mirror of
https://github.com/bigscience-workshop/petals
synced 2024-10-31 09:20:41 +00:00
f0c7383181
- finish renaming RemoteSequenceInfo -> RemoteSequenceManager (why: if it was an *Info, user would expect it to be similar - to a dataclass; whereas in actuality, the class is doing heavy network interactions on its own) - implement RemoteSequenceManager.make_sequence (from https://pastebin.com/uXgy2U8B ) - make RemoteSequentialInferenceSession use RemoteSequenceManager.make_sequence - make tests pass again - make it possible to create inference session without RemoteTransformerBlock - make a standalone test for RemoteSequential - rollback convert-model Co-authored-by: Tim Dettmers <tim.dettmers@gmail.com>
40 lines
1.5 KiB
Python
40 lines
1.5 KiB
Python
import random
|
|
|
|
import hivemind
|
|
import pytest
|
|
import torch
|
|
import transformers
|
|
from test_utils import *
|
|
|
|
from src.bloom.from_pretrained import load_pretrained_block
|
|
from src.client.remote_block 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 = transformers.AutoConfig.from_pretrained(MODEL_NAME)
|
|
|
|
for block_index in random.sample(range(config.n_layer), 3):
|
|
block_uid = f"{MODEL_NAME}{UID_DELIMITER}{block_index}"
|
|
remote_block = get_remote_module(dht, block_uid)
|
|
assert remote_block is not None, f"Could not find {block_uid} in DHT"
|
|
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() as sess:
|
|
for i in range(inputs.shape[1]):
|
|
outputs_inference.append(sess.step(inputs[:, i : i + 1, :]))
|
|
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)
|