mirror of
https://github.com/huggingface/diffusers.git
synced 2026-03-01 14:20:41 +08:00
Compare commits
3 Commits
update-ker
...
update-mod
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
e34f085d2e | ||
|
|
a123e95ee2 | ||
|
|
944a478989 |
@@ -856,7 +856,7 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
|
||||
)
|
||||
state_dict = {k: v for k, v in state_dict.items() if not k.startswith("text_encoders.t5xxl.transformer.")}
|
||||
|
||||
has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_", "lora_te1_")) for k in state_dict)
|
||||
has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_")) for k in state_dict)
|
||||
if has_diffb:
|
||||
zero_status_diff_b = state_dict_all_zero(state_dict, ".diff_b")
|
||||
if zero_status_diff_b:
|
||||
@@ -895,7 +895,7 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
|
||||
state_dict = {
|
||||
_custom_replace(k, limit_substrings): v
|
||||
for k, v in state_dict.items()
|
||||
if k.startswith(("lora_unet_", "lora_te_", "lora_te1_"))
|
||||
if k.startswith(("lora_unet_", "lora_te_"))
|
||||
}
|
||||
|
||||
if any("text_projection" in k for k in state_dict):
|
||||
|
||||
@@ -38,7 +38,6 @@ from ..utils import (
|
||||
is_flash_attn_available,
|
||||
is_flash_attn_version,
|
||||
is_kernels_available,
|
||||
is_kernels_version,
|
||||
is_sageattention_available,
|
||||
is_sageattention_version,
|
||||
is_torch_npu_available,
|
||||
@@ -319,7 +318,6 @@ class _HubKernelConfig:
|
||||
repo_id: str
|
||||
function_attr: str
|
||||
revision: str | None = None
|
||||
version: int | None = None
|
||||
kernel_fn: Callable | None = None
|
||||
wrapped_forward_attr: str | None = None
|
||||
wrapped_backward_attr: str | None = None
|
||||
@@ -329,34 +327,31 @@ class _HubKernelConfig:
|
||||
|
||||
# Registry for hub-based attention kernels
|
||||
_HUB_KERNELS_REGISTRY: dict["AttentionBackendName", _HubKernelConfig] = {
|
||||
# TODO: temporary revision for now. Remove when merged upstream into `main`.
|
||||
AttentionBackendName._FLASH_3_HUB: _HubKernelConfig(
|
||||
repo_id="kernels-community/flash-attn3",
|
||||
function_attr="flash_attn_func",
|
||||
revision="fake-ops-return-probs",
|
||||
wrapped_forward_attr="flash_attn_interface._flash_attn_forward",
|
||||
wrapped_backward_attr="flash_attn_interface._flash_attn_backward",
|
||||
version=1,
|
||||
),
|
||||
AttentionBackendName._FLASH_3_VARLEN_HUB: _HubKernelConfig(
|
||||
repo_id="kernels-community/flash-attn3",
|
||||
function_attr="flash_attn_varlen_func",
|
||||
version=1,
|
||||
# revision="fake-ops-return-probs",
|
||||
),
|
||||
AttentionBackendName.FLASH_HUB: _HubKernelConfig(
|
||||
repo_id="kernels-community/flash-attn2",
|
||||
function_attr="flash_attn_func",
|
||||
revision=None,
|
||||
wrapped_forward_attr="flash_attn_interface._wrapped_flash_attn_forward",
|
||||
wrapped_backward_attr="flash_attn_interface._wrapped_flash_attn_backward",
|
||||
version=1,
|
||||
),
|
||||
AttentionBackendName.FLASH_VARLEN_HUB: _HubKernelConfig(
|
||||
repo_id="kernels-community/flash-attn2",
|
||||
function_attr="flash_attn_varlen_func",
|
||||
version=1,
|
||||
repo_id="kernels-community/flash-attn2", function_attr="flash_attn_varlen_func", revision=None
|
||||
),
|
||||
AttentionBackendName.SAGE_HUB: _HubKernelConfig(
|
||||
repo_id="kernels-community/sage-attention",
|
||||
function_attr="sageattn",
|
||||
version=1,
|
||||
repo_id="kernels-community/sage_attention", function_attr="sageattn", revision=None
|
||||
),
|
||||
}
|
||||
|
||||
@@ -526,10 +521,6 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
|
||||
raise RuntimeError(
|
||||
f"Backend '{backend.value}' is not usable because the `kernels` package isn't available. Please install it with `pip install kernels`."
|
||||
)
|
||||
if not is_kernels_version(">=", "0.12"):
|
||||
raise RuntimeError(
|
||||
f"Backend '{backend.value}' needs to be used with a `kernels` version of at least 0.12. Please update with `pip install -U kernels`."
|
||||
)
|
||||
|
||||
elif backend == AttentionBackendName.AITER:
|
||||
if not _CAN_USE_AITER_ATTN:
|
||||
|
||||
@@ -1836,6 +1836,7 @@ class ModularPipeline(ConfigMixin, PushToHubMixin):
|
||||
create_pr = kwargs.pop("create_pr", False)
|
||||
token = kwargs.pop("token", None)
|
||||
repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
|
||||
update_model_card = kwargs.pop("update_model_card", False)
|
||||
repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id
|
||||
|
||||
# Generate modular pipeline card content
|
||||
@@ -1848,6 +1849,7 @@ class ModularPipeline(ConfigMixin, PushToHubMixin):
|
||||
is_pipeline=True,
|
||||
model_description=MODULAR_MODEL_CARD_TEMPLATE.format(**card_content),
|
||||
is_modular=True,
|
||||
update_model_card=update_model_card,
|
||||
)
|
||||
model_card = populate_model_card(model_card, tags=card_content["tags"])
|
||||
|
||||
|
||||
@@ -50,11 +50,7 @@ This modular pipeline is composed of the following blocks:
|
||||
|
||||
{components_description} {configs_section}
|
||||
|
||||
## Input/Output Specification
|
||||
|
||||
### Inputs {inputs_description}
|
||||
|
||||
### Outputs {outputs_description}
|
||||
{io_specification_section}
|
||||
"""
|
||||
|
||||
|
||||
@@ -799,6 +795,46 @@ def format_output_params(output_params, indent_level=4, max_line_length=115):
|
||||
return format_params(output_params, "Outputs", indent_level, max_line_length)
|
||||
|
||||
|
||||
def format_params_markdown(params, header="Inputs"):
|
||||
"""Format a list of InputParam or OutputParam objects as a markdown bullet-point list.
|
||||
|
||||
Suitable for model cards rendered on Hugging Face Hub.
|
||||
|
||||
Args:
|
||||
params: list of InputParam or OutputParam objects to format
|
||||
header: Header text (e.g. "Inputs" or "Outputs")
|
||||
|
||||
Returns:
|
||||
A formatted markdown string, or empty string if params is empty.
|
||||
"""
|
||||
if not params:
|
||||
return ""
|
||||
|
||||
def get_type_str(type_hint):
|
||||
if isinstance(type_hint, UnionType) or get_origin(type_hint) is Union:
|
||||
type_strs = [t.__name__ if hasattr(t, "__name__") else str(t) for t in get_args(type_hint)]
|
||||
return " | ".join(type_strs)
|
||||
return type_hint.__name__ if hasattr(type_hint, "__name__") else str(type_hint)
|
||||
|
||||
lines = [f"**{header}:**\n"] if header else []
|
||||
for param in params:
|
||||
type_str = get_type_str(param.type_hint) if param.type_hint != Any else ""
|
||||
name = f"**{param.kwargs_type}" if param.name is None and param.kwargs_type is not None else param.name
|
||||
param_str = f"- `{name}` (`{type_str}`"
|
||||
|
||||
if hasattr(param, "required") and not param.required:
|
||||
param_str += ", *optional*"
|
||||
if param.default is not None:
|
||||
param_str += f", defaults to `{param.default}`"
|
||||
param_str += ")"
|
||||
|
||||
desc = param.description if param.description else "No description provided"
|
||||
param_str += f": {desc}"
|
||||
lines.append(param_str)
|
||||
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
def format_components(components, indent_level=4, max_line_length=115, add_empty_lines=True):
|
||||
"""Format a list of ComponentSpec objects into a readable string representation.
|
||||
|
||||
@@ -1055,8 +1091,7 @@ def generate_modular_model_card_content(blocks) -> dict[str, Any]:
|
||||
- blocks_description: Detailed architecture of blocks
|
||||
- components_description: List of required components
|
||||
- configs_section: Configuration parameters section
|
||||
- inputs_description: Input parameters specification
|
||||
- outputs_description: Output parameters specification
|
||||
- io_specification_section: Input/Output specification (per-workflow or unified)
|
||||
- trigger_inputs_section: Conditional execution information
|
||||
- tags: List of relevant tags for the model card
|
||||
"""
|
||||
@@ -1075,15 +1110,6 @@ def generate_modular_model_card_content(blocks) -> dict[str, Any]:
|
||||
if block_desc:
|
||||
blocks_desc_parts.append(f" - {block_desc}")
|
||||
|
||||
# add sub-blocks if any
|
||||
if hasattr(block, "sub_blocks") and block.sub_blocks:
|
||||
for sub_name, sub_block in block.sub_blocks.items():
|
||||
sub_class = sub_block.__class__.__name__
|
||||
sub_desc = sub_block.description.split("\n")[0] if getattr(sub_block, "description", "") else ""
|
||||
blocks_desc_parts.append(f" - *{sub_name}*: `{sub_class}`")
|
||||
if sub_desc:
|
||||
blocks_desc_parts.append(f" - {sub_desc}")
|
||||
|
||||
blocks_description = "\n".join(blocks_desc_parts) if blocks_desc_parts else "No blocks defined."
|
||||
|
||||
components = getattr(blocks, "expected_components", [])
|
||||
@@ -1109,63 +1135,76 @@ def generate_modular_model_card_content(blocks) -> dict[str, Any]:
|
||||
if configs_description:
|
||||
configs_section = f"\n\n## Configuration Parameters\n\n{configs_description}"
|
||||
|
||||
inputs = blocks.inputs
|
||||
outputs = blocks.outputs
|
||||
# Branch on whether workflows are defined
|
||||
has_workflows = getattr(blocks, "_workflow_map", None) is not None
|
||||
|
||||
# format inputs as markdown list
|
||||
inputs_parts = []
|
||||
required_inputs = [inp for inp in inputs if inp.required]
|
||||
optional_inputs = [inp for inp in inputs if not inp.required]
|
||||
if has_workflows:
|
||||
workflow_map = blocks._workflow_map
|
||||
parts = []
|
||||
|
||||
if required_inputs:
|
||||
inputs_parts.append("**Required:**\n")
|
||||
for inp in required_inputs:
|
||||
if hasattr(inp.type_hint, "__name__"):
|
||||
type_str = inp.type_hint.__name__
|
||||
elif inp.type_hint is not None:
|
||||
type_str = str(inp.type_hint).replace("typing.", "")
|
||||
else:
|
||||
type_str = "Any"
|
||||
desc = inp.description or "No description provided"
|
||||
inputs_parts.append(f"- `{inp.name}` (`{type_str}`): {desc}")
|
||||
# If blocks overrides outputs (e.g. to return just "images" instead of all intermediates),
|
||||
# use that as the shared output for all workflows
|
||||
blocks_outputs = blocks.outputs
|
||||
blocks_intermediate = getattr(blocks, "intermediate_outputs", None)
|
||||
shared_outputs = (
|
||||
blocks_outputs if blocks_intermediate is not None and blocks_outputs != blocks_intermediate else None
|
||||
)
|
||||
|
||||
if optional_inputs:
|
||||
if required_inputs:
|
||||
inputs_parts.append("")
|
||||
inputs_parts.append("**Optional:**\n")
|
||||
for inp in optional_inputs:
|
||||
if hasattr(inp.type_hint, "__name__"):
|
||||
type_str = inp.type_hint.__name__
|
||||
elif inp.type_hint is not None:
|
||||
type_str = str(inp.type_hint).replace("typing.", "")
|
||||
else:
|
||||
type_str = "Any"
|
||||
desc = inp.description or "No description provided"
|
||||
default_str = f", default: `{inp.default}`" if inp.default is not None else ""
|
||||
inputs_parts.append(f"- `{inp.name}` (`{type_str}`){default_str}: {desc}")
|
||||
parts.append("## Workflow Input Specification\n")
|
||||
|
||||
inputs_description = "\n".join(inputs_parts) if inputs_parts else "No specific inputs defined."
|
||||
# Per-workflow details: show trigger inputs with full param descriptions
|
||||
for wf_name, trigger_inputs in workflow_map.items():
|
||||
trigger_input_names = set(trigger_inputs.keys())
|
||||
try:
|
||||
workflow_blocks = blocks.get_workflow(wf_name)
|
||||
except Exception:
|
||||
parts.append(f"<details>\n<summary><strong>{wf_name}</strong></summary>\n")
|
||||
parts.append("*Could not resolve workflow blocks.*\n")
|
||||
parts.append("</details>\n")
|
||||
continue
|
||||
|
||||
# format outputs as markdown list
|
||||
outputs_parts = []
|
||||
for out in outputs:
|
||||
if hasattr(out.type_hint, "__name__"):
|
||||
type_str = out.type_hint.__name__
|
||||
elif out.type_hint is not None:
|
||||
type_str = str(out.type_hint).replace("typing.", "")
|
||||
else:
|
||||
type_str = "Any"
|
||||
desc = out.description or "No description provided"
|
||||
outputs_parts.append(f"- `{out.name}` (`{type_str}`): {desc}")
|
||||
wf_inputs = workflow_blocks.inputs
|
||||
# Show only trigger inputs with full parameter descriptions
|
||||
trigger_params = [p for p in wf_inputs if p.name in trigger_input_names]
|
||||
|
||||
outputs_description = "\n".join(outputs_parts) if outputs_parts else "Standard pipeline outputs."
|
||||
parts.append(f"<details>\n<summary><strong>{wf_name}</strong></summary>\n")
|
||||
|
||||
trigger_inputs_section = ""
|
||||
if hasattr(blocks, "trigger_inputs") and blocks.trigger_inputs:
|
||||
trigger_inputs_list = sorted([t for t in blocks.trigger_inputs if t is not None])
|
||||
if trigger_inputs_list:
|
||||
trigger_inputs_str = ", ".join(f"`{t}`" for t in trigger_inputs_list)
|
||||
trigger_inputs_section = f"""
|
||||
inputs_str = format_params_markdown(trigger_params, header=None)
|
||||
parts.append(inputs_str if inputs_str else "No additional inputs required.")
|
||||
parts.append("")
|
||||
|
||||
parts.append("</details>\n")
|
||||
|
||||
# Common Inputs & Outputs section (like non-workflow pipelines)
|
||||
all_inputs = blocks.inputs
|
||||
all_outputs = shared_outputs if shared_outputs is not None else blocks.outputs
|
||||
|
||||
inputs_str = format_params_markdown(all_inputs, "Inputs")
|
||||
outputs_str = format_params_markdown(all_outputs, "Outputs")
|
||||
inputs_description = inputs_str if inputs_str else "No specific inputs defined."
|
||||
outputs_description = outputs_str if outputs_str else "Standard pipeline outputs."
|
||||
|
||||
parts.append(f"\n## Input/Output Specification\n\n{inputs_description}\n\n{outputs_description}")
|
||||
|
||||
io_specification_section = "\n".join(parts)
|
||||
# Suppress trigger_inputs_section when workflows are shown (it's redundant)
|
||||
trigger_inputs_section = ""
|
||||
else:
|
||||
# Unified I/O section (original behavior)
|
||||
inputs = blocks.inputs
|
||||
outputs = blocks.outputs
|
||||
inputs_str = format_params_markdown(inputs, "Inputs")
|
||||
outputs_str = format_params_markdown(outputs, "Outputs")
|
||||
inputs_description = inputs_str if inputs_str else "No specific inputs defined."
|
||||
outputs_description = outputs_str if outputs_str else "Standard pipeline outputs."
|
||||
io_specification_section = f"## Input/Output Specification\n\n{inputs_description}\n\n{outputs_description}"
|
||||
|
||||
trigger_inputs_section = ""
|
||||
if hasattr(blocks, "trigger_inputs") and blocks.trigger_inputs:
|
||||
trigger_inputs_list = sorted([t for t in blocks.trigger_inputs if t is not None])
|
||||
if trigger_inputs_list:
|
||||
trigger_inputs_str = ", ".join(f"`{t}`" for t in trigger_inputs_list)
|
||||
trigger_inputs_section = f"""
|
||||
### Conditional Execution
|
||||
|
||||
This pipeline contains blocks that are selected at runtime based on inputs:
|
||||
@@ -1178,7 +1217,18 @@ This pipeline contains blocks that are selected at runtime based on inputs:
|
||||
if hasattr(blocks, "model_name") and blocks.model_name:
|
||||
tags.append(blocks.model_name)
|
||||
|
||||
if hasattr(blocks, "trigger_inputs") and blocks.trigger_inputs:
|
||||
if has_workflows:
|
||||
# Derive tags from workflow names
|
||||
workflow_names = set(blocks._workflow_map.keys())
|
||||
if any("inpainting" in wf for wf in workflow_names):
|
||||
tags.append("inpainting")
|
||||
if any("image2image" in wf for wf in workflow_names):
|
||||
tags.append("image-to-image")
|
||||
if any("controlnet" in wf for wf in workflow_names):
|
||||
tags.append("controlnet")
|
||||
if any("text2image" in wf for wf in workflow_names):
|
||||
tags.append("text-to-image")
|
||||
elif hasattr(blocks, "trigger_inputs") and blocks.trigger_inputs:
|
||||
triggers = blocks.trigger_inputs
|
||||
if any(t in triggers for t in ["mask", "mask_image"]):
|
||||
tags.append("inpainting")
|
||||
@@ -1206,8 +1256,7 @@ This pipeline uses a {block_count}-block architecture that can be customized and
|
||||
"blocks_description": blocks_description,
|
||||
"components_description": components_description,
|
||||
"configs_section": configs_section,
|
||||
"inputs_description": inputs_description,
|
||||
"outputs_description": outputs_description,
|
||||
"io_specification_section": io_specification_section,
|
||||
"trigger_inputs_section": trigger_inputs_section,
|
||||
"tags": tags,
|
||||
}
|
||||
|
||||
@@ -86,7 +86,6 @@ from .import_utils import (
|
||||
is_inflect_available,
|
||||
is_invisible_watermark_available,
|
||||
is_kernels_available,
|
||||
is_kernels_version,
|
||||
is_kornia_available,
|
||||
is_librosa_available,
|
||||
is_matplotlib_available,
|
||||
|
||||
@@ -107,6 +107,7 @@ def load_or_create_model_card(
|
||||
widget: list[dict] | None = None,
|
||||
inference: bool | None = None,
|
||||
is_modular: bool = False,
|
||||
update_model_card: bool = False,
|
||||
) -> ModelCard:
|
||||
"""
|
||||
Loads or creates a model card.
|
||||
@@ -133,6 +134,9 @@ def load_or_create_model_card(
|
||||
`load_or_create_model_card` from a training script.
|
||||
is_modular: (`bool`, optional): Boolean flag to denote if the model card is for a modular pipeline.
|
||||
When True, uses model_description as-is without additional template formatting.
|
||||
update_model_card: (`bool`, optional): When True, regenerates the model card content even if one
|
||||
already exists on the remote repo. Existing card metadata (tags, license, etc.) is preserved. Only
|
||||
supported for modular pipelines (i.e., `is_modular=True`).
|
||||
"""
|
||||
if not is_jinja_available():
|
||||
raise ValueError(
|
||||
@@ -141,9 +145,17 @@ def load_or_create_model_card(
|
||||
" To install it, please run `pip install Jinja2`."
|
||||
)
|
||||
|
||||
if update_model_card and not is_modular:
|
||||
raise ValueError("`update_model_card=True` is only supported for modular pipelines (`is_modular=True`).")
|
||||
|
||||
try:
|
||||
# Check if the model card is present on the remote repo
|
||||
model_card = ModelCard.load(repo_id_or_path, token=token)
|
||||
# For modular pipelines, regenerate card content when requested (preserve existing metadata)
|
||||
if update_model_card and is_modular and model_description is not None:
|
||||
existing_data = model_card.data
|
||||
model_card = ModelCard(model_description)
|
||||
model_card.data = existing_data
|
||||
except (EntryNotFoundError, RepositoryNotFoundError):
|
||||
# Otherwise create a model card from template
|
||||
if from_training:
|
||||
|
||||
@@ -724,22 +724,6 @@ def is_transformers_version(operation: str, version: str):
|
||||
return compare_versions(parse(_transformers_version), operation, version)
|
||||
|
||||
|
||||
@cache
|
||||
def is_kernels_version(operation: str, version: str):
|
||||
"""
|
||||
Compares the current Kernels version to a given reference with an operation.
|
||||
|
||||
Args:
|
||||
operation (`str`):
|
||||
A string representation of an operator, such as `">"` or `"<="`
|
||||
version (`str`):
|
||||
A version string
|
||||
"""
|
||||
if not _kernels_available:
|
||||
return False
|
||||
return compare_versions(parse(_kernels_version), operation, version)
|
||||
|
||||
|
||||
@cache
|
||||
def is_hf_hub_version(operation: str, version: str):
|
||||
"""
|
||||
|
||||
@@ -454,8 +454,7 @@ class TestModularModelCardContent:
|
||||
"blocks_description",
|
||||
"components_description",
|
||||
"configs_section",
|
||||
"inputs_description",
|
||||
"outputs_description",
|
||||
"io_specification_section",
|
||||
"trigger_inputs_section",
|
||||
"tags",
|
||||
]
|
||||
@@ -552,18 +551,19 @@ class TestModularModelCardContent:
|
||||
blocks = self.create_mock_blocks(inputs=inputs)
|
||||
content = generate_modular_model_card_content(blocks)
|
||||
|
||||
assert "**Required:**" in content["inputs_description"]
|
||||
assert "**Optional:**" in content["inputs_description"]
|
||||
assert "prompt" in content["inputs_description"]
|
||||
assert "num_steps" in content["inputs_description"]
|
||||
assert "default: `50`" in content["inputs_description"]
|
||||
io_section = content["io_specification_section"]
|
||||
assert "**Inputs:**" in io_section
|
||||
assert "prompt" in io_section
|
||||
assert "num_steps" in io_section
|
||||
assert "*optional*" in io_section
|
||||
assert "defaults to `50`" in io_section
|
||||
|
||||
def test_inputs_description_empty(self):
|
||||
"""Test handling of pipelines without specific inputs."""
|
||||
blocks = self.create_mock_blocks(inputs=[])
|
||||
content = generate_modular_model_card_content(blocks)
|
||||
|
||||
assert "No specific inputs defined" in content["inputs_description"]
|
||||
assert "No specific inputs defined" in content["io_specification_section"]
|
||||
|
||||
def test_outputs_description_formatting(self):
|
||||
"""Test that outputs are correctly formatted."""
|
||||
@@ -573,15 +573,16 @@ class TestModularModelCardContent:
|
||||
blocks = self.create_mock_blocks(outputs=outputs)
|
||||
content = generate_modular_model_card_content(blocks)
|
||||
|
||||
assert "images" in content["outputs_description"]
|
||||
assert "Generated images" in content["outputs_description"]
|
||||
io_section = content["io_specification_section"]
|
||||
assert "images" in io_section
|
||||
assert "Generated images" in io_section
|
||||
|
||||
def test_outputs_description_empty(self):
|
||||
"""Test handling of pipelines without specific outputs."""
|
||||
blocks = self.create_mock_blocks(outputs=[])
|
||||
content = generate_modular_model_card_content(blocks)
|
||||
|
||||
assert "Standard pipeline outputs" in content["outputs_description"]
|
||||
assert "Standard pipeline outputs" in content["io_specification_section"]
|
||||
|
||||
def test_trigger_inputs_section_with_triggers(self):
|
||||
"""Test that trigger inputs section is generated when present."""
|
||||
|
||||
Reference in New Issue
Block a user