lars
8019a80563
Every metric name used to exist in four places: the dict keys each
StageTrainer returned, the hardcoded _metrics_fields() column list, the
~110-line metrics_row assembly in train(), and the tqdm/summary
formatting. The two had to be kept in exact correspondence by hand or
csv.DictWriter would raise.
Each metric is now declared once, as a MetricSpec on the trainer that
computes it. MetricsCollector derives the CSV header and W&B payload from
those declarations and owns all accumulation, so train() no longer carries
a running sum, and every isinstance(tr, WGANStageTrainer) branch is gone —
replaced by four trainer hooks (batch_loss, summary, val_objective,
supports_val_loss).
giant/train.py (1875 lines) becomes giant/training/:
trainers.py StageSpec + shared StageTrainer base + the two subclasses
metrics.py MetricSpec, MetricsCollector
stage2_inputs.py the pure AR/teacher-forcing tensor helpers, moved verbatim
loop.py train() (225 lines, was ~514) + graceful shutdown
checkpoint.py build/load, lifted out of train()'s closures
The trainers shared ~15 identical constructor arguments and copy-pasted
their cosine-warmup lambda, EMA setup, state_dict/load_state_dict,
resume_lr and train_mode/eval_mode. StageSpec resolves one stage's config
once (constructors go from 24 and 22 keyword arguments to (spec, model,
device)), the base class holds the rest, and build_stage_trainers drops
from ~100 lines to 15.
Metric columns are renamed to a uniform stage/split/metric scheme
(stage1/train/loss, stage2/train/d_loss, stage1/lr, stage1/router/entropy,
val/loss, ...). Old metrics.csv files and W&B history are not comparable.
The checkpoint format is unchanged.
BEHAVIOR CHANGE — WGAN best-checkpoint selection. The old code meant to
score a WGAN stage on its marginal KL, but the guard
`{n: kl for n in wgan_names if n not in val_loss_per_stage}` could never
fire: val_loss_per_stage was pre-seeded with 0.0 for every stage, so a
WGAN stage contributed a flat 0.0 and the KL was written to metrics.csv
without ever influencing best.pt. val_objective now returns it as
intended. On the test harness's default flow+wgan config val_loss went
from 2.182 (stage 1 only) to 15.137 (stage 1 + KL 12.954), and which epoch
won changed. Runs before this commit picked their best checkpoint on the
non-adversarial stages alone. Written up in docs/v0.3.0-followups.md.
Verified: 699 tests pass; ruff, ruff format and ty clean. Baseline-vs-
refactor metrics.csv compared across five configs (flow+wgan, AR+onehot,
routed, both-flow, AR-flow) — every comparable value bit-identical except
val/loss where the fix applies. Resume appends without a duplicate header
and reproduces a HEAD worktree's per-epoch losses and LRs exactly across
the resume boundary. A refactored last.pt loads through
cli.py:_load_model_weights in both raw and ema modes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2.9 KiB
v0.3.0 — post-implementation audit: open discrepancies
Status: steps 1–7 of docs/v0.3.0-design.md §12 are implemented (branch
v0.3.0-stage2-autoregressive, commits eb6dd27..200c6d2). This document
tracked discrepancies found between that implementation and the design contract
during a 2026-08-07 audit, as concrete work items. All items (1-9) are now
resolved — either implemented (1-5, 7-9) or explicitly deferred into
docs/v0.3.0-design.md §11.2 (6: stage2_model.generator = "ddpm"). Step 8
(estimate_batch_size recalibration) was intentionally still outstanding per
§12 and was never tracked here.
Confirmed correct during the audit (no action needed)
For reference — these were explicitly checked against the design doc and match it, including two spots the doc itself flagged as likely stale that turned out fine:
- Config schema,
DEFAULT_CONFIG,migrate_configtable (§4),save_config/merge_cli_overridesrecursion,default_out_dir_name,Conditioningenum,n_sec.mode = "stop_token"error (§9, §11.2). network.py's full class decomposition (§5.3), dict-returningbuild_models/build_critics(§5.4),ExpertTrunkseparate in/out dims, ST-Gumbel wiring (§2.1), AR token layout (§6.1), Markov/Attention history encoders (§6.2).- Shared PDG top-N type map (one map, not two, per §8),
other_policysample/modal/drop (§11.1), embedding L1-nearest decode, and the L1-distance diagnostic surfaced ingiant analyze(§11.3). analysis/render.py/analysis/router_gating.pycorrectly branch old-flat vs new-nestedmodel_config["router"]location — doc flagged this as a likely stale spot (§10) but it's actually fine.giant/training/'s per-stage trainers, mixed flow+wgan runs, WGAN critic cadence, per-stage router auxiliary losses, stage-prefixed metrics, stage-2-only training via ground-truthx1_s1(§7).
Behavior change: WGAN best-checkpoint selection
The giant/training/ split fixed a dead guard in WGAN validation scoring. A
WGAN stage was meant to contribute its marginal KL to the val_loss that
drives best.pt, but the guard if n not in val_loss_per_stage could never
fire (every stage was pre-seeded to 0.0), so the stage contributed a flat
0.0 and the KL was written to metrics.csv without ever being used.
WGANStageTrainer.val_objective now returns the KL as intended.
Consequence: any checkpoint selected before this commit under a config
with a WGAN stage — including the v0.3.0 default (stage2_model.generator = "wgan") — picked its best epoch on the non-adversarial stages alone. Measured
on the test harness's default flow+wgan config, val_loss went from 2.182
(stage 1 only) to 15.137 (stage 1 + KL 12.954), and which epoch won
changed. Do not compare val/loss or best.pt choice across this commit.
pipeline.py's deleted wgan+router rejection, per-stagecenters_initseeding, removed stale expert-size warning (§9, §10).