Pin strands==1.34.1, use python:3.13-slim to match local

Switch to watsonx backend: Granite 4 H-Small on GPU, no local model needed

Revert to working python:3.12-slim + ollama CPU build

Use system python3 from ollama base, no version pinning

Use ollama base image directly with Python installed on top for GPU

Minimal Dockerfile: python:3.12-slim + ollama binary, cpu-upgrade

python:3.12-slim + CUDA runtime libs via apt (no deadsnakes, no nvidia base)

Fix: ubuntu22.04 not 24.04

Base on nvidia/cuda runtime for actual GPU access in ollama

Force NVIDIA_VISIBLE_DEVICES=all in entrypoint to override HF void setting

Copy ollama CUDA runners for GPU acceleration on T4

Copy ollama binary into python:3.13-slim — simplest approach

Multi-stage: Python 3.13 deps + Ollama CUDA runtime

Pull model at runtime instead of build time

Fix: wait for Ollama to start before pulling model

Switch to Ollama with GPU: pre-pulled granite4:micro, T4 tier

Revert to CPU build on cpu-upgrade — CUDA build stuck

Use pre-built CUDA binary instead of compiling from source

CUDA build + GPU offload for T4 tier

Disable GGML_NATIVE to avoid illegal instruction on different CPU

Revert to granite-4.0-micro (dense) — hybrid arch not supported in llama.cpp b5580

Switch to granite-4.0-h-micro (3B hybrid) — h-small OOM on cpu-basic

Upgrade to Granite 4 H-Small (8B) for better tool calling

Deploy Amanat: Granite 4 Micro + llama-server + Chainlit + Auth0