Skip to content

Fix flakey XNNPACK tests #19653

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
metascroy wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix flakey XNNPACK tests #19653

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 18 additions & 1 deletion backends/xnnpack/test/models/llama2_et_example.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,9 +17,11 @@ def setUp(self):
torch._dynamo.reset()

def test_f32(self):
torch.manual_seed(0)
self._test()

def test_f16(self):
torch.manual_seed(0)
self._test(torch.float16)

# TODO - dynamic shape
Expand All @@ -31,7 +33,22 @@ def _test(self, dtype: torch.dtype = torch.float):
], f"Only fp32 and fp16 are supported, but got dtype: {dtype}"

llama2 = Llama2Model()
model = llama2.get_eager_model().to(dtype)
model = llama2.get_eager_model()
# The example uses a dummy small model with random weights for demo
# purposes only. Default torch init (e.g. nn.Embedding ~ N(0, 1))
# combined with the model dim produces intermediate activations that
# overflow in fp16 (max ~65504), yielding nan/-inf and making the
# output comparison flaky. Re-init parameters AND float buffers (RoPE
# tables, causal mask, etc.) to a small bounded range so activations
# stay representable; this still exercises the export + lowering
# pipeline.
with torch.no_grad():
for p in model.parameters():
p.uniform_(-0.02, 0.02)
for b in model.buffers():
if b.is_floating_point():
b.uniform_(-0.02, 0.02)
model = model.to(dtype)

# Only convert fp32 inputs to dtype
example_inputs = tuple(
Expand Down
2 changes: 2 additions & 0 deletions backends/xnnpack/test/ops/test_linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -547,6 +547,7 @@ def get_qnode_checks(quant_node_checks, dialect):
# )

def test_qd8_f32_per_channel_shared_dq_chain(self):
torch.manual_seed(42)
for use_bias in (False, True):
module = SharedDQChain(
input_size=13,
Expand All @@ -561,6 +562,7 @@ def test_qd8_f32_per_channel_shared_dq_chain(self):
is_per_channel=True,
linear_count=2,
uses_bias=use_bias,
atol=1.5e-1, # TODO(T212995726): Investigate right atol for rand[n] inputs
)

def _test_qd8_per_channel_linear(self, dtype: torch.dtype = torch.float):
Expand Down
Loading