Skip to content
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

Jump to bottom

Support QLayerNorm without weights #341

Merged
merged 2 commits into from
Oct 29, 2024
Merged

Support QLayerNorm without weights #341

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
500e77d
test(QLayerNorm): add non-affine configurations
dacorvo Oct 29, 2024
97a6266
fix(QModule): support modules without weights
dacorvo Oct 29, 2024
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
5 changes: 3 additions & 2 deletions optimum/quanto/nn/qmodule.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -217,7 +217,8 @@ def from_module(
if qmodule is None:
return None
# Move the quantized module to the target device, but with empty weights
qmodule = qmodule.to_empty(device=module.weight.device)
device = torch.device("cpu") if module.weight is None else module.weight.device
qmodule = qmodule.to_empty(device=device)
# Set scales that were initialized to empty values
qmodule.input_scale = torch.ones_like(qmodule.input_scale)
qmodule.output_scale = torch.ones_like(qmodule.output_scale)
Expand All @@ -226,7 +227,7 @@ def from_module(
if module.bias is not None:
qmodule.bias = module.bias

return qmodule.to(module.weight.device)
return qmodule.to(device)

@classmethod
def qcreate(
Expand Down
24 changes: 14 additions & 10 deletions test/nn/test_qlayernorm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,9 +20,9 @@
from optimum.quanto.nn import QLayerNorm


def _test_quantize_layernorm(batch_size, tokens, embeddings, dtype, activations, device):
def _test_quantize_layernorm(batch_size, tokens, embeddings, affine, dtype, activations, device):
# Instantiate a normalization layer
norm = torch.nn.LayerNorm(embeddings).to(dtype).to(device)
norm = torch.nn.LayerNorm(embeddings, elementwise_affine=affine).to(dtype).to(device)
qnorm = QLayerNorm.from_module(norm, activations=activations)
qinputs = random_qactivation((batch_size,) + (tokens, embeddings), qtype=activations, dtype=dtype).to(device)
# Calibrate to avoid clipping and to set the correct dtype
Expand All @@ -43,38 +43,42 @@ def _test_quantize_layernorm(batch_size, tokens, embeddings, dtype, activations,

@pytest.mark.parametrize("batch_size", [1, 10])
@pytest.mark.parametrize("tokens, embeddings", [(32, 32), (10, 32)])
def test_quantize_layernorm_float16_activations_int8(batch_size, tokens, embeddings, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, torch.float16, qint8, device)
@pytest.mark.parametrize("affine", [True, False], ids=["affine", "non-affine"])
def test_quantize_layernorm_float16_activations_int8(batch_size, tokens, embeddings, affine, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, affine, torch.float16, qint8, device)


@pytest.mark.parametrize("batch_size", [1, 10])
@pytest.mark.parametrize("tokens, embeddings", [(32, 32), (10, 32)])
def test_quantize_layernorm_float32_activations_int8(batch_size, tokens, embeddings, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, torch.float32, qint8, device)
@pytest.mark.parametrize("affine", [True, False], ids=["affine", "non-affine"])
def test_quantize_layernorm_float32_activations_int8(batch_size, tokens, embeddings, affine, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, affine, torch.float32, qint8, device)


@pytest.mark.parametrize("batch_size", [1, 10])
@pytest.mark.parametrize("tokens, embeddings", [(32, 32), (10, 32)])
@pytest.mark.parametrize("affine", [True, False], ids=["affine", "non-affine"])
@pytest.mark.parametrize(
"activations",
[qfloat8_e5m2, qfloat8_e4m3fn, qfloat8_e4m3fnuz],
ids=["a-float8-e5m2", "a-float8-e4m3", "a-float8-e4m3-uz"],
)
@pytest.mark.skip_device("mps")
def test_quantize_layernorm_float16_activations_float8(batch_size, tokens, embeddings, activations, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, torch.float16, activations, device)
def test_quantize_layernorm_float16_activations_float8(batch_size, tokens, embeddings, affine, activations, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, affine, torch.float16, activations, device)


@pytest.mark.parametrize("batch_size", [1, 10])
@pytest.mark.parametrize("tokens, embeddings", [(32, 32), (10, 32)])
@pytest.mark.parametrize("affine", [True, False], ids=["affine", "non-affine"])
@pytest.mark.parametrize(
"activations",
[qfloat8_e5m2, qfloat8_e4m3fn, qfloat8_e4m3fnuz],
ids=["a-float8-e5m2", "a-float8-e4m3", "a-float8-e4m3-uz"],
)
@pytest.mark.skip_device("mps")
def test_quantize_layernorm_float32_activations_float8(batch_size, tokens, embeddings, activations, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, torch.float32, activations, device)
def test_quantize_layernorm_float32_activations_float8(batch_size, tokens, embeddings, affine, activations, device):
_test_quantize_layernorm(batch_size, tokens, embeddings, affine, torch.float32, activations, device)


def test_quantize_layernom_no_activation():
Expand Down
Loading