PR #1992 [inference] test suite for ds-kernels (bert, roberta, gpt2, gpt-neo, gpt-j)

add fill masks tests

jeffra committed 3 years ago

add a temporary fix for the Bert-Type models; TODO: Fix this for the inference pipeline

Reza Yazdani committed 3 years ago

add cuda graph to tests

jeffra committed 3 years ago

add fp16/32

jeffra committed 3 years ago

add q&a tests and set hf framework to pt

jeffra committed 3 years ago

Merge branch 'master' into jeffra/fill-mask-tests

jeffra committed 3 years ago

fixed assignment of DS pipeline model

mrwyattii committed 3 years ago

generalized the model/task testing and added pytest fixtures

mrwyattii committed 3 years ago

fixed syntax error

mrwyattii committed 3 years ago

keyerror with text-gen assert_fn

mrwyattii committed 3 years ago

added back gptj model

mrwyattii committed 3 years ago

Merge branch 'master' into jeffra/fill-mask-tests

RezaYazdaniAminabadi committed 3 years ago

added temporary fix to get gpt models passing

mrwyattii committed 3 years ago

fixing some issue for the Bert-Postln model

Reza Yazdani committed 3 years ago

Merge branch 'jeffra/fill-mask-tests' of github.com:microsoft/DeepSpeed into jeffra/fill-mask-tests

Reza Yazdani committed 3 years ago

formatting

mrwyattii committed 3 years ago

added fp16 tests

mrwyattii committed 3 years ago

fixed bad comparison

mrwyattii committed 3 years ago

avoid CUDA init when getting GPU memory

mrwyattii committed 3 years ago

revert to not checking GPU memory

mrwyattii committed 3 years ago

extra import

mrwyattii committed 3 years ago

refactored validation of test config, add skip for AMD w/CUDA Graph

mrwyattii committed 3 years ago

cuda graph requires 1.10+

jeffra committed 3 years ago

ensure torch > 1.10 if cuda graph enabled, also fix typo

jeffra committed 3 years ago

prune incompatible model+task pairs

jeffra committed 3 years ago

add pretty print names for test fixtures

jeffra committed 3 years ago

use fp16 revision of gpt-j

jeffra committed 3 years ago

convert pipe model to half for non-DS case

jeffra committed 3 years ago

Merge branch 'master' into jeffra/fill-mask-tests

jeffra committed 3 years ago

remove dep on hf-accelerate

jeffra committed 3 years ago

manual convert gpt-j fp32->fp16, seeing issues with fp16 revision

jeffra committed 3 years ago

force inference tests to be seq

jeffra committed 3 years ago

added correctness tests for text-generation gpt models

mrwyattii committed 3 years ago

fixed device string for lm tests

mrwyattii committed 3 years ago

during injection reshape data on cpu to avoid extra GPU memory

jeffra committed 3 years ago

increased acceptable ppl diff val

mrwyattii committed 3 years ago

added perf test and fix for gpt-j-6B OOM on V100 16GB

mrwyattii committed 3 years ago

remove sequential and move nv-nightly to nightly

jeffra committed 3 years ago

Merge branch 'master' into jeffra/fill-mask-tests

jeffra committed 3 years ago

Merge branch 'master' into jeffra/fill-mask-tests

jeffra committed 3 years ago

skip inference tests on rocm for now

jeffra committed 3 years ago

add missing import

jeffra committed 3 years ago

make performance asserts less strict

mrwyattii committed 3 years ago

fix for datasets error

mrwyattii committed 3 years ago

seperate inference tests from regular unit tests

mrwyattii committed 3 years ago

added new req file for inference workflow

mrwyattii committed 3 years ago

skip test if deps aren't installed

jeffra committed 3 years ago

fix for inf reqs install

mrwyattii committed 3 years ago

DeepSpeed [inference] test suite for ds-kernels (bert, roberta, gpt2, gpt-neo, gpt-j) #1992 Merged

DeepSpeed
[inference] test suite for ds-kernels (bert, roberta, gpt2, gpt-neo, gpt-j)
#1992

Merged