[CODE] port HF bloom to hivemind

[justheuristic]

Why: so we can play with in in inference mode

Original bloom code: huggingface/transformers#17474

The quest is to

• implement bloom transformer layer as a hivemind expert.
• prepare a huggingface model that only has bloom embeddings and logits, but runs all transformer layers via hivemind.RemoteExpert

[justheuristic]

Naively quantized bloom 6b on CPU: http://nora:8800/notebooks/decentralized/jheuristic/bloom_test/cpu-qint-matmul.ipynb
source: https://gist.github.com/justheuristic/47830c9ddfd45889894e69d4f45ce233

[justheuristic]

On client-side computations

Since we plan to run embeddings/logits on client side, we need to compute them efficiently.
Embedding computation is cheap AF, but logits are more complicated

Computing final logits on colab CPU costs over a minute per one token

Solution 1: use fast KNN

• top-99% probabilities are held by the 100 most likely tokens
• use HNSW to find top-100 tokens that have highest dot product
• FAISS of ScaNN for fast nearest neighbor search

Solution 2: just use GPU

• colab T4 in fp16: 30ms per token (no longer a bottleneck)
• kudesnik m40 (~2x colab k80): 67ms, still very much acceptable
  
• gpus might not be available

Current opinion: use GPUs, think about fast CPU mode later.

[justheuristic]

Copied bloom version huggingface/transformers@ca2a55e
from huggingface

Their attention code is spectacularly bad, see #1

[justheuristic]

Next steps:

• push individual bloom layers to huggingface hub
• implement BloomBlock as hivemind.moe.server.ExpertBackend

[justheuristic]

As of 8221469

• Pushing model to hub is handled via python -m cli.convert_model --many_args_here, see README for usage example
• Server can run forward, backward and inference of bloom blocks, see README for instructions on how to start a server