BL-23: clip_scope: global default paired with unreviewed clipping_norm: 1.0 may over-clip cache_model_0 reference weights

Geschlossen
  • Stimmt etwas nicht?
#247 1 Kommentar 0 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen

Maintainer antworten meist innerhalb von 1 Tag

Dieses Issue hat noch niemand übernommen.

Bewertung

Schwierigkeit
4/5
Geschätzter Aufwand
3-5 Tage
Anfängerfreundlichkeit
48/100
Issue-Typ
Bug
Klarheit
Größtenteils klar
Aktivitätsstatus
Aktiv
Tech-Stack
machine-learning, python
Bereich
backend, machine-learning

Rechercherichtung

Read cid_services/client.py, cache_model_0/manifest.json, and tests/test_cache_client_dp.py, then inspect how apply_dp_mechanism and clip_state_dict handle the global clipping bound. Run the mechanism on cache_model_0's trained state_dict and measure its combined L2 norm. If the bound clips weights aggressively, retune the manifest value and add a test for post-clip norm retention; done when the real weights are verified and the test passes.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Beschreibung

Approved bug

Summary

PR #109 (fix/dp-sensitivity) correctly defaults clip_scope to global so that noise_scale = noise_multiplier * clipping_norm is calibrated against the actual L2 bound on the combined gradient norm — the sensitivity-calibration fix itself is sound. But cache_model_0/manifest.json's clipping_norm: 1.0 was never revisited after this change, and no test in PR #109 exercises the new default against cache_model_0's real trained state_dict.

Why this matters

Under the old per_layer scope, clipping_norm: 1.0 bounded each tensor independently to norm 1.0.

Under global, clipping_norm: 1.0 bounds the combined L2 norm of all tensors to 1.0:

‖[tensor_0, tensor_1, …, tensor_n]‖₂ = sqrt(Σ ‖tensor_i‖²) ≤ 1.0

For a model with multiple floating tensors this is a much tighter constraint — sqrt(Σ ‖tensor_i‖²) can easily exceed 1.0 even if no individual tensor's norm changes. The result is that clip_state_dict clips more aggressively than before, which degrades model quality proportionally to how much the combined norm exceeds the bound.

PR #109's new tests (tests/test_cache_client_dp.py) use build_small_state_dict (synthetic tensors), not cache_model_0's real trained weights, so this is unverified.

Fix

  1. Run apply_dp_mechanism / clip_state_dict against cache_model_0's actual trained state_dict under the new global default.
  2. Measure the combined L2 norm of the real weights before clipping.
  3. If 1.0 clips aggressively (combined norm >> 1.0), retune clipping_norm in cache_model_0/manifest.json to a value that allows the real weights to pass through unclipped or only lightly clipped.
  4. Add a test in tests/test_cache_client_dp.py that loads the real state_dict and asserts the combined norm post-clip is within a tolerance of the pre-clip norm.

References

  • Developer/BACK_LOG.md BL-23
  • PR #109 (fix/dp-sensitivity) — sensitivity fix and clip_scope: global default
  • cache_model_0/manifest.json — clipping_norm: 1.0, clip_scope: global
  • tests/test_cache_client_dp.py — build_small_state_dict (synthetic, not real weights)
  • cid_services/client.py — clip_state_dict / apply_dp_mechanism
Vorherrschende Sprache
Python
Sterne
4
Forks
7
Ø Merge
3 T. 7 Std.
Gemergte PRs (30 T.)
33

Entwicklungsumgebung

Dieses Projekt bietet weder Dev-Container noch Dockerfile noch Beitragsleitfaden – die Einrichtung liegt bei Ihnen. Beginnen Sie mit der README; die allgemeinen Schritte stehen in unserem Leitfaden für den ersten Beitrag.

Erste Schritte

  1. Lesen Sie das ganze Issue und danach den Beitragsleitfaden des Projekts.
  2. Schreiben Sie ins Issue, dass Sie es übernehmen — das erspart doppelte Arbeit.
  3. Forken Sie das Repository und arbeiten Sie in einem Branch.
  4. Öffnen Sie einen Pull Request, der die Issue-Nummer nennt.

Mehr aus InfiniteZeroFoundation/DevNet

Alle Issues in InfiniteZeroFoundation/DevNet

Ähnliche Issues

Weitere Issues zu Python

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.