optimum-rbln 0.7.4a7__tar.gz → 0.7.4a8__tar.gz

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/.github/ISSUE_TEMPLATE/model_request.md

@@ -8,7 +8,7 @@ assignees: ''
 
 ### Model Information
 - Model name: 
-- Model link (Hugging Face/Paper/GitHub): 
+- Model link (HuggingFace/Paper/GitHub): 
 - Task type (e.g., Text Generation, Image Classification): 
 
 ### Model Details

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/.github/scripts/auto_code_review.py

@@ -36,7 +36,7 @@ def get_pr_diff():
 
 
 def get_prompt(diff, pr):
-    system_prompt = """You are an experienced software engineer specializing in code reviews for deep learning libraries. Your task is to review code changes and related pull request (PR) information for `optimum-rbln`, a Python library that optimizes Hugging Face models for execution on RBLN NPUs.
+    system_prompt = """You are an experienced software engineer specializing in code reviews for deep learning libraries. Your task is to review code changes and related pull request (PR) information for `optimum-rbln`, a Python library that optimizes HuggingFace models for execution on RBLN NPUs.
 
 Focus on providing actionable and constructive feedback. Don't make generalized suggestions."""
 

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.4a7
-Summary: Optimum RBLN is the interface between the Hugging Face Transformers and Diffusers libraries and RBLN accelerators. It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
+Version: 0.7.4a8
+Summary: Optimum RBLN is the interface between the HuggingFace Transformers and Diffusers libraries and RBLN accelerators. It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
 Project-URL: Homepage, https://rebellions.ai
 Project-URL: Documentation, https://docs.rbln.ai
 Project-URL: Repository, https://github.com/rebellions-sw/optimum-rbln

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "optimum-rbln"
 description = """
-Optimum RBLN is the interface between the Hugging Face Transformers and Diffusers libraries and RBLN accelerators.
+Optimum RBLN is the interface between the HuggingFace Transformers and Diffusers libraries and RBLN accelerators.
 It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
 """
 authors = [

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/src/optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.4a7'
-__version_tuple__ = version_tuple = (0, 7, 4, 'a7')
+__version__ = version = '0.7.4a8'
+__version_tuple__ = version_tuple = (0, 7, 4, 'a8')

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/src/optimum/rbln/configuration_utils.py

@@ -174,6 +174,14 @@ class RBLNAutoConfig:
         cls = getattr(importlib.import_module("optimum.rbln"), cls_name)
         return cls(**kwargs)
 
+    @staticmethod
+    def load_from_dict(config_dict: Dict[str, Any]) -> "RBLNModelConfig":
+        cls_name = config_dict.get("cls_name")
+        if cls_name is None:
+            raise ValueError("`cls_name` is required.")
+        cls = getattr(importlib.import_module("optimum.rbln"), cls_name)
+        return cls(**config_dict)
+
     @staticmethod
     def load(
         path: str,
@@ -195,8 +203,9 @@ class RBLNAutoConfig:
         cls, config_file = load_config(path)
 
         rbln_keys = [key for key in kwargs.keys() if key.startswith("rbln_")]
-
         rbln_runtime_kwargs = {key[5:]: kwargs.pop(key) for key in rbln_keys if key[5:] in RUNTIME_KEYWORDS}
+        rbln_submodule_kwargs = {key[5:]: kwargs.pop(key) for key in rbln_keys if key[5:] in cls.submodules}
+
         rbln_kwargs = {
             key[5:]: kwargs.pop(key)
             for key in rbln_keys
@@ -206,6 +215,14 @@ class RBLNAutoConfig:
         if len(rbln_kwargs) > 0:
             raise ValueError(f"Cannot set the following arguments: {list(rbln_kwargs.keys())}")
 
+        # Process submodule's rbln_config
+        for submodule in cls.submodules:
+            if submodule not in config_file:
+                raise ValueError(f"Submodule {submodule} not found in rbln_config.json.")
+            submodule_config = config_file[submodule]
+            submodule_config.update(rbln_submodule_kwargs.pop(submodule, {}))
+            config_file[submodule] = RBLNAutoConfig.load_from_dict(submodule_config)
+
         if passed_rbln_config is not None:
             config_file.update(passed_rbln_config._runtime_options)
             # TODO(jongho): Reject if the passed_rbln_config has different attributes from the config_file

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/src/optimum/rbln/diffusers/modeling_diffusers.py

@@ -47,17 +47,11 @@ class RBLNDiffusionMixin:
 
     1. Create a new pipeline class that inherits from both this mixin and the original StableDiffusionPipeline.
     2. Define the required _submodules class variable listing the components to be compiled.
-    3. If needed, implement get_default_rbln_config for custom configuration of submodules.
 
     Example:
         ```python
         class RBLNStableDiffusionPipeline(RBLNDiffusionMixin, StableDiffusionPipeline):
             _submodules = ["text_encoder", "unet", "vae"]
-
-            @classmethod
-            def get_default_rbln_config(cls, model, submodule_name, rbln_config):
-                # Configuration for other submodules...
-                pass
         ```
 
     Class Variables:
@@ -77,14 +71,6 @@ class RBLNDiffusionMixin:
     _prefix = {}
     _rbln_config_class = None
 
-    @classmethod
-    def is_img2img_pipeline(cls):
-        return "Img2Img" in cls.__name__
-
-    @classmethod
-    def is_inpaint_pipeline(cls):
-        return "Inpaint" in cls.__name__
-
     @staticmethod
     def _maybe_apply_and_fuse_lora(
         model: torch.nn.Module,
@@ -151,7 +137,47 @@ class RBLNDiffusionMixin:
         lora_weights_names: Optional[Union[str, List[str]]] = None,
         lora_scales: Optional[Union[float, List[float]]] = None,
         **kwargs,
-    ) -> RBLNModel:
+    ) -> "RBLNDiffusionMixin":
+        """
+        Load a pretrained diffusion pipeline from a model checkpoint, with optional compilation for RBLN NPUs.
+
+        This method has two distinct operating modes:
+        - When `export=True`: Takes a PyTorch-based diffusion model, compiles it for RBLN NPUs, and loads the compiled model
+        - When `export=False`: Loads an already compiled RBLN model from `model_id` without recompilation
+
+        It supports various diffusion pipelines including Stable Diffusion, Kandinsky, ControlNet, and other diffusers-based models.
+
+        Args:
+            model_id (`str`):
+                The model ID or path to the pretrained model to load. Can be either:
+                - A model ID from the HuggingFace Hub
+                - A local path to a saved model directory
+            export (`bool`, *optional*, defaults to `False`):
+                If True, takes a PyTorch model from `model_id` and compiles it for RBLN NPU execution.
+                If False, loads an already compiled RBLN model from `model_id` without recompilation.
+            model_save_dir (`os.PathLike`, *optional*):
+                Directory to save the compiled model artifacts. Only used when `export=True`.
+                If not provided and `export=True`, a temporary directory is used.
+            rbln_config (`Dict[str, Any]`, *optional*, defaults to `{}`):
+                Configuration options for RBLN compilation. Can include settings for specific submodules
+                such as `text_encoder`, `unet`, and `vae`. Configuration can be tailored to the specific
+                pipeline being compiled.
+            lora_ids (`str` or `List[str]`, *optional*):
+                LoRA adapter ID(s) to load and apply before compilation. LoRA weights are fused
+                into the model weights during compilation. Only used when `export=True`.
+            lora_weights_names (`str` or `List[str]`, *optional*):
+                Names of specific LoRA weight files to load, corresponding to lora_ids. Only used when `export=True`.
+            lora_scales (`float` or `List[float]`, *optional*):
+                Scaling factor(s) to apply to the LoRA adapter(s). Only used when `export=True`.
+            **kwargs:
+                Additional arguments to pass to the underlying diffusion pipeline constructor or the
+                RBLN compilation process. These may include parameters specific to individual submodules
+                or the particular diffusion pipeline being used.
+
+        Returns:
+            `RBLNDiffusionMixin`: A compiled or loaded diffusion pipeline that can be used for inference on RBLN NPU.
+            The returned object is an instance of the class that called this method, inheriting from RBLNDiffusionMixin.
+        """
         rbln_config, kwargs = cls.get_rbln_config_class().initialize_from_kwargs(rbln_config, **kwargs)
 
         if export:

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/src/optimum/rbln/modeling_base.py

@@ -216,6 +216,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
             if isinstance(rbln_config, dict):
                 rbln_config_as_kwargs = {f"rbln_{key}": value for key, value in rbln_config.items()}
                 kwargs.update(rbln_config_as_kwargs)
+                rbln_config = None
             elif isinstance(rbln_config, RBLNModelConfig) and rbln_config.rbln_model_cls_name != cls.__name__:
                 raise ValueError(
                     f"Cannot use the passed rbln_config. Its model class name ({rbln_config.rbln_model_cls_name}) "
@@ -392,13 +393,13 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
     @classmethod
     def get_hf_class(cls):
         """
-        Lazily loads and caches the corresponding Hugging Face model class.
+        Lazily loads and caches the corresponding HuggingFace model class.
         Removes 'RBLN' prefix from the class name to get the original class name
         (e.g., RBLNLlamaForCausalLM -> LlamaForCausalLM) and imports it from
         the transformers/diffusers module.
 
         Returns:
-            type: The original Hugging Face model class
+            type: The original HuggingFace model class
         """
         if cls._hf_class is None:
             hf_cls_name = cls.__name__[4:]
@@ -478,7 +479,7 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
             save_directory (`Union[str, Path]`):
                 Directory where to save the model file.
             push_to_hub (`bool`, *optional*, defaults to `False`):
-                Whether or not to push your model to the Hugging Face model hub after saving it.
+                Whether or not to push your model to the HuggingFace model hub after saving it.
 
         """
         if os.path.isfile(save_directory):

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/src/optimum/rbln/transformers/models/auto/auto_factory.py

@@ -48,7 +48,7 @@ class _BaseAutoModelClass:
 
         Args:
             pretrained_model_name_or_path (str): Identifier or path to the pretrained model.
-            export (bool): Whether to infer the class based on Hugging Face (HF) architecture.
+            export (bool): Whether to infer the class based on HuggingFace (HF) architecture.
             kwargs: Additional arguments for configuration and loading.
 
         Returns:
@@ -86,14 +86,14 @@ class _BaseAutoModelClass:
         **kwargs,
     ):
         """
-        Infer the Hugging Face model class based on the configuration or model name.
+        Infer the HuggingFace model class based on the configuration or model name.
 
         Args:
             pretrained_model_name_or_path (str): Identifier or path to the pretrained model.
             kwargs: Additional arguments for configuration and loading.
 
         Returns:
-            PretrainedModel: The inferred Hugging Face model class.
+            PretrainedModel: The inferred HuggingFace model class.
         """
 
         # Try to load configuration if provided or retrieve it from the model ID

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/src/optimum/rbln/utils/hub.py

@@ -63,7 +63,7 @@ def pull_compiled_model_from_hub(
     force_download: bool,
     local_files_only: bool,
 ) -> Path:
-    """Pull model files from the Hugging Face Hub."""
+    """Pull model files from the HuggingFace Hub."""
     huggingface_token = _get_huggingface_token(use_auth_token)
     repo_files = list(
         map(
@@ -119,4 +119,4 @@ def _get_huggingface_token(use_auth_token: Union[bool, str]) -> str:
     elif use_auth_token:
         return HfFolder.get_token()
     else:
-        raise ValueError("`use_auth_token` must be provided to interact with the Hugging Face Hub.")
+        raise ValueError("`use_auth_token` must be provided to interact with the HuggingFace Hub.")

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/src/optimum/rbln/utils/model_utils.py

@@ -18,10 +18,10 @@ RBLN_PREFIX = "RBLN"
 
 def convert_hf_to_rbln_model_name(hf_model_name: str):
     """
-    Convert Hugging Face model name to RBLN model name.
+    Convert HuggingFace model name to RBLN model name.
 
     Args:
-        hf_model_name (str): The Hugging Face model name.
+        hf_model_name (str): The HuggingFace model name.
 
     Returns:
         str: The corresponding RBLN model name.
@@ -31,13 +31,13 @@ def convert_hf_to_rbln_model_name(hf_model_name: str):
 
 def convert_rbln_to_hf_model_name(rbln_model_name: str):
     """
-    Convert RBLN model name to Hugging Face model name.
+    Convert RBLN model name to HuggingFace model name.
 
     Args:
         rbln_model_name (str): The RBLN model name.
 
     Returns:
-        str: The corresponding Hugging Face model name.
+        str: The corresponding HuggingFace model name.
     """
 
     return rbln_model_name.removeprefix(RBLN_PREFIX)

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/tests/test_base.py

@@ -12,6 +12,8 @@ import transformers
 from diffusers import DiffusionPipeline
 from transformers import AutoConfig, CLIPConfig
 
+from optimum.rbln.configuration_utils import ContextRblnConfig
+
 
 class TestLevel(Enum):
     ESSENTIAL = 1
@@ -184,40 +186,44 @@ class BaseTest:
             if self.EXPECTED_OUTPUT:
                 self.assertEqual(output, self.EXPECTED_OUTPUT)
 
+        def _inner_test_save_load(self, tmpdir):
+            with self.subTest():
+                self.model.save_pretrained(tmpdir)
+                config_path = os.path.join(tmpdir, self.RBLN_CLASS.config_name)
+                self.assertTrue(os.path.exists(config_path), "save_pretrained does not work.")
+
+            with self.subTest():
+                # Test load
+                _ = self.RBLN_CLASS.from_pretrained(
+                    tmpdir,
+                    export=False,
+                    rbln_create_runtimes=False,
+                    **self.HF_CONFIG_KWARGS,
+                )
+
+            with self.subTest():
+                # Test saving from exported pipe
+                self.model.save_pretrained(tmpdir)
+                _ = self.RBLN_CLASS.from_pretrained(
+                    tmpdir,
+                    export=False,
+                    rbln_create_runtimes=False,
+                    **self.HF_CONFIG_KWARGS,
+                )
+
         def test_save_load(self):
             with tempfile.TemporaryDirectory() as tmpdir:
-                with self.subTest():
-                    self.model.save_pretrained(tmpdir)
-                    config_path = os.path.join(tmpdir, self.RBLN_CLASS.config_name)
-                    self.assertTrue(os.path.exists(config_path), "save_pretrained does not work.")
-
-                with self.subTest():
-                    # Test load
-                    _ = self.RBLN_CLASS.from_pretrained(
-                        tmpdir,
-                        export=False,
-                        rbln_create_runtimes=False,
-                        **self.HF_CONFIG_KWARGS,
-                    )
-
-                with self.subTest():
-                    # Test saving from exported pipe
-                    self.model.save_pretrained(tmpdir)
-                    _ = self.RBLN_CLASS.from_pretrained(
-                        tmpdir,
-                        export=False,
-                        rbln_create_runtimes=False,
-                        **self.HF_CONFIG_KWARGS,
-                    )
+                self._inner_test_save_load(tmpdir)
 
         def test_model_save_dir_load(self):
-            # Test model_save_dir
-            _ = self.RBLN_CLASS.from_pretrained(
-                self.get_rbln_local_dir(),
-                export=False,
-                rbln_create_runtimes=False,
-                **self.HF_CONFIG_KWARGS,
-            )
+            with ContextRblnConfig(create_runtimes=False):
+                # Test model_save_dir
+                _ = self.RBLN_CLASS.from_pretrained(
+                    self.get_rbln_local_dir(),
+                    export=False,
+                    rbln_create_runtimes=False,
+                    **self.HF_CONFIG_KWARGS,
+                )
 
         def test_automap(self):
             if self.RBLN_AUTO_CLASS is None:

{optimum_rbln-0.7.4a7 → optimum_rbln-0.7.4a8}/tests/test_llm.py

@@ -24,6 +24,7 @@ from optimum.rbln import (
     RBLNQwen2ForCausalLM,
     RBLNT5ForConditionalGeneration,
 )
+from optimum.rbln.configuration_utils import ContextRblnConfig
 
 from .test_base import BaseTest, DisallowedTestBase, TestLevel
 
@@ -224,7 +225,6 @@ class TestBartModel(LLMTest.TestLLM):
 class TestLlavaNextForConditionalGeneration(LLMTest.TestLLM):
     RBLN_AUTO_CLASS = RBLNAutoModelForVision2Seq
     RBLN_CLASS = RBLNLlavaNextForConditionalGeneration
-    TEST_LEVEL = TestLevel.FULL
     HF_MODEL_ID = "llava-hf/llava-v1.6-mistral-7b-hf"  # No tiny model yet.
     PROMPT = "[INST] <image>\nWhat’s shown in this image? [/INST]"
     RBLN_CLASS_KWARGS = {"rbln_config": {"language_model": {"use_inputs_embeds": True}}}
@@ -256,6 +256,18 @@ class TestLlavaNextForConditionalGeneration(LLMTest.TestLLM):
         inputs["do_sample"] = False
         return inputs
 
+    def _inner_test_save_load(self, tmpdir):
+        with ContextRblnConfig(create_runtimes=False):
+            super()._inner_test_save_load(tmpdir)
+
+            # Test loading from nested config
+            _ = self.RBLN_CLASS.from_pretrained(
+                tmpdir,
+                export=False,
+                rbln_config={"language_model": {"create_runtimes": False}},
+                **self.HF_CONFIG_KWARGS,
+            )
+
 
 class TestIdefics3ForConditionalGeneration(LLMTest.TestLLM):
     RBLN_AUTO_CLASS = RBLNAutoModelForVision2Seq