Skip to content

Refactor GroupedLinear quantization dispatch #3326

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

Refactor GroupedLinear quantization dispatch #3326

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
41 changes: 34 additions & 7 deletions tests/pytorch/test_custom_recipe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,7 +129,7 @@ def quantizer_factory(role):
assert inp.grad is not None


def test_custom_recipe_grouped_linear_sanity():
def test_custom_recipe_grouped_linear_matches_current_scaling():
available, reason = te.is_fp8_available(return_reason=True)
if not torch.cuda.is_available() or not available:
pytest.skip(f"FP8 unsupported on this device: {reason}")
Expand All @@ -144,8 +144,25 @@ def test_custom_recipe_grouped_linear_sanity():
m_splits = [16] * num_gemms
batch = sum(m_splits)

model = GroupedLinear(num_gemms, in_features, out_features, params_dtype=torch.bfloat16).cuda()
inp = torch.randn(batch, in_features, device="cuda", dtype=torch.bfloat16, requires_grad=True)
model_ref = GroupedLinear(
num_gemms, in_features, out_features, params_dtype=torch.bfloat16
).cuda()
model_custom = GroupedLinear(
num_gemms, in_features, out_features, params_dtype=torch.bfloat16
).cuda()
model_custom.load_state_dict(model_ref.state_dict())

base_inp = torch.randn(batch, in_features, device="cuda", dtype=torch.bfloat16)
inp_ref = base_inp.clone().detach().requires_grad_(True)
inp_custom = base_inp.clone().detach().requires_grad_(True)

with autocast(enabled=True, recipe=recipe.Float8CurrentScaling()):
out_ref = model_ref(inp_ref, m_splits)

scale = torch.ones(out_features, device="cuda", dtype=torch.float32)
scale[0] = 1e8
scale[1] = 1e-8
(out_ref.float() * scale.view(1, -1)).sum().backward()

def quantizer_factory(role):
if role is None:
Expand All @@ -157,11 +174,21 @@ def quantizer_factory(role):
custom_recipe = recipe.CustomRecipe(qfactory=quantizer_factory)

with autocast(enabled=True, recipe=custom_recipe):
out = model(inp, m_splits)
loss = out.float().sum()
loss.backward()
out_custom = model_custom(inp_custom, m_splits)
(out_custom.float() * scale.view(1, -1)).sum().backward()

assert inp.grad is not None
torch.testing.assert_close(out_ref, out_custom, rtol=0, atol=0)

assert inp_ref.grad is not None and inp_custom.grad is not None
torch.testing.assert_close(inp_ref.grad, inp_custom.grad, rtol=0, atol=0)

ref_params = dict(model_ref.named_parameters())
custom_params = dict(model_custom.named_parameters())
assert ref_params.keys() == custom_params.keys()
for name, ref_param in ref_params.items():
custom_param = custom_params[name]
assert ref_param.grad is not None and custom_param.grad is not None
torch.testing.assert_close(ref_param.grad, custom_param.grad, rtol=0, atol=0)


def test_custom_recipe_matches_current_scaling():
Expand Down
Loading
Loading