Skip to content

docs(bedrock): Bedrock fine-tuning reference + PT/Reserved Tier guidance - #8

Open
rsmets wants to merge 4 commits into
mainfrom
feat/bedrock-fine-tuning-reference
Open

docs(bedrock): Bedrock fine-tuning reference + PT/Reserved Tier guidance#8
rsmets wants to merge 4 commits into
mainfrom
feat/bedrock-fine-tuning-reference

Conversation

@rsmets

@rsmets rsmets commented Jun 8, 2026

Copy link
Copy Markdown
Contributor

Summary

Adds a field-tested Bedrock fine-tuning reference and aligns the capacity-reservation guidance across the aws-dev-toolkit Bedrock skill (SKILL.md, references/fine-tuning.md, references/cost-modeling.md).

The core of the customization guidance — when to reach for Provisioned Throughput (PT) vs. Reserved Tier vs. on-demand custom deployment — was refined per feedback from Justin Muller on the Bedrock service team, shared on the internal SA Slack thread about PT governance and its successor.

What Justin's input changed

  • PT vs. Reserved Tier are different infrastructures, not interchangeable options. PT reserves capacity on traditional Bedrock; Reserved Tier reserves capacity on Mantle. A given model is usually on one or the other — so it's a per-model availability question.
  • Lean toward Reserved Tier where a model offers it (better UX/capabilities), and fall back to PT when RT isn't available for that model.
  • Bring Your Own Model (BYOM) is the Mantle successor to Custom Model Import — added as a custom-model path.
  • No hard "PT is deprecated" claim. There's no published deprecation notice, so the docs avoid asserting one; they recommend on-demand custom deployment first and verifying the per-model matrix against current AWS docs.

Files

  • plugins/aws-dev-toolkit/skills/bedrock/references/fine-tuning.md — fine-tuning reference + PT/RT/BYOM framing
  • plugins/aws-dev-toolkit/skills/bedrock/references/cost-modeling.md — PT/Reserved Tier capacity-reservation entries
  • plugins/aws-dev-toolkit/skills/bedrock/SKILL.md — gotcha pointers updated to match

Note on public publication

This plugin ships publicly (outside AWS), so the content is kept to publicly-documented facts. Mantle, Reserved Tier, and BYOM are public/launched; the only thing intentionally held back is unofficial deprecation speculation.

🙏 Thanks to Justin Muller (Bedrock service team) for the clarifications that shaped this guidance.

Ray Smets added 4 commits June 4, 2026 12:07
Refine the fine-tuning, cost-modeling, and skill references for capacity
reservation on customized models:

- Frame PT vs Reserved Tier as different infrastructures (PT = traditional
  Bedrock, RT = Mantle) with per-model availability, not interchangeable options
- Lean toward Reserved Tier where a model offers it (better UX/capabilities),
  fall back to PT when RT isn't available for that model
- Add Bring Your Own Model (BYOM) as the Mantle successor to Custom Model Import
- Drop hard "PT is deprecated" verbiage — no published deprecation notice exists

Guidance refined per feedback from Justin Muller (Bedrock service team) on the
internal SA Slack thread re: Provisioned Throughput governance and successor.
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.

1 participant