build AWS AMI for radeon gpu testing

[soulaway]

We need to provide juno-deploy support for Radeon instances g4ad same as for g4dn

So that we may compare +- same hardware against each other.

[soulaway]

Currently I stuck at running an inference.
AMI Bake seems fine, but needs assist of @1nk1

ubuntu@ip-172-31-43-199:~$ /bin/sudo journalctl -u juno-node --no-pager -n 400
Jun 08 15:40:01 ip-172-31-43-199 systemd[1]: Started Juno Node node-1.
Jun 08 15:40:02 ip-172-31-43-199 juno-node[2242]: [0.863s][info][jfr,startup] Started recording 1. The result will be written to:
Jun 08 15:40:02 ip-172-31-43-199 juno-node[2242]: [0.863s][info][jfr,startup]
Jun 08 15:40:02 ip-172-31-43-199 juno-node[2242]: [0.863s][info][jfr,startup] /opt/juno/jfr/juno-tinyllama-1.1b-chat-v1.0-q4_k_m-20260608-154001.jfr
Jun 08 15:40:03 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:03 PM cab.ml.juno.health.HealthReporter startBackground
Jun 08 15:40:03 ip-172-31-43-199 juno-node[2242]: INFO: HealthReporter started for [node-1] → http://172.31.43.199:8080/health/probe  interval=5000 ms
Jun 08 15:40:03 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:03 PM cab.ml.juno.node.NodeMain main
Jun 08 15:40:03 ip-172-31-43-199 juno-node[2242]: INFO: Health reporter started → http://172.31.43.199:8080
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: WARNING: A terminally deprecated method in sun.misc.Unsafe has been called
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: WARNING: sun.misc.Unsafe::objectFieldOffset has been called by io.grpc.netty.shaded.io.netty.util.internal.PlatformDependent0$4 (file:/opt/juno/juno-node/target/juno-node.jar)
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: WARNING: Please consider reporting this to the maintainers of class io.grpc.netty.shaded.io.netty.util.internal.PlatformDependent0$4
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: WARNING: sun.misc.Unsafe::objectFieldOffset will be removed in a future release
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:04 PM cab.ml.juno.node.EmbeddedNodeServer start
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: INFO: Node [node-1] gRPC server started on port 19092
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: READY:node-1:19092
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:04 PM cab.ml.juno.node.NodeMain main
Jun 08 15:40:04 ip-172-31-43-199 juno-node[2242]: INFO: Node [node-1] running on port 19092 — waiting for requests
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:21 PM cab.ml.juno.node.EmbeddedNodeServer$NodeServiceImpl loadShard
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: INFO: Loading real model from: /opt/juno/models/tinyllama-1.1b-chat-v1.0-q4_k_m.gguf (useGpu=true)
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:21 PM cab.ml.juno.node.EmbeddedNodeServer$NodeServiceImpl loadShard
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: INFO: Shard context: layers 0-22 embeddings=true outputProj=true
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:21 PM cab.ml.juno.node.CudaBindings <clinit>
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: INFO: CudaBindings unavailable — Could not load CUDA library — tried: [libcudart.so.12, libcudart.so]. Ensure CUDA 12.x is installed and libcudart.so.12 is on LD_LIBRARY_PATH.
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:21 PM cab.ml.juno.node.CudaAvailability detect
Jun 08 15:40:21 ip-172-31-43-199 juno-node[2242]: INFO: CUDA not available — CudaBindings did not load
Jun 08 15:40:42 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:42 PM cab.ml.juno.node.RocmBindings probeHSSgemv
Jun 08 15:40:42 ip-172-31-43-199 juno-node[2242]: WARNING: rocblas_hssgemv_strided_batched: gfx1010/gfx1011 kernel absent in ROCm 7.x — disabling FP16 GEMV path
Jun 08 15:40:42 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:42 PM cab.ml.juno.node.RocmBindings <init>
Jun 08 15:40:42 ip-172-31-43-199 juno-node[2242]: INFO: RocmBindings ready — Panama FFI (libamdhip64 + librocblas) hssgemv=false
Jun 08 15:40:42 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:42 PM cab.ml.juno.node.RocmAvailability detect
Jun 08 15:40:42 ip-172-31-43-199 juno-node[2242]: INFO: ROCm available — 1 device(s)
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.GpuContext create
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: GpuContext ready — [rocm] device 0: AMD Radeon Pro V520, 8.4 GB VRAM (shared=true)
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.GpuContext shared
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: GpuContext.shared(0) — process-wide GPU context installed
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.GgufReader open
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: GGUF v3 — tensors=201 metadata=32
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.GgufReader open
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: Data section starts at byte 736800
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.ForwardPassHandlerLoader load
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: Detected architecture: llama  backend=RocmMatVec  file=/opt/juno/models/tinyllama-1.1b-chat-v1.0-q4_k_m.gguf  lora=none
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.ForwardPassHandlerLoader load
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: Routing to LlamaTransformerHandler (LLaMA-family architecture: llama)
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.LlamaTransformerHandler load
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: Loading GGUF shard: layers 0–22  embd=true  outProj=true  backend=RocmMatVec  file=/opt/juno/models/tinyllama-1.1b-chat-v1.0-q4_k_m.gguf
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.GgufReader open
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: GGUF v3 — tensors=201 metadata=32
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.GgufReader open
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: Data section starts at byte 736800
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:43 PM cab.ml.juno.node.LlamaTransformerHandler load
Jun 08 15:40:43 ip-172-31-43-199 juno-node[2242]: INFO: Model: LlamaConfig{arch=llama hidden=2048 layers=22 heads=32 kvHeads=4 headDim=64 ffn=5632 vocab=32000 eps=1.0e-05 ropeTheta=10000}
Jun 08 15:40:44 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:44 PM cab.ml.juno.node.LlamaTransformerHandler <init>
Jun 08 15:40:44 ip-172-31-43-199 juno-node[2242]: INFO: Shard loaded — 22 layers, with embeddings, with output projection
Jun 08 15:40:44 ip-172-31-43-199 juno-node[2242]: Jun 08, 2026 3:40:44 PM cab.ml.juno.node.LlamaTransformerHandler uploadFp32Resident
Jun 08 15:40:44 ip-172-31-43-199 juno-node[2242]: INFO: Uploading dequantized weights to GPU (FP32 device-resident — FP16 GEMV unsupported on this GFX target)…
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: #
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # A fatal error has been detected by the Java Runtime Environment:
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: #
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: #  SIGSEGV (0xb) at pc=0x000073c6829e9e24, pid=2242, tid=2581
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: #
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # JRE version: OpenJDK Runtime Environment (25.0.3+9) (build 25.0.3+9-2-22.04.2-Ubuntu)
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # Java VM: OpenJDK 64-Bit Server VM (25.0.3+9-2-22.04.2-Ubuntu, mixed mode, sharing, tiered, compressed oops, compressed class ptrs, g1 gc, linux-amd64)
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # Problematic frame:
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # C  [libamdhip64.so+0x3e9e24]
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: #
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # Core dump will be written. Default location: Determined by the following: "/usr/share/apport/apport -p%p -s%s -c%c -d%d -P%P -u%u -g%g -F%F -- %E" (alternatively, falling back to /opt/juno/core.2242)
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: #
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # JFR recording file will be written. Location: /opt/juno/hs_err_pid2242.jfr
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: #
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # An error report file with more information is saved as:
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: # /opt/juno/hs_err_pid2242.log
Jun 08 15:40:54 ip-172-31-43-199 juno-node[2242]: [52.946s][warning][os         ] Loading hsdis library failed
Jun 08 15:40:55 ip-172-31-43-199 juno-node[2242]: #
Jun 08 15:40:55 ip-172-31-43-199 juno-node[2242]: # If you would like to submit a bug report, please visit:
Jun 08 15:40:55 ip-172-31-43-199 juno-node[2242]: #   https://bugs.launchpad.net/ubuntu/+source/openjdk-25
Jun 08 15:40:55 ip-172-31-43-199 juno-node[2242]: # The crash happened outside the Java Virtual Machine in native code.
Jun 08 15:40:55 ip-172-31-43-199 juno-node[2242]: # See problematic frame for where to report the bug.
Jun 08 15:40:55 ip-172-31-43-199 juno-node[2242]: #