`convert_to_hf_gptneox` adds support to other neox models

[LorrinWWW]

Previous conversion code assumes the model is neox-20b, and some config is hard-coded, i.e. 44 layers. (also added a fp16 option.)

example:

python tools/convert_to_hf_gptneox.py \
    --config-name EleutherAI/pythia-6.9b-deduped \
    --ckpt-path ${PATH_CKPTS}/checkpoint_100 --save-path pythia-6.9b-finetuned \
    --n-stages 4 --n-layer-per-stage 8 --fp16

[ETERNALBLUEbullrun]

Would use (plus help with issues,) but for the lack of support for regular GPT models.

[csris]

We need to update the README as well. @LorrinWWW, can you do that?

[LorrinWWW]

@csris Sure I will do that

[LorrinWWW]

Done

[814007450]

谢谢，我已收到您的来信。Thank you！I have received your letter.ありがとう！電子メールを受信しました！