Changelog¶

Changelog for the Amazon Linux 2023-based strands-decider images.

Strands Decider 1.0.0 (strands-decider 0.3.1.dev1) — 2026-11-10¶

Tags: serve-sagemaker-cuda · serve-sagemaker-cuda-v1 · serve-sagemaker-cuda-v1.0 · serve-sagemaker-cuda-v1.0.0

strands-decider source: 9a7470e

Initial release of the strands-decider inference container on Amazon Linux 2023, for Amazon SageMaker AI real-time endpoints on NVIDIA GPUs (port 8080).
Serves strands-decider serve: POST /invocations and POST /v1/systemone answer noul, choice, and score questions with calibrated probabilities, and GET /ping reports readiness.
Finds the checkpoint under /opt/ml/model and loads the base model from base/ with no network access.
Warms up before /ping succeeds, so the first request does not wait for kernel compiles.
Runs one request on the model at a time by default (DECIDER_MAX_CONCURRENCY=1), so bursts of large requests queue instead of running out of GPU memory.
Splits a model larger than one GPU across the instance's GPUs with DECIDER_DEVICE_MAP=auto.
PyTorch 2.7.1 (CUDA 12.8) on the CUDA 12.9 base image, Python 3.13, transformers 5.19.0, flash-linear-attention, and a compiled causal-conv1d 1.7.0.

strands-decider-2B-qwen3.5-v1-2610, strands-decider-E2B-gemma4-v1-2610, and strands-decider-E4B-gemma4-v1-2610, tested on one L4.
strands-decider-12B-gemma4-v1-2610, tested on one L40S.
strands-decider-26B-A4B-gemma4-v1-2610, tested across L40S GPUs with DECIDER_DEVICE_MAP=auto.