AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

Slight performance improvement to Euler, EDMEuler, FlowMatchHeun, KDPM2Ancestral (#9616)

Browse Source 
* Slight performance improvement to Euler

* Slight performance improvement to EDMEuler

* Slight performance improvement to FlowMatchHeun

* Slight performance improvement to KDPM2Ancestral

* Update KDPM2AncestralDiscreteSchedulerTest

---------

Co-authored-by: YiYi Xu <yixu310@gmail.com>
This commit is contained in:
hlky
2024-10-15 06:34:25 +01:00
committed by GitHub
parent 5f0df17703
commit 9d0616189e
5 changed files with 23 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/diffusers/schedulers/scheduling_edm_euler.py
View File

@@ -333,14 +333,13 @@ class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
noise = randn_tensor(
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
)
eps = noise * s_noise
sigma_hat = sigma * (gamma + 1)
if gamma > 0:
noise = randn_tensor(
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
)
eps = noise * s_noise
sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise

9
src/diffusers/schedulers/scheduling_euler_discrete.py
View File

@@ -638,14 +638,13 @@ class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
noise = randn_tensor(
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
)
eps = noise * s_noise
sigma_hat = sigma * (gamma + 1)
if gamma > 0:
noise = randn_tensor(
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
)
eps = noise * s_noise
sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise

9
src/diffusers/schedulers/scheduling_flow_match_heun_discrete.py
View File

@@ -266,14 +266,13 @@ class FlowMatchHeunDiscreteScheduler(SchedulerMixin, ConfigMixin):
gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
noise = randn_tensor(
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
)
eps = noise * s_noise
sigma_hat = sigma * (gamma + 1)
if gamma > 0:
noise = randn_tensor(
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
)
eps = noise * s_noise
sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
if self.state_in_first_order:

6
src/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py
View File

@@ -524,9 +524,6 @@ class KDPM2AncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
gamma = 0
sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now
device = model_output.device
noise = randn_tensor(model_output.shape, dtype=model_output.dtype, device=device, generator=generator)
# 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
if self.config.prediction_type == "epsilon":
sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol
@@ -564,6 +561,9 @@ class KDPM2AncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
self.sample = None
prev_sample = sample + derivative * dt
noise = randn_tensor(
model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
)
prev_sample = prev_sample + noise * sigma_up
# upon completion increase step index by one

16
tests/schedulers/test_scheduler_kdpm2_ancestral.py
View File

@@ -59,8 +59,8 @@ class KDPM2AncestralDiscreteSchedulerTest(SchedulerCommonTest):
result_sum = torch.sum(torch.abs(sample))
result_mean = torch.mean(torch.abs(sample))
assert abs(result_sum.item() - 13849.3877) < 1e-2
assert abs(result_mean.item() - 18.0331) < 5e-3
assert abs(result_sum.item() - 13979.9433) < 1e-2
assert abs(result_mean.item() - 18.2030) < 5e-3
def test_prediction_type(self):
for prediction_type in ["epsilon", "v_prediction"]:
@@ -92,8 +92,8 @@ class KDPM2AncestralDiscreteSchedulerTest(SchedulerCommonTest):
result_sum = torch.sum(torch.abs(sample))
result_mean = torch.mean(torch.abs(sample))
assert abs(result_sum.item() - 328.9970) < 1e-2
assert abs(result_mean.item() - 0.4284) < 1e-3
assert abs(result_sum.item() - 331.8133) < 1e-2
assert abs(result_mean.item() - 0.4320) < 1e-3
def test_full_loop_device(self):
if torch_device == "mps":
@@ -119,8 +119,8 @@ class KDPM2AncestralDiscreteSchedulerTest(SchedulerCommonTest):
result_sum = torch.sum(torch.abs(sample))
result_mean = torch.mean(torch.abs(sample))
assert abs(result_sum.item() - 13849.3818) < 1e-1
assert abs(result_mean.item() - 18.0331) < 1e-3
assert abs(result_sum.item() - 13979.9433) < 1e-1
assert abs(result_mean.item() - 18.2030) < 1e-3
def test_full_loop_with_noise(self):
if torch_device == "mps":
@@ -154,5 +154,5 @@ class KDPM2AncestralDiscreteSchedulerTest(SchedulerCommonTest):
result_sum = torch.sum(torch.abs(sample))
result_mean = torch.mean(torch.abs(sample))
assert abs(result_sum.item() - 93087.0312) < 1e-2, f" expected result sum 93087.0312, but get {result_sum}"
assert abs(result_mean.item() - 121.2071) < 5e-3, f" expected result mean 121.2071, but get {result_mean}"
assert abs(result_sum.item() - 93087.3437) < 1e-2, f" expected result sum 93087.3437, but get {result_sum}"
assert abs(result_mean.item() - 121.2074) < 5e-3, f" expected result mean 121.2074, but get {result_mean}"