Fix per file ruff ignores related to line length (#26262)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -132,10 +132,14 @@ def test_simple_piecewise_compile(use_inductor):
|
||||
splitting_ops=["silly.attention"],
|
||||
use_inductor_graph_partition=False,
|
||||
use_inductor=use_inductor,
|
||||
expected_num_piecewise_graphs_seen=5, # 2 * num_layers + 1
|
||||
expected_num_piecewise_capturable_graphs_seen=3, # 1 + num_layers
|
||||
expected_num_backend_compilations=3, # num_piecewise_capturable_graphs_seen
|
||||
expected_num_cudagraph_captured=6, # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
|
||||
# 2 * num_layers + 1
|
||||
expected_num_piecewise_graphs_seen=5,
|
||||
# 1 + num_layers
|
||||
expected_num_piecewise_capturable_graphs_seen=3,
|
||||
# num_piecewise_capturable_graphs_seen
|
||||
expected_num_backend_compilations=3,
|
||||
# num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
|
||||
expected_num_cudagraph_captured=6,
|
||||
)
|
||||
|
||||
|
||||
@@ -147,14 +151,16 @@ def test_simple_inductor_graph_partition(splitting_ops):
|
||||
pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
|
||||
|
||||
_run_simple_model(
|
||||
# inductor graph partition automatically resets splitting_ops
|
||||
# to be an empty list
|
||||
# Inductor graph partition automatically resets splitting_ops to an empty list
|
||||
splitting_ops=splitting_ops,
|
||||
use_inductor_graph_partition=True,
|
||||
use_inductor=True,
|
||||
expected_num_piecewise_graphs_seen=1, # since not splitting at fx graph level
|
||||
expected_num_piecewise_capturable_graphs_seen=1, # since not splitting at fx graph level
|
||||
expected_num_backend_compilations=1, # since not splitting at fx graph level
|
||||
expected_num_cudagraph_captured=6, # inductor graph partition still captures 6
|
||||
# graph, same as fx graph partition.
|
||||
# Since not splitting at fx graph level
|
||||
expected_num_piecewise_graphs_seen=1,
|
||||
# Since not splitting at fx graph level
|
||||
expected_num_piecewise_capturable_graphs_seen=1,
|
||||
# Since not splitting at fx graph level
|
||||
expected_num_backend_compilations=1,
|
||||
# Inductor graph partition still captures 6 graph, same as fx graph partition
|
||||
expected_num_cudagraph_captured=6,
|
||||
)
|
||||
|
||||
@@ -367,11 +367,14 @@ def test_toy_llama(use_inductor: bool):
|
||||
kwargs = {"num_eager_compiles": 1, "num_inductor_compiles": 0}
|
||||
|
||||
with compilation_counter.expect(
|
||||
num_graphs_seen=1, # one graph for the model
|
||||
# One graph for the model
|
||||
num_graphs_seen=1,
|
||||
num_piecewise_graphs_seen=1,
|
||||
num_piecewise_capturable_graphs_seen=1,
|
||||
num_backend_compilations=1, # num_piecewise_capturable_graphs_seen
|
||||
num_cudagraph_captured=2, # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
|
||||
# num_piecewise_capturable_graphs_seen
|
||||
num_backend_compilations=1,
|
||||
# num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
|
||||
num_cudagraph_captured=2,
|
||||
**kwargs,
|
||||
):
|
||||
outputs.append(
|
||||
@@ -478,9 +481,10 @@ def benchmark():
|
||||
# it is fine here, because we only use the lambda function once.
|
||||
runtime = do_bench(
|
||||
lambda: graphs[b][0]( # noqa
|
||||
input_ids[:b], positions[:b]
|
||||
input_ids[:b], # noqa
|
||||
positions[:b], # noqa
|
||||
)
|
||||
) # noqa
|
||||
)
|
||||
piecewise_cudagraph_time[b] = runtime
|
||||
else:
|
||||
runtime = do_bench(lambda: graphs[b][0].replay()) # noqa
|
||||
|
||||
@@ -243,7 +243,7 @@ def test_fix_functionalization(model_class: torch.nn.Module, do_fusion: bool):
|
||||
# check if the functionalization pass is applied
|
||||
for op in model.ops_in_model(do_fusion):
|
||||
find_auto_fn(backend_no_func.graph_post_pass.nodes, op)
|
||||
assert find_auto_fn_maybe(backend_func.graph_post_pass.nodes, op) is None # noqa: E501
|
||||
assert find_auto_fn_maybe(backend_func.graph_post_pass.nodes, op) is None
|
||||
|
||||
# make sure the ops were all de-functionalized
|
||||
found = dict()
|
||||
|
||||
@@ -565,7 +565,7 @@ def test_attention_quant_pattern(
|
||||
elif quant_key.dtype == FP4_DTYPE:
|
||||
assert attn_nodes_post[0].kwargs.get("output_block_scale") is not None, (
|
||||
"Attention should have output_block_scale after FP4 fusion"
|
||||
) # noqa: E501
|
||||
)
|
||||
|
||||
# Check that results are close
|
||||
torch.testing.assert_close(result_unfused, result_fused_1, atol=1e-2, rtol=1e-2)
|
||||
|
||||
@@ -186,7 +186,7 @@ class TestQuantModel(torch.nn.Module):
|
||||
):
|
||||
# If fusion happens, the fused op is the one
|
||||
# we check for (de)functionalization
|
||||
return [torch.ops._C.fused_add_rms_norm_static_fp8_quant.default] # noqa: E501
|
||||
return [torch.ops._C.fused_add_rms_norm_static_fp8_quant.default]
|
||||
else:
|
||||
# If no fusion, the original ops are checked
|
||||
return [
|
||||
@@ -322,7 +322,7 @@ def sequence_parallelism_pass_on_test_model(
|
||||
# check if the functionalization pass is applied
|
||||
for op in model.ops_in_model():
|
||||
find_auto_fn(backend_no_func.graph_post_pass.nodes, op)
|
||||
assert find_auto_fn_maybe(backend_func.graph_post_pass.nodes, op) is None # noqa: E501
|
||||
assert find_auto_fn_maybe(backend_func.graph_post_pass.nodes, op) is None
|
||||
|
||||
# make sure the ops were all de-functionalized
|
||||
found = dict()
|
||||
|
||||
Reference in New Issue
Block a user