Skip to content

docs: remove leftover hiding-group and SingleSwitch references - #144

Merged
yairallouche merged 1 commit into
mainfrom
bugfix/docs-stale-hiding-groups
Oct 14, 2026
Merged

yairallouche merged 1 commit into
mainfrom
bugfix/docs-stale-hiding-groups

Conversation

@aviv1ron1

Copy link
Copy Markdown
Collaborator

Summary

Follow-up to #143. CLAUDE.md now describes MultiSwitch, but some content still covered the removed group-based KV-hiding mechanism or didn't match the code.

CLAUDE.md

  • Control Tokens: describe the token exchange (each control token is rewritten to its substitute id) instead of K=finfo.min group hiding.
  • Switch Configuration example: removed hiding_groups, hiding_policy, control_dims, lora_rank and lora_alpha. GraniteSwitchConfig reads none of them, and transformers silently keeps unknown keys. The example now shows adapter_substitute_token_ids, max_lora_rank, adapter_ranks and the ms_* fields.
  • Gotcha 7 removed: the hidden_count offset only applied to hiding groups. The other gotchas keep their numbers because tests cite "Gotcha clarify_query intrinsic enters repetition loop on embedded ALORA switch model #9".
  • Project tree: removed the deleted hiding_constant_report.py. tests/audio/ only contains test1.wav, so it's now described as fixtures and dropped from the test-file placement list; the audio-marked tests are in unit/, composer/, integration/ and vllm/.
  • <|base_reset|>: noted that it exists only when the model is composed with --base-reset-token.
  • HF backend: removed the training and Trainer claims, since nothing is trainable.
  • uv sync --group dev: no longer labelled "everything"; mentions dev-latest and the tutorials extra.
  • Documentation index: added MULTISWITCH_EXPLAINED.md, AUDIO.md and CICD.md.

Code comments: two comments still mentioned SingleSwitch (token_exchange.py, hf/modeling_granite_switch.py).

Testing

Docs and comments only. Pre-commit (including the link validator) passes.

#143 moved CLAUDE.md to MultiSwitch but left content describing the
removed group-based KV-hiding mechanism, plus some wording that does not
match the code:

- Control Tokens: describe the token exchange (control token rewritten to
  its substitute id) instead of K=finfo.min group hiding.
- Switch Configuration example: drop hiding_groups, hiding_policy,
  control_dims, lora_rank and lora_alpha, none of which GraniteSwitchConfig
  reads; show adapter_substitute_token_ids, max_lora_rank, adapter_ranks
  and the ms_* knobs instead.
- Remove Gotcha 7 (hidden_count offset), which only applied to hiding
  groups. The remaining gotchas keep their numbers because tests cite them.
- Project tree: drop the deleted hiding_constant_report.py; tests/audio/
  holds only a .wav fixture, so describe it as such and remove it from the
  test-file placement list (audio tests live in unit/composer/integration/vllm).
- <|base_reset|> exists only when composed with --base-reset-token.
- HF backend: drop the Trainer/training claims; nothing is trainable.
- `uv sync --group dev` is not "everything": note dev-latest and the
  tutorials extra.
- Documentation index: add MULTISWITCH_EXPLAINED, AUDIO and CICD.

Also fix two code comments that still referred to SingleSwitch
(token_exchange.py, hf/modeling_granite_switch.py).

Signed-off-by: aviv ron <rona@il.ibm.com>
@codecov-commenter

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

@aviv1ron1
aviv1ron1 marked this pull request as ready for review October 11, 2026 13:11
@yairallouche
yairallouche merged commit d3c612e into main Oct 14, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants