v0.3.0 step 1: new nested config schema, v0.2 migration shim

Replace the single global train.mode + [model] block with the four
top-level blocks docs/v0.3.0-design.md specifies ([conditioning],
[stage1_model], [stage2_model], [train]), so Stage 1 and Stage 2 can
run independent generative objectives and Stage 2 can train standalone.

- migrate_config translates old config.toml/checkpoint dicts on load,
  so nothing on /ceph goes dead; loudly rejects non-zero
  expert_hidden_dim/expert_n_blocks, which v0.3.0 no longer supports.
- merge_cli_overrides/save_config generalize from one hardcoded nesting
  level (model.router) to arbitrary recursive depth.
- default_out_dir_name candidates move to dotted paths against the new
  schema, with per-stage router/generator discriminators.
- validate_config adds cross-block checks the per-block schema can't
  express (particle_type.target=embedding needs a matching conditioning
  mode, tie_to_stage1 needs an active stage 1, etc).
- resolve_expert_dims is deleted (experts always inherit the stage's
  hidden_dim/n_res_blocks now) — pipeline.py/cli.py callers are left
  dangling on purpose, to be updated in the network.py/train.py steps
  that follow.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This commit is contained in:
2026-08-06 10:15:20 +02:00
parent a489991a3b
commit eb6dd27406
2 changed files with 1236 additions and 439 deletions
+661 -217
View File
@@ -1,3 +1,4 @@
import copy
import hashlib
import random
import subprocess
@@ -12,133 +13,287 @@ import torch
class Conditioning(str, Enum):
"""`model.conditioning` choices — shared by `giant.cli` and `scripts.dwarf`'s
Typer commands so the two CLIs can't silently drift apart on the option's
valid values (see DEFAULT_CONFIG["model"]["conditioning"] for what each
value means)."""
"""`conditioning.particle.type` / `conditioning.material.type` choices —
shared by `giant.cli` and `scripts.dwarf`'s Typer commands so the two
CLIs can't silently drift apart on the option's valid values (see
DEFAULT_CONFIG["conditioning"] for what each value means)."""
physical = "physical"
embedding = "embedding"
onehot = "onehot"
# Tags a config dict (config.toml, or a checkpoint's model_config) as the new
# v0.3 nested format. Absence of `[meta].config_version == CONFIG_VERSION` is
# read as "this is a v0.2 dict" by migrate_config below.
CONFIG_VERSION = 3
DEFAULT_CONFIG: dict = {
"train": {
"mode": "flow",
"epochs": 100,
"batch_size": 4096,
"lr": 3e-4,
"weight_decay": 0.01, # AdamW default — exposed so it can be tuned
"ema_decay": 0.9999, # EMA of model weights for sampling; 0 disables
# per-epoch val loss (not the marginal/KL validate_every pass) is
# capped to this many batches; 0 = full val set every epoch
"max_val_batches": 200,
"val_fraction": 0.1,
"num_workers": 4,
"seed": 0,
"validate_every": 10,
"validate_steps": 10,
"warmup_epochs": 5,
"lambda_nsec": 0.1,
"lambda_s2": 1.0,
# WGAN-GP-only knobs (mode == "wgan"; ignored by flow/ddpm). n_critic:
# critic updates per generator update. gp_weight: gradient-penalty
# coefficient (Gulrajani et al. 2017). critic_lr: 0.0 means "use
# `lr`" — not None, since save_config's TOML writer has no null
# literal to round-trip.
"n_critic": 5,
"gp_weight": 10.0,
"critic_lr": 0.0,
# Weights & Biases per-epoch metric logging (opt-in; see giant.train).
# "" for wandb_run_name means "use the checkpoint out_dir name" — not
# None, since save_config's TOML writer has no null literal to
# round-trip.
"wandb": False,
"wandb_project": "giant",
"wandb_run_name": "",
# Batch-granularity metrics (loss/grad_norm/lr) are logged every N
# optimizer steps, not every batch — a single epoch can be tens of
# thousands of steps (see steps_per_epoch above), and logging every
# one of them would flood the run with points the UI has to downsample
# anyway. Per-epoch metrics (the metrics.csv row) always log in full.
"wandb_log_every": 50,
"conditioning": {
# Width of the fused conditioning vector produced by the encoder's
# fusion MLP, consumed by every downstream trunk.
"out_dim": 128,
# false: stage 1 and stage 2 each construct their own ConditionEncoder
# with identical config but independent weights. true: one instance,
# shared by reference (halves the conditioning parameter count,
# forces a common representation).
"share_stages": False,
"particle": {
# "physical": a small MLP over log(mass)/charge, computable for
# any PDG code — generalizes beyond the training menu.
# "embedding": a learned nn.Embedding over a dense training-vocab
# index — memorizes the training menu; the generalization-
# comparison baseline, and the only mode compatible with
# stage2_model.particle_type.target = "embedding".
# "onehot": a fixed, unlearned vector — top (emb_dim - 1) PDG
# codes by training-set count, plus one "other" bin. NOT a
# reparameterization of "embedding": the vocabulary cap is the
# real difference.
"type": "physical",
# Width of this axis's vector. Under "onehot" this also sets the
# class count.
"emb_dim": 16,
# Depth of the sub-MLP under "physical". Ignored under
# "embedding"/"onehot".
"n_layers": 1,
},
"material": {
"type": "physical",
"emb_dim": 16,
"n_layers": 1,
},
},
"model": {
"stage1_model": {
# false skips building/training stage 1 entirely. The resulting
# checkpoint holds only stage 2 and cannot be rolled out.
"active": True,
# "flow": conditional flow matching (~10 ODE steps at inference).
# "ddpm": cosine-schedule diffusion baseline.
# "wgan": WGAN-GP, single forward pass at inference.
"generator": "flow",
# Trunk width — also the width of every expert under a routed trunk.
"hidden_dim": 256,
"n_blocks": 6,
"emb_dim": 16,
"dropout": 0.1,
# WGAN generator noise-vector width (mode == "wgan" only).
"noise_dim": 64,
# "physical" conditions on material/particle physical properties via
# a small MLP (giant.model.network.ConditionEncoder); "embedding"
# keeps the original learned pdg/material embedding tables — kept
# available as the generalization-comparison baseline. Checkpoints
# from before this option existed have no "conditioning" key and
# load as "embedding" (see giant.model.network.build_models).
"conditioning": "physical",
# Number of ResBlocks in the trunk, and in every expert under a
# routed trunk.
"n_res_blocks": 6,
"dropout": 0.0,
# Weight of this stage's loss in the total when both stages are
# active and non-adversarial. A WGAN stage's adversarial loss drives
# its own optimizer, so `lambda` scales only its non-adversarial
# auxiliary terms.
"lambda": 1.0,
"flow": {
# Width of the SinusoidalEmbedding for the flow time variable.
"time_dim": 64,
},
"ddpm": {
"time_dim": 64,
"n_steps": 1000,
},
"wgan": {
"noise_dim": 64,
"n_critic": 5,
"gp_weight": 10.0,
# 0.0 means "inherit train.lr" — not None, since the TOML writer
# has no null literal to round-trip.
"critic_lr": 0.0,
# 0 means "inherit stage1_model.hidden_dim/n_res_blocks" — same
# round-trip-friendly sentinel as critic_lr above.
"critic_hidden_dim": 0,
"critic_n_res_blocks": 0,
},
"router": {
"enabled": False,
"type": "energy", # selects the Router impl from ROUTER_REGISTRY
"n_experts": 4,
# 0 means "inherit model.hidden_dim/n_blocks" (see
# resolve_expert_dims below) — not a fixed 128/3, which silently
# ignored --hidden-dim/--n-blocks whenever routing was enabled.
# TOML has no null literal to round-trip (same pattern as
# critic_lr/wandb_run_name above), hence 0 rather than None.
"expert_hidden_dim": 0,
"expert_n_blocks": 0,
"temperature": 0.5, # energy/pdg-router kwarg
"learn_centers": True, # energy/pdg-router kwarg
# energy-router kwargs: mutually exclusive optional learnable
# gate-sharpness modes (see giant.model.network.EnergyRouter).
# learn_width generalizes the shared `temperature` to one
# learnable width per expert; learn_temperature instead makes
# the single shared `temperature` itself learnable. Both are
# bounded to [width_min_ratio, width_max_ratio] * temperature
# (sigmoid-parameterized, warm-started to reproduce `temperature`
# exactly at init) so gate sharpness can't run away to a
# collapse-inducing extreme during training.
# gate-sharpness modes. learn_width generalizes the shared
# `temperature` to one learnable width per expert;
# learn_temperature instead makes the single shared `temperature`
# itself learnable. Both are bounded to [width_min_ratio,
# width_max_ratio] * temperature so gate sharpness can't run away
# to a collapse-inducing extreme during training.
"learn_width": False,
"learn_temperature": False,
"width_min_ratio": 0.1,
"width_max_ratio": 10.0,
"lambda_balance": 0.0, # optional load-balance aux loss weight
# optional entropy-regularization aux loss weight (generic
# Router.entropy_loss, penalizes uniform/collapsed gating) — a
# secondary guard against all experts' widths/temperature
# co-inflating together, which lambda_balance alone can't see
# since per-expert usage shares stay even throughout that
# failure mode. Off by default; bounding above is the primary
# defense. See giant.model.network.Router.entropy_loss.
# Importance-CV^2 load-balancing aux loss weight (Shazeer et al.
# 2017).
"lambda_balance": 0.0,
# Entropy-regularization weight penalizing uniform/collapsed
# gating — a secondary guard against all experts' widths
# co-inflating together, which lambda_balance alone can't see.
"lambda_entropy": 0.0,
# Opt-in straight-through Gumbel-softmax train-time combine weights
# (see giant.model.network.Router.combine_weights): the training
# forward pass samples a hard one-hot combination — matching
# eval-time top-1 dispatch exactly — while the backward pass still
# flows a smooth gradient to every expert. Targets the train/eval
# mismatch identified as a likely contributor to experts
# overlapping instead of partitioning (see CLAUDE.md roadmap).
# gumbel_tau_start/_end are annealed linearly over training
# (giant.train._gumbel_tau); off by default, no effect unless
# gumbel = true.
# Opt-in straight-through Gumbel-softmax train-time combine
# weights: the training forward pass samples a hard one-hot
# combination (matching eval-time top-1 dispatch exactly) while
# the backward pass still flows smooth gradient to every expert.
"gumbel": False,
"gumbel_tau_start": 1.0,
"gumbel_tau_end": 0.1,
"emb_dim": 8, # process/pdg-router kwarg: own pdg(/mat) embedding width
"hidden_dim": 64, # process-router kwarg: its classifier's hidden width
"lambda_proc": 0.0, # process-router kwarg: supervised process-CE weight
# (0.0 still trains a working router — the gate gets gradient
# through the downstream flow loss like EnergyRouter's centers —
# but only lambda_proc > 0 grounds it in the true `process` label)
# type = "composed" routes on multiple axes at once (e.g. energy x
# pdg), each with its own expert count/hyperparameters. Axes are
# NOT in these defaults (there's no meaningful default axis list)
# — set them as flat axis{i}_{field} keys instead of "n_experts",
# e.g. axis0_type = "energy", axis0_n_experts = 4, axis1_type =
# "pdg", axis1_n_experts = 3, axis1_emb_dim = 8. See
# giant.model.network._parse_composed_axes / `--router-axis`.
# NOT in these defaults — set them as flat axis{i}_{field} keys
# instead of "n_experts", e.g. axis0_type = "energy",
# axis0_n_experts = 4, axis1_type = "pdg", axis1_n_experts = 3,
# axis1_emb_dim = 8. See giant.model.network._parse_composed_axes.
},
},
"stage2_model": {
# false trains stage 1 alone. giant rollout must then refuse the
# checkpoint; giant predict still works.
"active": True,
# "one_shot": predict all k_max slots simultaneously with padded
# slots masked from the loss (v0.2 behaviour).
# "autoregressive": emit one secondary at a time in descending-energy
# order.
"decoder": "autoregressive",
# As stage1_model.generator, but under "autoregressive" this is the
# objective for each token.
"generator": "wgan",
"hidden_dim": 256,
"n_res_blocks": 6,
"dropout": 0.0,
"lambda": 1.0,
# Maximum secondary slots. Under "one_shot" this is the fixed output
# width; under "autoregressive" it is a safety cap on the generation
# loop.
"k_max": 15,
# Width of the projected stage-1 outcome fed into stage 2's
# conditioning.
"context_dim": 64,
# "truth": the ground-truth stage-1 target vector, detached — stage-
# level teacher forcing (v0.2 behaviour). "sampled": stage 1's own
# sampled output, closing the train/inference gap at the cost of a
# sampling pass per batch and a moving target early in training.
"stage1_context": "truth",
"n_sec": {
# "head": a classifier over {0..k_max} on the condition encoding
# alone (no diffusion noise), callable independently at
# inference.
# "stop_token": an EOS-style implicit stop — accepted by the
# schema but not implemented in v0.3.0 (see validate_config).
# "truth": take n_sec from ground truth — standalone stage-2
# evaluation only, never for rollout.
"mode": "head",
"lambda": 0.1, # cross-entropy weight for the head
},
"particle_type": {
# The three targets mirror the three conditioning.particle
# modes. "onehot": class logits over conditioning.particle.emb_dim
# classes. "physical": regressed (log mass, charge). "embedding":
# regressed against conditioning's own particle embedding table
# (requires conditioning.particle.type = "embedding").
"target": "onehot",
"lambda": 1.0,
# How a predicted "other" class becomes a concrete PDG code at
# rollout. "sample": draw from the empirical within-bucket
# distribution recorded at map-build time. "modal": always the
# most common member. "drop": discard the secondary. Read only
# under target = "onehot".
"other_policy": "sample",
},
"autoregressive": {
# Canonical generation order. Single-valued for now; the key
# exists so an alternative ordering is not a config break.
"order": "energy_desc",
# How token i+1 sees tokens <= i. "markov": previous token plus
# running scalars (remaining energy budget, slot index) — a
# fixed-width summary. "attention": causal self-attention over
# all emitted tokens.
"history": "markov",
# "always": condition on the ground-truth previous secondary
# throughout training. "scheduled": scheduled sampling —
# interpolate toward the model's own prediction. "never":
# free-running from the start.
"teacher_forcing": "always",
"tf_p_start": 1.0,
"tf_p_end": 1.0,
"attn_n_heads": 4,
"attn_n_layers": 2,
},
"flow": {
"time_dim": 64,
},
"ddpm": {
"time_dim": 64,
"n_steps": 1000,
},
"wgan": {
"noise_dim": 64,
"n_critic": 5,
"gp_weight": 10.0,
"critic_lr": 0.0,
"critic_hidden_dim": 0,
"critic_n_res_blocks": 0,
# Straight-through Gumbel temperature for the particle-type
# one-hot (distinct from router.gumbel_tau_start/_end, which
# anneal expert-combination weights). Read only under
# particle_type.target = "onehot".
"gumbel_tau_start": 1.0,
"gumbel_tau_end": 0.1,
},
"router": {
# true: stage 2 shares stage 1's Router module instance, so
# expert i in stage 1 and expert i in stage 2 gate on identical
# conditions by construction — every other key in this block is
# then ignored. false: an independent router.
"tie_to_stage1": False,
"enabled": False,
"type": "energy",
"n_experts": 4,
"temperature": 0.5,
"learn_centers": True,
"learn_width": False,
"learn_temperature": False,
"width_min_ratio": 0.1,
"width_max_ratio": 10.0,
"lambda_balance": 0.0,
"lambda_entropy": 0.0,
"lambda_proc": 0.0,
"gumbel": False,
"gumbel_tau_start": 1.0,
"gumbel_tau_end": 0.1,
"emb_dim": 8,
"hidden_dim": 64,
},
},
"train": {
"epochs": 100,
"batch_size": 4096,
"lr": 3e-4,
"weight_decay": 0.01, # AdamW default — exposed so it can be tuned
"ema_decay": 0.9999, # EMA of model weights for sampling; 0 disables
"warmup_epochs": 5,
"val_fraction": 0.1,
# per-epoch val loss (not the marginal/KL validate_every pass) is
# capped to this many batches; 0 = full val set every epoch
"max_val_batches": 200,
"num_workers": 4,
"seed": 0,
"validate_every": 10,
"validate_steps": 10,
# Weights & Biases per-epoch metric logging. Default true in v0.3.0
# (was opt-in false): the v0.3.0 work is a sequence of architecture
# comparisons, and a run that wasn't logged isn't comparable. Set
# false for throwaway/debug runs.
"wandb": True,
"wandb_project": "giant",
# "" means "use the checkpoint out_dir name" — not None, since the
# TOML writer has no null literal to round-trip.
"wandb_run_name": "",
# Batch-granularity metrics (loss/grad_norm/lr) are logged every N
# optimizer steps, not every batch — a single epoch can be tens of
# thousands of steps. Per-epoch metrics (the metrics.csv row) always
# log in full.
"wandb_log_every": 50,
},
}
@@ -169,6 +324,9 @@ def auto_device() -> torch.device:
# Activation memory is assumed to scale linearly with
# batch_size * hidden_dim * n_blocks (the ResBlock stack dominates), so this
# is a rough estimate rather than a guaranteed bound.
# NOTE: not yet recalibrated for the v0.3.0 autoregressive stage-2 trunk —
# see docs/v0.3.0-design.md §9/§11.3, deliberately last in the implementation
# order.
_REF_BYTES = 7683 * 1024**2
_REF_BATCH_SIZE = 29696
_REF_HIDDEN_DIM = 1024
@@ -254,13 +412,15 @@ def warn_if_git_hash_mismatch(file_cfg: dict, config_path: Path) -> None:
def load_checkpoint_config(ckpt_path: str | Path) -> dict:
"""Load the full ``[train]``/``[model]``/``[meta]`` config.toml written
alongside a checkpoint by ``save_config``.
"""Load the full config.toml written alongside a checkpoint by
``save_config``.
Returns ``{}`` if no config.toml sits next to the checkpoint (older runs,
or a checkpoint moved without its sidecar) — this is best-effort
provenance for threading into a rollout's YAML sidecar, not a hard
requirement for using the checkpoint itself.
requirement for using the checkpoint itself. Returned as-loaded (v0.2 or
v0.3 shape); callers that need the v0.3 shape should run it through
`migrate_config` themselves.
"""
config_path = Path(ckpt_path).parent / "config.toml"
if not config_path.exists():
@@ -283,130 +443,413 @@ def warn_if_checkpoint_config_mismatch(ckpt_path: str | Path) -> None:
warn_if_git_hash_mismatch(load_toml(config_path), config_path)
def _get_path(d: dict, dotted: str):
"""Read a dotted path (e.g. "stage1_model.router.enabled") out of a
nested dict. Returns None if any component along the path is missing."""
cur = d
for part in dotted.split("."):
if not isinstance(cur, dict) or part not in cur:
return None
cur = cur[part]
return cur
def _set_path(d: dict, dotted: str, value) -> None:
"""Write a dotted path into a nested dict, creating intermediate dicts as
needed."""
parts = dotted.split(".")
cur = d
for part in parts[:-1]:
cur = cur.setdefault(part, {})
cur[parts[-1]] = value
def _deep_merge(base: dict, override: dict) -> dict:
"""Recursively merge `override` onto a copy of `base`.
Dict-valued keys recurse instead of being replaced wholesale, so
overriding one leaf (e.g. stage1_model.router.enabled) never drops
untouched siblings — the rest of stage1_model.router, or of
stage1_model — the same property v0.2's router-only bespoke merge had,
generalized here to arbitrary depth.
"""
result = dict(base)
for k, v in override.items():
if isinstance(v, dict) and isinstance(result.get(k), dict):
result[k] = _deep_merge(result[k], v)
else:
result[k] = v
return result
# v0.2 [train] keys that pass through to v0.3 [train] unchanged (same name,
# same meaning) when present in the loaded file — everything model-shaped
# moved to the stage/conditioning blocks instead (see the rest of
# migrate_config below).
_V02_TRAIN_PASSTHROUGH = (
"epochs",
"batch_size",
"lr",
"weight_decay",
"ema_decay",
"max_val_batches",
"val_fraction",
"num_workers",
"seed",
"validate_every",
"validate_steps",
"warmup_epochs",
"wandb",
"wandb_project",
"wandb_run_name",
"wandb_log_every",
)
# v0.2 model.hidden_dim/n_blocks/dropout applied identically to both stages
# (there was only ever one trunk shape) -> copied to both stage{1,2}_model.
_V02_MODEL_TO_BOTH_STAGES = (
("hidden_dim", "hidden_dim"),
("n_blocks", "n_res_blocks"),
("dropout", "dropout"),
)
# v0.2 train.{n_critic,gp_weight,critic_lr} applied identically to both
# stages' wgan sub-table (there was only ever one wgan objective, shared).
_V02_TRAIN_TO_BOTH_STAGES_WGAN = (
("n_critic", "n_critic"),
("gp_weight", "gp_weight"),
("critic_lr", "critic_lr"),
)
def migrate_config(cfg: dict) -> dict:
"""Translate a v0.2 config dict (single [train] + [model]) into the v0.3
nested format ([conditioning]/[stage1_model]/[stage2_model]/[train]).
Called on every config.toml load (see merge_cli_overrides) so old
training configs on disk keep working under new code without hand-
editing (decision 3, docs/v0.3.0-design.md §4). `[meta].config_version ==
CONFIG_VERSION` marks a dict as already-v0.3; its absence is read as
"this is v0.2" (the design doc's stated rule), so an already-migrated
dict is returned unchanged (deep-copied).
Only keys actually present in `cfg` are translated — `cfg` may be a
partial file (e.g. `[train]\\nepochs = 5\\n` with no [model] section at
all, relying on v0.2 defaults for everything else). Separately, a fixed
set of v0.2 architectural facts that were never exposed as config keys at
all (e.g. the conditioning MLP was always 2 layers deep, not the v0.3
default of 1) are injected unconditionally whenever this function decides
it is migrating a v0.2 dict, regardless of which keys the file happened
to set.
Operates on the config.toml shape. A checkpoint's `model_config` dict
(which additionally carries n_sec_head ownership, §4.1, and needs
`network.build_models`'s cooperation) is a separate migration surface,
deferred to the network.py refactor.
"""
if _get_path(cfg, "meta.config_version") == CONFIG_VERSION:
return copy.deepcopy(cfg)
cfg = copy.deepcopy(cfg)
old_train = cfg.pop("train", {})
old_model = cfg.pop("model", {})
old_router = dict(old_model.pop("router", {}))
new: dict = {}
for key in _V02_TRAIN_PASSTHROUGH:
if key in old_train:
_set_path(new, f"train.{key}", old_train[key])
if "mode" in old_train:
_set_path(new, "stage1_model.generator", old_train["mode"])
_set_path(new, "stage2_model.generator", old_train["mode"])
if "lambda_nsec" in old_train:
_set_path(new, "stage2_model.n_sec.lambda", old_train["lambda_nsec"])
if "lambda_s2" in old_train:
_set_path(new, "stage2_model.lambda", old_train["lambda_s2"])
for old_key, new_key in _V02_TRAIN_TO_BOTH_STAGES_WGAN:
if old_key in old_train:
_set_path(new, f"stage1_model.wgan.{new_key}", old_train[old_key])
_set_path(new, f"stage2_model.wgan.{new_key}", old_train[old_key])
for old_key, new_key in _V02_MODEL_TO_BOTH_STAGES:
if old_key in old_model:
_set_path(new, f"stage1_model.{new_key}", old_model[old_key])
_set_path(new, f"stage2_model.{new_key}", old_model[old_key])
if "emb_dim" in old_model:
_set_path(new, "conditioning.particle.emb_dim", old_model["emb_dim"])
_set_path(new, "conditioning.material.emb_dim", old_model["emb_dim"])
if "conditioning" in old_model:
_set_path(new, "conditioning.particle.type", old_model["conditioning"])
_set_path(new, "conditioning.material.type", old_model["conditioning"])
if "noise_dim" in old_model:
_set_path(new, "stage1_model.wgan.noise_dim", old_model["noise_dim"])
_set_path(new, "stage2_model.wgan.noise_dim", old_model["noise_dim"])
if "k_max" in old_model:
_set_path(new, "stage2_model.k_max", old_model["k_max"])
if old_router:
expert_hidden_dim = old_router.pop("expert_hidden_dim", 0)
expert_n_blocks = old_router.pop("expert_n_blocks", 0)
if expert_hidden_dim or expert_n_blocks:
raise ValueError(
"v0.2 config sets model.router.expert_hidden_dim/"
f"expert_n_blocks to a non-default value "
f"({expert_hidden_dim!r}, {expert_n_blocks!r}); v0.3.0 removed "
"per-expert sizing (experts always inherit the stage's "
"hidden_dim/n_res_blocks), so this config's routed experts "
"have a different width/depth than the monolith and its "
"checkpoint can only be loaded by v0.2 code."
)
_set_path(new, "stage1_model.router", dict(old_router))
stage2_router = dict(old_router)
stage2_router["tie_to_stage1"] = False
_set_path(new, "stage2_model.router", stage2_router)
# v0.2 architectural facts with no corresponding config key at all —
# always set once we've determined we're migrating a v0.2 dict,
# independent of what the file did/didn't specify. NOTE: n_layers here
# (2) differs from the v0.3 *default* (1) — this is not a typo, see the
# docstring above.
_set_path(new, "conditioning.out_dim", 128)
_set_path(new, "conditioning.particle.n_layers", 2)
_set_path(new, "conditioning.material.n_layers", 2)
_set_path(new, "stage1_model.active", True)
_set_path(new, "stage1_model.flow.time_dim", 64)
_set_path(new, "stage1_model.ddpm.time_dim", 64)
_set_path(new, "stage2_model.active", True)
_set_path(new, "stage2_model.flow.time_dim", 64)
_set_path(new, "stage2_model.ddpm.time_dim", 64)
_set_path(new, "stage2_model.context_dim", 64)
_set_path(new, "stage2_model.decoder", "one_shot")
_set_path(new, "stage2_model.particle_type.target", "physical")
new_meta = dict(cfg.pop("meta", {}))
new_meta["config_version"] = CONFIG_VERSION
new["meta"] = new_meta
# Anything else in the original dict (unrecognized top-level sections)
# carries through untouched rather than being silently dropped.
for k, v in cfg.items():
new.setdefault(k, v)
return new
def merge_cli_overrides(
defaults: dict,
config_path: Path | None,
train_overrides: dict,
model_overrides: dict,
overrides: dict,
) -> dict:
"""Resolve config as defaults -> TOML file -> explicit CLI flags.
"""Resolve config as defaults -> TOML file -> explicit overrides.
`model.router` is deep-merged one level (rather than replaced wholesale)
at each stage, so a TOML file or CLI flag only overriding e.g.
`router.enabled` doesn't drop the rest of the router defaults.
`overrides` is keyed by top-level section name (e.g. "stage1_model",
"train"), each value an arbitrarily nested dict of overrides to
deep-merge (see `_deep_merge`) — the shape stage-prefixed CLI flags
naturally produce. A v0.2-shaped TOML file is transparently migrated
(`migrate_config`) before merging, so old configs on disk keep working
under the new schema.
"""
cfg = {"train": dict(defaults["train"]), "model": dict(defaults["model"])}
cfg["model"]["router"] = dict(defaults["model"]["router"])
cfg = copy.deepcopy(defaults)
if config_path is not None:
file_cfg = load_toml(config_path)
cfg["train"].update(file_cfg.get("train", {}))
file_model = dict(file_cfg.get("model", {}))
file_router = file_model.pop("router", None)
cfg["model"].update(file_model)
if file_router:
cfg["model"]["router"].update(file_router)
file_cfg = migrate_config(load_toml(config_path))
for section, values in file_cfg.items():
if section == "meta":
continue
if isinstance(values, dict):
cfg[section] = _deep_merge(cfg.get(section, {}), values)
else:
cfg[section] = values
warn_if_git_hash_mismatch(file_cfg, config_path)
model_overrides = dict(model_overrides)
router_overrides = model_overrides.pop("router", None)
cfg["train"].update(train_overrides)
cfg["model"].update(model_overrides)
if router_overrides:
cfg["model"]["router"].update(router_overrides)
for section, values in overrides.items():
if isinstance(values, dict):
cfg[section] = _deep_merge(cfg.get(section, {}), values)
else:
cfg[section] = values
return cfg
def resolve_expert_dims(
router_cfg: dict, hidden_dim: int, n_blocks: int
) -> tuple[int, int]:
"""Resolve a router's expert hidden_dim/n_blocks, inheriting from the
monolith's when left at the 0 ("unset") sentinel.
def validate_config(cfg: dict) -> None:
"""Cross-block validation the per-block schema can't express on its own.
Used by both `giant.pipeline` (to build the checkpoint's `model_config`)
and `giant.cli`'s batch-size auto-estimate, so `--hidden-dim`/`--n-blocks`
size the experts the same way in both places unless
`router.expert_hidden_dim`/`expert_n_blocks` are explicitly overridden.
Raises ValueError with a clear message on the first violation found. Call
after `merge_cli_overrides` has produced a fully-merged v0.3 config —
these checks need to see across blocks, so they don't belong in
`migrate_config` (which only ever sees one dict's own keys) or in any
single block's defaults.
"""
expert_hidden_dim = router_cfg.get("expert_hidden_dim") or hidden_dim
expert_n_blocks = router_cfg.get("expert_n_blocks") or n_blocks
return expert_hidden_dim, expert_n_blocks
particle_type = _get_path(cfg, "conditioning.particle.type")
pt_target = _get_path(cfg, "stage2_model.particle_type.target")
if pt_target == "embedding" and particle_type != "embedding":
raise ValueError(
"stage2_model.particle_type.target = 'embedding' requires "
"conditioning.particle.type = 'embedding' (there is no embedding "
"table to regress against under conditioning.particle.type = "
f"{particle_type!r})"
)
for stage_name in ("stage1_model", "stage2_model"):
router = _get_path(cfg, f"{stage_name}.router") or {}
if (
router.get("enabled")
and router.get("type") in ("pdg", "process")
and particle_type == "physical"
):
raise ValueError(
f"{stage_name}.router.type = {router['type']!r} builds its "
"own training-vocab-scoped embedding, incompatible with "
"conditioning.particle.type = 'physical' (defeats "
"generalization beyond the training menu) — pick a "
"different router type or a different "
"conditioning.particle.type"
)
if _get_path(cfg, "stage2_model.router.tie_to_stage1") and not _get_path(
cfg, "stage1_model.active"
):
raise ValueError(
"stage2_model.router.tie_to_stage1 = true requires "
"stage1_model.active = true (there is no stage-1 router to tie to)"
)
if _get_path(cfg, "stage2_model.n_sec.mode") == "stop_token":
raise ValueError(
"stage2_model.n_sec.mode = 'stop_token' is accepted by the schema "
"but not implemented in v0.3.0 — use 'head' (default) or 'truth' "
"(standalone stage-2 evaluation only, never for rollout)"
)
_CONDITIONING_CODE = {"physical": "phys", "embedding": "emb"}
_CONDITIONING_CODE = {"physical": "phys", "embedding": "emb", "onehot": "oh"}
# Priority-ordered candidate fields for default_out_dir_name: (label, getter,
# formatter). `getter(train, model)` returns None when the field is at its
# default (and so should be omitted); otherwise formatter(value) renders the
# name token. The router is a single unit gated on `router.enabled` rather
# than one candidate per router key, since its type/n_experts are meaningless
# while disabled.
def _mode_candidate(train, model):
return None if train["mode"] == DEFAULT_CONFIG["train"]["mode"] else train["mode"]
def _path_candidate(dotted_path: str, prefix: str, formatter=str):
"""Candidate factory: show `prefix + formatter(value)` when the value at
`dotted_path` differs from its DEFAULT_CONFIG value, else omit."""
def _router_candidate(train, model):
router = model["router"]
if router["enabled"] == DEFAULT_CONFIG["model"]["router"]["enabled"]:
return None
return f"r-{router['type']}{router['n_experts']}"
def _router_flag_candidate(field, token_map):
"""Candidate factory for a boolean `model.router` sub-field.
Gated on `router.enabled` like `_router_candidate` (a disabled router's
sub-fields are meaningless), then omitted unless `field` differs from
its DEFAULT_CONFIG value — same "only show non-default" rule as every
other candidate. `token_map` need only cover the non-default value(s),
since the default value always yields None.
"""
def _candidate(train, model):
router = model["router"]
default_router = DEFAULT_CONFIG["model"]["router"]
if router["enabled"] == default_router["enabled"]:
def _candidate(cfg):
value = _get_path(cfg, dotted_path)
default = _get_path(DEFAULT_CONFIG, dotted_path)
if value == default:
return None
value = router[field]
if value == default_router[field]:
return None
return token_map[value]
return f"{prefix}{formatter(value)}"
return _candidate
def _conditioning_candidate(train, model):
if model["conditioning"] == DEFAULT_CONFIG["model"]["conditioning"]:
return None
code = _CONDITIONING_CODE.get(model["conditioning"], model["conditioning"])
return f"c{code}"
def _default_field_candidate(section_key, field, prefix):
def _candidate(train, model):
section = train if section_key == "train" else model
value = section[field]
if value == DEFAULT_CONFIG[section_key][field]:
def _conditioning_candidate(axis: str, short: str):
def _candidate(cfg):
value = _get_path(cfg, f"conditioning.{axis}.type")
default = _get_path(DEFAULT_CONFIG, f"conditioning.{axis}.type")
if value == default:
return None
return f"{prefix}{value}"
code = _CONDITIONING_CODE.get(value, value)
return f"{short}{code}"
return _candidate
def _router_candidate(stage_key: str, short: str):
"""Candidate for a stage's router as a single unit, gated on
`router.enabled` (a disabled router's type/n_experts are meaningless)."""
def _candidate(cfg):
router = _get_path(cfg, f"{stage_key}.router") or {}
default_router = _get_path(DEFAULT_CONFIG, f"{stage_key}.router") or {}
if router.get("enabled") == default_router.get("enabled"):
return None
return f"{short}r-{router['type']}{router['n_experts']}"
return _candidate
def _router_flag_candidate(stage_key: str, short: str, field: str, token_map: dict):
"""Candidate factory for a boolean field inside a stage's router block.
Gated on `router.enabled` like `_router_candidate`, then omitted unless
`field` differs from its DEFAULT_CONFIG value. `token_map` need only
cover the non-default value(s), since the default value always yields
None.
"""
def _candidate(cfg):
router = _get_path(cfg, f"{stage_key}.router") or {}
default_router = _get_path(DEFAULT_CONFIG, f"{stage_key}.router") or {}
if router.get("enabled") == default_router.get("enabled"):
return None
value = router.get(field)
if value == default_router.get(field):
return None
return f"{short}{token_map[value]}"
return _candidate
# Priority-ordered candidate fields for default_out_dir_name: (label,
# candidate(cfg) -> str | None). Beyond _OUT_DIR_NAME_MAX_FIELDS non-default
# fields, the remainder collapse into a hash suffix (see
# default_out_dir_name). Candidates read the whole nested cfg via dotted
# paths — there is no single "model" dict anymore now that architecture is
# split across conditioning/stage1_model/stage2_model.
_OUT_DIR_NAME_CANDIDATES = [
("mode", _mode_candidate),
("router", _router_candidate),
("gumbel", _router_flag_candidate("gumbel", {True: "gum"})),
("learn_centers", _router_flag_candidate("learn_centers", {False: "nolc"})),
("learn_width", _router_flag_candidate("learn_width", {True: "lw"})),
("learn_temperature", _router_flag_candidate("learn_temperature", {True: "lt"})),
("conditioning", _conditioning_candidate),
("hidden_dim", _default_field_candidate("model", "hidden_dim", "h")),
("n_blocks", _default_field_candidate("model", "n_blocks", "b")),
("emb_dim", _default_field_candidate("model", "emb_dim", "e")),
("lr", _default_field_candidate("train", "lr", "lr")),
("batch_size", _default_field_candidate("train", "batch_size", "bs")),
("seed", _default_field_candidate("train", "seed", "seed")),
("epochs", _default_field_candidate("train", "epochs", "ep")),
("stage1_generator", _path_candidate("stage1_model.generator", "")),
("stage2_generator", _path_candidate("stage2_model.generator", "s2-")),
("stage2_decoder", _path_candidate("stage2_model.decoder", "dec-")),
(
"stage2_history",
_path_candidate("stage2_model.autoregressive.history", "hist-"),
),
(
"particle_type_target",
_path_candidate("stage2_model.particle_type.target", "pt-"),
),
("stage1_router", _router_candidate("stage1_model", "s1")),
("stage2_router", _router_candidate("stage2_model", "s2")),
(
"stage1_gumbel",
_router_flag_candidate("stage1_model", "s1", "gumbel", {True: "gum"}),
),
(
"stage2_gumbel",
_router_flag_candidate("stage2_model", "s2", "gumbel", {True: "gum"}),
),
(
"stage1_learn_centers",
_router_flag_candidate("stage1_model", "s1", "learn_centers", {False: "nolc"}),
),
(
"stage2_learn_centers",
_router_flag_candidate("stage2_model", "s2", "learn_centers", {False: "nolc"}),
),
(
"stage1_learn_width",
_router_flag_candidate("stage1_model", "s1", "learn_width", {True: "lw"}),
),
(
"stage2_learn_width",
_router_flag_candidate("stage2_model", "s2", "learn_width", {True: "lw"}),
),
(
"stage1_learn_temperature",
_router_flag_candidate("stage1_model", "s1", "learn_temperature", {True: "lt"}),
),
(
"stage2_learn_temperature",
_router_flag_candidate("stage2_model", "s2", "learn_temperature", {True: "lt"}),
),
("particle_conditioning", _conditioning_candidate("particle", "c")),
("material_conditioning", _conditioning_candidate("material", "m")),
("stage1_hidden_dim", _path_candidate("stage1_model.hidden_dim", "h")),
("stage2_hidden_dim", _path_candidate("stage2_model.hidden_dim", "s2h")),
("stage1_n_res_blocks", _path_candidate("stage1_model.n_res_blocks", "b")),
("stage2_n_res_blocks", _path_candidate("stage2_model.n_res_blocks", "s2b")),
("particle_emb_dim", _path_candidate("conditioning.particle.emb_dim", "e")),
("lr", _path_candidate("train.lr", "lr")),
("batch_size", _path_candidate("train.batch_size", "bs")),
("seed", _path_candidate("train.seed", "seed")),
("epochs", _path_candidate("train.epochs", "ep")),
]
_OUT_DIR_NAME_MAX_FIELDS = 6
@@ -424,11 +867,10 @@ def default_out_dir_name(cfg: dict, now: datetime | None = None) -> str:
giant.train), which is the reason a timestamp is always included.
"""
now = now or datetime.now()
train, model = cfg["train"], cfg["model"]
tokens = []
overflow = []
for label, candidate in _OUT_DIR_NAME_CANDIDATES:
token = candidate(train, model)
token = candidate(cfg)
if token is None:
continue
if len(tokens) < _OUT_DIR_NAME_MAX_FIELDS:
@@ -476,30 +918,31 @@ def _toml_value(v) -> str:
return str(v)
def save_config(cfg: dict, out_dir: Path, meta: dict) -> None:
lines = []
# One-level-nested dict values (e.g. model.router) are rendered as their
# own [section.subsection] table after the parent section, since TOML
# doesn't accept a bare dict as a `key = value` scalar line.
nested_sections: list[tuple[str, dict]] = []
for section, values in cfg.items():
lines.append(f"[{section}]")
for k, v in values.items():
if isinstance(v, dict):
nested_sections.append((f"{section}.{k}", v))
continue
lines.append(f"{k:<14} = {_toml_value(v)}")
lines.append("")
def _write_section(lines: list[str], path: str, values: dict) -> None:
"""Write one TOML table (`[path]`) and recurse depth-first into any
dict-valued keys as `[path.subkey]` — handles the v0.3 schema's 2-3 level
nesting (e.g. stage1_model.router, stage2_model.n_sec) with no depth
limit, unlike the one-level-only writer this replaces."""
lines.append(f"[{path}]")
nested: list[tuple[str, dict]] = []
for k, v in values.items():
if isinstance(v, dict):
nested.append((f"{path}.{k}", v))
else:
lines.append(f"{k:<18} = {_toml_value(v)}")
lines.append("")
for sub_path, sub_values in nested:
_write_section(lines, sub_path, sub_values)
for name, values in nested_sections:
lines.append(f"[{name}]")
for k, v in values.items():
lines.append(f"{k:<14} = {_toml_value(v)}")
lines.append("")
def save_config(cfg: dict, out_dir: Path, meta: dict) -> None:
lines: list[str] = []
for section, values in cfg.items():
_write_section(lines, section, values)
lines.append("[meta]")
for k, v in meta.items():
lines.append(f"{k:<14} = {_toml_value(v)}")
lines.append(f"{k:<18} = {_toml_value(v)}")
(out_dir / "config.toml").write_text("\n".join(lines))
@@ -514,6 +957,7 @@ def build_run_meta(
n_train_steps: int,
) -> dict:
return {
"config_version": CONFIG_VERSION,
"git_hash": git_hash(),
"seed": seed,
"timestamp_utc": datetime.now(timezone.utc).isoformat(timespec="seconds"),
+575 -222
View File
@@ -1,46 +1,314 @@
from datetime import datetime
from pathlib import Path
from giant import config as gconfig
_CONFIGS_DIR = Path(__file__).resolve().parents[1] / "configs"
def _write_config(path, git_hash):
path.write_text(
f"""
[train]
epochs = 5
[model]
hidden_dim = 64
def _write_toml(path, git_hash=None, extra=""):
meta = f'\n[meta]\ngit_hash = "{git_hash}"\n' if git_hash is not None else ""
path.write_text(extra + meta)
[meta]
git_hash = "{git_hash}"
"""
# ---------------------------------------------------------------------------
# Conditioning enum
# ---------------------------------------------------------------------------
def test_conditioning_enum_has_onehot():
assert gconfig.Conditioning.onehot == "onehot"
assert {c.value for c in gconfig.Conditioning} == {
"physical",
"embedding",
"onehot",
}
# ---------------------------------------------------------------------------
# _deep_merge
# ---------------------------------------------------------------------------
def test_deep_merge_leaf_override_keeps_untouched_siblings():
base = {"a": 1, "b": {"c": 2, "d": 3}}
result = gconfig._deep_merge(base, {"b": {"c": 99}})
assert result == {"a": 1, "b": {"c": 99, "d": 3}}
def test_deep_merge_recurses_at_multiple_levels():
base = {
"stage1_model": {
"hidden_dim": 256,
"router": {"enabled": False, "type": "energy", "n_experts": 4},
}
}
result = gconfig._deep_merge(base, {"stage1_model": {"router": {"enabled": True}}})
assert result["stage1_model"]["hidden_dim"] == 256
assert result["stage1_model"]["router"] == {
"enabled": True,
"type": "energy",
"n_experts": 4,
}
def test_deep_merge_does_not_mutate_base():
base = {"a": {"b": 1}}
gconfig._deep_merge(base, {"a": {"b": 2}})
assert base == {"a": {"b": 1}}
def test_deep_merge_non_dict_override_replaces_wholesale():
base = {"a": {"b": 1}}
result = gconfig._deep_merge(base, {"a": 5})
assert result == {"a": 5}
# ---------------------------------------------------------------------------
# migrate_config
# ---------------------------------------------------------------------------
def test_migrate_config_already_v3_returned_unchanged():
cfg = {"meta": {"config_version": 3}, "stage1_model": {"generator": "flow"}}
result = gconfig.migrate_config(cfg)
assert result == cfg
result["stage1_model"]["generator"] = "wgan"
assert cfg["stage1_model"]["generator"] == "flow" # deep-copied, not aliased
def test_migrate_config_empty_dict_still_injects_hardcoded_v02_facts():
# No [train]/[model] at all still counts as "v0.2" (config_version
# absent) — the hardcoded architectural facts fire unconditionally.
new = gconfig.migrate_config({})
assert "train" not in new
assert new["conditioning"]["out_dim"] == 128
assert new["conditioning"]["particle"]["n_layers"] == 2
assert new["conditioning"]["material"]["n_layers"] == 2
assert new["stage1_model"]["active"] is True
assert new["stage1_model"]["flow"]["time_dim"] == 64
assert new["stage1_model"]["ddpm"]["time_dim"] == 64
assert new["stage2_model"]["active"] is True
assert new["stage2_model"]["flow"]["time_dim"] == 64
assert new["stage2_model"]["ddpm"]["time_dim"] == 64
assert new["stage2_model"]["context_dim"] == 64
assert new["stage2_model"]["decoder"] == "one_shot"
assert new["stage2_model"]["particle_type"]["target"] == "physical"
assert new["meta"] == {"config_version": 3}
def test_migrate_config_mode_maps_to_both_stage_generators():
new = gconfig.migrate_config({"train": {"mode": "wgan"}})
assert new["stage1_model"]["generator"] == "wgan"
assert new["stage2_model"]["generator"] == "wgan"
def test_migrate_config_lambda_nsec_and_lambda_s2():
new = gconfig.migrate_config({"train": {"lambda_nsec": 0.2, "lambda_s2": 2.0}})
assert new["stage2_model"]["n_sec"]["lambda"] == 0.2
assert new["stage2_model"]["lambda"] == 2.0
def test_migrate_config_wgan_knobs_map_to_both_stages():
new = gconfig.migrate_config(
{"train": {"n_critic": 3, "gp_weight": 5.0, "critic_lr": 1e-4}}
)
for stage in ("stage1_model", "stage2_model"):
assert new[stage]["wgan"]["n_critic"] == 3
assert new[stage]["wgan"]["gp_weight"] == 5.0
assert new[stage]["wgan"]["critic_lr"] == 1e-4
def test_merge_cli_overrides_applies_file_then_cli(tmp_path, monkeypatch):
def test_migrate_config_model_hidden_dim_n_blocks_dropout_map_to_both_stages():
new = gconfig.migrate_config(
{"model": {"hidden_dim": 128, "n_blocks": 4, "dropout": 0.2}}
)
for stage in ("stage1_model", "stage2_model"):
assert new[stage]["hidden_dim"] == 128
assert new[stage]["n_res_blocks"] == 4
assert new[stage]["dropout"] == 0.2
def test_migrate_config_emb_dim_and_conditioning_map_to_both_axes():
new = gconfig.migrate_config(
{"model": {"emb_dim": 32, "conditioning": "embedding"}}
)
for axis in ("particle", "material"):
assert new["conditioning"][axis]["emb_dim"] == 32
assert new["conditioning"][axis]["type"] == "embedding"
def test_migrate_config_noise_dim_maps_to_both_stages_wgan():
new = gconfig.migrate_config({"model": {"noise_dim": 128}})
assert new["stage1_model"]["wgan"]["noise_dim"] == 128
assert new["stage2_model"]["wgan"]["noise_dim"] == 128
def test_migrate_config_k_max_maps_to_stage2_only():
new = gconfig.migrate_config({"model": {"k_max": 20}})
assert new["stage2_model"]["k_max"] == 20
assert "k_max" not in new.get("stage1_model", {})
def test_migrate_config_router_copied_to_both_stages_with_tie_to_stage1_false():
new = gconfig.migrate_config(
{
"model": {
"router": {
"enabled": True,
"type": "energy",
"n_experts": 10,
"temperature": 0.05,
}
}
}
)
assert new["stage1_model"]["router"] == {
"enabled": True,
"type": "energy",
"n_experts": 10,
"temperature": 0.05,
}
assert new["stage2_model"]["router"] == {
"enabled": True,
"type": "energy",
"n_experts": 10,
"temperature": 0.05,
"tie_to_stage1": False,
}
def test_migrate_config_router_nonzero_expert_dims_raises():
cfg = {
"model": {
"router": {"enabled": True, "expert_hidden_dim": 128, "expert_n_blocks": 0}
}
}
try:
gconfig.migrate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "expert_hidden_dim" in str(e)
def test_migrate_config_router_zero_expert_dims_dropped_silently():
new = gconfig.migrate_config(
{
"model": {
"router": {
"enabled": True,
"type": "energy",
"n_experts": 4,
"expert_hidden_dim": 0,
"expert_n_blocks": 0,
}
}
}
)
assert "expert_hidden_dim" not in new["stage1_model"]["router"]
assert "expert_n_blocks" not in new["stage1_model"]["router"]
def test_migrate_config_train_passthrough_is_exact():
new = gconfig.migrate_config({"train": {"epochs": 7, "batch_size": 999, "seed": 3}})
assert new["train"] == {"epochs": 7, "batch_size": 999, "seed": 3}
def test_migrate_config_preserves_meta_git_hash():
new = gconfig.migrate_config({"meta": {"git_hash": "abc123"}})
assert new["meta"] == {"git_hash": "abc123", "config_version": 3}
def test_migrate_config_real_router_fixture_raises_on_nonzero_expert_dims():
cfg = gconfig.load_toml(_CONFIGS_DIR / "router_energy_n10_embedding.toml")
try:
gconfig.migrate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "expert_hidden_dim" in str(e)
def test_migrate_config_real_wgan_fixture():
cfg = gconfig.load_toml(_CONFIGS_DIR / "wgan_h128_b4_physical.toml")
new = gconfig.migrate_config(cfg)
assert new["stage1_model"]["generator"] == "wgan"
assert new["stage2_model"]["generator"] == "wgan"
for stage in ("stage1_model", "stage2_model"):
assert new[stage]["hidden_dim"] == 128
assert new[stage]["n_res_blocks"] == 4
assert new[stage]["dropout"] == 0.0
assert new["conditioning"]["particle"]["type"] == "physical"
assert new["conditioning"]["material"]["type"] == "physical"
assert new["train"]["epochs"] == 30
assert new["train"]["warmup_epochs"] == 3
# ---------------------------------------------------------------------------
# merge_cli_overrides
# ---------------------------------------------------------------------------
def test_merge_cli_overrides_defaults_only_matches_default_config():
cfg = gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, None, {})
assert cfg == gconfig.DEFAULT_CONFIG
assert cfg is not gconfig.DEFAULT_CONFIG
def test_merge_cli_overrides_nested_override_keeps_siblings():
cfg = gconfig.merge_cli_overrides(
gconfig.DEFAULT_CONFIG,
None,
{"stage1_model": {"router": {"enabled": True}}},
)
assert cfg["stage1_model"]["router"]["enabled"] is True
assert cfg["stage1_model"]["router"]["type"] == "energy" # default preserved
assert cfg["stage1_model"]["hidden_dim"] == 256 # untouched sibling section
def test_merge_cli_overrides_file_then_explicit_override_precedence(
tmp_path, monkeypatch
):
monkeypatch.setattr(gconfig, "git_hash", lambda: "abc123")
path = tmp_path / "config.toml"
_write_config(path, "abc123")
_write_toml(
path,
git_hash="abc123",
extra="[train]\nepochs = 5\n\n[model]\nhidden_dim = 64\n",
)
cfg = gconfig.merge_cli_overrides(
gconfig.DEFAULT_CONFIG,
path,
train_overrides={},
model_overrides={"hidden_dim": 128},
{"stage1_model": {"hidden_dim": 128}},
)
assert cfg["train"]["epochs"] == 5 # from file
assert cfg["model"]["hidden_dim"] == 128 # CLI override wins over file
assert cfg["stage1_model"]["hidden_dim"] == 128 # explicit override wins over file
assert cfg["stage2_model"]["hidden_dim"] == 64 # migrated from file, not overridden
def test_merge_cli_overrides_migrates_v2_file_transparently(tmp_path, monkeypatch):
monkeypatch.setattr(gconfig, "git_hash", lambda: "abc123")
path = tmp_path / "config.toml"
_write_toml(
path,
git_hash="abc123",
extra='[train]\nmode = "wgan"\n\n[model]\nconditioning = "embedding"\n',
)
cfg = gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {})
assert cfg["stage1_model"]["generator"] == "wgan"
assert cfg["stage2_model"]["generator"] == "wgan"
assert cfg["conditioning"]["particle"]["type"] == "embedding"
# hardcoded v0.2 fact still applied even though it's not a CLI-settable key
assert cfg["conditioning"]["particle"]["n_layers"] == 2
def test_merge_cli_overrides_warns_on_git_hash_mismatch(tmp_path, monkeypatch, capsys):
monkeypatch.setattr(gconfig, "git_hash", lambda: "current999")
path = tmp_path / "config.toml"
_write_config(path, "old111")
_write_toml(path, git_hash="old111", extra="[train]\nepochs = 5\n")
cfg = gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {}, {})
assert cfg["train"]["epochs"] == 5 # does not fail, config still applied
gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {})
captured = capsys.readouterr()
assert "warning" in captured.err
assert "old111" in captured.err
@@ -52,9 +320,9 @@ def test_merge_cli_overrides_no_warning_on_matching_git_hash(
):
monkeypatch.setattr(gconfig, "git_hash", lambda: "same123")
path = tmp_path / "config.toml"
_write_config(path, "same123")
_write_toml(path, git_hash="same123", extra="[train]\nepochs = 5\n")
gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {}, {})
gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {})
assert capsys.readouterr().err == ""
@@ -63,9 +331,9 @@ def test_merge_cli_overrides_no_warning_when_git_hash_unknown(
):
monkeypatch.setattr(gconfig, "git_hash", lambda: "unknown")
path = tmp_path / "config.toml"
_write_config(path, "abc123")
_write_toml(path, git_hash="abc123", extra="[train]\nepochs = 5\n")
gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {}, {})
gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {})
assert capsys.readouterr().err == ""
@@ -76,17 +344,295 @@ def test_merge_cli_overrides_no_warning_when_meta_section_absent(
path = tmp_path / "config.toml"
path.write_text("[train]\nepochs = 5\n")
gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {}, {})
gconfig.merge_cli_overrides(gconfig.DEFAULT_CONFIG, path, {})
assert capsys.readouterr().err == ""
def test_merge_cli_overrides_real_default_toml_fixture(monkeypatch):
monkeypatch.setattr(
gconfig, "git_hash", lambda: "c3bf3abebfe29a10fe42b9cbafbb3460ab78d243"
)
cfg = gconfig.merge_cli_overrides(
gconfig.DEFAULT_CONFIG, _CONFIGS_DIR / "default.toml", {}
)
assert cfg["stage1_model"]["generator"] == "flow"
assert cfg["stage1_model"]["hidden_dim"] == 256
assert cfg["stage2_model"]["hidden_dim"] == 256
assert cfg["conditioning"]["particle"]["emb_dim"] == 16
assert cfg["conditioning"]["particle"]["n_layers"] == 2 # migrated hardcoded fact
assert cfg["train"]["epochs"] == 100
assert cfg["stage2_model"]["decoder"] == "one_shot"
# ---------------------------------------------------------------------------
# save_config
# ---------------------------------------------------------------------------
def test_save_config_round_trips_multi_level_nesting(tmp_path):
cfg = {
"stage1_model": {
"hidden_dim": 256,
"router": {
"enabled": True,
"type": "energy",
"n_experts": 4,
},
},
"train": {"epochs": 100},
}
meta = {"config_version": 3, "git_hash": "abc123"}
gconfig.save_config(cfg, tmp_path, meta)
loaded = gconfig.load_toml(tmp_path / "config.toml")
assert loaded["stage1_model"]["hidden_dim"] == 256
assert loaded["stage1_model"]["router"] == {
"enabled": True,
"type": "energy",
"n_experts": 4,
}
assert loaded["train"] == {"epochs": 100}
assert loaded["meta"] == meta
def test_save_config_round_trips_three_level_nesting(tmp_path):
cfg = {
"stage2_model": {
"decoder": "autoregressive",
"n_sec": {"mode": "head", "lambda": 0.1},
"router": {"tie_to_stage1": True},
}
}
gconfig.save_config(cfg, tmp_path, {"config_version": 3})
loaded = gconfig.load_toml(tmp_path / "config.toml")
assert loaded["stage2_model"]["decoder"] == "autoregressive"
assert loaded["stage2_model"]["n_sec"] == {"mode": "head", "lambda": 0.1}
assert loaded["stage2_model"]["router"] == {"tie_to_stage1": True}
# ---------------------------------------------------------------------------
# default_out_dir_name
# ---------------------------------------------------------------------------
_NOW = datetime(2026, 7, 29, 14, 30)
def _cfg_with(**dotted_overrides):
"""Build a full DEFAULT_CONFIG-shaped dict with dotted-path overrides
applied via _deep_merge, e.g. _cfg_with(**{"stage1_model.hidden_dim": 512})."""
overrides: dict = {}
for dotted, value in dotted_overrides.items():
gconfig._set_path(overrides, dotted, value)
return gconfig._deep_merge(gconfig.DEFAULT_CONFIG, overrides)
def test_default_out_dir_name_all_defaults_is_just_the_timestamp():
assert (
gconfig.default_out_dir_name(gconfig.DEFAULT_CONFIG, now=_NOW)
== "20260729_1430"
)
def test_default_out_dir_name_stage1_generator_shown_bare_no_prefix():
cfg = _cfg_with(**{"stage1_model.generator": "wgan"})
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_wgan"
def test_default_out_dir_name_stage2_decoder_shown():
cfg = _cfg_with(**{"stage2_model.decoder": "one_shot"})
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_dec-one_shot"
def test_default_out_dir_name_particle_type_target_shown():
cfg = _cfg_with(**{"stage2_model.particle_type.target": "physical"})
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_pt-physical"
def test_default_out_dir_name_particle_conditioning_embedding_abbreviated():
cfg = _cfg_with(**{"conditioning.particle.type": "embedding"})
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_cemb"
def test_default_out_dir_name_stage1_router_shown_as_unit():
cfg = _cfg_with(
**{
"stage1_model.router.enabled": True,
"stage1_model.router.type": "energy",
"stage1_model.router.n_experts": 8,
}
)
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_s1r-energy8"
def test_default_out_dir_name_stage2_router_shown_as_unit_distinct_from_stage1():
cfg = _cfg_with(
**{
"stage2_model.router.enabled": True,
"stage2_model.router.type": "pdg",
"stage2_model.router.n_experts": 3,
}
)
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_s2r-pdg3"
def test_default_out_dir_name_router_disabled_omitted_even_if_subfields_nondefault():
cfg = _cfg_with(
**{
"stage1_model.router.enabled": False,
"stage1_model.router.type": "pdg",
"stage1_model.router.n_experts": 8,
}
)
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430"
def test_default_out_dir_name_router_gumbel_shown_when_enabled():
cfg = _cfg_with(
**{
"stage1_model.router.enabled": True,
"stage1_model.router.type": "energy",
"stage1_model.router.n_experts": 8,
"stage1_model.router.gumbel": True,
}
)
assert (
gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_s1r-energy8_s1gum"
)
def test_default_out_dir_name_overflow_caps_and_hashes_remainder():
cfg = _cfg_with(
**{
"stage1_model.generator": "wgan",
"stage2_model.generator": "flow",
"stage2_model.decoder": "one_shot",
"stage2_model.autoregressive.history": "attention",
"stage2_model.particle_type.target": "physical",
"stage1_model.router.enabled": True,
"stage1_model.router.type": "energy",
"stage1_model.router.n_experts": 8,
"stage2_model.router.enabled": True,
"stage2_model.router.type": "pdg",
"stage2_model.router.n_experts": 3,
}
)
name = gconfig.default_out_dir_name(cfg, now=_NOW)
# First 6 by priority: stage1_generator, stage2_generator, stage2_decoder,
# stage2_history, particle_type_target, stage1_router — stage2_router
# overflows into the hash suffix.
assert name.startswith(
"20260729_1430_wgan_s2-flow_dec-one_shot_hist-attention_pt-physical_s1r-energy8_+"
)
def test_default_out_dir_name_overflow_hash_is_deterministic_and_value_sensitive():
overrides = {
"stage1_model.generator": "wgan",
"stage2_model.generator": "flow",
"stage2_model.decoder": "one_shot",
"stage2_model.autoregressive.history": "attention",
"stage2_model.particle_type.target": "physical",
"stage1_model.router.enabled": True,
"stage1_model.router.type": "energy",
"stage1_model.router.n_experts": 8,
"train.seed": 3,
}
name_a = gconfig.default_out_dir_name(_cfg_with(**overrides), now=_NOW)
name_b = gconfig.default_out_dir_name(_cfg_with(**overrides), now=_NOW)
assert name_a == name_b
changed = dict(overrides, **{"train.seed": 99})
name_c = gconfig.default_out_dir_name(_cfg_with(**changed), now=_NOW)
assert name_c != name_a
# ---------------------------------------------------------------------------
# validate_config
# ---------------------------------------------------------------------------
def test_validate_config_default_config_passes():
gconfig.validate_config(gconfig.DEFAULT_CONFIG) # must not raise
def test_validate_config_embedding_target_requires_embedding_conditioning():
cfg = _cfg_with(
**{
"stage2_model.particle_type.target": "embedding",
"conditioning.particle.type": "physical",
}
)
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "embedding" in str(e)
def test_validate_config_embedding_target_passes_with_embedding_conditioning():
cfg = _cfg_with(
**{
"stage2_model.particle_type.target": "embedding",
"conditioning.particle.type": "embedding",
}
)
gconfig.validate_config(cfg) # must not raise
def test_validate_config_pdg_router_incompatible_with_physical_conditioning():
cfg = _cfg_with(
**{
"stage1_model.router.enabled": True,
"stage1_model.router.type": "pdg",
"conditioning.particle.type": "physical",
}
)
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "pdg" in str(e)
def test_validate_config_tie_to_stage1_requires_stage1_active():
cfg = _cfg_with(
**{
"stage2_model.router.tie_to_stage1": True,
"stage1_model.active": False,
}
)
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "tie_to_stage1" in str(e)
def test_validate_config_stop_token_not_implemented():
cfg = _cfg_with(**{"stage2_model.n_sec.mode": "stop_token"})
try:
gconfig.validate_config(cfg)
assert False, "expected ValueError"
except ValueError as e:
assert "stop_token" in str(e)
# ---------------------------------------------------------------------------
# checkpoint config-mismatch warnings (unchanged surface, still exercised)
# ---------------------------------------------------------------------------
def test_warn_if_checkpoint_config_mismatch_finds_sibling_toml(
tmp_path, monkeypatch, capsys
):
monkeypatch.setattr(gconfig, "git_hash", lambda: "current999")
ckpt_path = tmp_path / "best.pt"
ckpt_path.write_bytes(b"") # contents irrelevant, only its directory is used
_write_config(tmp_path / "config.toml", "old111")
ckpt_path.write_bytes(b"")
_write_toml(
tmp_path / "config.toml", git_hash="old111", extra="[train]\nepochs = 5\n"
)
gconfig.warn_if_checkpoint_config_mismatch(ckpt_path)
@@ -113,202 +659,9 @@ def test_warn_if_checkpoint_config_mismatch_no_warning_when_hashes_match(
monkeypatch.setattr(gconfig, "git_hash", lambda: "same123")
ckpt_path = tmp_path / "best.pt"
ckpt_path.write_bytes(b"")
_write_config(tmp_path / "config.toml", "same123")
_write_toml(
tmp_path / "config.toml", git_hash="same123", extra="[train]\nepochs = 5\n"
)
gconfig.warn_if_checkpoint_config_mismatch(ckpt_path)
assert capsys.readouterr().err == ""
def test_resolve_expert_dims_default_config_inherits_hidden_dim_and_n_blocks():
# The unset sentinel (expert_hidden_dim/n_blocks == 0 in DEFAULT_CONFIG)
# is exactly the bug fixed by resolve_expert_dims: it must not silently
# fall back to some other hardcoded default, only to the monolith's own
# hidden_dim/n_blocks, so --hidden-dim/--n-blocks reach the experts too.
router_cfg = dict(gconfig.DEFAULT_CONFIG["model"]["router"])
assert router_cfg["expert_hidden_dim"] == 0
assert router_cfg["expert_n_blocks"] == 0
hidden_dim, n_blocks = gconfig.resolve_expert_dims(router_cfg, 512, 6)
assert (hidden_dim, n_blocks) == (512, 6)
def test_resolve_expert_dims_missing_keys_also_inherit():
hidden_dim, n_blocks = gconfig.resolve_expert_dims({}, 512, 6)
assert (hidden_dim, n_blocks) == (512, 6)
def test_default_config_gumbel_router_defaults_off():
# Straight-through Gumbel-softmax combine weights (giant.model.network.
# Router.combine_weights) must be opt-in — existing routed configs and
# checkpoints should be unaffected unless gumbel is explicitly enabled.
router_cfg = gconfig.DEFAULT_CONFIG["model"]["router"]
assert router_cfg["gumbel"] is False
assert router_cfg["gumbel_tau_start"] == 1.0
assert router_cfg["gumbel_tau_end"] == 0.1
def test_resolve_expert_dims_explicit_override_wins():
router_cfg = {"expert_hidden_dim": 128, "expert_n_blocks": 3}
hidden_dim, n_blocks = gconfig.resolve_expert_dims(router_cfg, 512, 6)
assert (hidden_dim, n_blocks) == (128, 3)
def test_resolve_expert_dims_partial_override_mixes_explicit_and_inherited():
router_cfg = {"expert_hidden_dim": 128, "expert_n_blocks": 0}
hidden_dim, n_blocks = gconfig.resolve_expert_dims(router_cfg, 512, 6)
assert (hidden_dim, n_blocks) == (128, 6)
def _default_cfg(**overrides):
train_overrides = {
k: v for k, v in overrides.items() if k in gconfig.DEFAULT_CONFIG["train"]
}
model_overrides = {
k: v for k, v in overrides.items() if k in gconfig.DEFAULT_CONFIG["model"]
}
router_overrides = overrides.get("router")
if router_overrides:
model_overrides["router"] = router_overrides
return gconfig.merge_cli_overrides(
gconfig.DEFAULT_CONFIG, None, train_overrides, model_overrides
)
_NOW = datetime(2026, 7, 29, 14, 30)
def test_default_out_dir_name_all_defaults_is_just_the_timestamp():
cfg = _default_cfg()
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430"
def test_default_out_dir_name_single_non_default_field():
cfg = _default_cfg(hidden_dim=512)
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_h512"
def test_default_out_dir_name_conditioning_embedding_shown_abbreviated():
cfg = _default_cfg(conditioning="embedding")
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_cemb"
def test_default_out_dir_name_conditioning_default_omitted():
cfg = _default_cfg(conditioning="physical")
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430"
def test_default_out_dir_name_router_enabled_shown_as_unit():
cfg = _default_cfg(router={"enabled": True, "type": "energy", "n_experts": 8})
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_r-energy8"
def test_default_out_dir_name_router_disabled_omitted_even_if_subfields_nondefault():
cfg = _default_cfg(router={"enabled": False, "type": "pdg", "n_experts": 8})
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430"
def test_default_out_dir_name_router_gumbel_shown_when_enabled():
cfg = _default_cfg(
router={"enabled": True, "type": "energy", "n_experts": 8, "gumbel": True}
)
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_r-energy8_gum"
def test_default_out_dir_name_router_gumbel_omitted_when_router_disabled():
cfg = _default_cfg(
router={"enabled": False, "type": "energy", "n_experts": 8, "gumbel": True}
)
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430"
def test_default_out_dir_name_router_learn_centers_shown_only_when_disabled():
cfg_default = _default_cfg(
router={"enabled": True, "type": "energy", "n_experts": 8}
)
assert (
gconfig.default_out_dir_name(cfg_default, now=_NOW) == "20260729_1430_r-energy8"
)
cfg_off = _default_cfg(
router={
"enabled": True,
"type": "energy",
"n_experts": 8,
"learn_centers": False,
}
)
assert (
gconfig.default_out_dir_name(cfg_off, now=_NOW)
== "20260729_1430_r-energy8_nolc"
)
def test_default_out_dir_name_router_learn_width_and_temperature_shown():
cfg = _default_cfg(
router={
"enabled": True,
"type": "energy",
"n_experts": 8,
"learn_width": True,
}
)
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_r-energy8_lw"
cfg2 = _default_cfg(
router={
"enabled": True,
"type": "energy",
"n_experts": 8,
"learn_temperature": True,
}
)
assert gconfig.default_out_dir_name(cfg2, now=_NOW) == "20260729_1430_r-energy8_lt"
def test_default_out_dir_name_mode_shown_bare_no_prefix():
cfg = _default_cfg(mode="wgan")
assert gconfig.default_out_dir_name(cfg, now=_NOW) == "20260729_1430_wgan"
def test_default_out_dir_name_overflow_caps_and_hashes_remainder():
cfg = _default_cfg(
mode="wgan",
router={"enabled": True, "type": "energy", "n_experts": 8},
conditioning="embedding",
hidden_dim=512,
n_blocks=8,
emb_dim=32,
lr=1e-3,
batch_size=2048,
seed=3,
epochs=200,
)
name = gconfig.default_out_dir_name(cfg, now=_NOW)
# First 6 by priority: mode, router, conditioning, hidden_dim, n_blocks, emb_dim.
assert name.startswith("20260729_1430_wgan_r-energy8_cemb_h512_b8_e32_+4more-")
digest = name.split("-")[-1]
assert len(digest) == 6
def test_default_out_dir_name_overflow_hash_is_deterministic_and_value_sensitive():
base = dict(
mode="wgan",
router={"enabled": True, "type": "energy", "n_experts": 8},
conditioning="embedding",
hidden_dim=512,
n_blocks=8,
emb_dim=32,
lr=1e-3,
batch_size=2048,
seed=3,
epochs=200,
)
name_a = gconfig.default_out_dir_name(_default_cfg(**base), now=_NOW)
name_b = gconfig.default_out_dir_name(_default_cfg(**base), now=_NOW)
assert name_a == name_b # stable across calls with the same overflow set
changed = dict(base, epochs=999)
name_c = gconfig.default_out_dir_name(_default_cfg(**changed), now=_NOW)
assert (
name_c != name_a
) # differs when an overflowed value changes # n_blocks inherited, hidden_dim not