exllama-runpod-serverless

LLaMA GPTQ models with fast ExLlama inference on RunPod Serverless GPUs

Summary

This Docker image runs a Llama model on a serverless RunPod instance using the optimized turboderp's exllama repo.

Set Up

1. Create a RunPod account and navigate to the RunPod Serverless Console.
2. (Optional) Create a Network Volume to cache your model to speed up cold starts (but will incur some cost per hour for storage).
   • Note: Only certain Network Volume regions are compatible with certain instance types on RunPod, so try out if your Network Volume makes your desired instance type Unavailable, try other regions for your Network Volume.

3. Navigate to My Templates and click on the New Template button.

4. Enter in the following fields and click on the Save Template button:

   Template Field	Value
   Template Name	exllama-runpod-serverless
   Container Image	monotykamary/exllama-runpod-serverless:latest
   Container Disk	A size large enough to store your libraries + your desired model in 4bit.

   • Container Disk Size Guide:

     Model Parameters	Storage & VRAM
     7B	6GB
     13B	9GB
     33B	19GB
     65B	35GB
     70B	38GB

   • Environment Variables:

     Environment Variable	Example Value
     (Required) MODEL_REPO	TheBloke/airoboros-7B-gpt4-1.4-GPTQ or any other repo for GPTQ Llama model. See https://huggingface.co/models?other=llama&sort=trending&search=thebloke+gptq for other models. Must have .safetensors file(s).
     (Optional) PROMPT_PREFIX	"USER: "
     (Optional) PROMPT_SUFFIX	"ASSISTANT: "
     (Optional) MAX_SEQ_LEN	4096
     (Optional) ALPHA_VALUE	1
     (If using Network Volumes) HUGGINGFACE_HUB_CACHE	/runpod-volume/hub
     (If using Network Volumes) TRANSFORMERS_CACHE	/runpod-volume/hub

4. Now click on My Endpoints and click on the New Endpoint button.
5. Fill in the following fields and click on the Create button:

   Endpoint Field	Value
   Endpoint Name	exllama-runpod-serverless
   Select Template	exllama-runpod-serverless
   Min Provisioned Workers	0
   Max Workers	1
   Idle Timeout	5 seconds
   FlashBoot	Checked/Enabled
   GPU Type(s)	Use the Container Disk section of step 3 to determine the smallest GPU that can load the entire 4 bit model. In our example's case, use 16 GB GPU. Make smaller if using Network Volume instead.
   (Optional) Network Volume	airoboros-7b

Inference Usage

See the predict.py file for an example. For convenience we also copy the code below.

import os
import requests
from time import sleep
import logging
import argparse
import sys
import json

endpoint_id = os.environ["RUNPOD_ENDPOINT_ID"]
URI = f"https://api.runpod.ai/v2/{endpoint_id}/run"


def run(prompt, params={}, stream=False):
    request = {
        'prompt': prompt,
        'max_new_tokens': 1800,
        'temperature': 0.3,
        'top_k': 50,
        'top_p': 0.7,
        'repetition_penalty': 1.2,
        'batch_size': 8,
        'stream': stream
    }

    request.update(params)

    response = requests.post(URI, json=dict(input=request), headers = {
        "Authorization": f"Bearer {os.environ['RUNPOD_AI_API_KEY']}"
    })

    if response.status_code == 200:
        data = response.json()
        task_id = data.get('id')
        return stream_output(task_id, stream=stream)


def stream_output(task_id, stream=False):
    # try:
    url = f"https://api.runpod.ai/v2/{endpoint_id}/stream/{task_id}"
    headers = {
        "Authorization": f"Bearer {os.environ['RUNPOD_AI_API_KEY']}"
    }

    previous_output = ''

    try:
        while True:
            response = requests.get(url, headers=headers)
            if response.status_code == 200:
                data = response.json()
                if len(data['stream']) > 0:
                    new_output = data['stream'][0]['output']

                    if stream:
                        sys.stdout.write(new_output[len(previous_output):])
                        sys.stdout.flush()
                    previous_output = new_output

                if data.get('status') == 'COMPLETED':
                    if not stream:
                        return previous_output
                    break

            elif response.status_code >= 400:
                print(response)
            # Sleep for 0.1 seconds between each request
            sleep(0.1 if stream else 1)
    except Exception as e:
        print(e)
        cancel_task(task_id)


def cancel_task(task_id):
    url = f"https://api.runpod.ai/v2/{endpoint_id}/cancel/{task_id}"
    headers = {
        "Authorization": f"Bearer {os.environ['RUNPOD_AI_API_KEY']}"
    }
    response = requests.get(url, headers=headers)
    return response


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='Runpod AI CLI')
    parser.add_argument('-s', '--stream', action='store_true', help='Stream output')
    parser.add_argument('-p', '--params_json', type=str, help='JSON string of generation params')

    prompt = """Given the following clinical notes, what tests, diagnoses, and recommendations should the I give? Provide your answer as a detailed report with labeled sections "Diagnostic Tests", "Possible Diagnoses", and "Patient Recommendations".

17-year-old male, has come to the student health clinic complaining of heart pounding. Mr. Cleveland's mother has given verbal consent for a history, physical examination, and treatment
-began 2-3 months ago,sudden,intermittent for 2 days(lasting 3-4 min),worsening,non-allev/aggrav
-associated with dispnea on exersion and rest,stressed out about school
-reports fe feels like his heart is jumping out of his chest
-ros:denies chest pain,dyaphoresis,wt loss,chills,fever,nausea,vomiting,pedal edeam
-pmh:non,meds :aderol (from a friend),nkda
-fh:father had MI recently,mother has thyroid dz
-sh:non-smoker,mariguana 5-6 months ago,3 beers on the weekend, basketball at school
-sh:no std,no other significant medical conditions."""
    args = parser.parse_args()
    params = json.loads(args.params_json) if args.params_json else "{}"
    import time
    start = time.time()
    print(run(prompt, params=params, stream=args.stream))
    print("Time taken: ", time.time() - start, " seconds")

Run the above code using the following command in terminal with the runpoint endpoint id assigned to your endpoint in step 5.

RUNPOD_AI_API_KEY='**************' RUNPOD_ENDPOINT_ID='*******' python predict.py

To run with streaming enabled, use the --stream option. To set generation parameters, use the --params_json option to pass a JSON string of parameters:

RUNPOD_AI_API_KEY='**************' RUNPOD_ENDPOINT_ID='*******' python predict.py --params_json '{"temperature": 0.3, "max_tokens": 1000, "prompt_prefix": "USER: ", "prompt_suffix": "ASSISTANT: "}'

You can generate the API key here under API Keys.