optimum-rbln 0.8.4a3__tar.gz → 0.8.4a5__tar.gz

optimum_rbln-0.8.4a5/.github/renovate.json

@@ -0,0 +1,42 @@
+{
+    "$schema": "https://docs.renovatebot.com/renovate-schema.json",
+    "prHourlyLimit": 0,
+    "extends": [
+        ":semanticCommitsDisabled"
+    ],
+    "commitMessagePrefix": "dependency:",
+    "repositories": [
+        "rebellions-sw/optimum-rbln"
+    ],
+    "lockFileMaintenance": {
+        "enabled": true
+    },
+    "packageRules": [
+        {
+            "matchPackageNames": [
+                "*"
+            ],
+            "enabled": false
+        },
+        {
+            "matchDatasources": [
+                "pypi"
+            ],
+            "matchPackageNames": [
+                "diffusers"
+            ],
+            "enabled": true,
+            "groupName": "huggingface-diffusers"
+        },
+        {
+            "matchDatasources": [
+                "pypi"
+            ],
+            "matchPackageNames": [
+                "transformers"
+            ],
+            "enabled": true,
+            "groupName": "huggingface-transformers"
+        }
+    ]
+}

optimum_rbln-0.8.4a5/.github/workflows/auto_dependency_bot.yml

@@ -0,0 +1,22 @@
+name: Auto Dependency Bot
+on:
+  schedule:
+    - cron: '0 0 * * *'  # UTC 00:00 = KST 09:00
+  workflow_dispatch:
+
+jobs:
+  renovate:
+    runs-on: rebel-k8s-runner
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Run Renovate
+        uses: renovatebot/github-action@v39.0.2
+        with:
+          configurationFile: .github/renovate.json
+          token: ${{ secrets.RENOVATE_TOKEN }}
+        env:
+          LOG_LEVEL: debug
+          RENOVATE_PLATFORM: github
+          RENOVATE_TOKEN: ${{ secrets.RENOVATE_TOKEN }}

{optimum_rbln-0.8.4a3 → optimum_rbln-0.8.4a5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.4a3
+Version: 0.8.4a5
 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

{optimum_rbln-0.8.4a3 → optimum_rbln-0.8.4a5}/src/optimum/rbln/__version__.py

@@ -28,7 +28,7 @@ version_tuple: VERSION_TUPLE
 commit_id: COMMIT_ID
 __commit_id__: COMMIT_ID
 
-__version__ = version = '0.8.4a3'
-__version_tuple__ = version_tuple = (0, 8, 4, 'a3')
+__version__ = version = '0.8.4a5'
+__version_tuple__ = version_tuple = (0, 8, 4, 'a5')
 
 __commit_id__ = commit_id = None

{optimum_rbln-0.8.4a3 → optimum_rbln-0.8.4a5}/src/optimum/rbln/diffusers/modeling_diffusers.py

@@ -142,8 +142,8 @@ class 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
+            - 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.
 

{optimum_rbln-0.8.4a3 → optimum_rbln-0.8.4a5}/src/optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_text2world.py

@@ -89,6 +89,36 @@ class RBLNCosmosTextToWorldPipeline(RBLNDiffusionMixin, CosmosTextToWorldPipelin
         rbln_config: Dict[str, Any] = {},
         **kwargs: Any,
     ):
+        """
+        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:
+                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.
+            safety_checker:
+                Optional custom safety checker to use instead of the default one. Only used when `export=True`.
+            rbln_config:
+                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.
+            **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.
+        """
+
         rbln_config, kwargs = cls.get_rbln_config_class().initialize_from_kwargs(rbln_config, **kwargs)
         if safety_checker is None and export:
             safety_checker = RBLNCosmosSafetyChecker(rbln_config=rbln_config.safety_checker)

{optimum_rbln-0.8.4a3 → optimum_rbln-0.8.4a5}/src/optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_video2world.py

@@ -89,6 +89,36 @@ class RBLNCosmosVideoToWorldPipeline(RBLNDiffusionMixin, CosmosVideoToWorldPipel
         rbln_config: Dict[str, Any] = {},
         **kwargs: Any,
     ):
+        """
+        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:
+                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.
+            safety_checker:
+                Optional custom safety checker to use instead of the default one. Only used when `export=True`.
+            rbln_config:
+                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.
+            **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.
+        """
+
         rbln_config, kwargs = cls.get_rbln_config_class().initialize_from_kwargs(rbln_config, **kwargs)
         if safety_checker is None and export:
             safety_checker = RBLNCosmosSafetyChecker(rbln_config=rbln_config.safety_checker)

{optimum_rbln-0.8.4a3 → optimum_rbln-0.8.4a5}/src/optimum/rbln/transformers/models/qwen2_vl/modeling_qwen2_vl.py

@@ -254,8 +254,7 @@ class RBLNQwen2VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
                     "device": 0,
                 },
                 "tensor_parallel_size": 8,
-                "kvcache_partition_len": 16_384,
-                "max_seq_len": 114_688,
+                "max_seq_len": 32_768,
                 "device": [0, 1, 2, 3, 4, 5, 6, 7],
             },
         )
@@ -407,7 +406,6 @@ class RBLNQwen2VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
                 input_id,
                 image_grid_thw[image_idx : image_idx + image_nums] if image_grid_thw is not None else None,
                 video_grid_thw[video_idx : video_idx + video_nums] if video_grid_thw is not None else None,
-                attention_mask=attention_mask[b_idx : b_idx + 1] if attention_mask is not None else None,
             )
             image_idx += image_nums
             video_idx += video_nums

{optimum_rbln-0.8.4a3 → optimum_rbln-0.8.4a5}/src/optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py

@@ -83,7 +83,7 @@ class RBLNRuntimeDecoder(RBLNPytorchRuntime):
                 decoding_step = cache_position[b_idx].item()
                 if not (0 <= decoding_step < self.dec_max_seq_len):
                     raise ValueError(
-                        f"Decoding step {decoding_step} out of bounds for attention mask with shape {self.dec_attn_mask.shape}."
+                        f"Decoding step {decoding_step} out of bounds for decoder_max_seq_len ({self.dec_max_seq_len})."
                     )
                 decoder_attention_mask[b_idx, : decoding_step + 1] = 1