Giters

ravenscroftj / turbopilot

Turbopilot is an open source large-language-model based code completion engine that runs locally on CPU

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Wasn't able to get the docker image running (Illegal instruction (core dumped)) (no AVX support?)

voarsh2 opened this issue · comments

commented 
main: seed = 1681324505

Illegal instruction (core dumped)

Wasn't able to get the docker image running.

apiVersion: apps/v1
kind: Deployment
metadata:
  annotations:
    deployment.kubernetes.io/revision: "4"
    field.cattle.io/publicEndpoints: '[{"addresses":["192.168.100.103"],"port":32435,"protocol":"TCP","serviceName":"serge:turbopilot-nodeport","allNodes":true}]'
  labels:
    workload.user.cattle.io/workloadselector: apps.deployment-serge-turbopilot
  name: turbopilot
  namespace: serge
spec:
  progressDeadlineSeconds: 600
  replicas: 1
  revisionHistoryLimit: 10
  selector:
    matchLabels:
      workload.user.cattle.io/workloadselector: apps.deployment-serge-turbopilot
  strategy:
    rollingUpdate:
      maxSurge: 25%
      maxUnavailable: 25%
    type: RollingUpdate
  template:
    metadata:
      creationTimestamp: null
      labels:
        workload.user.cattle.io/workloadselector: apps.deployment-serge-turbopilot
    spec:
      containers:
      - env:
        - name: MODEL
          value: /models/codegen-2B-multi-ggml-4bit-quant.bin
        image: ghcr.io/ravenscroftj/turbopilot/turbopilot:latest
        imagePullPolicy: IfNotPresent
        name: container-0
        ports:
        - containerPort: 18080
          name: turbopilot
          protocol: TCP
        terminationMessagePath: /dev/termination-log
        terminationMessagePolicy: File
        volumeMounts:
        - mountPath: /models
          name: turbopilot
      dnsPolicy: ClusterFirst
      restartPolicy: Always
      schedulerName: default-scheduler
      terminationGracePeriodSeconds: 30
      volumes:
      - name: turbopilot
        persistentVolumeClaim:
          claimName: turbopilot
commented

Hi there thanks for your ticket. Can I ask what operating system and processor architecture you are running your k8s cluster on?

commented

Hi there thanks for your ticket. Can I ask what operating system and processor architecture you are running your k8s cluster on?

I am running Ubuntu 22.04 LTS.
Kubernetes V1.23
The CPU's support AVX (not AVX2) - some example of my node CPU's:
(48 x Intel(R) Xeon(R) CPU E5-2697 v2 @ 2.70GHz (2 Sockets), 32 x Intel(R) Xeon(R) CPU E5-2650 v2 @ 2.60GHz (2 Sockets))

  • I haven't tried the container on my Ryzen system that has AVX2 support..... will try that now......

Update. No complaints on the K8 node that has (what I assume) is AVX2 support. You might want to add AVX support (plenty of powerful enough CPU's don't have AVX2 support) - gotta love bulky RAM. :D

commented

It happened to me, too
when i run
docker run --rm -it -v /home/yang/models:/models -e MODEL="/models/codegen-2B-multi-ggml-4bit-quant.bin" -p 80:18080 ghcr.io/ravenscroftj/turbopilot/turbopilot:latest
terminal output
main: seed = 1681348410
gptj_model_load: loading model from '/models/codegen-2B-multi-ggml-4bit-quant.bin' - please wait ...
gptj_model_load: n_vocab = 51200
gptj_model_load: n_ctx = 2048
gptj_model_load: n_embd = 2560
gptj_model_load: n_head = 32
gptj_model_load: n_layer = 32
gptj_model_load: n_rot = 64
gptj_model_load: f16 = 3
Illegal instruction (core dumped)

My Operating Environment :
docker 20.10.18 Community
Ubuntu 22.04.1 LTS
Intel(R) Celeron(R) CPU J1900 @ 1.99GHz
I pulled the docker image directly

commented

Great thanks for looking into the AVX thing - in hindsight that was the obvious problem - I should have known 🤦

So I think the short term solution would be to build the image on the system you're targetting and the C preprocessor should pick up what instructionsets are supported. The slightly longer term solution is for me to add CI builds that build with different sets of CPU instructions and make them available as part of the release

commented 
docker run --rm -it -v ./models:/models -e THREADS=6 -e MODEL="/models/codegen-6B-multi-ggml-4bit-quant.bin" -p 18080:18080 ghcr.io/ravenscroftj/turbopilot/turbopilot:latest

WARNING: The requested image's platform (linux/amd64) does not match the detected host platform (linux/arm64/v8) and no specific platform was requested

main: seed = 1681194918
gptj_model_load: loading model from '/models/codegen-6B-multi-ggml-4bit-quant.bin' - please wait ...
gptj_model_load: n_vocab = 51200
gptj_model_load: n_ctx   = 2048
gptj_model_load: n_embd  = 4096
gptj_model_load: n_head  = 16
gptj_model_load: n_layer = 33
gptj_model_load: n_rot   = 64
gptj_model_load: f16     = 2
gptj_model_load: ggml ctx size = 6325.92 MB
Illegal instruction (core dumped)

Apple M1 Pro
Mac 13.2
colima 0.5.4

commented

hi @oshanz - please try changing the image uri to ghcr.io/ravenscroftj/turbopilot:latest - the path from your log above is the old build.

commented

@ravenscroftj same issue with above image on mac m1.

commented

hi @4j4y - just to clarify when you say image above you are referring to ghcr.io/ravenscroftj/turbopilot:latest (as opposed to the old image with /turbopilot/turbopilot in it)?

I don't have an apple silicon device so I'm struggling to work out what is working and what isn't. Are you running docker desktop on your m1 in order to run the image? Don't suppose you can share the output from your docker pull or docker run command (I'm after the image hash that you are using).

Finally - have you tried downloading the binary zip file from the releases page and running that? Does it give the same error?

commented

support for older instruction sets is provided in release 0.0.5