shanto268 commited on
Commit
f76c3bf
·
verified ·
1 Parent(s): 142609e

v4 flagship Space (3 tiers) from registry v2 (earlier contents: 906bc45ae05c)

Browse files
.dockerignore CHANGED
@@ -1,2 +1,3 @@
1
  numba_cache
2
  __pycache__
 
 
1
  numba_cache
2
  __pycache__
3
+ bake
Dockerfile CHANGED
@@ -3,10 +3,10 @@ FROM python:3.12-slim
3
  RUN useradd -m -u 1000 user
4
  USER user
5
  ENV SYSTEM=spaces HOME=/home/user PATH=/home/user/.local/bin:$PATH NUMBA_CPU_NAME=generic NUMBA_CACHE_DIR=/home/user/app/numba_cache \
6
- GRADIO_SERVER_NAME=0.0.0.0 PORT=7860 PYTHONUNBUFFERED=1
7
  WORKDIR /home/user/app
8
  COPY --chown=user requirements.txt .
9
- RUN pip install --no-cache-dir --user -r requirements.txt "gradio==6.29.1"
10
  COPY --chown=user . .
11
  # OAuth Spaces (hf_oauth): Gradio builds real login routes only when SYSTEM=spaces (set above: HF does not set it in Docker
12
  # Spaces, but it does pass SPACE_ID and the OAUTH_* variables at run time). At build time those are absent, so warm.py gets
 
3
  RUN useradd -m -u 1000 user
4
  USER user
5
  ENV SYSTEM=spaces HOME=/home/user PATH=/home/user/.local/bin:$PATH NUMBA_CPU_NAME=generic NUMBA_CACHE_DIR=/home/user/app/numba_cache \
6
+ GRADIO_SERVER_NAME=0.0.0.0 PORT=7860 PYTHONUNBUFFERED=1 OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 MKL_NUM_THREADS=2
7
  WORKDIR /home/user/app
8
  COPY --chown=user requirements.txt .
9
+ RUN pip install --no-cache-dir --user -r requirements.txt "gradio==6.29.1" && pip freeze > IMAGE_FREEZE.txt
10
  COPY --chown=user . .
11
  # OAuth Spaces (hf_oauth): Gradio builds real login routes only when SYSTEM=spaces (set above: HF does not set it in Docker
12
  # Spaces, but it does pass SPACE_ID and the OAUTH_* variables at run time). At build time those are absent, so warm.py gets
app.py CHANGED
@@ -5,6 +5,8 @@ Balanced, Fast) retrained on all on-stack data, through geomgnn.flagship. Names,
5
  import os
6
  import platform
7
  import sys
 
 
8
  from pathlib import Path
9
 
10
  if platform.machine() in ("x86_64", "AMD64"): # OpenBLAS SkylakeX Cholesky kernel segfaulted on a v3 MoM system (LIMITS.md)
@@ -17,6 +19,7 @@ os.environ.setdefault("TLSQ_RESPONSE_CACHE", "64")
17
  os.environ.setdefault("TLSQ_RESPONSE_CACHE_MB", "1024")
18
  os.environ.setdefault("TLSQ_EMBED_CACHE", os.path.join(os.environ.get("TMPDIR", "/tmp"), "tlsq_embed_cache"))
19
  os.environ.setdefault("TLSQ_EMBED_CACHE_MB", "1024")
 
20
  os.environ.pop("TLSQ_GNN_PRIOR_MODE", None) # every tier names its own solver mode (registry flagship_hub.prior_stacks)
21
 
22
  HERE = Path(__file__).resolve().parent
@@ -35,6 +38,7 @@ from tlsq_infer.registry import load_registry # noqa: E402
35
 
36
  os.environ.setdefault("TLSQ_NUM_THREADS", str(set_num_threads()))
37
 
 
38
  from geomgnn.app_flagship import build_app # noqa: E402
39
 
40
  reg = load_registry()
@@ -49,12 +53,41 @@ if not ex.exists(): # repository checkout: the bundle's examples (sync.py --fla
49
  f = HERE.parents[2] / "data_local" / "release_core" / "gds" / "ref" / "grid" / src
50
  if f.exists():
51
  shutil.copy(f, ex / name)
52
- if os.environ.get("TLSQ_PRELOAD", "1") != "0": # load every tier (SHA checks) + numba warmup + one small prediction before the queue opens
53
- from geomgnn import flagship as FS # noqa: E402
 
 
 
 
 
54
 
55
  FS.preload(reg, example=ex / "interdigital.gds")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
56
  demo = build_app(reg["flagship_hub"]["title"], examples_dir=ex)
57
  demo.queue(default_concurrency_limit=int(os.environ.get("TLSQ_CONCURRENCY", 1)), max_size=32)
 
 
 
 
 
 
58
 
59
  if __name__ == "__main__":
60
  demo.launch(server_name=os.environ.get("GRADIO_SERVER_NAME", "0.0.0.0"), server_port=int(os.environ.get("PORT", 7860)))
 
5
  import os
6
  import platform
7
  import sys
8
+ import threading
9
+ import traceback
10
  from pathlib import Path
11
 
12
  if platform.machine() in ("x86_64", "AMD64"): # OpenBLAS SkylakeX Cholesky kernel segfaulted on a v3 MoM system (LIMITS.md)
 
19
  os.environ.setdefault("TLSQ_RESPONSE_CACHE_MB", "1024")
20
  os.environ.setdefault("TLSQ_EMBED_CACHE", os.path.join(os.environ.get("TMPDIR", "/tmp"), "tlsq_embed_cache"))
21
  os.environ.setdefault("TLSQ_EMBED_CACHE_MB", "1024")
22
+ os.environ.setdefault("TLSQ_PRIOR_LEVEL_CACHE_MB", "256") # solved prior levels (Accurate and Balanced share 1.5) and meshes (substrate changes)
23
  os.environ.pop("TLSQ_GNN_PRIOR_MODE", None) # every tier names its own solver mode (registry flagship_hub.prior_stacks)
24
 
25
  HERE = Path(__file__).resolve().parent
 
38
 
39
  os.environ.setdefault("TLSQ_NUM_THREADS", str(set_num_threads()))
40
 
41
+ from geomgnn import app_flagship as G # noqa: E402
42
  from geomgnn.app_flagship import build_app # noqa: E402
43
 
44
  reg = load_registry()
 
53
  f = HERE.parents[2] / "data_local" / "release_core" / "gds" / "ref" / "grid" / src
54
  if f.exists():
55
  shutil.copy(f, ex / name)
56
+ BAKE = Path(os.environ.get("TLSQ_RESPONSE_BAKE", HERE / "bake" / "respcache.pkl")) # the examples' results, baked at image build
57
+
58
+
59
+ def preload():
60
+ """Load every tier (SHA checks) + numba warmup + one small prediction; then the baked example results, if they verify here."""
61
+ from geomgnn import flagship as FS
62
+ from geomgnn import paper as PA
63
 
64
  FS.preload(reg, example=ex / "interdigital.gds")
65
+ if BAKE.exists() and int(os.environ.get("TLSQ_RESPONSE_CACHE", "0")) > 0:
66
+ PA.STARTUP["bake"] = G.load_bake(BAKE, ex)
67
+
68
+
69
+ def _preload_background():
70
+ try:
71
+ preload()
72
+ except BaseException: # noqa: BLE001 (a Space that cannot load its models must not serve)
73
+ traceback.print_exc()
74
+ os._exit(1)
75
+ G.READY.set()
76
+
77
+
78
+ def bake():
79
+ """Image build (warm.py, after the import's preload): every example x tier into the response cache, pickled to BAKE."""
80
+ return G.bake(ex, BAKE)
81
+
82
+
83
  demo = build_app(reg["flagship_hub"]["title"], examples_dir=ex)
84
  demo.queue(default_concurrency_limit=int(os.environ.get("TLSQ_CONCURRENCY", 1)), max_size=32)
85
+ if os.environ.get("TLSQ_PRELOAD", "1") != "0":
86
+ if __name__ == "__main__": # the server: the page is up while the tiers load; predictions wait for them (G.READY)
87
+ G.READY.clear()
88
+ threading.Thread(target=_preload_background, name="preload", daemon=True).start()
89
+ else: # imported (warm.py, tests): load before returning
90
+ preload()
91
 
92
  if __name__ == "__main__":
93
  demo.launch(server_name=os.environ.get("GRADIO_SERVER_NAME", "0.0.0.0"), server_port=int(os.environ.get("PORT", 7860)))
geomgnn/SOURCES.json CHANGED
@@ -21,10 +21,13 @@
21
  "rewrites": []
22
  },
23
  "_tlsq/embed/proxy.py": {
24
- "source": "/Users/shanto/Documents/ChatGPT/transfer_learning/geomfast_stage/vendor/tlsq/embed/proxy.py",
25
- "source_sha256": "78d8dffd2fac75632a91ebee151247afe4157716ffb11e624e6b2932c618c86e",
26
- "sha256": "78d8dffd2fac75632a91ebee151247afe4157716ffb11e624e6b2932c618c86e",
27
- "rewrites": []
 
 
 
28
  },
29
  "_tlsq/embed/fastprior.py": {
30
  "source": "/Users/shanto/Documents/ChatGPT/transfer_learning/geomfast_stage/vendor/tlsq/embed/fastprior.py",
@@ -162,8 +165,8 @@
162
  },
163
  "fastprior_pkg.py": {
164
  "source": "/Users/shanto/LFL/tl_squadds/scripts/fastprior_pkg/fastprior_pkg.py",
165
- "source_sha256": "956f2b62c308ec1561bbc033716b52c24dc63a916b9373cc44090eb6514e5dde",
166
- "sha256": "956f2b62c308ec1561bbc033716b52c24dc63a916b9373cc44090eb6514e5dde",
167
  "rewrites": [],
168
  "note": "gated fast-solve package (agents/fastprior_pkg/RESULTS.md), verbatim; bound with bind('geomgnn'); vendored 2026-10-17"
169
  },
 
21
  "rewrites": []
22
  },
23
  "_tlsq/embed/proxy.py": {
24
+ "source": "/Users/shanto/LFL/tl_squadds/src/tlsq/embed/proxy.py",
25
+ "source_sha256": "f87a0f1614ca93b73aac0747d0acdafbd9b8c243f69c68c98d8919e158a57b28",
26
+ "sha256": "f87a0f1614ca93b73aac0747d0acdafbd9b8c243f69c68c98d8919e158a57b28",
27
+ "rewrites": [],
28
+ "note": "phase B speed-up (2026-11-15): the training-source copy src/tlsq/embed/proxy.py (byte-identical); row-blocked image table, bounded image-table cache, compiled bitwise interpolation (np.interp fallback); every prior bit-identical to the frozen stage file",
29
+ "frozen_stage_source": "/Users/shanto/Documents/ChatGPT/transfer_learning/geomfast_stage/vendor/tlsq/embed/proxy.py",
30
+ "frozen_stage_sha256": "78d8dffd2fac75632a91ebee151247afe4157716ffb11e624e6b2932c618c86e"
31
  },
32
  "_tlsq/embed/fastprior.py": {
33
  "source": "/Users/shanto/Documents/ChatGPT/transfer_learning/geomfast_stage/vendor/tlsq/embed/fastprior.py",
 
165
  },
166
  "fastprior_pkg.py": {
167
  "source": "/Users/shanto/LFL/tl_squadds/scripts/fastprior_pkg/fastprior_pkg.py",
168
+ "source_sha256": "4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a",
169
+ "sha256": "4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a",
170
  "rewrites": [],
171
  "note": "gated fast-solve package (agents/fastprior_pkg/RESULTS.md), verbatim; bound with bind('geomgnn'); vendored 2026-10-17"
172
  },
geomgnn/_tlsq/embed/proxy.py CHANGED
@@ -17,6 +17,7 @@ from __future__ import annotations
17
  import os
18
  import queue
19
  import threading
 
20
  from concurrent.futures import ThreadPoolExecutor
21
  from functools import partial
22
 
@@ -440,8 +441,97 @@ def _near_pairs(c, hm):
440
 
441
 
442
  SUBSTRATE = {"model": "slab_open", "t_um": 350.0} # MoM Green's function (PRIOR_SCHEME.md 14.6): "slab_open" (default since 2026-10-10: 350 um substrate, vacuum below), "halfspace" (released models' priors), "slab_pec"
443
- _IMG_CACHE: dict = {}
444
  _IMG_GRID: dict = {}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
445
 
446
 
447
  class substrate:
@@ -474,7 +564,9 @@ def _image_correction(d, eps_r: float, exact: bool = False):
474
  if m == "halfspace":
475
  return None
476
  key = (m, float(SUBSTRATE["t_um"]), float(eps_r))
477
- if key not in _IMG_CACHE:
 
 
478
  K, t = (eps_r - 1) / (eps_r + 1), float(SUBSTRATE["t_um"])
479
  n = np.arange(1, 400)
480
  nmax = int(np.argmax(K ** (2 * n) < 1e-4)) + 1
@@ -486,13 +578,20 @@ def _image_correction(d, eps_r: float, exact: bool = False):
486
  w, D = np.r_[w, total - w.sum()], np.r_[2 * n * t, 2 * (nmax + 1) * t]
487
  du = t / 1000
488
  u = np.arange(0.0, 1e5 + 2 * du, du)
489
- T = (w[None, :] / np.sqrt(u[:, None] ** 2 + D[None, :] ** 2)).sum(1)
 
 
 
490
  _IMG_CACHE[key] = (du, T, w, D)
 
491
  du, T = _IMG_CACHE[key][:2]
492
  if exact:
493
  w, D = _IMG_CACHE[key][2:]
494
  d2 = np.asarray(d, float) ** 2
495
  return sum(wn / np.sqrt(d2 + Dn * Dn) for wn, Dn in zip(w, D))
 
 
 
496
  if key not in _IMG_GRID:
497
  _IMG_GRID[key] = du * np.arange(len(T))
498
  return np.interp(d, _IMG_GRID[key], T)
 
17
  import os
18
  import queue
19
  import threading
20
+ from collections import OrderedDict
21
  from concurrent.futures import ThreadPoolExecutor
22
  from functools import partial
23
 
 
441
 
442
 
443
  SUBSTRATE = {"model": "slab_open", "t_um": 350.0} # MoM Green's function (PRIOR_SCHEME.md 14.6): "slab_open" (default since 2026-10-10: 350 um substrate, vacuum below), "halfspace" (released models' priors), "slab_pec"
444
+ _IMG_CACHE: OrderedDict = OrderedDict() # LRU of image tables (model, t_um, eps_r) -> (du, T, w, D)
445
  _IMG_GRID: dict = {}
446
+ _IMG_KEEP = (4, 512 << 20) # at most 4 tables and 512 MB (the newest is always kept)
447
+ _INTERP: dict = {}
448
+
449
+
450
+ def _img_evict():
451
+ while len(_IMG_CACHE) > 1 and (len(_IMG_CACHE) > _IMG_KEEP[0] or sum(v[1].nbytes for v in _IMG_CACHE.values())
452
+ + sum(g.nbytes for g in _IMG_GRID.values()) > _IMG_KEEP[1]):
453
+ _IMG_GRID.pop(_IMG_CACHE.popitem(last=False)[0], None)
454
+
455
+
456
+ try: # compiled np.interp on the uniform image table (_interp_uniform)
457
+ import numba as _nb
458
+ from numba.core import types as _nbt
459
+ from numba.extending import intrinsic as _intrinsic
460
+
461
+ @_intrinsic
462
+ def _fma(typingctx, a, b, c):
463
+ def gen(context, builder, sig, args):
464
+ return builder.fma(*args)
465
+ return _nbt.float64(_nbt.float64, _nbt.float64, _nbt.float64), gen
466
+
467
+ @_nb.njit(nogil=True, cache=True)
468
+ def _interp_run(x, du, fp, out, fused):
469
+ """numpy's arr_interp on xp_j = du * j: the index by division, corrected with numpy's comparisons, then numpy's slope
470
+ and slope * (x - xp_j) + fp_j, plain or fused (fma) as this numpy build's C compiler does it."""
471
+ n = fp.shape[0]
472
+ last = du * (n - 1)
473
+ for i in range(x.shape[0]):
474
+ v = x[i]
475
+ if v != v:
476
+ out[i] = v
477
+ continue
478
+ if v > last:
479
+ out[i] = fp[n - 1]
480
+ continue
481
+ if v < 0.0:
482
+ out[i] = fp[0]
483
+ continue
484
+ j = min(int(v / du), n - 1)
485
+ while j > 0 and du * j > v:
486
+ j -= 1
487
+ while j < n - 1 and du * (j + 1) <= v:
488
+ j += 1
489
+ xj = du * j
490
+ if j == n - 1 or xj == v:
491
+ out[i] = fp[j]
492
+ continue
493
+ x1 = du * (j + 1)
494
+ s = (fp[j + 1] - fp[j]) / (x1 - xj)
495
+ r = _fma(s, v - xj, fp[j]) if fused else s * (v - xj) + fp[j]
496
+ if r != r:
497
+ r = _fma(s, v - x1, fp[j + 1]) if fused else s * (v - x1) + fp[j + 1]
498
+ if r != r and fp[j] == fp[j + 1]:
499
+ r = fp[j]
500
+ out[i] = r
501
+ except ImportError:
502
+ _interp_run = None
503
+
504
+
505
+ def _interp_uniform(d, du, T):
506
+ """np.interp(d, du * arange(len(T)), T), bit for bit: the compiled kernel (plain or fused multiply-add, whichever matches this
507
+ numpy build) when a probe at first use matches np.interp bit for bit, else None (the caller uses np.interp)."""
508
+ if "fn" not in _INTERP:
509
+ _INTERP["fn"] = None
510
+ if _interp_run is not None:
511
+ try:
512
+ rng, ok = np.random.default_rng(0), []
513
+ for fp in (rng.normal(size=4096), T): # a rough table tells the two apart, then the real one
514
+ xp = du * np.arange(len(fp))
515
+ x = np.concatenate([rng.uniform(-1, 1.01 * xp[-1], 50000), rng.uniform(0, 64 * du, 20000), xp[rng.integers(0, len(fp), 2000)],
516
+ np.nextafter(xp[rng.integers(1, len(fp), 2000)], 0), xp[[0, -1]], [np.nan, np.inf, 1e300]])
517
+ ref, got = np.interp(x, xp, fp).tobytes(), []
518
+ for fused in (False, True):
519
+ y = np.empty_like(x)
520
+ _interp_run(x, du, fp, y, fused)
521
+ got.append(y.tobytes())
522
+ ok.append([g == ref for g in got])
523
+ if ok[0][0] != ok[0][1] and ok[1][ok[0][1]]:
524
+ _INTERP["fn"] = ok[0][1]
525
+ except Exception: # noqa: BLE001 (llvm.fma or the compile unavailable: np.interp)
526
+ pass
527
+ fused = _INTERP["fn"]
528
+ d = np.asarray(d, float)
529
+ if fused is None or d.ndim == 0:
530
+ return None
531
+ x = np.ascontiguousarray(d)
532
+ out = np.empty(x.shape)
533
+ _interp_run(x.reshape(-1), du, T, out.reshape(-1), fused)
534
+ return out
535
 
536
 
537
  class substrate:
 
564
  if m == "halfspace":
565
  return None
566
  key = (m, float(SUBSTRATE["t_um"]), float(eps_r))
567
+ if key in _IMG_CACHE:
568
+ _IMG_CACHE.move_to_end(key)
569
+ else:
570
  K, t = (eps_r - 1) / (eps_r + 1), float(SUBSTRATE["t_um"])
571
  n = np.arange(1, 400)
572
  nmax = int(np.argmax(K ** (2 * n) < 1e-4)) + 1
 
578
  w, D = np.r_[w, total - w.sum()], np.r_[2 * n * t, 2 * (nmax + 1) * t]
579
  du = t / 1000
580
  u = np.arange(0.0, 1e5 + 2 * du, du)
581
+ T = np.empty(len(u))
582
+ for i in range(0, len(u), 1 << 18): # row blocks: the same sums, without len(u) x images temporaries (~8 GB at 10 um)
583
+ ub = u[i:i + (1 << 18)]
584
+ T[i:i + len(ub)] = (w[None, :] / np.sqrt(ub[:, None] ** 2 + D[None, :] ** 2)).sum(1)
585
  _IMG_CACHE[key] = (du, T, w, D)
586
+ _img_evict()
587
  du, T = _IMG_CACHE[key][:2]
588
  if exact:
589
  w, D = _IMG_CACHE[key][2:]
590
  d2 = np.asarray(d, float) ** 2
591
  return sum(wn / np.sqrt(d2 + Dn * Dn) for wn, Dn in zip(w, D))
592
+ out = _interp_uniform(d, du, T)
593
+ if out is not None:
594
+ return out
595
  if key not in _IMG_GRID:
596
  _IMG_GRID[key] = du * np.arange(len(T))
597
  return np.interp(d, _IMG_GRID[key], T)
geomgnn/app_flagship.py CHANGED
@@ -20,6 +20,7 @@ from __future__ import annotations
20
  import functools
21
  import html
22
  import json
 
23
  import time
24
  from pathlib import Path
25
 
@@ -34,12 +35,18 @@ from tlsq_infer.registry import load_registry
34
  from . import __version__
35
  from . import access as AC
36
  from . import api as A
 
37
  from . import ground_ui as GU
38
  from . import flagship as FS
39
  from . import ui as UI
40
  from . import usage as U
41
 
 
42
  ORDER = ["accurate", "balanced", "fast"]
 
 
 
 
43
 
44
 
45
  def served(reg=None) -> list[str]:
@@ -74,10 +81,25 @@ def expected(reg, k) -> str | None:
74
  return f"Expect an error of {UI.pct(b5)} on a new design of a trained family. We measured this on 39 layouts the model never saw." if b5 else None
75
 
76
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77
  def _predict(lay, model, substrate, thickness_um, w=None, own=False):
78
  from . import respcache
 
79
  reg = load_registry()
80
  name = resolve(reg, model)
 
81
  AC.check_size(lay.n_conductors)
82
  fn = lambda: FS.predict(lay, name, reg, substrate=substrate, thickness_um=thickness_um) # noqa: E731
83
  if lay.n_conductors > 8:
@@ -87,10 +109,10 @@ def _predict(lay, model, substrate, thickness_um, w=None, own=False):
87
  with HEAVY:
88
  return fn0()
89
  w = w or AC.Who()
90
- settings = ("flagship", name, str(substrate), float(thickness_um), reg["version"])
91
- if not respcache.has(lay, settings):
92
  AC.QUOTA.check(w, own)
93
- res = respcache.cached(lay, settings, fn)
94
  if not res["timing_s"].get("response_cache_hit"):
95
  AC.QUOTA.charge(w, own)
96
  return res
@@ -187,19 +209,65 @@ def inspect_gfn(f, profile: gr.OAuthProfile | None = None, oauth_token: gr.OAuth
187
  return inspect_fn(f)
188
 
189
 
190
- def api_predict_gds(f, ground_layers="1/0", conductor_layers="1/10", margin_gaps=0.0, substrate="Si (cryogenic, training)",
 
 
 
 
 
 
191
  custom_eps=11.45, thickness_um=350.0, model="accurate", ground_pieces="", profile: gr.OAuthProfile | None = None,
192
  oauth_token: gr.OAuthToken | None = None, request: gr.Request = None):
193
  t = time.perf_counter()
194
- lay, res = _served(AC.who(profile, request, oauth_token), f, lambda: prepare(
195
- _path(f), ground_layers=_parse_layers(ground_layers), conductor_layers=_parse_layers(conductor_layers),
196
- margin_gaps=_margin_api(margin_gaps), max_conductors=LIMITS["max_conductors"], ground_pieces=str(ground_pieces or "").strip() or None),
197
- model, _eps(substrate, custom_eps), float(thickness_um))
198
  res["timing_s"]["prepare_and_total_s"] = time.perf_counter() - t
199
  res["ground_detection"] = lay.info.get("ground_detection")
200
  return res
201
 
202
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
203
  def models_fn():
204
  reg = load_registry()
205
  h = reg["flagship_hub"]
@@ -213,7 +281,19 @@ def models_fn():
213
  from . import paper as PA
214
  return out + [{"geomgnn_version": __version__, "tlsq_infer_version": TLSQ_VERSION, "registry_version": reg["version"],
215
  "training_layouts": h.get("training_layouts"),
216
- "startup": dict(PA.STARTUP)}]
 
 
 
 
 
 
 
 
 
 
 
 
217
 
218
 
219
  def limits() -> dict:
@@ -530,27 +610,29 @@ def build_app(title: str, examples_dir=None):
530
  ui_predict = functools.wraps(predict_fn)(UI.predict_ui(predict_fn, lambda r: expected(reg, r["tier"]), note=True))
531
  ui_predict.__name__ = "predict_fn_ui"
532
  uouts = pouts + [status, result_box, mat_note]
533
- hidden = dict(api_visibility="private", show_progress="hidden")
 
 
534
  substrate.change(lambda s: gr.update(visible=s == CUSTOM), substrate, custom_eps, **hidden)
535
  for ev in (substrate.change, custom_eps.change, thickness.change):
536
  ev(lambda s, e, t: stack_note(reg, s, e, t), [substrate, custom_eps, thickness], sub_note, **hidden)
537
  crop.change(lambda c: gr.update(visible=bool(c)), crop, crop_row, **hidden)
538
  up = gds.change(upload_fn, [gds, view], [layers, layer_acc, layer_help, upload_msg, pick, role, img, svg_btn, viewbar, vst, view_msg],
539
- api_visibility="private")
540
- view.input(show_view, vin, vout, api_visibility="private")
541
- i_btn.click(_api(inspect_gfn, 2), gds, [layers, layer_plot], api_name="inspect")
542
  edit = UI.wire_editor(layers, pick, role)
543
  edit.then(lambda t: gr.update(label=_roles_label(t)), layers, layer_acc, **hidden)
544
  r_btn.click(_api(prepare_gfn, 4), [gds, layers, crop, margin, margin_um, select, gst], [prep_plot, warn_md, summary, prepared_gds],
545
- api_name="prepare")
546
  after = lambda ev: ev.then(refresh_view, vin, vout, **hidden) # noqa: E731
547
- GU.wire(gds, layers, pred_btn, gparts, [up, edit], after=after).then(ui_predict, pins, uouts, api_visibility="private")
548
  for ev in (crop.change, margin.change, margin_um.change, select.change):
549
  ev(preview_fn, vin, [view] + vout, trigger_mode="always_last", **hidden)
550
- pred_btn.click(ui_predict, pins, uouts, api_visibility="private")
551
- p_btn.click(_api(predict_fn, 5), pins, pouts, api_name="predict")
552
  UI.mark_stale(res_state, status, mat, report, [gds.change, model.change, substrate.change, custom_eps.change, thickness.change,
553
  select.change, crop.change, margin.change, margin_um.change, edit.then])
554
- a_btn.click(_api(api_predict_gds, 1), [gds, a_g, a_c, a_m, substrate, custom_eps, thickness, model, a_gp], a_out, api_name="predict_gds")
555
- m_btn.click(models_fn, None, m_out, api_name="models")
556
  return demo
 
20
  import functools
21
  import html
22
  import json
23
+ import threading
24
  import time
25
  from pathlib import Path
26
 
 
35
  from . import __version__
36
  from . import access as AC
37
  from . import api as A
38
+ from . import bundle
39
  from . import ground_ui as GU
40
  from . import flagship as FS
41
  from . import ui as UI
42
  from . import usage as U
43
 
44
+ bundle.install()
45
  ORDER = ["accurate", "balanced", "fast"]
46
+ MIN_THICKNESS_UM = 1.0
47
+ SI = "Si (cryogenic, training)"
48
+ READY = threading.Event() # cleared by the Space entry while the tiers load in the background; every predict waits for it
49
+ READY.set()
50
 
51
 
52
  def served(reg=None) -> list[str]:
 
81
  return f"Expect an error of {UI.pct(b5)} on a new design of a trained family. We measured this on 39 layouts the model never saw." if b5 else None
82
 
83
 
84
+ def check_thickness(thickness_um) -> float:
85
+ t = float(thickness_um)
86
+ if not t >= MIN_THICKNESS_UM:
87
+ raise ValueError(f"The substrate must be at least {MIN_THICKNESS_UM:g} µm thick (got {thickness_um!r} µm). The models are "
88
+ "tested on substrates of 10-1000 µm.")
89
+ return t
90
+
91
+
92
+ def settings(reg, name, substrate, thickness_um) -> tuple:
93
+ """The response-cache settings of a request (with the prepared layout, the cache key)."""
94
+ return ("flagship", name, str(substrate), float(thickness_um), reg["version"])
95
+
96
+
97
  def _predict(lay, model, substrate, thickness_um, w=None, own=False):
98
  from . import respcache
99
+ READY.wait()
100
  reg = load_registry()
101
  name = resolve(reg, model)
102
+ check_thickness(thickness_um)
103
  AC.check_size(lay.n_conductors)
104
  fn = lambda: FS.predict(lay, name, reg, substrate=substrate, thickness_um=thickness_um) # noqa: E731
105
  if lay.n_conductors > 8:
 
109
  with HEAVY:
110
  return fn0()
111
  w = w or AC.Who()
112
+ st = settings(reg, name, substrate, thickness_um)
113
+ if not respcache.has(lay, st):
114
  AC.QUOTA.check(w, own)
115
+ res = respcache.cached(lay, st, fn)
116
  if not res["timing_s"].get("response_cache_hit"):
117
  AC.QUOTA.charge(w, own)
118
  return res
 
209
  return inspect_fn(f)
210
 
211
 
212
+ def _api_layout(f, ground_layers="1/0", conductor_layers="1/10", margin_gaps=0.0, ground_pieces=""):
213
+ return prepare(_path(f), ground_layers=_parse_layers(ground_layers), conductor_layers=_parse_layers(conductor_layers),
214
+ margin_gaps=_margin_api(margin_gaps), max_conductors=LIMITS["max_conductors"],
215
+ ground_pieces=str(ground_pieces or "").strip() or None)
216
+
217
+
218
+ def api_predict_gds(f, ground_layers="1/0", conductor_layers="1/10", margin_gaps=0.0, substrate=SI,
219
  custom_eps=11.45, thickness_um=350.0, model="accurate", ground_pieces="", profile: gr.OAuthProfile | None = None,
220
  oauth_token: gr.OAuthToken | None = None, request: gr.Request = None):
221
  t = time.perf_counter()
222
+ lay, res = _served(AC.who(profile, request, oauth_token), f, lambda: _api_layout(f, ground_layers, conductor_layers, margin_gaps,
223
+ ground_pieces),
224
+ model, _eps(substrate, custom_eps), float(thickness_um))
 
225
  res["timing_s"]["prepare_and_total_s"] = time.perf_counter() - t
226
  res["ground_detection"] = lay.info.get("ground_detection")
227
  return res
228
 
229
 
230
+ def bake(examples_dir, path) -> dict:
231
+ """Image build: every bundled example x tier on the training stack, through /predict_gds and the UI Predict path (an example click),
232
+ into the response cache, pickled to path. Both paths prepare the same layout, so they share one entry per example and tier."""
233
+ from . import respcache
234
+ reg = load_registry()
235
+ t = time.perf_counter()
236
+ for f in sorted(Path(examples_dir).glob("*.gds")):
237
+ table = inspect_fn(str(f))[0]
238
+ gstate = GU.check_fn(str(f), table)[4]
239
+ for tier in served(reg):
240
+ api_predict_gds(str(f), model=tier)
241
+ predict_fn(str(f), table, False, 0, None, "", SI, 11.45, 350, tier, gstate)
242
+ out = {"entries": respcache.dump(path), "s": round(time.perf_counter() - t, 1)}
243
+ print("bake " + json.dumps(out), flush=True)
244
+ return out
245
+
246
+
247
+ def _same(a: dict, b: dict) -> bool:
248
+ return {k: v for k, v in a.items() if k != "timing_s"} == {k: v for k, v in b.items() if k != "timing_s"}
249
+
250
+
251
+ def load_bake(path, examples_dir) -> dict:
252
+ """Server start: the baked example results join the response cache only if this machine recomputes the interdigital example
253
+ bit for bit on every tier (same packages, CPU code paths and threads as the image build); otherwise they are dropped."""
254
+ from . import respcache
255
+ reg = load_registry()
256
+ t = time.perf_counter()
257
+ items = respcache.load_items(path)
258
+ baked = dict(items)
259
+ for tier in served(reg):
260
+ lay = _api_layout(str(Path(examples_dir) / "interdigital.gds"))
261
+ old = baked.get(respcache.key(lay, *settings(reg, tier, SI, 350.0)), (None,))[0]
262
+ if old is None or not _same(old, FS.predict(lay, tier, reg, substrate=SI, thickness_um=350.0)):
263
+ out = {"bake": "dropped", "reason": f"interdigital / {tier} differs from the baked result", "s": time.perf_counter() - t}
264
+ break
265
+ else:
266
+ out = {"bake": "loaded", "entries": len(items), "cache_entries": respcache.put_many(items), "s": time.perf_counter() - t}
267
+ print("bake " + json.dumps(out, default=float), flush=True)
268
+ return out
269
+
270
+
271
  def models_fn():
272
  reg = load_registry()
273
  h = reg["flagship_hub"]
 
281
  from . import paper as PA
282
  return out + [{"geomgnn_version": __version__, "tlsq_infer_version": TLSQ_VERSION, "registry_version": reg["version"],
283
  "training_layouts": h.get("training_layouts"),
284
+ "startup": dict(PA.STARTUP), "packages": packages()}]
285
+
286
+
287
+ def packages() -> dict:
288
+ """Versions of the main packages (the image's full list: IMAGE_FREEZE.txt next to app.py, Docker variant)."""
289
+ from importlib.metadata import PackageNotFoundError, version
290
+ out = {}
291
+ for p in ("torch", "numpy", "scipy", "numba", "shapely", "gdstk", "gradio"):
292
+ try:
293
+ out[p] = version(p)
294
+ except PackageNotFoundError:
295
+ out[p] = None
296
+ return out
297
 
298
 
299
  def limits() -> dict:
 
610
  ui_predict = functools.wraps(predict_fn)(UI.predict_ui(predict_fn, lambda r: expected(reg, r["tier"]), note=True))
611
  ui_predict.__name__ = "predict_fn_ui"
612
  uouts = pouts + [status, result_box, mat_note]
613
+ light = dict(concurrency_id="light", concurrency_limit=1) # previews and file handling: never queued behind a solve
614
+ solve = dict(concurrency_id="predict", concurrency_limit=1) # every predict path: one at a time (process-global prior stack)
615
+ hidden = dict(api_visibility="private", show_progress="hidden", **light)
616
  substrate.change(lambda s: gr.update(visible=s == CUSTOM), substrate, custom_eps, **hidden)
617
  for ev in (substrate.change, custom_eps.change, thickness.change):
618
  ev(lambda s, e, t: stack_note(reg, s, e, t), [substrate, custom_eps, thickness], sub_note, **hidden)
619
  crop.change(lambda c: gr.update(visible=bool(c)), crop, crop_row, **hidden)
620
  up = gds.change(upload_fn, [gds, view], [layers, layer_acc, layer_help, upload_msg, pick, role, img, svg_btn, viewbar, vst, view_msg],
621
+ api_visibility="private", **light)
622
+ view.input(show_view, vin, vout, api_visibility="private", **light)
623
+ i_btn.click(_api(inspect_gfn, 2), gds, [layers, layer_plot], api_name="inspect", **light)
624
  edit = UI.wire_editor(layers, pick, role)
625
  edit.then(lambda t: gr.update(label=_roles_label(t)), layers, layer_acc, **hidden)
626
  r_btn.click(_api(prepare_gfn, 4), [gds, layers, crop, margin, margin_um, select, gst], [prep_plot, warn_md, summary, prepared_gds],
627
+ api_name="prepare", **light)
628
  after = lambda ev: ev.then(refresh_view, vin, vout, **hidden) # noqa: E731
629
+ GU.wire(gds, layers, pred_btn, gparts, [up, edit], after=after).then(ui_predict, pins, uouts, api_visibility="private", **solve)
630
  for ev in (crop.change, margin.change, margin_um.change, select.change):
631
  ev(preview_fn, vin, [view] + vout, trigger_mode="always_last", **hidden)
632
+ pred_btn.click(ui_predict, pins, uouts, api_visibility="private", **solve)
633
+ p_btn.click(_api(predict_fn, 5), pins, pouts, api_name="predict", **solve)
634
  UI.mark_stale(res_state, status, mat, report, [gds.change, model.change, substrate.change, custom_eps.change, thickness.change,
635
  select.change, crop.change, margin.change, margin_um.change, edit.then])
636
+ a_btn.click(_api(api_predict_gds, 1), [gds, a_g, a_c, a_m, substrate, custom_eps, thickness, model, a_gp], a_out, api_name="predict_gds", **solve)
637
+ m_btn.click(models_fn, None, m_out, api_name="models", **light)
638
  return demo
geomgnn/bundle.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Integrity check of a Space bundle's checkpoint copies. deploy/sync.py --flagship ships each checkpoint without its train_keys (unused at
2
+ inference) and records both hashes in models/<tier>/info.json: <key>_sha256 = the published checkpoint (= registry, the provenance) and
3
+ <key>_bundle_sha256 = the shipped copy. install() makes geomgnn.paper check the copy against the latter; folders without
4
+ *_bundle_sha256 and the registry's source paths keep the registry check unchanged."""
5
+ from __future__ import annotations
6
+
7
+ import json
8
+ from pathlib import Path
9
+
10
+ from . import paper as PA
11
+
12
+ _REGISTRY_CHECK = PA._check
13
+
14
+
15
+ def _check(reg, name, model_dir=None):
16
+ md = PA._model_dir(model_dir)
17
+ info = Path(md) / name / "info.json" if md else None
18
+ z = json.loads(info.read_text()) if info is not None and info.exists() else {}
19
+ if "checkpoints_bundle_sha256" not in z:
20
+ return _REGISTRY_CHECK(reg, name, model_dir)
21
+ a = PA.arms(reg)[name]
22
+ for key in ("checkpoints", "surrogate"):
23
+ fs, hs = PA.files(reg, name, key, model_dir), z.get(key + "_bundle_sha256") or []
24
+ if z.get(key + "_sha256", []) != PA._hashes(a, key) or len(fs) != len(hs):
25
+ raise RuntimeError(f"{name}: {info} does not list the registry's {key} (sha256)")
26
+ for f, h in zip(fs, hs):
27
+ if PA._sha(f) != h:
28
+ raise RuntimeError(f"{name}: {f.name} SHA256 differs from the bundle manifest ({h[:12]})")
29
+ return a
30
+
31
+
32
+ def install() -> None:
33
+ PA._check = _check
geomgnn/fastprior_pkg.py CHANGED
@@ -823,7 +823,7 @@ def _mixed_chol(A, B):
823
  try:
824
  cf = cho_factor(A.astype(np.float32, order="F"), lower=True, overwrite_a=True, check_finite=False)
825
  except np.linalg.LinAlgError:
826
- return _ORIG["chol_solve"](A, B)
827
  B = np.asfortranarray(B, dtype=float)
828
  X = cho_solve(cf, B.astype(np.float32), check_finite=False).astype(float)
829
  bn = np.linalg.norm(B, axis=0)
@@ -832,7 +832,7 @@ def _mixed_chol(A, B):
832
  if (np.linalg.norm(R, axis=0) <= CFG["refine_tol"] * bn).all():
833
  return X
834
  X += cho_solve(cf, R.astype(np.float32), check_finite=False)
835
- return _ORIG["chol_solve"](A, B)
836
 
837
 
838
  def _chol_solve(A, B):
@@ -845,7 +845,13 @@ def _chol_solve(A, B):
845
 
846
  def _sym_solve(A, B):
847
  if not isinstance(A, _Dev):
848
- return _ORIG["sym_solve"](A, B)
 
 
 
 
 
 
849
  if A.t.shape[0] > FROZEN_DENSE_MAX:
850
  return _gpu_chol(A, B)
851
  try:
 
823
  try:
824
  cf = cho_factor(A.astype(np.float32, order="F"), lower=True, overwrite_a=True, check_finite=False)
825
  except np.linalg.LinAlgError:
826
+ return _ORIG["chol_solve"](A.copy(order="F"), B)
827
  B = np.asfortranarray(B, dtype=float)
828
  X = cho_solve(cf, B.astype(np.float32), check_finite=False).astype(float)
829
  bn = np.linalg.norm(B, axis=0)
 
832
  if (np.linalg.norm(R, axis=0) <= CFG["refine_tol"] * bn).all():
833
  return X
834
  X += cho_solve(cf, R.astype(np.float32), check_finite=False)
835
+ return _ORIG["chol_solve"](A.copy(order="F"), B)
836
 
837
 
838
  def _chol_solve(A, B):
 
845
 
846
  def _sym_solve(A, B):
847
  if not isinstance(A, _Dev):
848
+ if "mixed" not in STATE["components"]:
849
+ return _ORIG["sym_solve"](A, B)
850
+ try: # = _ORIG["sym_solve"] without its copy of A: _mixed_chol only reads A (its float64 fallbacks copy it)
851
+ return _mixed_chol(A, B)
852
+ except np.linalg.LinAlgError:
853
+ from scipy.linalg import solve
854
+ return solve(A, B, lower=True, assume_a="sym", overwrite_a=True, check_finite=False)
855
  if A.t.shape[0] > FROZEN_DENSE_MAX:
856
  return _gpu_chol(A, B)
857
  try:
geomgnn/paper.py CHANGED
@@ -19,6 +19,7 @@ import hashlib
19
  import json
20
  import os
21
  import time
 
22
  from pathlib import Path
23
 
24
  import numpy as np
@@ -186,13 +187,82 @@ def apply_stack(reg, a, threads=2):
186
  A._thread_limits(threads)
187
 
188
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
189
  def solve_levels(sc, levels, threads=2, stack=None) -> tuple[dict, dict]:
190
  """Positive-mutual prior matrices P = -C at each level under the installed stack (the statements of
191
  geomgnn.api.predict_scene's prior loop, any level list); stack = (eps_r, thickness_um) of the substrate the prior is solved at
192
- (default the scene's, i.e. the training stack)."""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
193
  from ._tlsq.embed.fastprior_loc import micro_mom_loc
194
  from . import levelpar
195
- eps_r, t_um = stack or (sc.stack["substrate_eps_r"], sc.stack["substrate_thickness_um"])
196
  if levelpar.enabled(sc, levels, threads): # the levels concurrently (spawned workers, same stack and statements)
197
  return levelpar.solve(sc, list(levels), threads, ALPHA, (eps_r, t_um))
198
  proxy.NUM_THREADS = threads
 
19
  import json
20
  import os
21
  import time
22
+ from collections import OrderedDict
23
  from pathlib import Path
24
 
25
  import numpy as np
 
187
  A._thread_limits(threads)
188
 
189
 
190
+ _LEVELS: OrderedDict = OrderedDict() # solved prior levels: (layout, level, substrate, stack) -> P; TLSQ_PRIOR_LEVEL_CACHE_MB (0 = off)
191
+ _MESHES: OrderedDict = OrderedDict() # the prior's window meshes (geometry and lattice only, no substrate), same bound
192
+
193
+
194
+ def _geom_hash(geoms) -> str:
195
+ import shapely
196
+ h = hashlib.sha256()
197
+ for g in geoms:
198
+ h.update(shapely.to_wkb(g))
199
+ h.update(b"|")
200
+ return h.hexdigest()
201
+
202
+
203
+ def _level_key(sc, threads, eps_r, t_um) -> tuple:
204
+ return (_geom_hash([c.geom for c in sc.conductors]), float(eps_r), float(t_um), int(threads), ALPHA, A._STATE["prior_impl"],
205
+ repr(A._STATE.get("stack")), repr(A._STATE.get("pr")))
206
+
207
+
208
+ def _put(cache, k, v, cap) -> None:
209
+ cache[k] = v
210
+ nb = lambda x: sum(a.nbytes for a in (x if isinstance(x, tuple) else (x,)) if isinstance(a, np.ndarray)) # noqa: E731
211
+ while len(cache) > 1 and sum(nb(x) for x in cache.values()) > cap:
212
+ cache.popitem(last=False)
213
+
214
+
215
+ def _memo_mesh(f, cap):
216
+ """f(metals, *args) -> mesh arrays, memoised on the metals' WKB and the arguments; hits return copies."""
217
+ def g(metals, *args, **kw):
218
+ k = (f.__module__, f.__name__, _geom_hash(metals), repr(args), repr(sorted(kw.items())))
219
+ v = _MESHES.get(k)
220
+ if v is None:
221
+ v = tuple(x.copy() if isinstance(x, np.ndarray) else x for x in f(metals, *args, **kw))
222
+ _put(_MESHES, k, v, cap)
223
+ else:
224
+ _MESHES.move_to_end(k)
225
+ return tuple(x.copy() if isinstance(x, np.ndarray) else x for x in v)
226
+ return g
227
+
228
+
229
  def solve_levels(sc, levels, threads=2, stack=None) -> tuple[dict, dict]:
230
  """Positive-mutual prior matrices P = -C at each level under the installed stack (the statements of
231
  geomgnn.api.predict_scene's prior loop, any level list); stack = (eps_r, thickness_um) of the substrate the prior is solved at
232
+ (default the scene's, i.e. the training stack). With TLSQ_PRIOR_LEVEL_CACHE_MB > 0 every solved level is kept (LRU, bounded
233
+ in MB) and reused by any later call with the same layout, level, substrate and prior stack (tiers sharing a level), and the
234
+ window meshes are reused across substrates (same layout and level)."""
235
+ eps_r, t_um = stack or (sc.stack["substrate_eps_r"], sc.stack["substrate_thickness_um"])
236
+ cap = float(os.environ.get("TLSQ_PRIOR_LEVEL_CACHE_MB", "0")) * 2**20
237
+ key = _level_key(sc, threads, eps_r, t_um) if cap > 0 and levels else None
238
+ out, tm = {}, {}
239
+ for p in levels if key else ():
240
+ v = _LEVELS.get(key + (float(p),))
241
+ if v is not None:
242
+ _LEVELS.move_to_end(key + (float(p),))
243
+ out[p], tm[f"{p:g}"] = v.copy(), 0.0
244
+ todo = [p for p in levels if p not in out]
245
+ if todo:
246
+ from ._tlsq.embed import fastprior_loc as FL
247
+ from . import priorv2
248
+ hooks = [(m, n, getattr(m, n)) for m, n in ((priorv2, "_mesh"), (FL, "mesh_local"))] if key else []
249
+ try:
250
+ for m, n, f in hooks:
251
+ setattr(m, n, _memo_mesh(f, cap))
252
+ P, t = _solve_levels(sc, todo, threads, eps_r, t_um)
253
+ finally:
254
+ for m, n, f in hooks:
255
+ setattr(m, n, f)
256
+ out.update(P)
257
+ tm.update(t)
258
+ for p in todo if key else ():
259
+ _put(_LEVELS, key + (float(p),), P[p].copy(), cap)
260
+ return {p: out[p] for p in levels}, {f"{p:g}": tm[f"{p:g}"] for p in levels}
261
+
262
+
263
+ def _solve_levels(sc, levels, threads, eps_r, t_um) -> tuple[dict, dict]:
264
  from ._tlsq.embed.fastprior_loc import micro_mom_loc
265
  from . import levelpar
 
266
  if levelpar.enabled(sc, levels, threads): # the levels concurrently (spawned workers, same stack and statements)
267
  return levelpar.solve(sc, list(levels), threads, ALPHA, (eps_r, t_um))
268
  proxy.NUM_THREADS = threads
geomgnn/respcache.py CHANGED
@@ -10,6 +10,7 @@ import os
10
  import pickle
11
  import threading
12
  from collections import OrderedDict
 
13
 
14
  _LRU: OrderedDict = OrderedDict()
15
  _LOCK = threading.Lock()
@@ -62,6 +63,34 @@ def cached(lay, settings: tuple, fn):
62
  return res
63
 
64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
65
  def stats() -> dict:
66
  with _LOCK:
67
  return {"entries": len(_LRU), "mb": _BYTES[0] / 2**20}
 
10
  import pickle
11
  import threading
12
  from collections import OrderedDict
13
+ from pathlib import Path
14
 
15
  _LRU: OrderedDict = OrderedDict()
16
  _LOCK = threading.Lock()
 
63
  return res
64
 
65
 
66
+ def dump(path) -> int:
67
+ """Pickle the entries (image build: the bundled examples, loaded back by load_items + put_many at start)."""
68
+ with _LOCK:
69
+ items = list(_LRU.items())
70
+ Path(path).parent.mkdir(parents=True, exist_ok=True)
71
+ Path(path).write_bytes(pickle.dumps(items, protocol=5))
72
+ return len(items)
73
+
74
+
75
+ def load_items(path) -> list:
76
+ return pickle.loads(Path(path).read_bytes())
77
+
78
+
79
+ def put_many(items) -> int:
80
+ n, mb = _limits()
81
+ if n <= 0:
82
+ return 0
83
+ with _LOCK:
84
+ for k, (res, size) in items:
85
+ if k not in _LRU:
86
+ _LRU[k] = (res, size)
87
+ _BYTES[0] += size
88
+ while len(_LRU) > n or _BYTES[0] > mb:
89
+ _, (_, s) = _LRU.popitem(last=False)
90
+ _BYTES[0] -= s
91
+ return len(_LRU)
92
+
93
+
94
  def stats() -> dict:
95
  with _LOCK:
96
  return {"entries": len(_LRU), "mb": _BYTES[0] / 2**20}
models/accurate/acc_v2m_sf_s0.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:cc69c5818f5028c26177159b88239116520abbb290743d6a6c091657e8dbf1c5
3
- size 8410606
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1c7a9e3be100f083ee9418695525ae54c8469f8a82e8da4d8c728da5e0cd2ca3
3
+ size 381223
models/accurate/acc_v2m_sf_s1.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2f88606dbec5c599550d088213a4c37af3b3ad5134ebad2e60519fb534553a57
3
- size 8410606
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:857b75095a29d8fa0b12d27a161750cdc89611c2227a1545e1ac447518d0c414
3
+ size 381223
models/accurate/acc_v2m_sf_s2.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d13bafc6b2eb7cadcd97a387d791c0210976db1a9d618f9acc1b614e3701daed
3
- size 8410606
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e86e6d1e42df20fc7255974b7f739d59317148d2b16be93f012425819f309ac
3
+ size 381223
models/accurate/info.json CHANGED
@@ -11,6 +11,11 @@
11
  "2f88606dbec5c599550d088213a4c37af3b3ad5134ebad2e60519fb534553a57",
12
  "d13bafc6b2eb7cadcd97a387d791c0210976db1a9d618f9acc1b614e3701daed"
13
  ],
 
 
 
 
 
14
  "checkpoints_source": [
15
  "data_local/flagship_v4/hpc/runs/acc_v2m_sf_s0/model.pt",
16
  "data_local/flagship_v4/hpc/runs/acc_v2m_sf_s1/model.pt",
@@ -18,5 +23,6 @@
18
  ],
19
  "surrogate": [],
20
  "surrogate_sha256": [],
 
21
  "surrogate_source": []
22
  }
 
11
  "2f88606dbec5c599550d088213a4c37af3b3ad5134ebad2e60519fb534553a57",
12
  "d13bafc6b2eb7cadcd97a387d791c0210976db1a9d618f9acc1b614e3701daed"
13
  ],
14
+ "checkpoints_bundle_sha256": [
15
+ "1c7a9e3be100f083ee9418695525ae54c8469f8a82e8da4d8c728da5e0cd2ca3",
16
+ "857b75095a29d8fa0b12d27a161750cdc89611c2227a1545e1ac447518d0c414",
17
+ "6e86e6d1e42df20fc7255974b7f739d59317148d2b16be93f012425819f309ac"
18
+ ],
19
  "checkpoints_source": [
20
  "data_local/flagship_v4/hpc/runs/acc_v2m_sf_s0/model.pt",
21
  "data_local/flagship_v4/hpc/runs/acc_v2m_sf_s1/model.pt",
 
23
  ],
24
  "surrogate": [],
25
  "surrogate_sha256": [],
26
+ "surrogate_bundle_sha256": [],
27
  "surrogate_source": []
28
  }
models/balanced/bal_v2m_sf_s0.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:aa742ab93b1e3ffcd86a753bda5cc4a68812acbe8140a6b0e311c11d1f36b4b2
3
- size 8410606
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7a0353f2e8875c73c4188eccb641d15e8279a65870535a78cabf94f2e410f3ab
3
+ size 381223
models/balanced/bal_v2m_sf_s1.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:486466473eba61c090f4e3fa7988bf880fc183570c5d02bd78f4c1356691f926
3
- size 8410606
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:79654f3fe6a647e0e678eb027ae5f11590ba2ec588deb127d3d1e0c468283b4e
3
+ size 381223
models/balanced/bal_v2m_sf_s2.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1b55ded04fd624585eeeccf2ab6e4a8e9354a9b96694131e1737dde43ad3ff58
3
- size 8410606
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f4fa6514b7f2740c376f664105c6fe723aca1c7c76a5eadec625e9685d721a30
3
+ size 381223
models/balanced/info.json CHANGED
@@ -11,6 +11,11 @@
11
  "486466473eba61c090f4e3fa7988bf880fc183570c5d02bd78f4c1356691f926",
12
  "1b55ded04fd624585eeeccf2ab6e4a8e9354a9b96694131e1737dde43ad3ff58"
13
  ],
 
 
 
 
 
14
  "checkpoints_source": [
15
  "data_local/flagship_v4/hpc/runs/bal_v2m_sf_s0/model.pt",
16
  "data_local/flagship_v4/hpc/runs/bal_v2m_sf_s1/model.pt",
@@ -18,5 +23,6 @@
18
  ],
19
  "surrogate": [],
20
  "surrogate_sha256": [],
 
21
  "surrogate_source": []
22
  }
 
11
  "486466473eba61c090f4e3fa7988bf880fc183570c5d02bd78f4c1356691f926",
12
  "1b55ded04fd624585eeeccf2ab6e4a8e9354a9b96694131e1737dde43ad3ff58"
13
  ],
14
+ "checkpoints_bundle_sha256": [
15
+ "7a0353f2e8875c73c4188eccb641d15e8279a65870535a78cabf94f2e410f3ab",
16
+ "79654f3fe6a647e0e678eb027ae5f11590ba2ec588deb127d3d1e0c468283b4e",
17
+ "f4fa6514b7f2740c376f664105c6fe723aca1c7c76a5eadec625e9685d721a30"
18
+ ],
19
  "checkpoints_source": [
20
  "data_local/flagship_v4/hpc/runs/bal_v2m_sf_s0/model.pt",
21
  "data_local/flagship_v4/hpc/runs/bal_v2m_sf_s1/model.pt",
 
23
  ],
24
  "surrogate": [],
25
  "surrogate_sha256": [],
26
+ "surrogate_bundle_sha256": [],
27
  "surrogate_source": []
28
  }
models/fast/acc_v1_s0.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9dff20ade26619e7713c2d145dada65e4554a623780290673cfe9452cf9b64ba
3
- size 8410094
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e2909bd0bf06d3cd07be27160050e323a69700de732ef54c9afa00c38b7521e0
3
+ size 380559
models/fast/acc_v1_s1.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f1ad2ca065dd650ccabcad73a3c6ad2ac1042f4f68bb828893bbdecc47411ffe
3
- size 8410094
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:085a7819d86f66962a840e1279017f23487d8d3b12fa9773a1fd0f884621b241
3
+ size 380559
models/fast/acc_v1_s2.pt CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8cbada211724df5cc05a7b4fdd232607de047b248a43993d3844bd68ff8060c4
3
- size 8410094
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b607329a83a938911d4ebbd8d801b9c58306ce3c96b34cd82ed502a8fb588fc
3
+ size 380559
models/fast/info.json CHANGED
@@ -11,6 +11,11 @@
11
  "f1ad2ca065dd650ccabcad73a3c6ad2ac1042f4f68bb828893bbdecc47411ffe",
12
  "8cbada211724df5cc05a7b4fdd232607de047b248a43993d3844bd68ff8060c4"
13
  ],
 
 
 
 
 
14
  "checkpoints_source": [
15
  "data_local/flagship_v4/hpc/runs/acc_v1_s0/model.pt",
16
  "data_local/flagship_v4/hpc/runs/acc_v1_s1/model.pt",
@@ -22,6 +27,9 @@
22
  "surrogate_sha256": [
23
  "132d6f389eda798d31bd5654a96e24cce18fa750134a9fb2a9725eafed339dac"
24
  ],
 
 
 
25
  "surrogate_source": [
26
  "data_local/flagship_v4/hpc/sur/sall_all_s0/model.pt"
27
  ]
 
11
  "f1ad2ca065dd650ccabcad73a3c6ad2ac1042f4f68bb828893bbdecc47411ffe",
12
  "8cbada211724df5cc05a7b4fdd232607de047b248a43993d3844bd68ff8060c4"
13
  ],
14
+ "checkpoints_bundle_sha256": [
15
+ "e2909bd0bf06d3cd07be27160050e323a69700de732ef54c9afa00c38b7521e0",
16
+ "085a7819d86f66962a840e1279017f23487d8d3b12fa9773a1fd0f884621b241",
17
+ "5b607329a83a938911d4ebbd8d801b9c58306ce3c96b34cd82ed502a8fb588fc"
18
+ ],
19
  "checkpoints_source": [
20
  "data_local/flagship_v4/hpc/runs/acc_v1_s0/model.pt",
21
  "data_local/flagship_v4/hpc/runs/acc_v1_s1/model.pt",
 
27
  "surrogate_sha256": [
28
  "132d6f389eda798d31bd5654a96e24cce18fa750134a9fb2a9725eafed339dac"
29
  ],
30
+ "surrogate_bundle_sha256": [
31
+ "132d6f389eda798d31bd5654a96e24cce18fa750134a9fb2a9725eafed339dac"
32
+ ],
33
  "surrogate_source": [
34
  "data_local/flagship_v4/hpc/sur/sall_all_s0/model.pt"
35
  ]
registry.yaml CHANGED
@@ -729,7 +729,7 @@ gnn_models:
729
  patches: [priorv2m] # VALUE-CHANGING: geomgnn/priorv2.py = scripts/priorv2/priorv2.py, installed(policy='fc', dip=True, margin=True)
730
  module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
731
  mode: exact_compiled # geomgnn.fastprior_pkg solver mode (fv + hmat_vec + mixed + numba compiled H-matrix); TLSQ_GNN_PRIOR_MODE overrides
732
- speed_package_sha256: 956f2b62c308ec1561bbc033716b52c24dc63a916b9373cc44090eb6514e5dde # scripts/fastprior_pkg/fastprior_pkg.py, vendored verbatim
733
  speed_approved: "user 2026-10-17 ('faster at equal accuracy is a win'): fastprior_pkg 'Ready to deploy' (agents/fastprior_pkg/RESULTS.md); exact_compiled VAL tie on all 13 families; deploy-path parity gate in deploy/tests/test_gnn_space.py (<= 1e-6 vs the v2m exact path)"
734
  certified_opt_in: "mode certified (+ Rayleigh-Ritz windows with an error certificate) is gated on VAL (tie) but not enabled here; set TLSQ_GNN_PRIOR_MODE=certified (Space variable) to opt in"
735
  value_changing_approved: "user 2026-10-17 ('do whatever to make the models faster and better'); prior v2 PASSED its pre-declared protocol, arm Bm (frozen nets + v2m) recommended for deployment (agents/priorv2/RESULTS.md verdict 4, FINALIZATION_V2.json)"
@@ -971,7 +971,7 @@ paper_hub:
971
  mode: exact_compiled
972
  module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
973
  farfield_sha256: 6c877b47c2894fe78a38ad4199b29e450b610c2c45a261c8cf4d53efc1912836 # scripts/farfield/farfield.py (agents/farfield), vendored verbatim
974
- speed_package_sha256: 956f2b62c308ec1561bbc033716b52c24dc63a916b9373cc44090eb6514e5dde
975
  value_changing_approved: "user 2026-10-17 (v2m, gnn_models.GNN); user 2026-10-20: adopted arm (b) of agents/farfield (substrate dipole for far pairs in different ground pockets; pre-declared rule met, agents/farfield/RESULTS.md; adoption agents/farfield/ADOPTION.md)"
976
  speed_approved: "user 2026-10-17 (gnn_models.GNN)"
977
  previous: corrected_halfspace # the same stack with the half-space far field (before 2026-10-20); process-wide override: TLSQ_FAR_FIELD=halfspace
@@ -981,7 +981,7 @@ paper_hub:
981
  patches: [priorv2m]
982
  mode: exact_compiled
983
  module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
984
- speed_package_sha256: 956f2b62c308ec1561bbc033716b52c24dc63a916b9373cc44090eb6514e5dde
985
  value_changing_approved: "user 2026-10-17 (gnn_models.GNN)"
986
  speed_approved: "user 2026-10-17 (gnn_models.GNN)"
987
  uncorrected: # prior v1, frozen solver (+ the value-preserving rejection-tail fix, bit-identical)
@@ -993,7 +993,7 @@ paper_hub:
993
  uncorrected_certified: # prior v1, certified solve (Rayleigh-Ritz windows with an error estimate)
994
  patches: []
995
  mode: certified
996
- speed_package_sha256: 956f2b62c308ec1561bbc033716b52c24dc63a916b9373cc44090eb6514e5dde
997
  speed_approved: "VAL tie on all 13 families (agents/fastprior_pkg/RESULTS.md); paper tab:routes (d)"
998
 
999
  _paper_ckpt:
@@ -1391,7 +1391,7 @@ paper_arms:
1391
  flagship_hub:
1392
  public_link: SQuADDS/planar-capacitance # static Space, printed in the paper; opens the demo top-level (deploy/spaces/SQuADDS-planar-capacitance); never rename
1393
  deploy_20261114: # public-demo limits + short card (no paper model); Docker SYSTEM=spaces so HF OAuth works
1394
- commit: 4d44ef6cd727d4bd905c579e1882b2a99ec47423
1395
  live: deploy/tests/results/flagship_space_live_limits20261114.json # examples x tiers vs Mac <= 1.1e-3 rel; anon non-example refused; oauth_token call ran
1396
  space: shanto268/PlanarCapacitanceModel # the URL is kept; its earlier v3 tiers are archived (archive below)
1397
  title: Planar Capacitance Model
@@ -1403,7 +1403,7 @@ flagship_hub:
1403
  checkpoints: data_local/flagship_v4/checkpoints.json # sha256 of every flagship checkpoint; sync.py refuses any other file
1404
  results: agents/flagship_v4/RESULTS.md
1405
  hf_cards_20261114: # README-only rewrite (re-pushed same day: /ml-paper/ link, then no links to the private paper model) (handoff/hf_cards/), private; package pins unchanged (manifest bb9d0ce / 5d4e094, dataset 13d46953)
1406
- SQuADDS/SQuADDS_PlanarCapacitance: b96e3f693ea50ed55040d2e4599a1cde220d6869
1407
  SQuADDS/SQuADDS_PlanarCapacitance-Synthetic: 3179ea7aed6f0ba92b549cc29aa645feea937513
1408
  SQuADDS/SQuADDS_PlanarCapacitance-heldout: c538385b3e8d48f4183b68d00e376a06762b0705
1409
  SQuADDS/PlanarCapacitance-flagship: a89f30850b638da1f693042b7d7cf14f20b19098 # + balanced/bal_v2m_sf_s{0,1,2}.pt (same sha256 as the Space) and SHA256SUMS
 
729
  patches: [priorv2m] # VALUE-CHANGING: geomgnn/priorv2.py = scripts/priorv2/priorv2.py, installed(policy='fc', dip=True, margin=True)
730
  module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
731
  mode: exact_compiled # geomgnn.fastprior_pkg solver mode (fv + hmat_vec + mixed + numba compiled H-matrix); TLSQ_GNN_PRIOR_MODE overrides
732
+ speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a # scripts/fastprior_pkg/fastprior_pkg.py, vendored verbatim
733
  speed_approved: "user 2026-10-17 ('faster at equal accuracy is a win'): fastprior_pkg 'Ready to deploy' (agents/fastprior_pkg/RESULTS.md); exact_compiled VAL tie on all 13 families; deploy-path parity gate in deploy/tests/test_gnn_space.py (<= 1e-6 vs the v2m exact path)"
734
  certified_opt_in: "mode certified (+ Rayleigh-Ritz windows with an error certificate) is gated on VAL (tie) but not enabled here; set TLSQ_GNN_PRIOR_MODE=certified (Space variable) to opt in"
735
  value_changing_approved: "user 2026-10-17 ('do whatever to make the models faster and better'); prior v2 PASSED its pre-declared protocol, arm Bm (frozen nets + v2m) recommended for deployment (agents/priorv2/RESULTS.md verdict 4, FINALIZATION_V2.json)"
 
971
  mode: exact_compiled
972
  module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
973
  farfield_sha256: 6c877b47c2894fe78a38ad4199b29e450b610c2c45a261c8cf4d53efc1912836 # scripts/farfield/farfield.py (agents/farfield), vendored verbatim
974
+ speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a
975
  value_changing_approved: "user 2026-10-17 (v2m, gnn_models.GNN); user 2026-10-20: adopted arm (b) of agents/farfield (substrate dipole for far pairs in different ground pockets; pre-declared rule met, agents/farfield/RESULTS.md; adoption agents/farfield/ADOPTION.md)"
976
  speed_approved: "user 2026-10-17 (gnn_models.GNN)"
977
  previous: corrected_halfspace # the same stack with the half-space far field (before 2026-10-20); process-wide override: TLSQ_FAR_FIELD=halfspace
 
981
  patches: [priorv2m]
982
  mode: exact_compiled
983
  module_sha256: 1a0bdcb71500c7f11e63238afe531c18caabd5c9f6f7aaa73d9b7dcd0ad61f8c
984
+ speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a
985
  value_changing_approved: "user 2026-10-17 (gnn_models.GNN)"
986
  speed_approved: "user 2026-10-17 (gnn_models.GNN)"
987
  uncorrected: # prior v1, frozen solver (+ the value-preserving rejection-tail fix, bit-identical)
 
993
  uncorrected_certified: # prior v1, certified solve (Rayleigh-Ritz windows with an error estimate)
994
  patches: []
995
  mode: certified
996
+ speed_package_sha256: 4fff680f35cb8b098de6c9b1b6fc55780064362ffc829bb47bf69a77454ce38a
997
  speed_approved: "VAL tie on all 13 families (agents/fastprior_pkg/RESULTS.md); paper tab:routes (d)"
998
 
999
  _paper_ckpt:
 
1391
  flagship_hub:
1392
  public_link: SQuADDS/planar-capacitance # static Space, printed in the paper; opens the demo top-level (deploy/spaces/SQuADDS-planar-capacitance); never rename
1393
  deploy_20261114: # public-demo limits + short card (no paper model); Docker SYSTEM=spaces so HF OAuth works
1394
+ commit: 142609ed7021861aea2be8ee5f51594b70e5918b # + text in the author's voice, Fast in the selector, Model not Tier, footer (was 4d44ef6)
1395
  live: deploy/tests/results/flagship_space_live_limits20261114.json # examples x tiers vs Mac <= 1.1e-3 rel; anon non-example refused; oauth_token call ran
1396
  space: shanto268/PlanarCapacitanceModel # the URL is kept; its earlier v3 tiers are archived (archive below)
1397
  title: Planar Capacitance Model
 
1403
  checkpoints: data_local/flagship_v4/checkpoints.json # sha256 of every flagship checkpoint; sync.py refuses any other file
1404
  results: agents/flagship_v4/RESULTS.md
1405
  hf_cards_20261114: # README-only rewrite (re-pushed same day: /ml-paper/ link, then no links to the private paper model) (handoff/hf_cards/), private; package pins unchanged (manifest bb9d0ce / 5d4e094, dataset 13d46953)
1406
+ SQuADDS/SQuADDS_PlanarCapacitance: c79953313874e9c116ca2f39a414d92025d18484 # physics_prior rename (24de414) + data/legacy_v1 removed
1407
  SQuADDS/SQuADDS_PlanarCapacitance-Synthetic: 3179ea7aed6f0ba92b549cc29aa645feea937513
1408
  SQuADDS/SQuADDS_PlanarCapacitance-heldout: c538385b3e8d48f4183b68d00e376a06762b0705
1409
  SQuADDS/PlanarCapacitance-flagship: a89f30850b638da1f693042b7d7cf14f20b19098 # + balanced/bal_v2m_sf_s{0,1,2}.pt (same sha256 as the Space) and SHA256SUMS
requirements.txt CHANGED
@@ -1,17 +1,15 @@
1
  --extra-index-url https://download.pytorch.org/whl/cpu
2
- torch
3
- numpy
4
- scipy
5
- shapely>=2.0
6
- gdstk>=0.9
7
- scikit-learn==1.9.1
8
- joblib
9
- pandas
10
- pyyaml
11
- matplotlib
12
- huggingface_hub
13
  numba==0.68.0
14
  llvmlite==0.50.0
15
- threadpoolctl
16
  authlib==1.6.5
17
  itsdangerous==2.2.0
 
1
  --extra-index-url https://download.pytorch.org/whl/cpu
2
+ torch==2.14.1
3
+ numpy==2.5.3
4
+ scipy==1.18.1
5
+ shapely==2.1.2
6
+ gdstk==1.0.1
7
+ pandas==3.0.6
8
+ pyyaml==6.0.3
9
+ matplotlib==3.11.2
10
+ huggingface_hub==2.2.0
 
 
11
  numba==0.68.0
12
  llvmlite==0.50.0
13
+ threadpoolctl==3.7.0
14
  authlib==1.6.5
15
  itsdangerous==2.2.0
warm.py CHANGED
@@ -1,10 +1,22 @@
1
  """Build-time warm-up (Docker variant): the Space's own preload at import of app.py (every arm / tier on the interdigital
2
- example: the prior's numba kernels, the NN kernel, the surrogates), so the image carries the compiled cache."""
 
 
3
  import os
 
 
 
4
  import time
 
5
 
6
- t = time.perf_counter()
7
  os.environ["TLSQ_PRELOAD"] = "1"
8
- import app # noqa: E402,F401 (app.py preloads at import; the server starts only under __main__)
9
-
10
  print("warm.py: preload and cache population", round(time.perf_counter() - t, 1), "s", flush=True)
 
 
 
 
 
 
 
 
1
  """Build-time warm-up (Docker variant): the Space's own preload at import of app.py (every arm / tier on the interdigital
2
+ example: the prior's numba kernels, the NN kernel, the surrogates), so the image carries the compiled cache; then (flagship) the
3
+ examples' results baked for the response cache. Each step runs in its own process: the preload must succeed; the bake may fail
4
+ (e.g. the builder runs out of memory), and the image is then built without it (the Space computes the examples on demand)."""
5
  import os
6
+ import resource
7
+ import subprocess
8
+ import sys
9
  import time
10
+ from pathlib import Path
11
 
 
12
  os.environ["TLSQ_PRELOAD"] = "1"
13
+ t = time.perf_counter()
14
+ subprocess.run([sys.executable, "-c", "import app"], check=True) # noqa: S603 (app.py preloads at import; no server)
15
  print("warm.py: preload and cache population", round(time.perf_counter() - t, 1), "s", flush=True)
16
+ t = time.perf_counter()
17
+ r = subprocess.run([sys.executable, "-c", "import app\nif hasattr(app, 'bake'):\n app.bake()"]) # noqa: S603
18
+ peak = resource.getrusage(resource.RUSAGE_CHILDREN).ru_maxrss / 2**10
19
+ if r.returncode:
20
+ Path(os.environ.get("TLSQ_RESPONSE_BAKE", "bake/respcache.pkl")).unlink(missing_ok=True)
21
+ print(f"warm.py: bake FAILED (exit {r.returncode}, e.g. out of memory): the image is built without baked example results", flush=True)
22
+ print("warm.py: bake step", round(time.perf_counter() - t, 1), "s, peak RSS of a step", round(peak), "MB", flush=True)