fix: count parameters shared across modules once (#322, #377, #358, #303)

[Mikyx-1]

Summary

Fixes parameter over-counting for any model that shares parameters, so summary(...).total_params always equals sum(p.numel() for p in model.parameters()) — what PyTorch itself reports.

Important

Stacked on #396 (fix for #327). This branch contains the #327 commit as its base, so the diff currently shows both commits. Please merge #396 first; afterwards this PR's diff reduces to just the second commit (fix: count parameters shared across modules once). The two fixes are complementary and only together resolve all the linked issues — see below.

The two complementary bugs

Shared parameters reach torchinfo two ways, and each was counted multiple times:

	Mechanism	Fixed by
A	One module instance reused under several parents (e.g. YOLO detection layers)	#327 / #396
B	One parameter tensor referenced by distinct modules — weight tying: tied embeddings / lm_head, shared projection heads (T5, SD, LLMs)	this PR

#327 dedups by id(module), so case B slips through (tied tensors live on different module objects). This PR closes case B.

Root cause (B)

total_params was built bottom-up by summing each row's num_params, with no dedup across parameter tensors. A tensor referenced by N modules was added N times.

Fix

Take parameter totals from the root module. Module.named_parameters() already deduplicates shared tensors (remove_duplicate=True) and includes submodules not run in the forward pass, so this matches sum(p.numel() for p in model.parameters()) exactly. A module whose parameters were all already counted by an earlier row is marked (recursive), so the per-row column still sums to the total.

Issues resolved (verified)

Issue	Symptom	Result on this branch
#322	overcount 1.78–2.37× (YOLO, T5, SD)	✅ YOLOv8n 5,257,936 → 3,157,200; YOLOv10n 4,932,416 → 2,775,520 (matches model.info())
#377	tied embeddings / lm_head double-counted	✅ counted once by default
#358	flan-t5-small reports 128M, expected 76,961,152	✅ 76,961,152 (exact)
#303	flan-t5-small reports 109M, expected ~77M	✅ 76,961,152 (= model.parameters())

YOLO is fixed by the #327 half (module reuse); the tied-weight cases by this half. Both are needed.

Tests

• New TiedWeightsModel fixture + test_tied_weights regression test.
• flan_t5_small.out snapshot regenerated (Python 3.14): total 93,410,688 → 76,961,152, decoder embedding row now correctly (recursive).
• Full suite green (snapshots regenerated under Python 3.14 to match CI); zero regressions.

🤖 Generated with Claude Code

[TylerYep]

Looks good, thanks for the fix!