support torchrun and optimize gpu assignment

[lxning]

Description

Please read our CONTRIBUTING.md prior to creating your first pull request.

Please include a summary of the feature or issue being fixed. Please also include relevant motivation and context. List any dependencies that are required for this change.

• support torchrun parameters in model config yaml file.
• support gpu assignment "deviceIds" in model config yaml file.
• update doc about model config and gpu assignment algorithm.
• set torchrun log-dir under logs

Fixes #(issue)
#2207

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Feature/Issue validation/testing

Please describe the Unit or Integration tests that you ran to verify your changes and relevant result summary. Provide instructions so it can be reproduced.
Please also list any relevant details for your test configuration.

• test gpu assignment

cat model-store/t/model-config.yaml
minWorkers: 1
maxWorkers: 1
maxBatchDelay: 100
responseTimeout: 120
parallelType: "pp"
deviceType: "gpu"
deviceIds: [2,3]

torchrun:
    nproc-per-node: 2

pippy:
    rpc_timeout: 1800

1. model-config.yaml
2. log

2023-03-30T23:45:17,829 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - [PID]67915
2023-03-30T23:45:17,830 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - Torch worker started.
2023-03-30T23:45:17,830 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - Python runtime: 3.8.16
2023-03-30T23:45:17,830 [DEBUG] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - W-29500-bloom_1.0 State change null -> WORKER_STARTED
2023-03-30T23:45:17,830 [DEBUG] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - W-29500-bloom_1.0 State change null -> WORKER_STARTED
2023-03-30T23:45:17,834 [INFO ] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - Connecting to: /tmp/.ts.sock.29500
2023-03-30T23:45:17,834 [INFO ] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - Connecting to: /tmp/.ts.sock.29500
2023-03-30T23:45:17,840 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - Connection accepted: /tmp/.ts.sock.29500.
2023-03-30T23:45:17,840 [INFO ] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - Connecting to: /tmp/.ts.sock.29501
2023-03-30T23:45:17,840 [INFO ] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - Connecting to: /tmp/.ts.sock.29501
2023-03-30T23:45:17,842 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - Connection accepted: /tmp/.ts.sock.29501.
2023-03-30T23:45:17,843 [INFO ] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - Flushing req.cmd LOAD to backend at: 1680219917843
2023-03-30T23:45:17,843 [INFO ] W-29500-bloom_1.0 org.pytorch.serve.wlm.WorkerThread - Flushing req.cmd LOAD to backend at: 1680219917843
2023-03-30T23:45:17,858 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - model_name: bloom, batchSize: 1
2023-03-30T23:45:17,864 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - model_name: bloom, batchSize: 1
2023-03-30T23:45:18,439 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - Transformers version 4.27.4
2023-03-30T23:45:18,440 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - CUDA_VISIBLE_DEVICES=2,3
2023-03-30T23:45:18,440 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - Transformers version 4.27.4
2023-03-30T23:45:18,440 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - CUDA_VISIBLE_DEVICES=2,3
2023-03-30T23:45:18,470 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - n_devs=2
2023-03-30T23:45:18,471 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - worker0, 1: 0
2023-03-30T23:45:18,471 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - worker1, 1: 1
2023-03-30T23:45:18,471 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - rank = 1 pid/device = 67916/cuda:1
2023-03-30T23:45:18,472 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - n_devs=2
2023-03-30T23:45:18,472 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - worker0, 0: 0
2023-03-30T23:45:18,472 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - worker1, 0: 1
2023-03-30T23:45:18,473 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - rank = 0 pid/device = 67915/cuda:0
2023-03-30T23:45:18,859 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - rpc.timeout=1800
2023-03-30T23:45:18,859 [INFO ] W-29500-bloom_1.0-stdout MODEL_LOG - rpc.timeout=1800

Checklist:

• Did you have fun?
• Have you added tests that prove your fix is effective or that this feature works?
• Has code been commented, particularly in hard-to-understand areas?
• Have you made corresponding changes to the documentation?

[codecov]

Codecov Report

Merging #2209 (0cc3f6d) into master (4ca86f3) will increase coverage by 0.01%.
The diff coverage is n/a.

❗ Current head 0cc3f6d differs from pull request most recent head ddccf5d. Consider uploading reports for the commit ddccf5d to get more accurate results

@@            Coverage Diff             @@
##           master    #2209      +/-   ##
==========================================
+ Coverage   71.44%   71.46%   +0.01%     
==========================================
  Files          73       73              
  Lines        3338     3329       -9     
  Branches       57       57              
==========================================
- Hits         2385     2379       -6     
+ Misses        950      947       -3     
  Partials        3        3              

Impacted Files	Coverage Δ
ts/service.py	74.35% <ø> (+3.09%)	⬆️

... and 2 files with indirect coverage changes

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[HamidShojanazeri]

thanks @lxning added two minor comments

[HamidShojanazeri]

@lxning are we still using it?

[lxning]

yes, this is unit test

[HamidShojanazeri]

adding an example how model_yaml_config can be useful + adding one how to access it in the handler.