Skip to content

GH Task Runner (RWKV-x-dev focused suite) #95

GH Task Runner (RWKV-x-dev focused suite)

GH Task Runner (RWKV-x-dev focused suite) #95

name: GH Task Runner (RWKV-x-dev focused suite)
on:
workflow_dispatch:
inputs:
approval_notice:
description: 'WARNING: This will spin up a large number of tasks - get approval from admin before running'
required: false
default: 'NOT_APPROVED'
type: choice
options:
- NOT_APPROVED
- APPROVED
model_hf_repo:
description: 'Model Hugging Face Repository'
required: true
default: 'RWKV/rwkv-5-world-1b5'
model_args:
description: 'Model Arguments (ie: dtype="float16")'
required: false
default: 'dtype=bfloat16,trust_remote_code=True'
batch_size:
description: 'Batch Size'
required: true
default: 'auto'
# backend:
# description: 'Backend to use'
# required: true
# default: 'nvidia-gpu'
# type: choice
# options:
# - nvidia-gpu
# - intel-gpu
# - amd-gpu
# - any-gpu
gpu_vram:
description: 'Minimum GPU VRAM (ignored for MPS)'
required: true
default: '24'
type: choice
options:
- 16
- 24
# - 40
- 48
- 80
rwkv5_file_url:
description: 'Model file URL (for rwkv5 .pth eval)'
default: ''
required: false
rwkv5_test_name:
description: 'Model dev test name (for test)'
default: 'TEST_MODEL_FILE'
required: false
num_fewshot:
description: 'num_fewshot setting (ignored if < 0)'
required: true
default: -1
upload_output:
description: 'Upload to HF / B2'
required: false
default: true
type: boolean
env:
# HF repo to sync to
HF_REPO_SYNC: rwkv-x-dev/lm-eval-output
# Model HF repo
MODEL_HF_REPO: ${{ github.event.inputs.model_hf_repo }}
# HF / B2 sync settings
HUGGING_FACE_HUB_TOKEN: ${{secrets.HUGGING_FACE_HUB_TOKEN}}
B2_APPLICATION_KEY_ID: ${{secrets.B2_APPLICATION_KEY_ID}}
B2_APPLICATION_KEY: ${{secrets.B2_APPLICATION_KEY}}
B2_PATH_LM_EVAL_OUTPUT: ${{vars.B2_PATH_LM_EVAL_OUTPUT}}
jobs:
gh-task-runner:
# Check for approval notice
if: ${{ github.event.inputs.approval_notice == 'APPROVED' }}
# Strategy Matrix
strategy:
# Disable fail-fast behavior
fail-fast: false
matrix:
# NOTE: There is a matrix limit of 256 on github
run_task:
### Main suite metrics to include
### ----------------------------------------------------------------------------
- anli
- glue
- truthfulqa
- cmmlu
- lambada
- pythia
- mmlu
- blimp
- ai2_arc
# - arc_easy
# - arc_challenge
- logiqa
- winogrande
- openbookqa
- hellaswag
- copa
- sciq
- piqa
### Slow suite metrics to include
### ----------------------------------------------------------------------------
# - triviaqa
- record
### Various multilang tasks
### ----------------------------------------------------------------------------
- xcopa
# - xcopa_*
- xnli
# - xnli_*
- xstorycloze
# - xstorycloze_*
- xwinograd
# - xwinograd_*
- lambada_multilingual
# - lambada_openai_*
- pawsx
# - paws_*
# Name of the job
name: "[${{ matrix.run_task }}] ${{ github.event.inputs.rwkv5_file_url && github.event.inputs.rwkv5_test_name || github.event.inputs.model_hf_repo }} - ${{ github.event.inputs.model_args }}"
# Due to github worker hard limitation, of 24 hours
# we apply a timeout of 23 hours instead.
timeout-minutes: 1380
# Select the type of runner that the job will run on
runs-on:
- nvidia-gpu
- gpu-vram-${{ github.event.inputs.gpu_vram }}
# - gpu-count-any
# Actual task setup, and run steps
steps:
- name: Checkout repository
uses: actions/checkout@v3
- name: Run the task
uses: ./.github/actions/gh-task-runner-composite
with:
run_task: ${{ matrix.run_task }}
num_fewshot: ${{ github.event.inputs.num_fewshot }}
model_hf_repo: ${{ github.event.inputs.model_hf_repo }}
model_args: ${{ github.event.inputs.model_args }}
batch_size: ${{ github.event.inputs.batch_size }}
rwkv5_file_url: ${{ github.event.inputs.rwkv5_file_url }}
rwkv5_test_name: ${{ github.event.inputs.rwkv5_test_name }}
upload_output: ${{ github.event.inputs.upload_output }}
# upload_output:
# name: "Upload to HF / B2"
# needs: gh-task-runner
# runs-on: ubuntu-latest
# if: ${{ github.event.inputs.upload_output }}
# steps:
# - name: Checkout repository
# uses: actions/checkout@v3
# - name: Run the task
# uses: ./.github/actions/gh-upload-output
# with:
# num_fewshot: ${{ github.event.inputs.num_fewshot }}
# model_hf_repo: ${{ github.event.inputs.model_hf_repo }}
# model_args: ${{ github.event.inputs.model_args }}
# batch_size: ${{ github.event.inputs.batch_size }}
# backend: nvidia-gpu