sal-torch

Structurally Adaptive Learning for PyTorch

Training-time sparsification that makes neural networks structurally resilient to compression.

Install

pip install sal-torch            # core
pip install sal-torch[hf]        # + HuggingFace Trainer
pip install sal-torch[reports]   # + PDF/visual reports
pip install sal-torch[crypto]    # + commercial license verification
pip install sal-torch[all]       # everything

from sal import SALConfig, SALCallback

config = SALConfig.auto(model)
trainer = Trainer(model=model, callbacks=[SALCallback(config)])
trainer.train()

Three lines. Any transformer. Compression-resilient.

Fully fine-tune. SAL works by letting the model reorganize around silenced heads, so it needs the weights that do the reorganizing to be trainable. Under LoRA/QLoRA we measured it as actively harmful — see When to use SAL.

Know your model before you touch it

FIScanner — how fragile is this model?

The Fragility Index is a structural diagnostic, scored in [0, 1], that measures how much redundant pathway a model's attention graph has. Heads are compared by their activation signatures; an edge between two heads is fragile when they share no common neighbour, i.e. the function it carries has no backup. FI is the fraction of such edges.

• Low FI → heavily triangulated graph, lots of redundancy → robust.
• High FI → many unsupported edges → fragile under compression.

FI is purely diagnostic — it measures, it never perturbs. You can use it with or without SAL training.

from sal import FIScanner

result = FIScanner(model, probe_dataset).scan()

print(result.fi_score)         # 0.0 - 1.0; lower is more robust
print(result.summary)          # "FI=0.1842 | 3 immune, 2 buffer, 1 critical"
print(result.critical_layers)  # layers whose removal moves FI the most
print(result.immune_layers)    # layers you can compress with little effect

result.save("fragility.json")
result.save("fragility.pdf")   # per-head heatmap (needs sal-torch[reports])

Track it during training with FIMonitor, or call the primitives directly:

from sal import FIMonitor, compute_fi, extract_activation_graph

trainer = Trainer(model=model, callbacks=[FIMonitor(probe_dataset, interval=500)])

adjacency = extract_activation_graph(model, probe_dataset)
fi = compute_fi(adjacency)

PlasticityScanner — where can a model absorb compression?

FI tells you how fragile a model is. PlasticityScanner tells you how much room it has to reorganize, so you know where it is safe to compress. It scores three complementary axes per layer — routing flexibility (attention entropy), inter-layer redundancy (linear CKA), and intra-layer redundancy (an MI proxy) — and folds them into an absorption map that labels each layer ELASTIC (safe), SATURATED (bottleneck), or HUB (compensates when others are pruned).

from sal import PlasticityScanner

pmap = PlasticityScanner(model, probe_dataset).scan()
print(pmap.summary)              # "3 elastic, 1 saturated, 2 hub | mean routing=0.61 ..."

rec = pmap.recommend(target_compression=0.33)
rec.safe_to_prune                # [(layer, head), ...] — prune these first
rec.never_touch                  # heads in hub layers — leave alone
rec.expected_impact              # heuristic accuracy delta

pmap.save("plasticity.json")     # raw scores
pmap.save("plasticity.pdf")      # visual report (needs sal-torch[reports])

sal.compare() — SAL vs. other pruning methods

Benchmark SAL against post-hoc baselines at a matched compression level and see which keeps the most accuracy (or lowest loss) after heads are removed.

from sal import compare

result = compare(model, train_dataset, eval_dataset,
                 methods=["sal", "magnitude", "random_posthoc"],
                 compression=0.33, sal_epochs=3, metric="accuracy")
print(result.table)              # method | score | pruned_heads | time
print(result.winner)
result.save("comparison.pdf")    # bar chart + table

# plug in your own method
compare.register_method("my_pruner", lambda model, ds, eval_ds, ctx: my_score)

Does it survive real compression?

We polled practitioners on how they actually compress models. Of 33 responses, 39% quantize (INT8/INT4) — more than pruning and distillation. SAL was built against head pruning, so the honest question is whether the resilience it trains in generalizes to the compression people actually ship.

Short answer, measured over four runs: yes, if you fully fine-tune. A SAL-trained GPT-2 Medium at INT4 scores higher than the uncompressed standard model at a quarter of the size. Under LoRA the same setup loses. The numbers, including the runs SAL lost, are in What we measured.

RobustnessTest — one model, every degradation

from sal import RobustnessTest

test = RobustnessTest(model, eval_dataset, metric="accuracy")
report = test.run(methods=["int8", "int4", "head_pruning_33", "head_pruning_50",
                           "neuron_dropout_10", "neuron_dropout_20"])

print(report.table)
# real output — DistilBERT fine-tuned on SST-2, no SAL, 512 eval examples
# method              baseline     after     delta     std  survived
# ------------------------------------------------------------------
# int8                  0.8594    0.8359   -0.0234       -        OK
# int4                  0.8594    0.8672   +0.0078       -        OK
# head_pruning_33       0.8594    0.7656   -0.0938       -      FAIL
# head_pruning_50       0.8594    0.7363   -0.1230       -      FAIL
# neuron_dropout_10     0.8594    0.8392   -0.0202   0.015        OK
# neuron_dropout_20     0.8594    0.8320   -0.0273   0.022        OK

print(report.robustness_score)   # aggregate 0-1: mean quality retained
print(report.survival_rate)      # fraction of methods survived

report.save("robustness.json")
report.save("robustness.pdf")    # bars + retention radar (needs sal-torch[reports])

A method counts as survived when relative degradation stays within survival_threshold (default 5% of the clean baseline). Methods:

Method	What it does
int8	Dynamic INT8 over every nn.Linear (torch.ao.quantization, CPU)
int4	4-bit weight-only — bitsandbytes NF4 when available, otherwise a simulated per-channel INT4 round-trip (the backend used is recorded on each result)
head_pruning_<pct>	Silences <pct>% of attention heads using the shipped HeadMasker
neuron_dropout_<pct>	Zeroes <pct>% of FFN neurons at inference — dead units / noisy hardware. Repeated over several fault patterns; mean ± std reported

pip install sal-torch[quant] adds bitsandbytes for real NF4. Without it, int4 falls back to simulation rather than disappearing from your report — pass allow_simulated_quant=False if you would rather see the row skipped.

robustness_compare() — SAL-trained vs. standard

from sal import robustness_compare

result = robustness_compare(
    sal_model=sal_trained_model,
    baseline_model=standard_model,
    eval_dataset=eval_dataset,
    methods=["int8", "int4", "head_pruning_33"],
    metric="accuracy",
)

print(result.table)
print(result.summary)
result.save("robustness_comparison.pdf")

Each model is scored against its own clean baseline, so the comparison measures resilience rather than which model was better to begin with. The row winner is whichever model loses proportionally less.

What we measured

Every run below trains the same model twice from identical weights — once plain, once with SAL at prune_fraction=0.33 — then evaluates both dense under the full battery. Scripts are in scripts/; results in scripts/robustness_scale_results.json. Single seed each.

SAL wins per category, by absolute accuracy (which model scores higher — the deployment question):

run	model / task	training	clean cost	quantization	pruning	combined
v0.4.0	DistilBERT / SST-2	full FT	+1.17pp	1/2	2/2	not tested
scale	GPT-2 Medium / SST-2	LoRA r=16	-3.12pp	0/2	1/2	1/2
scale	GPT-2 Medium / SST-2	full FT	+0.39pp	2/2	2/2	2/2
scale	Phi-2 2.7B / MMLU	LoRA r=16	-1.17pp	0/2	1/2	0/2

The DistilBERT run also tested inference-time neuron dropout at 10% and 20%, which the standard model won both times — by 0.5pp and 0.6pp, inside that run's noise floor. The scale runs do not test dropout, so it has no column here.

The two GPT-2 rows are a controlled comparison: identical model, task, data, seed and battery. The only thing that changes is whether LoRA is in the way.

Full fine-tuning: SAL wins every variant

GPT-2 Medium, SST-2, 354.8M/354.8M trainable, 512 eval examples:

variant           baseline       SAL     delta  base_size  sal_size    winner
-----------------------------------------------------------------------------
dense               0.8848    0.8887   +0.0039     1419.3    1419.3       SAL
int8                0.8828    0.8887   +0.0059      513.3     513.3       SAL
int4                0.8750    0.8926   +0.0176      361.9     361.9       SAL
prune33             0.8359    0.8555   +0.0195     1419.3    1419.3       SAL
prune50             0.8145    0.8379   +0.0234     1419.3    1419.3       SAL
prune33+int8        0.8398    0.8555   +0.0156      513.3     513.3       SAL
prune33+int4        0.8340    0.8613   +0.0273      361.9     361.9       SAL

Seven for seven, and SAL costs nothing on the clean model (+0.39pp). One eval example is 0.195pp here, so int4, both pruning rows and both combined rows are clear of the noise floor; dense and int8 individually are not. All seven point the same way.

The Pareto result

SAL/int4 scores 0.8926 at 361.9MB. That beats the uncompressed baseline (0.8848 at 1419.3MB) — higher accuracy at a quarter of the size — and it is the only point on the accuracy-vs-size frontier. Nothing in the standard arm comes close at any size.

Under LoRA, the same setup fails

variant           baseline       SAL     delta    winner
--------------------------------------------------------
dense               0.8906    0.8594   -0.0312  baseline
int8                0.8828    0.8613   -0.0215  baseline
int4                0.8594    0.8301   -0.0293  baseline
prune33             0.8496    0.8418   -0.0078  baseline
prune50             0.8145    0.8496   +0.0352       SAL
prune33+int8        0.8496    0.8477   -0.0020  baseline
prune33+int4        0.8301    0.8359   +0.0059       SAL

Same model, same data, same seed. SAL loses four of six compressed variants and gives up 3.1 points of clean accuracy to get there. Only the heaviest structural damage (prune50) still favours it.

We are leaving this table in the README because it is the finding that explains the mechanism: SAL works by letting the model reorganize around silenced heads, and LoRA freezes the weights that would do the reorganizing. Rank-16 adapters on c_attn cannot absorb 126 silenced heads. The perturbation lands, the adaptation cannot.

What is not established

• Full fine-tuning is necessary, not automatically sufficient. The v0.4.0 DistilBERT run was also full fine-tuning and still showed no quantization effect — but its INT4 barely dented the baseline at all (it improved it, i.e. noise), so there was no headroom to win. Where quantization costs the standard model something, SAL has recovered it; where it costs nothing, there is nothing to recover.
• Scale is still open. Phi-2 2.7B was LoRA-only, so "LoRA starves it" and "SAL stops working above ~350M" remain confounded at that size. Phi-2 under full fine-tuning is the experiment that separates them.
• Single seed everywhere. Two GPT-2 baselines trained on identical data differ by 0.58pp, which is the run-to-run floor.

When to use SAL

your setup	recommendation
Full fine-tuning	Yes. Validated across quantization, head pruning, and combined compression on GPT-2 Medium; validated for head pruning on DistilBERT.
LoRA / QLoRA adapters	Not recommended. Measured worse than not using SAL at all, and it costs clean accuracy. The adapters are too small to redistribute what the masking removes.
Models above ~1B	Unvalidated. No full-fine-tuning result at that scale yet.
You only quantize, never prune	Worth testing on your model with RobustnessTest before committing — the size of the win tracks how much quantization costs your baseline.

If you are on LoRA and want compression resilience, the honest answer today is that SAL is not the tool; use RobustnessTest to measure what your compression actually costs and PlasticityScanner to choose where to cut.

Continual learning without replay buffers

StructuralGuard — protect what matters when you fine-tune

When you fine-tune a trained model on a new task, it quietly overwrites the structure that carried the old one. StructuralGuard reads the model's structural map and freezes the critical attention heads (hub layers, structural bottlenecks, and the functionally unique heads) while leaving the redundant heads free to absorb the new task. No EWC, no replay buffer, no distillation — the topology itself decides what to protect.

from sal import StructuralGuard

# After training on task A, build a guard from the model's structure.
guard = StructuralGuard.from_model(model, probe_dataset, protection_level=0.5)

print(guard.protected_heads)   # [(layer, head), ...] frozen during fine-tuning
print(guard.trainable_heads)   # [(layer, head), ...] free to absorb task B
print(guard.protection_map)    # {layer: [protected head indices]}

guard.protect(model)           # zero gradients for protected heads (backward hooks)
trainer.train()                # fine-tune on task B with ANY training loop
guard.release()

drift = guard.measure_drift(model, probe_dataset=probe_dataset)
print(drift.forgetting_score)      # 0 = nothing forgot, 1 = total reorganization
print(drift.protected_integrity)   # ~1.0 if the protected heads held

guard.save("model_guard.json")     # serialize; reload before task C, D, ...
guard = StructuralGuard.load("model_guard.json")

Protection is at the head level — some heads in a layer can be frozen while others in the same layer keep learning. protection_level (0.0–1.0) sets the fraction of the most critical heads to protect.

HuggingFace Trainer? Use the callback — it applies protection on train_begin, measures drift on train_end:

from sal import StructuralGuardCallback

guard = StructuralGuard.from_model(model, probe_dataset)
callback = StructuralGuardCallback(guard)
trainer = Trainer(model=model, callbacks=[callback])
trainer.train()
print(callback.drift_report.summary)

DriftMonitor — measure structural forgetting after any fine-tuning

DriftMonitor quantifies how much a model's structure moved, guarded or not. Snapshot before and after, then compare.

from sal import DriftMonitor

monitor = DriftMonitor(model, probe_dataset)
monitor.snapshot("before_task_b")
trainer.train()
monitor.snapshot("after_task_b")

drift = monitor.compare("before_task_b", "after_task_b")
print(drift.summary)
print(drift.layer_drift)             # per-layer activation retention (1 = identical)
print(drift.classification_changes)  # layers whose fragility class flipped
drift.save("drift_report.json")
drift.save("drift_report.pdf")       # visual before/after comparison

Snapshots are keyed, so you can track drift across many sequential tasks and compare any pair.

Examples

• examples/quickstart.py — 3-line SAL training on DistilBERT
• examples/standalone_fi.py — Fragility Index scan, no training
• examples/full_control.py — manual config + standalone trainer
• examples/compare_with_without_sal.py — SAL vs. baseline under compression

New here? Start with docs/getting_started.md.

Roadmap

See ROADMAP.md for what's shipped, what's next, and how to request features — including the full four-run evidence trail behind the v0.4.0 robustness claims, losses included. Next up is CompressionPipeline (v0.5.0), which turns the validated SAL + INT4 recipe into a single call and refuses to run silently on LoRA.

License

BSL 1.1 — free for research and evaluation. Commercial production requires a license.

Built by Cognitive Engineering in Switzerland.