thermoclaw 0.1.0

Thermodynamic diagnostics for neural network training — see what your optimizer is actually doing.

Thermoclaw

Thermodynamic diagnostics for neural network training — see what your optimiser is actually doing.

Thermoclaw wraps any PyTorch optimizer and exposes real-time thermodynamic quantities: entropy production, internal vs external entropy flow, weight-decay collapse detection, equilibrium detection, and actionable per-layer diagnostics.

---

It found the problem. Acting on it worked.

We ran Thermoclaw on a Pythia-410M fine-tuning run with weight_decay=0.3 on WikiText-103. At step 39, CollapseDetector flagged weight decay collapse in the attention layers:

[MEDIUM] Weight decay collapse in 5 attention layers: grad/param ratio
dropped 3.1x from early to late training. Reduce weight_decay.

We reduced weight_decay from 0.3 to 0.01 at that point and kept training.

Run	Final PPL (steps 280-300)
Unmodified (wd=0.3 throughout)	256.8
Thermoclaw intervention at step 39	222.9
Improvement	+33.8 PPL

No hyperparameter search. No manual inspection. One warning, one change.

That's what Thermoclaw does. The warning is causal — not correlational.

---

Install

pip install thermoclaw                    # core (torch + numpy only)
pip install thermoclaw[viz]               # + matplotlib dashboards
pip install thermoclaw[hf]                # + HuggingFace Trainer callback
pip install thermoclaw[all]               # everything

Quick Start

Observe any optimizer

from thermoclaw import Observer, diagnose

model = YourModel()
optimizer = torch.optim.AdamW(model.parameters(), lr=3e-4)
observer = Observer(model, optimizer)

for batch in loader:
    loss = criterion(model(batch))
    loss.backward()
    optimizer.step()
    observer.step(loss=loss.item())
    optimizer.zero_grad()

report = diagnose(observer)
print(report)
observer.plot_dashboard(save_path='dashboard.png')

d_iS / d_eS entropy split (the killer feature)

from thermoclaw import Observer, EntropySplit, diagnose

observer = Observer(model, optimizer)
splitter = EntropySplit(model, optimizer, observer)

for batch in loader:
    loss = criterion(model(batch))
    loss.backward()
    optimizer.step()
    observer.step(loss=loss.item())
    splitter.step()                        # ← decomposes entropy
    optimizer.zero_grad()

report = diagnose(observer, splitter)
print(report)
# Output:
#   [HIGH] Weight decay is the dominant entropy source for 12 attention
#   layers (mean R_ie=4.2). Consider reducing weight_decay for attention
#   layers by 4-8×, or excluding them.

splitter.plot_entropy_split(save_path='entropy_split.png')

CollapseDetector — catch weight-decay collapse in real time

Weight decay can silently erode parameter capacity faster than gradient signal can rebuild it. The observable signature is ‖g‖/‖θ‖ (grad/param ratio) dropping over time. CollapseDetector tracks this directly — no calibration period required, no incommensurable quantities.

from thermoclaw import CollapseDetector

detector = CollapseDetector(model, optimizer)

for batch in loader:
    loss.backward()
    optimizer.step()
    detector.step()                        # ← call before zero_grad
    optimizer.zero_grad()

    # Mid-training intervention trigger
    if detector.is_collapsing:
        for pg in optimizer.param_groups:
            pg['weight_decay'] *= 0.1

# Post-training report
recs = detector.get_recommendations()
# → ["[HIGH] Weight decay collapse in 5 mlp layers: grad/param ratio
#     dropped 4.2× from early to late training. Reduce weight_decay."]

is_collapsing returns True as soon as any HIGH or MEDIUM collapse signal fires — use it for automated mid-training interventions. Validated at Pythia-410M scale: firing correctly at step 39 with wd=0.3, PPL gap +33.8 vs unmodified arm.

ThermoScheduler (drop-in replacement)

One line to replace your cosine scheduler. Keeps your optimizer, just makes the schedule thermodynamically aware:

from thermoclaw import ThermoScheduler

optimizer = torch.optim.AdamW(model.parameters(), lr=3e-4)
scheduler = ThermoScheduler(optimizer, total_steps=10000)

for step, batch in enumerate(loader):
    loss = criterion(model(batch))
    loss.backward()
    optimizer.step()
    scheduler.step()          # ← replaces cosine_scheduler.step()
    optimizer.zero_grad()

HuggingFace integration (one line)

from thermoclaw.integrations.huggingface import ThermoclawCallback

trainer = Trainer(
    model=model,
    args=training_args,
    callbacks=[ThermoclawCallback()],      # ← that's it
)
trainer.train()
# Dashboard PNG, CSV, and diagnostic report saved to output_dir automatically

What Thermoclaw measures

Quantity	Symbol	What it means
Entropy production	σ_l = η‖g‖²	How much "thermodynamic work" each layer is doing
Entropy ratio	r_l = σ/σ*	1.0 = equilibrium. <0.85 = under-trained. >1.15 = over-trained
External entropy	d_eS	Entropy that reduces loss (productive learning)
Internal entropy	d_iS	Entropy from weight decay, momentum, noise (overhead)
R_ie = d_iS/d_eS	—	The diagnostic ratio. >2 = warning, >5 = critical
Grad/param ratio	‖g‖/‖θ‖	CollapseDetector signal. Dropping trend = weight decay eroding capacity
Dispersion	D = Var(r_l)	Inter-layer training uniformity
Gradient alignment	ρ = cos(g_t, g_{t-1})	Step coherence. Negative = oscillation
Parameter distance	E = ‖θ−θ₀‖²	How far weights have moved from init

A note on ρ and momentum (non-obvious)

High momentum (β₁ → 1) increases ρ, not decreases it — momentum smooths consecutive gradient vectors, making them more correlated. This means ρ is not a reliable signal for over-damped momentum. If β₁ is too high, use equilibrium fraction instead: over-damped runs show low eq_fraction despite high ρ, because damped updates can't reach homeostatic σ* levels. Validated at Pythia-410M scale: β₁=0.999 produced ρ=0.63 vs baseline ρ=0.39, with eq_fraction=0.15 vs 0.59.

The d_iS / d_eS decomposition

Standard training observes total loss and calls it a day. But total entropy production σ conflates two fundamentally different thermodynamic flows:

• d_eS (external) — gradient-driven parameter updates that reduce loss. This is productive work.
• d_iS (internal) — entropy from weight decay, momentum friction, stochastic noise. This is heat.

When d_iS >> d_eS, the optimizer is spending most of its entropy budget on overhead. Thermoclaw decomposes d_iS further into:

• d_iS_wd — weight decay contribution
• d_iS_momentum — momentum friction
• d_iS_noise — stochastic gradient noise

This tells you exactly which layers, at which step, are wasting compute — and why.

Per-layer param groups

For full per-layer resolution, use make_param_groups:

from thermoclaw import make_param_groups

groups = make_param_groups(model, lr=3e-4, weight_decay=0.01)
optimizer = torch.optim.AdamW(groups)
observer = Observer(model, optimizer)

Confidence scoring

Recommendations are conservative. Thermoclaw only flags issues where the physics signal is unambiguous:

• [HIGH] — Single dominant source (>60% of d_iS), R_ie > 5, consistent across regions. Safe to act on.
• [MEDIUM] — Clear signal but moderate R_ie (2-5). Worth investigating.
• [LOW] — Signal present but multiple sources contribute. Informational only.

Wrong recommendations that sound authoritative destroy trust faster than no recommendations at all. Thermoclaw would rather under-claim than over-claim.

Validated

Three-tier validation on H100 80GB (Pythia-410M, WikiText-103, bfloat16):

Tier	Test	Result
T1: Analytical	σ, ρ, d_iS_wd, d_eS+d_iS=σ, E, D — 8 ground-truth checks	8/8 PASS
T2A: High LR	lr=3e-2 → R_ie=1.6×10²⁰, eq=0.014, flagged HIGH	PASS
T2B: High WD	wd=5.0 → unhealthy, flagged MEDIUM	PASS
T2C: Overdamped	β₁=0.999 → ρ=0.63 (vs baseline 0.39), eq=0.15, flagged HIGH	PASS
T2D: Baseline	lr=3e-4/wd=0.01/β₁=0.9 → no collapse or WD pathology recs	PASS
T3: Intervention	CollapseDetector fires step 39 (wd=0.3), PPL gap +33.8 vs unmodified	PASS

Origin

Thermoclaw's thermodynamic framework comes from the EPTO (Entropy-Production Targeted Optimisation) research project. The key insight: neural network training is a non-equilibrium thermodynamic process, and the quantities that matter (entropy production, entropy ratios, equilibrium fraction) can be measured for any optimizer, not just EPTO.

License

Apache 2.0