Replaces train.py's single global training loop with a StageTrainer
hierarchy (FlowDDPMStageTrainer, WGANStageTrainer) — one per active
stage, each owning its own optimizer/LR schedule/EMA and reading only
the shared batch tuple (stage 2 always teacher-forces on the
ground-truth x1_s1, so stages never need each other's output at train
time). Supports every stage1/stage2 generator combination, including
the design doc's headline mixed case (stage1=flow + stage2=wgan) and
its reverse, plus stage1-only/stage2-only ablation runs, routed+gumbel
stages, and checkpoint save/resume. metrics.csv/wandb logging are
stage-prefixed. validate_marginals calls are guarded with a one-time
warning and a Wasserstein-magnitude fallback for wgan best-checkpoint
selection, since giant/sample.py still assumes stage1 always owns
n_sec_head (decision 1 moved it to stage 2 by default) — deferred to
design doc step 6, not silently papered over.
pipeline.py's run_setup_stage/run_train_job now read the new nested
config directly; the dangling resolve_expert_dims call and the
--mode wgan --router rejection are both gone (routed WGAN works).
cli.py's train/new-run build correctly-shaped config overrides
(architecture flags -> stage1_model only per the approved decision;
--mode/--n-critic/--gp-weight/--critic-lr broadcast to both stages,
matching migrate_config's own precedent and avoiding a regression on
the common --mode case); predict/rollout's dangling build_models
tuple-unpack is fixed; new-run now tags config_version, fixing a bug
where a re-loaded v0.3 config.toml would have been silently corrupted
by migrate_config mistaking it for v0.2.
config.py's validate_config rejects mixed particle/material
conditioning types for now (ConditionEncoder supports it, the data
pipeline in giant/data/transforms.py doesn't yet). analysis/render.py
and router_gating.py handle both the new nested model_config shape and
legacy flat checkpoints. scripts/warm_setup_cache.py updated for
run_setup_stage's new signature.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Decomposes the ten permutation classes in giant/model/network.py into
the reusable parts from docs/v0.3.0-design.md §5: ConditionEncoder (now
independently configurable per particle/material axis), ContextAdapter,
Trunk/MonolithicTrunk/RoutedTrunk/ExpertTrunk, and the stage classes
Stage1Model/Stage2OneShot/CriticModel (Stage2Autoregressive stubbed,
raises NotImplementedError until step 4/5). build_models/build_critics
now return a dict keyed by stage and accept the new nested config shape,
with routed WGAN reachable for the first time (the old --mode wgan
--router rejection is gone) and stage2_model.router.tie_to_stage1
sharing a literal Router instance.
A v0.2 checkpoint's flat model_config auto-migrates via
_migrate_legacy_model_config + migrate_legacy_state_dict, preserving the
n_sec_head's attachment to Stage1Model (legacy_owner="stage1", design
doc §4.1). tests/test_migration_v02_v03.py proves this bit-identical
against a frozen v0.2 snapshot (tests/legacy/network_v02_snapshot.py)
for both flow and wgan, both conditioning modes.
scripts/check_migration_v02_v03.py is the real-checkpoint counterpart
for a portal machine with /ceph access.
giant/model/schedule.py's flow-matching/DDPM loss helpers are updated
to the new model-call convention (t as a keyword). giant/sample.py,
giant/rollout.py, and giant/validate.py are not yet updated (deferred
to design doc step 6) — their exercising tests are marked xfail with
that reasoning rather than silently broken.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The whole point of conditioning="physical" is generalizing to a
species/material outside the training menu, but two independent code
paths still hard-required training-vocab membership:
- giant/data/transforms.py: build_cond_features unconditionally raised
KeyError on an out-of-vocab pdg/material. _vectorized_map_lookup
gains a strict=False mode (dummy index instead of raising), used only
under conditioning="physical" where ConditionEncoder never reads
cond_cat anyway; "embedding" mode is untouched and still raises,
since cond_cat IS the conditioning signal there.
- giant/rollout.py: the known_pdg termination gate still killed a track
on step 1 for any pdg outside pdg_map, regardless of conditioning
mode. Now skipped entirely under conditioning="physical".
- giant/model/network.py: PdgRouter/ProcessRouter always build their
own training-vocab nn.Embedding independent of conditioning, silently
reintroducing the same limitation at the routing layer. build_models
now raises loudly if conditioning="physical" is paired with either
router type, rather than silently building a model that can't
generalize the way it claims to.
This unblocks the held-out-species/material generalization experiment
against the multi-material dataset (see CLAUDE.md roadmap). Each fix
has a regression test, including an end-to-end rollout test seeded
with a resolvable-but-out-of-vocab PDG code.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Trains the routed trunk's forward combination as a hard one-hot sample
(matching eval-time top-1 dispatch exactly) while keeping a smooth gradient
on the backward pass, targeting the train/eval mismatch identified as a
likely contributor to experts overlapping instead of partitioning in the
first energy-router rollout benchmark. Off by default (model.router.gumbel);
existing routed configs/checkpoints are unaffected.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
EnergyRouter's gate sharpness was a single fixed temperature shared by
every expert, with no way for an expert to independently learn how much
of the energy axis it covers. Adds two mutually exclusive, default-off
modes: learn_width (per-expert learnable width) and learn_temperature
(single learnable shared scalar), both bounded via a sigmoid
interpolation warm-started to reproduce today's fixed-temperature gate
exactly at init, to compare against each other without risking the
unbounded-width collapse failure mode. Also promotes gate_stats's
entropy into a generic, optional Router.entropy_loss (lambda_entropy) as
a secondary guard against all experts' widths co-inflating together.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The 2026-07-22 rollout benchmark's router_gating diagnostic showed the
10-expert EnergyRouter's default linspace(-2, 2, n_experts) init assumes a
roughly uniform z-normalized energy distribution, leaving experts heavily
overlapping instead of partitioning the range. Add an optional
centers_init kwarg (backward compatible, defaults to the old linspace) and
have giant train estimate it from a reservoir sample of the real energy
column, collected during the existing normalizer-fitting pass.
Adds model.conditioning = "physical" | "embedding": physical mode routes
particle mass/charge and material Z_eff/A_eff/density/X0/lambda_int through
small MLPs to replace the learned PDG/material embedding tables, so the
surrogate generalizes to PDG codes/materials outside the training vocab
instead of memorizing it. "embedding" stays available as the comparison
baseline (old checkpoints without the key default to it).
Stage 2 now regresses a secondary's mass/charge directly against a fixed
physics-derived target instead of a learned/snapped embedding, and uses no
snapping at inference — the model's raw predicted (mass, charge) is the
secondary's physical identity, including for its own further rollout steps.
A separate reporting-only nearest-known-PDG lookup (never fed back into the
model) populates output pdg columns / the embedding-mode rollout fallback.
giant/materials.py's table is populated with Geant4's own built-in NIST
constants (Z_eff, A_eff, density, X0, lambda_int), extracted directly from
the Geant4 11.4.1 build vendored in minicalosim via G4NistManager rather
than hand-typed literature values. G4_LYSO is left unfilled: confirmed (both
by runtime lookup and by searching minicalosim's history) that it's never
actually a constructed Geant4 material there, only documentation/UI color-map
text.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Route on several independent axes at once (e.g. energy x pdg), each with
its own expert count and hyperparameters. The joint gate is the outer
product of per-axis softmax gates, so it stays a partition of unity and
top1/balance_loss factor per-axis. Config uses flat axis{i}_{field} keys
in model.router (TOML/CLI friendly), also settable via repeatable
--router-axis flags.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Routes on the pre-step PDG code, which — unlike ProcessRouter's process
label — is already known at gate time (a conditioning input), so no
supervision is needed and classify_loss falls back to the zero default.
Generalizes EnergyRouter's soft-turn-on-then-Voronoi trick from a 1-D
distance to a small learned PDG embedding space: its own embedding table
maps each PDG code to a point, and n_experts learnable centers partition
that space.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Routes on the physics process (Compton, phot, brems, ...) that ends a
step, supervised by a small classifier since process is a post-step
outcome unobservable at gate time. Threads a process label end-to-end
through the data pipeline (loader, build_features, dataset batches,
training loss/checkpointing) alongside the existing EnergyRouter.
Both stages can now route through a pluggable Router (EnergyRouter as the
first implementation, a soft turn-on gate over pre-step log-energy) into
several small ExpertTrunks instead of one monolithic trunk. Trains as a
differentiable soft mixture and dispatches to a single expert per row at
eval time, which is the source of the per-call speedup this prototype is
after (issue #5's ~10x native-Geant4 budget). Disabled by default, so
existing configs/checkpoints are unaffected; build_models() centralizes
routed-vs-monolith construction across train/predict/rollout.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
lars