[CI] Refactor Wan Model Tests #13082
Conversation
yiyixuxu
left a comment
yiyixuxu
left a comment
There was a problem hiding this comment.
really nice thanks!
should we start to add a guide for contributor some where, maybe https://huggingface.co/docs/diffusers/main/en/conceptual/contribution
dg845
left a comment
dg845
left a comment
There was a problem hiding this comment.
Thanks! I see there are two Wan model related failures from the CI:
tests/models/transformers/test_models_transformer_wan_animate.py::TestWanAnimateTransformer3DAttention::test_fuse_unfuse_qkv_projectionstests/models/transformers/test_models_transformer_wan_vace.py::TestWanVACETransformer3DAttention::test_fuse_unfuse_qkv_projections
If I try to run the new Wan tests locally, for example with
pytest tests/models/transformers/test_models_transformer_wan.pyI get some more test failures:
tests/models/transformers/test_models_transformer_wan.py::TestWanTransformer3Dtest_keep_in_fp32_modulestest_from_save_pretrained_dtype_inference[fp16,bf16]
tests/models/transformers/test_models_transformer_wan.py::TestWanTransformer3DGGUFtest_gguf_quantization_inferencetest_gguf_keep_modules_in_fp32test_gguf_quantization_dtype_assignmenttest_gguf_quantization_lora_inferencetest_gguf_dequantizetest_gguf_quantized_layers
tests/models/transformers/test_models_transformer_wan.py::TestWanTransformer3DGGUFCompiletest_gguf_torch_compiletest_gguf_torch_compile_with_group_offload
Are these test failures expected?
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
Thanks for flagging @dg845. I've fixed the test issues. There are some GGUF related fixes that should probably go in a different PR (will handle that later) |
| @@ -214,15 +214,18 @@ def __init__( | |||
| self.add_v_proj = torch.nn.Linear(added_kv_proj_dim, self.inner_dim, bias=True) | |||
| self.norm_added_k = torch.nn.RMSNorm(dim_head * heads, eps=eps) | |||
|
|
|||
| self.is_cross_attention = cross_attention_dim_head is not None | |||
| if is_cross_attention is not None: | |||
| self.is_cross_attention = is_cross_attention | |||
| else: | |||
| self.is_cross_attention = cross_attention_dim_head is not None | |||
There was a problem hiding this comment.
Seems like an unrelated change?
sayakpaul
left a comment
sayakpaul
left a comment
There was a problem hiding this comment.
Thanks, I left some comments!
| self.inner_dim = dim_head * heads | ||
| self.heads = heads | ||
| self.cross_attention_head_dim = cross_attention_dim_head | ||
| self.cross_attention_dim_head = cross_attention_dim_head |
There was a problem hiding this comment.
Same as above. Would be nice if you could explain these changes? Were these flagged by the newly written test suite?
| # Test with float16 | ||
| model.to(torch_device) | ||
| model.to(torch.float16) | ||
| # Save the model and reload with float16 dtype | ||
| # _keep_in_fp32_modules is only enforced during from_pretrained loading | ||
| model.save_pretrained(tmp_path) | ||
| model = self.model_class.from_pretrained(tmp_path, torch_dtype=torch.float16).to(torch_device) |
There was a problem hiding this comment.
Have we run this test for all the other models to ensure this isn't breaking anything?
| # Get model dtype from first parameter | ||
| model_dtype = next(model_quantized.parameters()).dtype | ||
|
|
||
| inputs = self.get_dummy_inputs() | ||
| # Cast inputs to model dtype | ||
| inputs = { | ||
| k: v.to(model_dtype) if isinstance(v, torch.Tensor) and v.is_floating_point() else v | ||
| for k, v in inputs.items() | ||
| } |
| def test_gguf_quantized_layers(self): | ||
| self._test_quantized_layers({"compute_dtype": torch.bfloat16}) |
There was a problem hiding this comment.
Why remove? Is it a duplicate?
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| import unittest |
There was a problem hiding this comment.
I am guessing the changes under tests/models/transformers/ were all auto-generated?
| class TestWanVACETransformer3DCompile(WanVACETransformer3DTesterConfig, TorchCompileTesterMixin): | ||
| """Torch compile tests for Wan VACE Transformer 3D.""" | ||
|
|
||
| def test_torch_compile_repeated_blocks(self): |
There was a problem hiding this comment.
I think we can further simplify this test by letting users pass a recompile_limit. I will open a PR.
5 participants
What does this PR do?
Update Wan tests with new format
Fixes # (issue)
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.