app.py

import torch
import torch.nn.functional as F
import transformers
import gradio as gr

from src.client import DistributedBloomForCausalLM

INITIAL_PEERS = ['/ip4/193.106.95.184/tcp/443/p2p/QmSXDXLeSMXjS4YerDrdn1zpGQaNzkZ9ogN2SoAEyAdDhs']

import hivemind  # test that DHT instances work on localhost
dht1 = hivemind.DHT(start=True)
dht2 = hivemind.DHT(start=True, initial_peers=dht1.get_visible_maddrs())


tokenizer = transformers.BloomTokenizerFast.from_pretrained("bigscience/test-bloomd-6b3")
model = DistributedBloomForCausalLM.from_pretrained("bigscience/test-bloomd-6b3", initial_peers=INITIAL_PEERS, low_cpu_mem_usage=True, torch_dtype=torch.float32)

def inference(text, seq_length=1):
    input_ids = tokenizer(text, return_tensors='pt')['input_ids']
    final_tokens = input_ids
    with torch.inference_mode(), model.transformer.h.inference_session() as remote_transformer:
        for i in range(seq_length):
            h = model.transformer.word_embeddings(input_ids)
            h = model.transformer.word_embeddings_layernorm(h)
            h = remote_transformer.step(h)
            h = model.transformer.ln_f(h)
            h = F.linear(h, weight=model.transformer.word_embeddings.weight)  # note: this line takes a while, will also be fixed
            next_token_ix = torch.multinomial((h[0, -1] / 0.8).softmax(-1), 1)
        
            final_tokens = torch.cat([final_tokens, next_token_ix.view(1, 1)], dim=-1)
            input_ids = next_token_ix.view(1, 1)
        return tokenizer.decode(final_tokens[0], skip_special_tokens=False)
 
iface = gr.Interface(
fn=inference, 
inputs=[
gr.Textbox(lines=10, label="Input text"),
gr.inputs.Slider(
        minimum=0,
        maximum=1000,
        step=1,
        default=42,
        label="Sequence length for generation"
    )
], 
outputs="text"
)
iface.launch()