Repository navigation
feat(rust): port predict_tournament and cached_embed_fn - #1388
kevin-gatimu wants to merge 6 commits into
Conversation
|
Held for integration, not on merit. This is one of nine PRs in the same chain that were each opened against The full conflict map and what I need is on #1377, so I am not repeating it here: short version, pick an order, rebase each onto the one before it, and regenerate For this one specifically, the conflicting paths against the integration branch were: |
24 PRs. Vietnamese kept off the English checkpoint (#1392), deciding-window attribution preserved through end hooks (#1369), RemoteRouter identifying itself with a configurable User-Agent (#1396), ten more inputs refused where they enter (#1372, #1394, #1397, #1398, #1400, #1401, #1402, #1403, #1404, #1405, #1406), per-language temperatures and per-call budgets in .NET and Rust (#1370, #1371), .NET calibration files (#1384), the Go router lifecycle, SHA-256 pins and routed inference (#1374, #1379, #1395), a C++ foundation (#1399, #1407), and parity-lane sharding that replaces the budget headroom added for 0.4.16 (#1373, #1375). Verified before tagging: all 92 suites pass, 640 tests, up from 612. ruff clean on both lists. laya-ts 1251 tests across 44 files, tsc --noEmit clean. Every generated-artifact gate clean on CPython 3.12 / unicodedata 15.0.0: Java 20 families match laya (9 unverified), UnicodeTables matches CPython 15.0.0, LanguageTables matches laya/lang.py, three Go tables_gen.go match, both laya-cpp .inc files match. Version agrees across pyproject.toml, laya.__version__ and nix/package.nix. README carries 0.4.17 only. Security audit of the merge diff, 362766 added lines over 480 files: zero obfuscation, zero secrets, zero boundary hits, zero binaries, no mode or symlink changes, 20 of 20 actions SHA-pinned including the new cpp.yml, npm install hooks unchanged, and no third-party host in any added line (example.com 90, github.com 2, download.pytorch.org 1, docs.rs 1). The 12 dynamic-exec hits are the C++ test harness invoking its own checker through sys.executable, the Go sample generator reading laya's own example by AST at development time, and one comment in rust.yml. Held: #1364 (unchanged, it moves a cross-port contract), #1381 (a #[doc(hidden)] but pub Rust hook that would ship as public API able to replace the forward pass), #1386 (route_batch switched from self.route to self._route, silently bypassing subclass overrides the docstring promises), and nine stacked port PRs that conflict with each other and need linearizing (#1377, #1378, #1380, #1382, #1383, #1385, #1387, #1388, #1389, #1390). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2b573fa to
47f2128
Compare
|
Rebased onto #1383 (now on 0.4.17) with goldens regenerated; conflicts resolved. Also fixed the tournament parity test to loop over engines::checkpoints(), so a sharded parity cell only checks the checkpoints it owns. |
47f2128 to
3cbd138
Compare
3cbd138 to
63c512a
Compare
LayaOptions::calibration names a calibration JSON that the engine loads last, after its tokenizer, sessions, checkpoint temperatures and language table, as ONNXAgent.load_calibration does: the payload is validated with laya/calibrate.py's texts, its temperatures replace the checkpoint's and are clamped with Python's warning, and its binning_map recalibrates answer_confidence (unrounded, then rounded to 4 places) per option-count bucket, skipped under a language override. The confidence gate and the predict_long window choice read the binned value. The router passes the file to every engine it builds. The identity warning for a version-2 file is not emitted: the SDK has no equivalent of Python's model_id_or_path. Goldens and the calibration probe come from tools/regen_golden.py with a fixture Python's own load_calibration reads; the case lists are shared verbatim with the .NET SDK.
…o 1.0 like Python 0.4.20 `clamp_temperature` and `temperature_rejected` catch the OverflowError of a JSON integer too wide for a double, so a calibration file's temperature or bucket written as one is the neutral 1.0 and the clamp warning names it digit for digit. A bin value still refuses it, as `apply_binning_map` does. The routing goldens are regenerated on Python 0.4.20.
Questions gain with_option_order and option_order (a permutation of the option slots). Options are encoded and marked in slot order, the softmax runs in slot order and Calibration::unpermute_probs maps the probabilities back to option order before the answer is built, as laya/common.py and Agent do. The order is validated with Python's text where Python validates it, joins the router's batch grouping key, and is dropped with Python's warning when the shortlist narrows the labels. Existing constructors are unchanged. Goldens come from tools/regen_golden.py; the case lists are shared verbatim with the .NET SDK.
Questions gain with_labels, with_noul_labels and labels: the false and true option labels Python's noul questions accept. They are validated and stripped as laya/common.py _resolve_noul_labels does, with Python's text, at prediction as Agent._check_question does; rendered as "<label>: <criterion or default text>"; part of the router's batch grouping key. The answer is unchanged (P(true)). Question::noul is unchanged. Goldens come from tools/regen_golden.py; the case lists are shared verbatim with the .NET SDK.
Python's shortlist copies the question dict and replaces only its criteria (laya/shortlist.py:129-131), so a labels entry on a choice survives narrowing and Agent._check_question still refuses it. The rebuilt choice now carries the labels; the stale option_order is still dropped.
LayaShortlist::predict_tournament narrows each large choice in rounds of
contiguous groups of group_size labels, one predict call per round with
the group winners kept in order, then answers the full request once with
each choice cut to its finalists, as laya/shortlist.py predict_tournament
does: same group slicing, round ids, errors and option_order warning; the
result is the final call's with tournament metadata {labels, n, rounds}.
predict_tournament_with takes a predict closure so per-call options reach every
round, as Python's **predict_kwargs.
CachedEmbed ports cached_embed_fn: an LRU of float32 rows keyed
on the exact text, misses embedded in one de-duplicated call, non-finite
values cached as 0.0, the dimension guard, cache_info and cache_clear.
Goldens, a tournament probe and a cache probe come from
tools/regen_golden.py and Python's own functions; the case lists are
shared verbatim with the .NET SDK.
63c512a to
32097a4
Compare
Depends on #1383 (noul labels, on #1378 option_order); only the last commit is new.
Ports Python's
predict_tournamentandcached_embed_fnto the Rust SDK.LayaShortlist::predict_tournamentandpredict_tournament_with(a predict closure, so per-call options reach every round as**predict_kwargsdo): rounds of contiguous groups, winners in group order, one final call with each choice cut to its finalists; result is the final call's plustournament(){labels, n, rounds}; same errors andoption_orderwarning aslaya/shortlist.py.CachedEmbed: LRU of f32 rows keyed on the exact text, misses embedded in one de-duplicated call, non-finite values cached as 0.0, dimension guard,cache_info/cache_clear.group_size/maxsize(usize), a predictor returning nothing.predict_tournamentover a scripted predictor, andcached_embed_fn), shared with the .NET PR.Tests: 418 passed with models, 0 failed, 0 skipped; fmt and clippy clean.