github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-12-06 00:08:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

[inference] streaming Linear 1D Row inference (#1874)

Browse Source
This commit is contained in:
Jiarui Fang
2022-11-10 17:03:21 +08:00
committed by GitHub
parent a141681260
commit c2947dadf1
4 changed files with 629 additions and 554 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
tests/test_fx/test_complete_workflow.py
View File

@@ -32,7 +32,7 @@ class MLP(torch.nn.Module):
return x
def run_workflow(world_size):
def run_workflow(world_size, dev):
# initailization
with LazyInitContext() as ctx:
model = MLP(16)
@@ -46,7 +46,7 @@ def run_workflow(world_size):
gm = torch.fx.GraphModule(model, graph, model.__class__.__name__)
# annotate
annotated_gm = transformer_mlp_pass(gm, process_group=ProcessGroup())
annotated_gm = transformer_mlp_pass(gm, process_group=ProcessGroup(tp_degree=world_size))
annotated_gm.recompile()
# materialization and sharding
@@ -61,22 +61,25 @@ def run_workflow(world_size):
# test forward to make sure that IR transform will produce the same results
# like how ColoTensor would do it normally
data = torch.rand(4, 16)
data = torch.rand(4, 16, device=dev)
non_fx_out = model(data)
fx_out = annotated_gm(data)
assert torch.equal(non_fx_out, fx_out), f'{non_fx_out} vs {fx_out}'
def run_dist(rank, world_size, port):
def run_dist(rank, world_size, dev, port):
colossalai.launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
run_workflow(world_size)
run_workflow(world_size, dev)
@pytest.mark.dist
@pytest.mark.parametrize('world_size', [1, 2])
@pytest.mark.parametrize('dev', ['cuda', 'cpu'])
@rerun_if_address_is_in_use()
def test_complete_workflow(world_size):
run_func = partial(run_dist, world_size=world_size, port=free_port())
def test_complete_workflow(world_size, dev):
if dev == 'cpu' and world_size > 1:
return
run_func = partial(run_dist, world_size=world_size, dev=dev, port=free_port())
mp.spawn(run_func, nprocs=world_size)