Make config dataclasses the single source of truth for DEFAULT_CONFIG
CI / Format (ruff format) (push) Successful in 31s
CI / Lint (ruff check) (push) Successful in 31s
CI / Sync project version with tag (push) Has been skipped
CI / Lint (ruff check) (pull_request) Successful in 44s
CI / Type check (ty) (push) Successful in 46s
CI / Format (ruff format) (pull_request) Successful in 39s
CI / Sync project version with tag (pull_request) Has been skipped
CI / Type check (ty) (pull_request) Successful in 39s
CI / Tests (pull_request) Successful in 3m38s
CI / Tests (push) Successful in 3m45s

DEFAULT_CONFIG and build_models/build_critics/StageSpec.from_config's
inline .get(key, default) fallbacks had already drifted: two keys
(stage2_model.decoder, stage2_model.particle_type.target) resolved
differently depending on whether a config dict came from
merge_cli_overrides (fully populated, correct) or was hand-built and
partial (fell back to stale v0.2-shaped literals). Introduce frozen
dataclasses (GiantConfig and its nested blocks) in giant/config.py as
the actual single declaration of every default; DEFAULT_CONFIG is now
generated from them instead of hand-maintained, and build_models,
build_critics, and StageSpec.from_config consume the dataclasses
instead of duplicating literal fallbacks, so this class of drift can't
recur. Router/n_sec sub-blocks keep an `extra` catch-all for their
genuinely dynamic keys (composed-router axes, runtime-seeded
centers_init, legacy_owner).

Fixing the fallback surfaced the same latent bug in two existing
partial-config callers that had been silently depending on it: a
test fixture in test_train.py and scripts/warm_setup_cache.py's
minimal cfg (now merged against DEFAULT_CONFIG instead of hand-rolled,
closing the gap for good). See issues.md Issue 1.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
2026-08-12 14:35:14 +02:00
parent 55332db67a
commit 9bf5874308
8 changed files with 963 additions and 369 deletions
+56
View File
@@ -12,6 +12,7 @@ from giant.model.network import (
Stage1Model,
Stage2Autoregressive,
Stage2OneShot,
build_critics,
build_models,
cat_col_layout,
stage2_trunk_sec_dim,
@@ -652,3 +653,58 @@ def test_build_models_share_stages_true_shared_params_are_in_both_stage_paramete
assert shared_ids
assert shared_ids <= {id(p) for p in stage1.parameters()}
assert shared_ids <= {id(p) for p in stage2.parameters()}
# ── build_models/build_critics: DEFAULT_CONFIG fallback drift (issues.md #1) ─
def _partial_model_config() -> dict:
"""A hand-built model_config that omits stage2_model.decoder and
stage2_model.particle_type — deliberately not derived from
DEFAULT_CONFIG, unlike _minimal_model_config above. Regression fixture
for issues.md Issue 1: build_models/build_critics/StageSpec.from_config's
own fallback defaults for these two keys must equal DEFAULT_CONFIG's
("autoregressive" / "onehot"), not the old, now-wrong v0.2-shaped
("one_shot" / "physical") literals that used to live in three separate
.get(key, default) call sites."""
return {
"pdg_vocab": 3,
"mat_vocab": 2,
"conditioning": {
"particle": {"type": "physical", "emb_dim": 4, "n_layers": 1},
"material": {"type": "physical", "emb_dim": 4, "n_layers": 1},
},
"stage1_model": {"active": False},
"stage2_model": {
"generator": "wgan",
"hidden_dim": 8,
"n_res_blocks": 1,
"k_max": 3,
# decoder and particle_type deliberately omitted
},
}
def test_build_models_omitted_decoder_and_particle_type_match_default_config():
built = build_models(_partial_model_config())
assert isinstance(built["stage2"], Stage2Autoregressive)
assert built["stage2"].particle_type_cfg["target"] == "onehot"
def test_build_critics_omitted_particle_type_matches_default_config():
cfg = _partial_model_config()
cfg["stage2_model"]["generator"] = "wgan"
onehot_critic = build_critics(cfg)["stage2"]
assert onehot_critic is not None
onehot_in_dim = onehot_critic.input_proj.in_features
cfg["stage2_model"]["particle_type"] = {"target": "physical"}
physical_critic = build_critics(cfg)["stage2"]
assert physical_critic is not None
physical_in_dim = physical_critic.input_proj.in_features
# onehot's per-slot type width is emb_dim classes vs. physical's fixed
# (log-mass, charge) pair — different unless emb_dim happens to be 2, so
# this also confirms the critic was actually built in onehot mode by
# default, not silently falling back to physical.
assert onehot_in_dim != physical_in_dim