[CI] add a big GPU marker to run memory-intensive tests separately on CI

.github/workflows/nightly_tests.yml

@@ -2,6 +2,7 @@ name: Nightly and release tests on main/release branch
 
 on:
   workflow_dispatch:
+  pull_request:
   schedule:
     - cron: "0 0 * * *" # every day at midnight
 
@@ -18,6 +19,7 @@ env:
 
 jobs:
   setup_torch_cuda_pipeline_matrix:
+    if: github.event_name == 'schedule'
     name: Setup Torch Pipelines CUDA Slow Tests Matrix
     runs-on:
       group: aws-general-8-plus
@@ -49,6 +51,7 @@ jobs:
           path: reports
 
   run_nightly_tests_for_torch_pipelines:
+    if: github.event_name == 'schedule'
     name: Nightly Torch Pipelines CUDA Tests
     needs: setup_torch_cuda_pipeline_matrix
     strategy:
@@ -106,6 +109,7 @@ jobs:
           python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_nightly_tests_for_other_torch_modules:
+    if: github.event_name == 'schedule'
     name: Nightly Torch CUDA Tests
     runs-on:
       group: aws-g4dn-2xlarge
@@ -180,6 +184,61 @@ jobs:
         pip install slack_sdk tabulate
         python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
+  run_big_gpu_torch_tests:
+    name: Torch tests on big GPU (24GB)
+    strategy:
+      fail-fast: false
+      max-parallel: 8
+    runs-on:
+      group: aws-g6e-xlarge-plus
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host --gpus 0
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+      - name: NVIDIA-SMI
+        run: nvidia-smi
+      - name: Install dependencies
+        run: |
+          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+          python -m uv pip install -e [quality,test]
+          python -m uv pip install peft@git+https://github.com/huggingface/peft.git
+          pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          python -m uv pip install pytest-reportlog
+      - name: Environment
+        run: |
+          python utils/print_env.py
+      - name: Selected Torch CUDA Test on big GPU
+        env:
+          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+          # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+          CUBLAS_WORKSPACE_CONFIG: :16:8
+        run: |
+          python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+            -m "big_gpu_with_torch_cuda" \
+            --make-reports=tests_big_gpu_torch_cuda \
+            --report-log=tests_big_gpu_torch_cuda.log \
+            tests/
+      - name: Failure short reports
+        if: ${{ failure() }}
+        run: |
+          cat reports/tests_big_gpu_torch_cuda_stats.txt
+          cat reports/tests_big_gpu_torch_cuda_failures_short.txt
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: torch_cuda_big_gpu_test_reports
+          path: reports
+      - name: Generate Report and Notify Channel
+        if: always()
+        run: |
+          pip install slack_sdk tabulate
+          python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+
   run_flax_tpu_tests:
     name: Nightly Flax TPU Tests
     runs-on: docker-tpu
@@ -237,6 +296,7 @@ jobs:
         python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
   run_nightly_onnx_tests:
+    if: github.event_name == 'schedule'
     name: Nightly ONNXRuntime CUDA tests on Ubuntu
     runs-on:
       group: aws-g4dn-2xlarge

src/diffusers/utils/testing_utils.py

@@ -54,6 +54,7 @@
 ) > version.parse("4.33")
 
 USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version
+BIG_GPU_MEMORY = 40
 
 if is_torch_available():
     import torch
@@ -307,6 +308,26 @@ def require_torch_accelerator_with_fp64(test_case):
     )
 
 
+def require_big_gpu_with_torch_cuda(test_case):
+    """
+    Decorator marking a test that requires a bigger GPU (24GB) for execution. Some example pipelines: Flux, SD3, Cog,
+    etc.
+    """
+    if not is_torch_available():
+        return unittest.skip("test requires PyTorch")(test_case)
+
+    import torch
+
+    if not torch.cuda.is_available():
+        return unittest.skip("test requires PyTorch CUDA")(test_case)
+
+    device_properties = torch.cuda.get_device_properties(0)
+    total_memory = device_properties.total_memory / (1024**3)
+    return unittest.skipUnless(
+        total_memory >= BIG_GPU_MEMORY, f"test requires a GPU with at least {BIG_GPU_MEMORY} GB memory"
+    )(test_case)
+
+
 def require_torch_accelerator_with_training(test_case):
     """Decorator marking a test that requires an accelerator with support for training."""
     return unittest.skipUnless(

tests/pipelines/controlnet_flux/test_controlnet_flux.py

@@ -17,6 +17,7 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
 
@@ -30,7 +31,7 @@
 from diffusers.utils import load_image
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
-    require_torch_gpu,
+    require_big_gpu_with_torch_cuda,
     slow,
     torch_device,
 )
@@ -180,7 +181,8 @@ def test_xformers_attention_forwardGenerator_pass(self):
 
 
 @slow
-@require_torch_gpu
+@require_big_gpu_with_torch_cuda
+@pytest.mark.big_gpu_with_torch_cuda
 class FluxControlNetPipelineSlowTests(unittest.TestCase):
     pipeline_class = FluxControlNetPipeline
 

tests/pipelines/controlnet_flux/test_controlnet_flux_img2img.py

@@ -2,6 +2,7 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
@@ -14,7 +15,7 @@
 )
 from diffusers.utils.testing_utils import (
     numpy_cosine_similarity_distance,
-    require_torch_gpu,
+    require_big_gpu_with_torch_cuda,
     slow,
     torch_device,
 )
@@ -225,7 +226,8 @@ def test_fused_qkv_projections(self):
 
 
 @slow
-@require_torch_gpu
+@require_big_gpu_with_torch_cuda
+@pytest.mark.big_gpu_with_torch_cuda
 class FluxControlNetImg2ImgPipelineSlowTests(unittest.TestCase):
     pipeline_class = FluxControlNetImg2ImgPipeline
     repo_id = "black-forest-labs/FLUX.1-schnell"
@@ -261,7 +263,6 @@ def get_inputs(self, device, seed=0):
             "generator": generator,
         }
 
-    @unittest.skip("We cannot run inference on this model with the current CI hardware")
     def test_flux_controlnet_img2img_inference(self):
         pipe = self.pipeline_class.from_pretrained(self.repo_id, torch_dtype=torch.bfloat16)
         pipe.enable_model_cpu_offload()

tests/pipelines/controlnet_sd3/test_controlnet_sd3.py

@@ -17,6 +17,7 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
 
@@ -30,7 +31,7 @@
 from diffusers.utils import load_image
 from diffusers.utils.testing_utils import (
     enable_full_determinism,
-    require_torch_gpu,
+    require_big_gpu_with_torch_cuda,
     slow,
     torch_device,
 )
@@ -195,7 +196,8 @@ def test_xformers_attention_forwardGenerator_pass(self):
 
 
 @slow
-@require_torch_gpu
+@require_big_gpu_with_torch_cuda
+@pytest.mark.big_gpu_with_torch_cuda
 class StableDiffusion3ControlNetPipelineSlowTests(unittest.TestCase):
     pipeline_class = StableDiffusion3ControlNetPipeline
 

tests/pipelines/flux/test_pipeline_flux.py

@@ -2,13 +2,14 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
 from diffusers.utils.testing_utils import (
     numpy_cosine_similarity_distance,
-    require_torch_gpu,
+    require_big_gpu_with_torch_cuda,
     slow,
     torch_device,
 )
@@ -191,7 +192,8 @@ def test_fused_qkv_projections(self):
 
 
 @slow
-@require_torch_gpu
+@require_big_gpu_with_torch_cuda
+@pytest.mark.big_gpu_with_torch_cuda
 class FluxPipelineSlowTests(unittest.TestCase):
     pipeline_class = FluxPipeline
     repo_id = "black-forest-labs/FLUX.1-schnell"
@@ -220,8 +222,6 @@ def get_inputs(self, device, seed=0):
             "generator": generator,
         }
 
-    # TODO: Dhruv. Move large model tests to a dedicated runner)
-    @unittest.skip("We cannot run inference on this model with the current CI hardware")
     def test_flux_inference(self):
         pipe = self.pipeline_class.from_pretrained(self.repo_id, torch_dtype=torch.bfloat16)
         pipe.enable_model_cpu_offload()

tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3.py

@@ -2,13 +2,14 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
 
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, SD3Transformer2DModel, StableDiffusion3Pipeline
 from diffusers.utils.testing_utils import (
     numpy_cosine_similarity_distance,
-    require_torch_gpu,
+    require_big_gpu_with_torch_cuda,
     slow,
     torch_device,
 )
@@ -226,7 +227,8 @@ def test_fused_qkv_projections(self):
 
 
 @slow
-@require_torch_gpu
+@require_big_gpu_with_torch_cuda
+@pytest.mark.big_gpu_with_torch_cuda
 class StableDiffusion3PipelineSlowTests(unittest.TestCase):
     pipeline_class = StableDiffusion3Pipeline
     repo_id = "stabilityai/stable-diffusion-3-medium-diffusers"

tests/pipelines/stable_diffusion_3/test_pipeline_stable_diffusion_3_img2img.py

@@ -3,6 +3,7 @@
 import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel
 
@@ -16,7 +17,7 @@
 from diffusers.utils.testing_utils import (
     floats_tensor,
     numpy_cosine_similarity_distance,
-    require_torch_gpu,
+    require_big_gpu_with_torch_cuda,
     slow,
     torch_device,
 )
@@ -194,7 +195,8 @@ def test_multi_vae(self):
 
 
 @slow
-@require_torch_gpu
+@require_big_gpu_with_torch_cuda
+@pytest.mark.big_gpu_with_torch_cuda
 class StableDiffusion3Img2ImgPipelineSlowTests(unittest.TestCase):
     pipeline_class = StableDiffusion3Img2ImgPipeline
     repo_id = "stabilityai/stable-diffusion-3-medium-diffusers"

utils/print_env.py

@@ -37,6 +37,10 @@
     print("Cuda version:", torch.version.cuda)
     print("CuDNN version:", torch.backends.cudnn.version())
     print("Number of GPUs available:", torch.cuda.device_count())
+    if torch.cuda.is_available():
+        device_properties = torch.cuda.get_device_properties(0)
+        total_memory = device_properties.total_memory / (1024**3)
+        print(f"CUDA memory: {total_memory} GB")
 except ImportError:
     print("Torch version:", None)