finalize the ci

tests/flux/test_flux_cache.py

@@ -8,7 +8,7 @@ from .utils import run_test
 @pytest.mark.parametrize(
     "cache_threshold,height,width,num_inference_steps,lora_name,lora_strength,expected_lpips",
     [
-        (0.12, 1024, 1024, 30, None, 1, 0.26),
+        (0.12, 1024, 1024, 30, None, 1, 0.212),
     ],
 )
 def test_flux_dev_loras(

tests/flux/test_flux_dev.py

@@ -8,8 +8,8 @@ from .utils import run_test
 @pytest.mark.parametrize(
     "height,width,num_inference_steps,attention_impl,cpu_offload,expected_lpips",
     [
-        (1024, 1024, 50, "flashattn2", False, 0.226),
-        (2048, 512, 25, "nunchaku-fp16", False, 0.243),
+        (1024, 1024, 50, "flashattn2", False, 0.139),
+        (2048, 512, 25, "nunchaku-fp16", False, 0.148),
     ],
 )
 def test_flux_dev(

tests/flux/test_flux_dev_loras.py

@@ -8,10 +8,10 @@ from .utils import run_test
 @pytest.mark.parametrize(
     "num_inference_steps,lora_name,lora_strength,cpu_offload,expected_lpips",
     [
-        (25, "realism", 0.9, True, 0.178),
-        (25, "ghibsky", 1, False, 0.164),
+        (25, "realism", 0.9, True, 0.136),
+        (25, "ghibsky", 1, False, 0.186),
         # (28, "anime", 1, False, 0.284),
-        (24, "sketch", 1, True, 0.223),
+        (24, "sketch", 1, True, 0.260),
         # (28, "yarn", 1, False, 0.211),
         # (25, "haunted_linework", 1, True, 0.317),
     ],
@@ -51,7 +51,7 @@ def test_flux_dev_hypersd8_1536x2048():
         lora_names="hypersd8",
         lora_strengths=0.125,
         cache_threshold=0,
-        expected_lpips=0.291,
+        expected_lpips=0.164,
     )
 
 
@@ -71,7 +71,7 @@ def test_flux_dev_turbo8_1024x1920():
         lora_names="turbo8",
         lora_strengths=1,
         cache_threshold=0,
-        expected_lpips=0.189,
+        expected_lpips=0.120,
     )
 
 
@@ -91,7 +91,7 @@ def test_flux_dev_turbo8_yarn_2048x1024():
         lora_names=["turbo8", "yarn"],
         lora_strengths=[1, 1],
         cache_threshold=0,
-        expected_lpips=0.252,
+        expected_lpips=0.255,
     )
 
 
@@ -111,5 +111,5 @@ def test_flux_dev_turbo8_yarn_1024x1024():
         lora_names=["realism", "ghibsky", "anime", "sketch", "yarn", "haunted_linework", "turbo8"],
         lora_strengths=[0, 0, 0, 0, 0, 1, 1],
         cache_threshold=0,
-        expected_lpips=0.44,
+        expected_lpips=0.310,
     )

tests/flux/test_flux_schnell.py

@@ -8,10 +8,10 @@ from .utils import run_test
 @pytest.mark.parametrize(
     "height,width,attention_impl,cpu_offload,expected_lpips",
     [
-        (1024, 1024, "flashattn2", False, 0.250),
-        (1024, 1024, "nunchaku-fp16", False, 0.255),
-        (1920, 1080, "nunchaku-fp16", False, 0.253),
-        (2048, 2048, "nunchaku-fp16", True, 0.274),
+        (1024, 1024, "flashattn2", False, 0.126),
+        (1024, 1024, "nunchaku-fp16", False, 0.126),
+        (1920, 1080, "nunchaku-fp16", False, 0.141),
+        (2048, 2048, "nunchaku-fp16", True, 0.166),
     ],
 )
 def test_int4_schnell(height: int, width: int, attention_impl: str, cpu_offload: bool, expected_lpips: float):

tests/flux/test_flux_tools.py

@@ -15,12 +15,12 @@ def test_flux_canny_dev():
         dtype=torch.bfloat16,
         height=1024,
         width=1024,
-        num_inference_steps=50,
+        num_inference_steps=30,
         guidance_scale=30,
         attention_impl="nunchaku-fp16",
         cpu_offload=False,
         cache_threshold=0,
-        expected_lpips=0.103 if get_precision() == "int4" else 0.164,
+        expected_lpips=0.076 if get_precision() == "int4" else 0.164,
     )
 
 
@@ -39,7 +39,7 @@ def test_flux_depth_dev():
         attention_impl="nunchaku-fp16",
         cpu_offload=False,
         cache_threshold=0,
-        expected_lpips=0.170 if get_precision() == "int4" else 0.120,
+        expected_lpips=0.137 if get_precision() == "int4" else 0.120,
     )
 
 
@@ -53,12 +53,12 @@ def test_flux_fill_dev():
         dtype=torch.bfloat16,
         height=1024,
         width=1024,
-        num_inference_steps=50,
+        num_inference_steps=30,
         guidance_scale=30,
         attention_impl="nunchaku-fp16",
         cpu_offload=False,
         cache_threshold=0,
-        expected_lpips=0.045,
+        expected_lpips=0.046,
     )
 
 
@@ -72,14 +72,14 @@ def test_flux_dev_canny_lora():
         dtype=torch.bfloat16,
         height=1024,
         width=1024,
-        num_inference_steps=50,
+        num_inference_steps=30,
         guidance_scale=30,
         attention_impl="nunchaku-fp16",
         cpu_offload=False,
         lora_names="canny",
         lora_strengths=0.85,
         cache_threshold=0,
-        expected_lpips=0.103,
+        expected_lpips=0.081,
     )
 
 
@@ -100,7 +100,7 @@ def test_flux_dev_depth_lora():
         cache_threshold=0,
         lora_names="depth",
         lora_strengths=0.85,
-        expected_lpips=0.163,
+        expected_lpips=0.181,
     )
 
 
@@ -121,7 +121,7 @@ def test_flux_fill_dev_turbo():
         cache_threshold=0,
         lora_names="turbo8",
         lora_strengths=1,
-        expected_lpips=0.048,
+        expected_lpips=0.036,
     )
 
 
@@ -135,10 +135,10 @@ def test_flux_dev_redux():
         dtype=torch.bfloat16,
         height=1024,
         width=1024,
-        num_inference_steps=50,
+        num_inference_steps=20,
         guidance_scale=2.5,
         attention_impl="nunchaku-fp16",
         cpu_offload=False,
         cache_threshold=0,
-        expected_lpips=(0.198 if get_precision() == "int4" else 0.198),
+        expected_lpips=(0.143 if get_precision() == "int4" else 0.198),
     )

tests/flux/test_shuttle_jaguar.py

@@ -6,7 +6,7 @@ from nunchaku.utils import get_precision, is_turing
 
 @pytest.mark.skipif(is_turing(), reason="Skip tests due to using Turing GPUs")
 @pytest.mark.parametrize(
-    "height,width,attention_impl,cpu_offload,expected_lpips", [(1024, 1024, "nunchaku-fp16", False, 0.25)]
+    "height,width,attention_impl,cpu_offload,expected_lpips", [(1024, 1024, "nunchaku-fp16", False, 0.186)]
 )
 def test_shuttle_jaguar(height: int, width: int, attention_impl: str, cpu_offload: bool, expected_lpips: float):
     run_test(

tests/flux/utils.py

@@ -134,7 +134,7 @@ def run_test(
     cache_threshold: float = 0,
     lora_names: str | list[str] | None = None,
     lora_strengths: float | list[float] = 1.0,
-    max_dataset_size: int = 8,
+    max_dataset_size: int = 4,
     i2f_mode: str | None = None,
     expected_lpips: float = 0.5,
 ):