mirror of
https://github.com/bigscience-workshop/petals
synced 2024-10-31 09:20:41 +00:00
7bd5916744
1. Petals can be now installed using `pip install git+https://github.com/bigscience-workshop/petals` - In case if you already cloned the repo, you can do `pip install .` or `pip install .[dev]` 2. Moved `src` => `src/petals` - Replaced `from src.smth import smth` with `from petals.smth import smth` 3. Moved `cli` => `src/petals/cli` - Replaced `python -m cli.run_smth` with `python -m petals.cli.run_smth` (all utilities are now available right after pip installation) 4. Moved the `requirements*.txt` contents to `setup.cfg` (`requirements.txt` for packages is not supported well by modern packaging utils) 5. Increased the package version from `0.2` to `1.0alpha1`
44 lines
1.8 KiB
Python
44 lines
1.8 KiB
Python
import random
|
|
|
|
import hivemind
|
|
import pytest
|
|
import torch
|
|
from test_utils import *
|
|
|
|
from petals.bloom.from_pretrained import load_pretrained_block
|
|
from petals.client import DistributedBloomConfig
|
|
from petals.client.remote_sequential import RemoteTransformerBlock
|
|
from petals.data_structures import UID_DELIMITER
|
|
from petals.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(ValueError, match=r"Maximum length exceeded") 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)
|