mirror of
https://github.com/bigscience-workshop/petals
synced 2024-11-18 03:25:33 +00:00
0be21775af
* remove transformer block, implement as sequence size 1 * reimplement get_remote_module * fix readme Co-authored-by: Alexander Borzunov <hxrussia@gmail.com> Co-authored-by: Aleksandr Borzunov <borzunov.alexander@gmail.com>
47 lines
1.8 KiB
Python
47 lines
1.8 KiB
Python
import random
|
|
|
|
import hivemind
|
|
import pytest
|
|
import torch
|
|
import transformers
|
|
from hivemind import P2PHandlerError
|
|
from test_utils import *
|
|
|
|
import src
|
|
from src import DistributedBloomConfig
|
|
from src.bloom.from_pretrained import load_pretrained_block
|
|
from src.client.remote_sequential import RemoteTransformerBlock
|
|
from src.data_structures import UID_DELIMITER
|
|
from src.dht_utils import get_remote_module
|
|
|
|
|
|
@pytest.mark.forked
|
|
def test_remote_block_exact_match(atol_forward=1e-5, atol_inference=1e-3):
|
|
dht = hivemind.DHT(initial_peers=INITIAL_PEERS, client_mode=True, start=True)
|
|
config = DistributedBloomConfig.from_pretrained(MODEL_NAME)
|
|
|
|
for block_index in random.sample(range(config.n_layer), 3):
|
|
remote_block = get_remote_module(dht, f"{MODEL_NAME}{UID_DELIMITER}{block_index}", config)
|
|
assert isinstance(remote_block, RemoteTransformerBlock)
|
|
|
|
inputs = torch.randn(1, 8, config.hidden_size)
|
|
outputs_forward = remote_block(inputs)
|
|
|
|
outputs_inference = []
|
|
with remote_block.inference_session(max_length=inputs.shape[1]) as sess:
|
|
for i in range(inputs.shape[1]):
|
|
outputs_inference.append(sess.step(inputs[:, i : i + 1, :]))
|
|
|
|
# test that max length is respected
|
|
with pytest.raises(P2PHandlerError) as exc_info:
|
|
sess.step(inputs[:, -1:, :])
|
|
assert "Maximum length exceeded" in repr(exc_info.value)
|
|
|
|
outputs_inference = torch.cat(outputs_inference, dim=1)
|
|
|
|
ref_block = load_pretrained_block(MODEL_NAME, block_index, torch_dtype=torch.float32)
|
|
(outputs_local,) = ref_block(inputs)
|
|
|
|
assert torch.allclose(outputs_local, outputs_forward, rtol=0, atol=atol_forward)
|
|
assert torch.allclose(outputs_local, outputs_inference, rtol=0, atol=atol_inference)
|