guldo-cli 0.1.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- guldo/__init__.py +0 -0
- guldo/_src/COMPILER.md +329 -0
- guldo/_src/DEPLOY.md +200 -0
- guldo/_src/LAUNCH.md +85 -0
- guldo/_src/PROTOCOL.md +294 -0
- guldo/_src/RULES.md +25 -0
- guldo/_src/SUBMIT.md +170 -0
- guldo/_src/__init__.py +0 -0
- guldo/_src/attest.py +257 -0
- guldo/_src/build_site.py +958 -0
- guldo/_src/bundle_templates/check.py +1382 -0
- guldo/_src/bundle_templates/qcheck.py +253 -0
- guldo/_src/bundle_templates/qscore.py +402 -0
- guldo/_src/bundle_templates/score.py +649 -0
- guldo/_src/certificate.py +828 -0
- guldo/_src/certs_site.py +108 -0
- guldo/_src/claims.py +343 -0
- guldo/_src/compile_domain.py +1312 -0
- guldo/_src/compile_quality.py +286 -0
- guldo/_src/count_runner.py +248 -0
- guldo/_src/counted.py +274 -0
- guldo/_src/declared.py +215 -0
- guldo/_src/deploy/Caddyfile +48 -0
- guldo/_src/deploy/ledger-backup.service +9 -0
- guldo/_src/deploy/ledger-backup.sh +73 -0
- guldo/_src/deploy/ledger-backup.timer +10 -0
- guldo/_src/deploy/ledger-export.service +12 -0
- guldo/_src/deploy/ledger-export.timer +9 -0
- guldo/_src/deploy/ledger-web.service +38 -0
- guldo/_src/deploy/pin_cpu.sh +34 -0
- guldo/_src/deploy/provision_eval.sh +128 -0
- guldo/_src/deploy_smoke.py +148 -0
- guldo/_src/dlog_experiment.py +165 -0
- guldo/_src/domains/aidecode/compiled.json +586 -0
- guldo/_src/domains/aidecode/decl.py +134 -0
- guldo/_src/domains/aidecode/reference.py +73 -0
- guldo/_src/domains/aiinfer/compiled.json +722 -0
- guldo/_src/domains/aiinfer/decl.py +91 -0
- guldo/_src/domains/aiinfer/reference.py +42 -0
- guldo/_src/domains/dlog/compiled.json +248 -0
- guldo/_src/domains/dlog/decl.py +41 -0
- guldo/_src/domains/dlog/reference.py +10 -0
- guldo/_src/domains/maxcut/compiled.json +291 -0
- guldo/_src/domains/maxcut/decl.py +46 -0
- guldo/_src/domains/maxcut/reference.py +16 -0
- guldo/_src/drift_check.py +89 -0
- guldo/_src/export_bundle.py +750 -0
- guldo/_src/extrapolate.py +326 -0
- guldo/_src/families.py +53 -0
- guldo/_src/gen_cheat_table.py +36 -0
- guldo/_src/gpu_meter.py +634 -0
- guldo/_src/gpu_runner.py +63 -0
- guldo/_src/harness.py +273 -0
- guldo/_src/instances.py +86 -0
- guldo/_src/instances_factor.py +40 -0
- guldo/_src/instances_mlwe.py +69 -0
- guldo/_src/judge.py +278 -0
- guldo/_src/limits_plan.py +158 -0
- guldo/_src/official.py +189 -0
- guldo/_src/ops.py +281 -0
- guldo/_src/portal.py +313 -0
- guldo/_src/portal_core.py +57 -0
- guldo/_src/power.py +174 -0
- guldo/_src/prereg.py +173 -0
- guldo/_src/public_portal.py +500 -0
- guldo/_src/quality_core.py +126 -0
- guldo/_src/recheck.py +236 -0
- guldo/_src/remeasure.py +88 -0
- guldo/_src/report_html.py +493 -0
- guldo/_src/results/.keep +0 -0
- guldo/_src/rounds.py +846 -0
- guldo/_src/runner.py +52 -0
- guldo/_src/sandbox/Dockerfile +11 -0
- guldo/_src/sandbox/Dockerfile.v2 +19 -0
- guldo/_src/sandbox/README.md +40 -0
- guldo/_src/sandbox/linux_selftest.py +117 -0
- guldo/_src/sandbox/meter.py +89 -0
- guldo/_src/sandbox/meter_core.py +28 -0
- guldo/_src/site_assets.py +176 -0
- guldo/_src/site_live_check.py +79 -0
- guldo/_src/solvers/aidecode_flops_kv.py +73 -0
- guldo/_src/solvers/aidecode_flops_lossy.py +73 -0
- guldo/_src/solvers/aidecode_flops_ref.py +64 -0
- guldo/_src/solvers/aidecode_kv.py +76 -0
- guldo/_src/solvers/aidecode_kv_f32.py +77 -0
- guldo/_src/solvers/aidecode_lossy.py +76 -0
- guldo/_src/solvers/aidecode_repeat.py +5 -0
- guldo/_src/solvers/aiinfer_batched.py +40 -0
- guldo/_src/solvers/aiinfer_f32.py +41 -0
- guldo/_src/solvers/aiinfer_gpu_batched.py +42 -0
- guldo/_src/solvers/aiinfer_gpu_ref.py +45 -0
- guldo/_src/solvers/aiinfer_lossy.py +42 -0
- guldo/_src/solvers/baseline_bkz.py +14 -0
- guldo/_src/solvers/cheat_lookup.py +24 -0
- guldo/_src/solvers/cheat_table.json +1 -0
- guldo/_src/solvers/cheat_table_v1_stale.json +1 -0
- guldo/_src/solvers/daemon_cpu_hider.py +17 -0
- guldo/_src/solvers/dlog_bsgs.py +19 -0
- guldo/_src/solvers/dlog_stride4.py +18 -0
- guldo/_src/solvers/egg_base.py +11 -0
- guldo/_src/solvers/egg_eightaxis.py +68 -0
- guldo/_src/solvers/egg_template.py +11 -0
- guldo/_src/solvers/factor_brent.py +34 -0
- guldo/_src/solvers/factor_floyd.py +19 -0
- guldo/_src/solvers/factor_trial_fast_py.py +21 -0
- guldo/_src/solvers/factor_trial_np.py +21 -0
- guldo/_src/solvers/factor_trial_py.py +12 -0
- guldo/_src/solvers/fixed_overhead_ref.py +21 -0
- guldo/_src/solvers/fork_bomb.py +18 -0
- guldo/_src/solvers/gpu_noop.py +6 -0
- guldo/_src/solvers/hostile_probe.py +36 -0
- guldo/_src/solvers/kill_meter.py +27 -0
- guldo/_src/solvers/memory_bomb.py +10 -0
- guldo/_src/solvers/noop.py +5 -0
- guldo/_src/solvers/orphan_cpu_hider.py +16 -0
- guldo/_src/solvers/planted_peek.py +23 -0
- guldo/_src/solvers/sleeping_daemon_hang.py +12 -0
- guldo/_src/solvers/small_n_trick.py +22 -0
- guldo/_src/solvers/starter_template.py +21 -0
- guldo/_src/solvers/stdout_flood.py +9 -0
- guldo/_src/solvers/tmp_fill.py +10 -0
- guldo/_src/solvers/tuned_bkz.py +25 -0
- guldo/_src/static_check.py +214 -0
- guldo/_src/submit_check.py +166 -0
- guldo/_src/svp.py +889 -0
- guldo/_src/tensorops.py +304 -0
- guldo/_src/tflops.py +453 -0
- guldo/_src/tflops_runner.py +134 -0
- guldo/cli.py +60 -0
- guldo_cli-0.1.0.dist-info/METADATA +30 -0
- guldo_cli-0.1.0.dist-info/RECORD +134 -0
- guldo_cli-0.1.0.dist-info/WHEEL +5 -0
- guldo_cli-0.1.0.dist-info/entry_points.txt +2 -0
- guldo_cli-0.1.0.dist-info/top_level.txt +1 -0
|
@@ -0,0 +1,73 @@
|
|
|
1
|
+
"""Reference for aidecode: recompute the whole sequence for every generated token (no KV cache), float64.
|
|
2
|
+
|
|
3
|
+
The model definition (the contract a submission must reproduce): SplitMix64 counter stream under `seed` (element k = finalise((k+1)*GOLD + seed), mapped to
|
|
4
|
+
[-1, 1) as (z >> 11) * 2^-52 - 1). Stream layout, in order: tok (64x128, scale sqrt(3)); for each of 2 layers Wq, Wk, Wv, Wo (128x128), W1 (128x512), W2 (512x128)
|
|
5
|
+
(scale W_SCALE*sqrt(3/fan_in), W_SCALE = 4); pos ((8+T)x128, scale POS_SCALE*sqrt(3), POS_SCALE = 3); then 8 prompt tokens floor((u+1)/2*64). Network on a
|
|
6
|
+
sequence: x = tok[t] + pos[i]; per layer: h = rms(x); causal 4-head attention softmax(q.k/sqrt(32)) v, x += o Wo; x += relu(rms(x) W1) W2; logits =
|
|
7
|
+
rms(x) tok^T with rms(x) = x / sqrt(mean(x^2) + 1e-6). Generation is greedy (lowest index on exact ties).
|
|
8
|
+
"""
|
|
9
|
+
import math
|
|
10
|
+
|
|
11
|
+
import numpy as np
|
|
12
|
+
|
|
13
|
+
V, D, H, DH, FF, L, PROMPT = 64, 128, 4, 32, 512, 2, 8
|
|
14
|
+
W_SCALE, POS_SCALE = 4.0, 3.0
|
|
15
|
+
_GOLD = 0x9E3779B97F4A7C15
|
|
16
|
+
_C1 = 0xBF58476D1CE4E5B9
|
|
17
|
+
_C2 = 0x94D049BB133111EB
|
|
18
|
+
|
|
19
|
+
|
|
20
|
+
def _uniform(seed, start, n):
|
|
21
|
+
u = np.uint64
|
|
22
|
+
with np.errstate(over="ignore"):
|
|
23
|
+
z = (np.arange(start, start + n, dtype=u) + u(1)) * u(_GOLD) + u(seed)
|
|
24
|
+
z = (z ^ (z >> u(30))) * u(_C1)
|
|
25
|
+
z = (z ^ (z >> u(27))) * u(_C2)
|
|
26
|
+
z = z ^ (z >> u(31))
|
|
27
|
+
return (z >> u(11)).astype(np.float64) * 2.0 ** -52 - 1.0
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
def _params(seed, T):
|
|
31
|
+
o = [0]
|
|
32
|
+
|
|
33
|
+
def take(shape, scale):
|
|
34
|
+
n = int(np.prod(shape))
|
|
35
|
+
a = (_uniform(seed, o[0], n) * scale).reshape(shape)
|
|
36
|
+
o[0] += n
|
|
37
|
+
return a
|
|
38
|
+
sd, sf = W_SCALE * math.sqrt(3.0 / D), W_SCALE * math.sqrt(3.0 / FF)
|
|
39
|
+
P = {"tok": take((V, D), math.sqrt(3.0)), "layers": []}
|
|
40
|
+
for _ in range(L):
|
|
41
|
+
P["layers"].append({"Wq": take((D, D), sd), "Wk": take((D, D), sd), "Wv": take((D, D), sd), "Wo": take((D, D), sd),
|
|
42
|
+
"W1": take((D, FF), sd), "W2": take((FF, D), sf)})
|
|
43
|
+
P["pos"] = take((PROMPT + T, D), POS_SCALE * math.sqrt(3.0))
|
|
44
|
+
P["prompt"] = np.minimum(((_uniform(seed, o[0], PROMPT) + 1.0) * 0.5 * V).astype(np.int64), V - 1)
|
|
45
|
+
return P
|
|
46
|
+
|
|
47
|
+
|
|
48
|
+
def _rms(x):
|
|
49
|
+
return x / np.sqrt((x * x).mean(axis=-1, keepdims=True) + 1e-6)
|
|
50
|
+
|
|
51
|
+
|
|
52
|
+
def _forward_last(P, toks):
|
|
53
|
+
n = len(toks)
|
|
54
|
+
x = P["tok"][toks] + P["pos"][:n]
|
|
55
|
+
mask = np.triu(np.full((n, n), -1e30), 1)
|
|
56
|
+
for l in P["layers"]:
|
|
57
|
+
h = _rms(x)
|
|
58
|
+
q, k, v = [(h @ l[w]).reshape(n, H, DH).transpose(1, 0, 2) for w in ("Wq", "Wk", "Wv")]
|
|
59
|
+
s = q @ k.transpose(0, 2, 1) / math.sqrt(DH) + mask
|
|
60
|
+
s -= s.max(-1, keepdims=True)
|
|
61
|
+
a = np.exp(s)
|
|
62
|
+
a /= a.sum(-1, keepdims=True)
|
|
63
|
+
x = x + (a @ v).transpose(1, 0, 2).reshape(n, D) @ l["Wo"]
|
|
64
|
+
x = x + np.maximum(_rms(x) @ l["W1"], 0.0) @ l["W2"]
|
|
65
|
+
return _rms(x[-1]) @ P["tok"].T
|
|
66
|
+
|
|
67
|
+
|
|
68
|
+
def solve(seed, prompt_len, T):
|
|
69
|
+
P = _params(seed, T)
|
|
70
|
+
seq = [int(t) for t in P["prompt"]]
|
|
71
|
+
for _ in range(T):
|
|
72
|
+
seq.append(int(np.argmax(_forward_last(P, np.asarray(seq)))))
|
|
73
|
+
return seq[PROMPT:]
|