_TOKEN`) for
   ESM-C / ESM-3 and `@spaces.GPU` (Nvidia H200 / A10G ZeroGPU) for ESM-2 8M–15B.
4. Launches via `demo.launch(...)` so ZeroGPU's worker daemon initializes natively while
   exposing `/api/*`, `/figures/*`, and `/studio` with CORS enabled for Vercel production.
"""

import gc
import os
import re
import sys
from pathlib import Path
from typing import Any, Dict, List, Optional

# ZeroGPU requirement: import `spaces` before anything that could import `torch`
try:
    import spaces
except ImportError:
    spaces = None

os.environ.setdefault("MPLCONFIGDIR", "/tmp")
os.environ.setdefault("HF_DATASET_REPO", "trizzhal9000/quantppi-embeddings")

REPO_ROOT = Path(__file__).resolve().parent
EXPLAINER_DIR = REPO_ROOT / "explainer"
if str(REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(REPO_ROOT))
if str(EXPLAINER_DIR) not in sys.path:
    sys.path.insert(0, str(EXPLAINER_DIR))

import numpy as np
import pandas as pd

import server as qserver


def _gpu_decorator(fn):
    if spaces is not None and hasattr(spaces, "GPU"):
        return spaces.GPU(duration=180)(fn)
    return fn


ESM2_HF_REPOS: Dict[str, str] = {
    "esm2_8m": "facebook/esm2_t6_8M_UR50D",
    "esm2_35m": "facebook/esm2_t12_35M_UR50D",
    "esm2_150m": "facebook/esm2_t30_150M_UR50D",
    "esm2_650m": "facebook/esm2_t33_650M_UR50D",
    "esm2_3b": "facebook/esm2_t36_3B_UR50D",
    "esm2_15b": "facebook/esm2_t48_15B_UR50D",
}

FORGE_MODEL_CANDIDATES: Dict[str, List[str]] = {
    "esmc_300m": ["esmc-300m-2024-12", "esmc_300m"],
    "esmc_600m": ["esmc-600m-2024-12", "esmc_600m"],
    "esmc_6b": ["esmc-6b-2024-12", "esmc_6b"],
    "esm3_1.4b": ["esm3-open-2024-03", "esm3-sm-open-v1", "esm3-medium-2024-08"],
}

_CPU_SMALL_MODEL_CACHE: Dict[str, Any] = {}


def _extract_via_forge_api(clean_seq: str, model_scale: str) -> Optional[List[float]]:
    """Extracts ESM-C (300M, 600M, 6B) and ESM-3 (1.4B) embeddings via EvolutionaryScale Forge API."""
    forge_token = (os.environ.get("ESM_FORGE_TOKEN") or "").strip()
    if not forge_token or model_scale not in FORGE_MODEL_CANDIDATES:
        return None

    try:
        from esm.sdk import client as esm_forge_client
        from esm.sdk.api import ESMProtein, LogitsConfig

        s_trunc = clean_seq[:1022]
        for model_name in FORGE_MODEL_CANDIDATES[model_scale]:
            try:
                forge_client = esm_forge_client(
                    model=model_name,
                    url="https://forge.evolutionaryscale.ai",
                    token=forge_token,
                )
                protein_tensor = forge_client.encode(ESMProtein(sequence=s_trunc))
                res = forge_client.logits(
                    protein_tensor,
                    LogitsConfig(sequence=True, return_embeddings=True),
                )
                emb_t = res.embeddings
                if hasattr(emb_t, "detach"):
                    arr = emb_t.detach().float().cpu().numpy()
                else:
                    arr = np.asarray(emb_t, dtype=np.float32)
                tok_slice = arr[0, 1 : 1 + len(s_trunc), :] if arr.ndim == 3 else arr[1 : 1 + len(s_trunc), :]
                vec = tok_slice.astype(np.float32).mean(axis=0)
                nrm = float(np.linalg.norm(vec))
                if nrm > 1e-12:
                    vec /= nrm
                return vec.tolist()
            except Exception as inner_exc:
                print(f"[QuantPPI Forge] Model {model_name} attempt notice: {inner_exc}")
                continue
    except Exception as exc:
        print(f"[QuantPPI Forge] SDK notice: {exc}")
    return None


@_gpu_decorator
def zerogpu_extract_sequence_embedding(clean_seq: str, model_scale: str) -> List[float]:
    """Runs PLM forward pass on Hugging Face ZeroGPU (Nvidia H200 / A10G) and returns unit-L2 vector."""
    import torch

    s_trunc = clean_seq[:1022]
    device = "cuda" if torch.cuda.is_available() else "cpu"
    hf_token = (os.environ.get("HF_TOKEN") or "").strip() or None

    if model_scale in ESM2_HF_REPOS:
        from transformers import AutoTokenizer, EsmModel

        hf_id = ESM2_HF_REPOS[model_scale]
        use_bf16 = device == "cuda" and model_scale in ("esm2_3b", "esm2_15b")
        dtype = torch.bfloat16 if use_bf16 else torch.float32

        tokenizer = AutoTokenizer.from_pretrained(hf_id, token=hf_token)
        model = EsmModel.from_pretrained(
            hf_id,
            add_pooling_layer=False,
            torch_dtype=dtype,
            token=hf_token,
        ).to(device).eval()
        try:
            with torch.inference_mode():
                toks = tokenizer([s_trunc], return_tensors="pt", truncation=True, max_length=1024).to(device)
                if use_bf16:
                    with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
                        hidden = model(**toks).last_hidden_state
                else:
                    hidden = model(**toks).last_hidden_state
                vec = hidden[0, 1 : 1 + len(s_trunc), :].to(torch.float32).mean(dim=0)
                vec = torch.nn.functional.normalize(vec, p=2, dim=0)
                return vec.cpu().numpy().tolist()
        finally:
            del model
            gc.collect()
            if torch.cuda.is_available():
                torch.cuda.empty_cache()

    if model_scale in ("esmc_300m", "esmc_600m", "esmc_6b"):
        try:
            from esm.models.esmc import ESMC
            from esm.sdk.api import ESMProtein, LogitsConfig

            client = ESMC.from_pretrained(model_scale).to(device)
            client = (client.to(torch.bfloat16) if (device == "cuda" and model_scale == "esmc_6b") else client.float()).eval()
            try:
                with torch.inference_mode():
                    pt = client.encode(ESMProtein(sequence=s_trunc))
                    out = client.logits(pt, LogitsConfig(sequence=True, return_embeddings=True))
                    emb_t = out.embeddings
                    tok_slice = emb_t[0, 1 : 1 + len(s_trunc), :].float() if emb_t.ndim == 3 else emb_t[1 : 1 + len(s_trunc), :].float()
                    vec = torch.nn.functional.normalize(tok_slice.mean(dim=0), p=2, dim=0)
                    return vec.cpu().numpy().tolist()
            finally:
                del client
                gc.collect()
                if torch.cuda.is_available():
                    torch.cuda.empty_cache()
        except Exception:
            if model_scale in ("esmc_300m", "esmc_600m"):
                from transformers import AutoModel, AutoTokenizer

                hf_repo = "Synthyra/ESMplusplus_small" if model_scale == "esmc_300m" else "Synthyra/ESMplusplus_large"
                tok = AutoTokenizer.from_pretrained(hf_repo, trust_remote_code=True, token=hf_token)
                mod = AutoModel.from_pretrained(hf_repo, trust_remote_code=True, torch_dtype=torch.float32, token=hf_token).to(device).eval()
                try:
                    with torch.inference_mode():
                        enc = tok(s_trunc, return_tensors="pt", truncation=True, max_length=1024).to(device)
                        outs = mod(**enc, output_hidden_states=True)
                        hid = outs.last_hidden_state if getattr(outs, "last_hidden_state", None) is not None else outs.hidden_states[-1]
                        vec = torch.nn.functional.normalize(hid[0, 1 : 1 + len(s_trunc), :].float().mean(dim=0), p=2, dim=0)
                        return vec.cpu().numpy().tolist()
                finally:
                    del mod
                    gc.collect()
                    if torch.cuda.is_available():
                        torch.cuda.empty_cache()
            raise

    if model_scale == "esm3_1.4b":
        if hf_token:
            os.environ.setdefault("HF_TOKEN", hf_token)
        from esm.models.esm3 import ESM3
        from esm.sdk.api import ESMProtein, LogitsConfig

        client = None
        for cand in ("esm3_sm_open_v1", "esm3-sm-open-v1"):
            try:
                client = ESM3.from_pretrained(cand).to(device)
                break
            except Exception:
                continue
        if client is None:
            raise RuntimeError("Could not load ESM-3 1.4B weights.")
        client = (client.to(torch.bfloat16) if device == "cuda" else client.float()).eval()
        try:
            with torch.inference_mode():
                pt = client.encode(ESMProtein(sequence=s_trunc))
                out = client.logits(pt, LogitsConfig(sequence=True, return_embeddings=True))
                emb_t = out.embeddings
                tok_slice = emb_t[0, 1 : 1 + len(s_trunc), :].float() if emb_t.ndim == 3 else emb_t[1 : 1 + len(s_trunc), :].float()
                vec = torch.nn.functional.normalize(tok_slice.mean(dim=0), p=2, dim=0)
                return vec.cpu().numpy().tolist()
        finally:
            del client
            gc.collect()
            if torch.cuda.is_available():
                torch.cuda.empty_cache()

    raise ValueError(f"Unsupported model_scale for ZeroGPU FASTA inference: {model_scale}")


def _cpu_stateless_fallback(clean_seq: str, model_scale: str) -> Optional[List[float]]:
    """Stateless CPU fallback on HF Space host RAM for small/medium checkpoints when called via external REST."""
    import torch

    s_trunc = clean_seq[:1022]
    hf_token = (os.environ.get("HF_TOKEN") or "").strip() or None

    if model_scale in ("esm2_8m", "esm2_35m", "esm2_150m", "esm2_650m"):
        from transformers import AutoTokenizer, EsmModel

        hf_id = ESM2_HF_REPOS[model_scale]
        if hf_id not in _CPU_SMALL_MODEL_CACHE:
            if len(_CPU_SMALL_MODEL_CACHE) >= 2:
                _CPU_SMALL_MODEL_CACHE.pop(next(iter(_CPU_SMALL_MODEL_CACHE.keys())), None)
            tok = AutoTokenizer.from_pretrained(hf_id, token=hf_token)
            mod = EsmModel.from_pretrained(hf_id, add_pooling_layer=False, torch_dtype=torch.float32, token=hf_token).eval()
            _CPU_SMALL_MODEL_CACHE[hf_id] = (tok, mod)
        tok, mod = _CPU_SMALL_MODEL_CACHE[hf_id]
        with torch.inference_mode():
            toks = tok([s_trunc], return_tensors="pt", truncation=True, max_length=1024)
            hid = mod(**toks).last_hidden_state
            vec = torch.nn.functional.normalize(hid[0, 1 : 1 + len(s_trunc), :].float().mean(dim=0), p=2, dim=0)
            return vec.numpy().tolist()

    if model_scale in ("esmc_300m", "esmc_600m"):
        from transformers import AutoModel, AutoTokenizer

        hf_repo = "Synthyra/ESMplusplus_small" if model_scale == "esmc_300m" else "Synthyra/ESMplusplus_large"
        if hf_repo not in _CPU_SMALL_MODEL_CACHE:
            if len(_CPU_SMALL_MODEL_CACHE) >= 2:
                _CPU_SMALL_MODEL_CACHE.pop(next(iter(_CPU_SMALL_MODEL_CACHE.keys())), None)
            tok = AutoTokenizer.from_pretrained(hf_repo, trust_remote_code=True, token=hf_token)
            mod = AutoModel.from_pretrained(hf_repo, trust_remote_code=True, torch_dtype=torch.float32, token=hf_token).eval()
            _CPU_SMALL_MODEL_CACHE[hf_repo] = (tok, mod)
        tok, mod = _CPU_SMALL_MODEL_CACHE[hf_repo]
        with torch.inference_mode():
            enc = tok(s_trunc, return_tensors="pt", truncation=True, max_length=1024)
            outs = mod(**enc, output_hidden_states=True)
            hid = outs.last_hidden_state if getattr(outs, "last_hidden_state", None) is not None else outs.hidden_states[-1]
            vec = torch.nn.functional.normalize(hid[0, 1 : 1 + len(s_trunc), :].float().mean(dim=0), p=2, dim=0)
            return vec.numpy().tolist()

    return None


def cloud_extract_fasta_dispatcher(clean_seq: str, model_scale: str) -> Optional[List[float]]:
    """Dispatches custom FASTA embedding extraction across Forge API -> ZeroGPU -> Space CPU."""
    if model_scale in FORGE_MODEL_CANDIDATES:
        forge_vec = _extract_via_forge_api(clean_seq, model_scale)
        if forge_vec is not None:
            return forge_vec

    try:
        return zerogpu_extract_sequence_embedding(clean_seq, model_scale)
    except Exception as gpu_exc:
        print(f"[QuantPPI Space] ZeroGPU call notice for {model_scale}: {gpu_exc}")
        return _cpu_stateless_fallback(clean_seq, model_scale)


# Register cloud FASTA extractor onto server module
qserver.EXTERNAL_CLOUD_FASTA_EXTRACTOR = cloud_extract_fasta_dispatcher


ALL_MODEL_SCALES_META = [
    {"id": "esm2_8m", "name": "ESM-2 8M (6-layer, D=320)", "dimension": 320, "fp32_bytes": 1280},
    {"id": "esm2_35m", "name": "ESM-2 35M (12-layer, D=480)", "dimension": 480, "fp32_bytes": 1920},
    {"id": "esm2_150m", "name": "ESM-2 150M (30-layer, D=640)", "dimension": 640, "fp32_bytes": 2560},
    {"id": "esmc_300m", "name": "ESM-C 300M (30-layer, D=960)", "dimension": 960, "fp32_bytes": 3840},
    {"id": "esmc_600m", "name": "ESM-C 600M (36-layer, D=1152)", "dimension": 1152, "fp32_bytes": 4608},
    {"id": "esm2_650m", "name": "ESM-2 650M (33-layer, D=1280)", "dimension": 1280, "fp32_bytes": 5120},
    {"id": "esm3_1.4b", "name": "ESM-3 1.4B (48-layer, D=1536)", "dimension": 1536, "fp32_bytes": 6144},
    {"id": "esm2_3b", "name": "ESM-2 3B (36-layer, D=2560)", "dimension": 2560, "fp32_bytes": 10240},
    {"id": "esmc_6b", "name": "ESM-C 6B (80-layer, D=2560)", "dimension": 2560, "fp32_bytes": 10240},
    {"id": "esm2_15b", "name": "ESM-2 15B (48-layer, D=5120)", "dimension": 5120, "fp32_bytes": 20480},
]


def _load_structure_entry_shared(target_raw: str) -> Optional[Dict[str, Any]]:
    tid = target_raw.strip()
    if not tid:
        return None
    if tid == "custom_fasta":
        if qserver.STATE._last_fasta_query:
            closest_scaff = qserver.STATE.match_closest_scope_by_sequence(qserver.STATE._last_fasta_query)["id"]
            return _load_structure_entry_shared(closest_scaff)
        return _load_structure_entry_shared("d1dlwa_")

    if qserver.has_pdb_structure(tid):
        ent_p = qserver.PDB_DIR / f"{tid}.ent"
        obj = qserver.parse_pdb_ca_atoms(ent_p)
        obj["ca_atoms"] = [dict(a) for a in obj["ca_atoms"]]
        if ent_p.exists():
            try:
                obj["raw_pdb_text"] = ent_p.read_text(encoding="utf-8", errors="replace")
            except Exception:
                pass
        r_idx = qserver.STATE.scope_id_to_raw_idx.get(tid)
        obj["metadata"] = dict(qserver.STATE.scope_entries[r_idx]) if r_idx is not None else {"id": tid}
        return obj

    if re.match(r"^[0-9][A-Za-z0-9]{3}$", tid):
        live = qserver.fetch_live_rcsb_pdb(tid)
        if live:
            mapped = (
                qserver.STATE.resolve_scope_entry(tid)
                if qserver.STATE.has_exact_or_alias_scope_match(tid)
                else qserver.STATE.match_closest_scope_by_sequence(live["metadata"].get("sequence", ""))
            )
            meta = dict(live["metadata"])
            if mapped:
                meta.setdefault("sccs", mapped.get("sccs", ""))
                meta.setdefault("class_name", mapped.get("class_name", ""))
            return {
                **live,
                "ca_atoms": [dict(a) for a in live["ca_atoms"]],
                "metadata": meta,
            }

    if tid.upper() in qserver.STATE.sp_by_acc:
        sp_entry = qserver.STATE.sp_by_acc[tid.upper()]
        scaff = qserver.STATE.match_closest_scope_by_sequence(sp_entry.get("sequence", ""))
        scaff_id = scaff["id"]
        if qserver.has_pdb_structure(scaff_id):
            ent_p = qserver.PDB_DIR / f"{scaff_id}.ent"
            obj = qserver.parse_pdb_ca_atoms(ent_p)
            obj["ca_atoms"] = [dict(a) for a in obj["ca_atoms"]]
            obj["metadata"] = {
                **dict(sp_entry),
                "sccs": scaff.get("sccs", ""),
                "pdb_id": scaff.get("pdb_id", ""),
                "description": f"{sp_entry.get('description', tid)} [Homologous crystal domain {scaff_id}]",
            }
            return obj

    res_entry = qserver.STATE.resolve_scope_entry(tid)
    resolved = qserver.STATE.scope_entries[res_entry["raw_idx"]]["id"] if "raw_idx" in res_entry else res_entry["id"]
    if qserver.has_pdb_structure(resolved):
        ent_p = qserver.PDB_DIR / f"{resolved}.ent"
        obj = qserver.parse_pdb_ca_atoms(ent_p)
        obj["ca_atoms"] = [dict(a) for a in obj["ca_atoms"]]
        if ent_p.exists():
            try:
                obj["raw_pdb_text"] = ent_p.read_text(encoding="utf-8", errors="replace")
            except Exception:
                pass
        r_idx = qserver.STATE.scope_id_to_raw_idx.get(resolved)
        meta = dict(qserver.STATE.scope_entries[r_idx]) if r_idx is not None else {"id": resolved}
        if re.match(r"^[0-9][A-Za-z0-9]{3}$", tid):
            meta["id"] = tid.upper()
            meta["pdb_id"] = tid.upper()
        obj["metadata"] = meta
        return obj
    return None


def build_quantppi_rest_app() -> Any:
    """Builds a dedicated FastAPI app with CORS for `/api/*`, `/figures/*`, `/studio`, and `/assets/*`."""
    from fastapi import FastAPI, Request
    from fastapi.middleware.cors import CORSMiddleware
    from fastapi.responses import FileResponse, JSONResponse

    rest_app = FastAPI(title="QuantPPI Cloud ZeroGPU API", version="2.0.0")
    rest_app.add_middleware(
        CORSMiddleware,
        allow_origins=["*"],
        allow_credentials=False,
        allow_methods=["*"],
        allow_headers=["*"],
    )

    @rest_app.get("/api/health")
    def api_health():
        return JSONResponse({
            "status": "ok",
            "zerogpu": spaces is not None,
            "hf_dataset_repo": os.environ.get("HF_DATASET_REPO", "trizzhal9000/quantppi-embeddings"),
            "forge_configured": bool(os.environ.get("ESM_FORGE_TOKEN")),
            "models": [m["id"] for m in ALL_MODEL_SCALES_META],
        })

    @rest_app.get("/api/overview")
    def api_overview(dataset: str = "scope", model_scale: str = "esm2_650m"):
        geom = qserver.STATE.get_corpus_geometry(dataset, model_scale)
        return JSONResponse({
            "geometry": geom,
            "codecs": qserver.CODEC_CATALOG,
            "benchmarks": qserver.STATE.benchmarks_payload,
            "datasets": [
                {
                    "id": "scope",
                    "name": "SCOPe 2.08 ASTRAL-40 Structural Domains",
                    "total_sequences": len(qserver.STATE.scope_entries),
                    "candidates": len(qserver.STATE.p1_cand_raw_ids),
                    "canonical_queries": len(qserver.STATE.p1_query_raw_ids),
                    "pdb_structures": 15118,
                    "source_fasta": "data/astral-scopedom-seqres-gd-sel-gs-bib-40-2.08.fa",
                },
                {
                    "id": "swissprot",
                    "name": "UniProtKB/Swiss-Prot Human Proteome",
                    "total_sequences": len(qserver.STATE.sp_entries),
                    "candidates": len(qserver.STATE.p5_cand_raw_ids),
                    "canonical_queries": len(qserver.STATE.p5_query_raw_ids),
                    "go_annotated": len(qserver.STATE.go_map),
                    "source_fasta": "data/swissprot_human_full.fasta",
                },
            ],
            "model_scales": ALL_MODEL_SCALES_META,
        })

    @rest_app.get("/api/proteins")
    def api_proteins(
        dataset: str = "scope",
        q: str = "",
        queries_only: str = "0",
        limit: int = 40,
        offset: int = 0,
    ):
        q_clean = q.strip().lower()
        q_only = str(queries_only).lower() in ("1", "true")
        lim = min(int(limit), 200)
        off = max(int(offset), 0)

        if dataset == "scope":
            source = qserver.STATE.scope_p1_queries if q_only else qserver.STATE.scope_p1_candidates
            if q_clean:
                filtered = [
                    e for e in source
                    if q_clean in e["id"].lower()
                    or q_clean in e["pdb_id"].lower()
                    or q_clean in e["sccs"].lower()
                    or q_clean in e["description"].lower()
                    or q_clean in e["organism"].lower()
                    or (len(q_clean) >= 4 and q_clean.upper() in e["sequence"])
                ]
            else:
                filtered = source
        else:
            source = qserver.STATE.sp_p5_queries if q_only else (list(qserver.STATE.sp_extra_entries.values()) + qserver.STATE.sp_entries)
            if q_clean:
                filtered = [
                    e for e in source
                    if q_clean in e["id"].lower()
                    or q_clean in e["gene"].lower()
                    or q_clean in e["entry_name"].lower()
                    or q_clean in e["description"].lower()
                    or any(q_clean in go.lower() for go in e.get("go_terms", []))
                    or (len(q_clean) >= 4 and q_clean.upper() in e["sequence"])
                ]
            else:
                filtered = source

        slice_items = [
            {k: v for k, v in item.items() if k != "sequence"}
            for item in filtered[off : off + lim]
        ]
        return JSONResponse({
            "dataset": dataset,
            "total_matches": len(filtered),
            "items": slice_items,
        })

    @rest_app.get("/api/quantize")
    def api_quantize_get(
        dataset: str = "scope",
        model_scale: str = "esm2_650m",
        protein_id: str = "d1dlwa_",
        codec_id: str = "track3_hybrid_k2_block224",
        transform_name: str = "circulant",
        custom_sequence: str = "",
    ):
        res = qserver.run_single_protein_quantization_pipeline(
            dataset=dataset,
            model_scale=model_scale,
            protein_id=protein_id,
            codec_id=codec_id,
            transform_name=transform_name,
            custom_sequence=custom_sequence or None,
        )
        return JSONResponse(res)

    @rest_app.post("/api/quantize")
    async def api_quantize_post(request: Request):
        body: Dict[str, Any] = await request.json()
        res = qserver.run_single_protein_quantization_pipeline(
            dataset=body.get("dataset", "scope"),
            model_scale=body.get("model_scale", "esm2_650m"),
            protein_id=body.get("protein_id", "d1dlwa_"),
            codec_id=body.get("codec_id", "track3_hybrid_k2_block224"),
            transform_name=body.get("transform_name", "circulant"),
            custom_sequence=body.get("custom_sequence"),
        )
        return JSONResponse(res)

    @rest_app.get("/api/search")
    def api_search(
        dataset: str = "scope",
        model_scale: str = "esm2_650m",
        query_id: str = "d1dlwa_",
        codec_id: str = "track3_hybrid_k2_block224",
        top_k: int = 30,
        fasta_sequence: str = "",
    ):
        res = qserver.run_live_search(
            dataset=dataset,
            model_scale=model_scale,
            query_id=query_id,
            codec_id=codec_id,
            top_k=min(int(top_k), 100),
            fasta_sequence=fasta_sequence or None,
        )
        return JSONResponse(res)

    @rest_app.get("/api/ablation")
    def api_ablation(
        query_id: str = "d1dlwa_",
        k: int = 4,
        model_scale: str = "esm2_650m",
    ):
        res = qserver.run_live_theorem_ablation(
            query_id=query_id,
            k=int(k),
            model_scale=model_scale,
        )
        return JSONResponse(res)

    @rest_app.get("/api/structure")
    def api_structure(id: str = "d6a1ia1", compare_id: str = ""):
        primary = _load_structure_entry_shared(id)
        if not primary:
            return JSONResponse({"error": f"3D PDB structure not found for {id}"}, status_code=404)

        comparison = _load_structure_entry_shared(compare_id) if compare_id else None
        if comparison and primary["ca_atoms"] and comparison["ca_atoms"]:
            ref_xyz = np.array([[a["cx"], a["cy"], a["cz"]] for a in primary["ca_atoms"]], dtype=np.float64)
            mov_xyz = np.array([[a["cx"], a["cy"], a["cz"]] for a in comparison["ca_atoms"]], dtype=np.float64)
            aligned_xyz, rmsd = qserver.kabsch_superpose(ref_xyz, mov_xyz)
            for i, a in enumerate(comparison["ca_atoms"]):
                a["ax"] = round(float(aligned_xyz[i, 0]), 3)
                a["ay"] = round(float(aligned_xyz[i, 1]), 3)
                a["az"] = round(float(aligned_xyz[i, 2]), 3)
            comparison["kabsch_rmsd_angstrom"] = round(rmsd, 3)

        return JSONResponse({"primary": primary, "comparison": comparison})

    @rest_app.get("/api/benchmarks")
    def api_benchmarks():
        return JSONResponse(qserver.STATE.benchmarks_payload)

    dist_dir = EXPLAINER_DIR / "dist"
    fig_dir = EXPLAINER_DIR / "public" / "figures"

    @rest_app.get("/figures/{fig_name:path}")
    def serve_figure(fig_name: str):
        for base in (dist_dir / "figures", fig_dir):
            cand = base / fig_name
            if cand.exists() and cand.is_file():
                return FileResponse(cand)
        return JSONResponse({"error": "Figure not found"}, status_code=404)

    @rest_app.get("/studio")
    @rest_app.get("/studio/{rest_of_path:path}")
    def serve_studio(rest_of_path: str = "index.html"):
        rel = rest_of_path.lstrip("/") or "index.html"
        target = dist_dir / rel
        if not target.exists() or target.is_dir():
            target = dist_dir / "index.html"
        return FileResponse(target)

    @rest_app.get("/assets/{asset_path:path}")
    def serve_studio_assets(asset_path: str):
        target = dist_dir / "assets" / asset_path
        if target.exists() and target.is_file():
            return FileResponse(target)
        return JSONResponse({"error": "Asset not found"}, status_code=404)

    return rest_app


class QuantPPIRestDispatchMiddleware:
    """Intercepts `/api/*`, `/figures/*`, `/studio*`, and `/assets/*` at the ASGI layer before Gradio's SPA router."""

    def __init__(self, app: Any, rest_app: Any) -> None:
        self.app = app
        self.rest_app = rest_app

    async def __call__(self, scope: Dict[str, Any], receive: Any, send: Any) -> None:
        if scope.get("type") == "http":
            path = scope.get("path", "")
            if (
                path.startswith("/api/")
                or path == "/api"
                or path.startswith("/figures/")
                or path == "/studio"
                or path.startswith("/studio/")
                or path.startswith("/assets/")
            ):
                await self.rest_app(scope, receive, send)
                return
        await self.app(scope, receive, send)


def install_gradio_asgi_dispatch_hook() -> None:
    """Hooks into `gradio.routes.App.create_app` BEFORE `demo.launch()` starts Uvicorn on ZeroGPU."""
    from gradio.routes import App as GradioServerApp

    if getattr(GradioServerApp, "_quantppi_hooked", False):
        return

    rest_app = build_quantppi_rest_app()
    orig_create_app = GradioServerApp.create_app

    def _patched_create_app(*args: Any, **kwargs: Any) -> Any:
        gradio_fastapi_app = orig_create_app(*args, **kwargs)
        gradio_fastapi_app.add_middleware(QuantPPIRestDispatchMiddleware, rest_app=rest_app)
        return gradio_fastapi_app

    GradioServerApp.create_app = staticmethod(_patched_create_app)
    GradioServerApp._quantppi_hooked = True


def build_gradio_blocks():
    import gradio as gr

    codec_choices = [(f"{c['name']} ({c['compression_1280']}x)", c["id"]) for c in qserver.CODEC_CATALOG]
    scale_choices = [(m["name"], m["id"]) for m in ALL_MODEL_SCALES_META]

    def gr_search(dataset: str, model_scale: str, query_id: str, fasta_seq: str, codec_id: str):
        f_clean = fasta_seq.strip()
        res = qserver.run_live_search(
            dataset=dataset,
            model_scale=model_scale,
            query_id=query_id.strip(),
            codec_id=codec_id,
            top_k=30,
            fasta_sequence=f_clean if f_clean else None,
        )
        m = res["metrics"]
        summary_md = (
            f"### Query `{res['query']['id']}` — {res['query']['description']} ({res['query']['seq_len']} aa)\n"
            f"- **Model Scale**: `{model_scale}` (`D={res['dimension']}`) · **Codec**: `{res['codec']['name']}` (`{res['codec']['actual_bytes']} B/protein`)\n"
            f"- **Recall@20@30**: **{m['recall_20_at_30']*100:.1f}%** · **Latency**: `{m['search_latency_ms']} ms`"
        )
        rows = [
            {
                "Quant Rank": h["quant_rank"],
                "FP32 Rank": h["fp32_rank"],
                "Protein ID": h["id"],
                "Fold / Gene": h.get("sccs") or h.get("gene") or "",
                "Homology / GO": h.get("homology_tier") or "",
                "Quant Score": h["quant_score"],
                "FP32 Score": h["fp32_score"],
                "Description": h["description"],
            }
            for h in res["hits"]
        ]
        return summary_md, pd.DataFrame(rows)

    with gr.Blocks(title="QuantPPI — ZeroGPU 10-PLM Protein Embedding Compression & Homology Search") as demo:
        gr.Markdown(
            "## QuantPPI — Bleed-Free Multi-Radix Quantization Across 10 PLMs (ESM-2, ESM-C & ESM-3)\n"
            "Live ZeroGPU + EvolutionaryScale Forge cloud backend powering all 10 PLMs (`ESM-2 8M–15B`, `ESM-C 300M–6B`, `ESM-3 1.4B`), "
            "the 108-codec QuantPPI registry, arbitrary raw FASTA sequence embedding extraction, and 3D PyMOL Kabsch superposition."
        )
        with gr.Tabs():
            with gr.TabItem("Interactive Web Showcase (/studio)"):
                gr.HTML(
                    '<iframe src="/studio" style="width:100%; height:1020px; border:1px solid #cbd5e1; border-radius:12px; background:#ffffff;" allow="fullscreen"></iframe>'
                )
            with gr.TabItem("ZeroGPU Direct Homology & FASTA Search"):
                with gr.Row():
                    ds_in = gr.Dropdown(
                        choices=[("SCOPe ASTRAL-40 (15,166)", "scope"), ("Swiss-Prot Human (18,102)", "swissprot")],
                        value="scope",
                        label="Corpus Database",
                    )
                    scale_in = gr.Dropdown(
                        choices=scale_choices,
                        value="esm2_650m",
                        label="Model Scale (10 PLMs)",
                    )
                    codec_in = gr.Dropdown(
                        choices=codec_choices,
                        value="track3_hybrid_k2_block224",
                        label="QuantPPI Codec",
                    )
                with gr.Row():
                    qid_in = gr.Textbox(value="d1dlwa_", label="Catalog Query ID (ignored if FASTA provided)")
                    fasta_in = gr.Textbox(
                        value="",
                        lines=2,
                        placeholder="Optional: paste raw FASTA sequence here to run live ZeroGPU / Forge inference...",
                        label="Raw FASTA Sequence (Optional)",
                    )
                run_btn = gr.Button("Execute Live QuantPPI Search", variant="primary")
                out_md = gr.Markdown()
                out_df = gr.Dataframe(interactive=False)
                run_btn.click(gr_search, inputs=[ds_in, scale_in, qid_in, fasta_in, codec_in], outputs=[out_md, out_df])

    return demo


if __name__ == "__main__":
    install_gradio_asgi_dispatch_hook()
    port = int(os.environ.get("GRADIO_SERVER_PORT", os.environ.get("PORT", "7860")))
    host = os.environ.get("GRADIO_SERVER_NAME", os.environ.get("HOST", "0.0.0.0"))
    demo = build_gradio_blocks()
    demo.queue()
    print(f"[QuantPPI Space] Launching ZeroGPU Gradio + REST /api/* on http://{host}:{port}")
    demo.launch(
        server_name=host,
        server_port=port,
    )
elif os.environ.get("SPACE_ID"):
    install_gradio_asgi_dispatch_hook()
    demo = build_gradio_blocks()