fireworks2bedrock-migrate

One command migrates a Python repository's Kimi K3 inference from Fireworks to Amazon Bedrock. It scans the repository, rewrites the inference calls, adds a small runtime helper, and opens a pull request. The model is unchanged — only the backend changes.

• Model ID: accounts/fireworks/models/kimi-k3 → global.moonshotai.kimi-k3 (or us.moonshotai.kimi-k3 with --us-only)
• Requests route through Bedrock's OpenAI-compatible Chat Completions API — no Bedrock SDK required
• Fine-tuned models, adapters, and non-inference APIs are detected and left untouched

Requirements

• Python 3.9+
• A git repository for the branch/commit/PR flow (or --no-git to skip git entirely)
• To open the PR: the gh CLI or a git credential helper with a GitHub token
• At deploy time: an AWS role with bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream for the Kimi K3 inference profile

Installation

pip install fireworks2bedrock-migrate

Or run it once without installing:

uvx fireworks2bedrock-migrate migrate --pr
# or
pipx run fireworks2bedrock-migrate migrate --pr

Quick start

cd path/to/your-repo
fireworks2bedrock-migrate migrate --pr

On a git repository with a clean working tree, this checks out the branch f2b/kimi-k3-bedrock, rewrites every base Kimi K3 call site, adds the helper/config/dependencies, writes a migration report, pushes the branch, and opens a pull request.

Commands

fireworks2bedrock-migrate scan    [PATH] [--json]
fireworks2bedrock-migrate migrate [PATH] [options]

PATH defaults to the current directory.

scan — report without changing anything

Lists every Fireworks touchpoint and how it would be handled:

fireworks2bedrock-migrate scan . --json

migrate — apply the migration

Option	Effect
--pr	push the branch and open a pull request
--us-only	target us.moonshotai.kimi-k3 instead of global.moonshotai.kimi-k3
--region <REGION>	default region recorded in f2b.json (default us-east-1)
--json	machine-readable output on stdout
--no-git	skip branch/commit/push (no git required)

What it migrates

Call site	Before	After
Model ID	accounts/fireworks/models/kimi-k3	KIMI_K3 (global.moonshotai.kimi-k3)
OpenAI client	OpenAI(base_url="https://api.fireworks.ai/inference/v1")	get_llm_client()
Async OpenAI	AsyncOpenAI(base_url=...)	get_async_llm_client()
Fireworks SDK	Fireworks(...) / AsyncFireworks(...)	get_llm_client() / get_async_llm_client()
Fireworks LLM	LLM(model=...)	get_llm_client().bind(KIMI_K3)
LangChain	ChatFireworks(...)	ChatOpenAI(..., **langchain_kwargs())
Raw HTTP	requests.post("https://api.fireworks.ai/...")	flagged for review; helper client suggested

The LiteLLM spelling fireworks_ai/accounts/fireworks/models/kimi-k3 is rewritten too.

What it leaves alone

These are detected, left untouched, and listed in the report and f2b.json:

• Fine-tuned / uploaded models — accounts/<account>/models/<name> where <account> is not fireworks
• Non-K3 Fireworks models — accounts/fireworks/models/<other>
• Non-inference APIs — fine_tuning, datasets, deployments, rerank, embeddings, batch, audio, image_generation
• LoRA adapters and dedicated deployments

Files it adds to your repository

File	Purpose
llm_client.py	runtime helper, co-located with the migrated files
f2b.json	per-model backend routing (repo root)
requirements.txt	adds openai and aws-bedrock-token-generator (boto3 only if a model routes to converse)
BEDROCK_MIGRATION.md	full change log, what was left alone, IAM policy, rollback

How the runtime helper works

llm_client.py keeps the OpenAI shape, so migrated call sites do not change:

from llm_client import get_llm_client, KIMI_K3

client = get_llm_client()
resp = client.chat.completions.create(model=KIMI_K3, messages=[...])

Each model is routed at runtime by f2b.json:

Backend	Meaning
openai-compatible	Bedrock's OpenAI-compatible Chat Completions API (default)
converse	Bedrock Converse API via boto3, OpenAI shape translated on the fly
fireworks	Fireworks — used for the BLOCK models and for rollback

Bedrock requests authenticate with a short-lived token generated from the ambient AWS credentials (aws-bedrock-token-generator). The default backend uses the OpenAI-compatible endpoint (/openai/v1) rather than the Converse API, so no Bedrock SDK is required. Unknown accounts/.../models/... IDs stay on Fireworks as a safety net.

Classification and exit codes

Each finding is classified MIGRATE, BLOCK, or REVIEW. REVIEW items — for example a client whose base_url is assembled at runtime, or a model ID built from an f-string — need a decision before deploy and are listed in the report and in the JSON review_items field.

Exit code	Meaning
0	migration applied (and PR opened with --pr)
20	nothing to migrate, or already migrated
30	environment problem (not a directory, dirty working tree, or checkout failed)

JSON output

scan --json and migrate --json each emit one JSON document:

{
  "version": "0.0.2",
  "repo": "/abs/path/to/repo",
  "language": "python",
  "branch": "f2b/kimi-k3-bedrock",
  "pr_url": "https://github.com/acme/repo/pull/12",
  "findings": [
    {
      "class": "MIGRATE",
      "file": "app/llm.py",
      "line": 14,
      "kind": "openai_client",
      "detail": "OpenAI(base_url=fireworks) -> get_llm_client()"
    }
  ],
  "checks": [],
  "review_items": [],
  "exit_code": 0
}

checks is reserved for future automatic verification and is empty today.

Rollback

Set F2B_BACKEND=fireworks to route back to Fireworks with no code change.

Idempotency

Running migrate on an already-migrated repository is a no-op and exits 20.

License

Apache-2.0