optimum-rbln 0.8.4a3__py3-none-any.whl → 0.8.4a5__py3-none-any.whl

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/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/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/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/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/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
 

{optimum_rbln-0.8.4a3.dist-info → optimum_rbln-0.8.4a5.dist-info}/METADATA

@@ -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.dist-info → optimum_rbln-0.8.4a5.dist-info}/RECORD

@@ -1,10 +1,10 @@
 optimum/rbln/__init__.py,sha256=DAJM5PWAYFiWVlyxVXUvj3CaFOEhX1yhEfhIt1LxL-A,18714
-optimum/rbln/__version__.py,sha256=y3K5Q-p3yh9vxRHMdoMoq6ATy6fBGj7-nnrEJehRK2c,712
+optimum/rbln/__version__.py,sha256=O3WB9nLWdjxzyq2yPcJiRu11qFUdi3QvSHtzLw_25A8,712
 optimum/rbln/configuration_utils.py,sha256=WNubd8EJIrdBkLOGT2UJJorgNL3lzhjg3a4bihAIptY,34761
 optimum/rbln/modeling.py,sha256=cAIPWEw5DGzUWeqjCbocRhU6OO3jyhVGW60AmBLh1Nw,14134
 optimum/rbln/modeling_base.py,sha256=97ju0uHJXB7PaorKaspf-FbLfsaHy0HwRVLJqtVscXA,27574
 optimum/rbln/diffusers/__init__.py,sha256=1tgU_xWA42BmInqu9bBz_5R_E9TGhhK3mI06YlaiTLg,7232
-optimum/rbln/diffusers/modeling_diffusers.py,sha256=3bzL0ZH7XyS8nGMWRSMIGjl9H3H2fhiZgmPaIF50mwg,20464
+optimum/rbln/diffusers/modeling_diffusers.py,sha256=SgunWDT_vAcCgoPLvHH4Vwc8sw8NvMLmkzys1fZcwJc,20472
 optimum/rbln/diffusers/configurations/__init__.py,sha256=vMRnPY4s-Uju43xP038D2EA18X_mhy2YfsZVpSU-VoA,1322
 optimum/rbln/diffusers/configurations/models/__init__.py,sha256=7q95gtgDzCeIBogGw8SLQoHT4Wch7vpLJVF2UQovuoo,567
 optimum/rbln/diffusers/configurations/models/configuration_autoencoder_kl.py,sha256=ADS4SGZbwY6fy3SVNhgo3Zg4KxzAAGq5_zsJ97Dezh4,3201
@@ -46,8 +46,8 @@ optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py
 optimum/rbln/diffusers/pipelines/cosmos/__init__.py,sha256=h2j6S8IJPVHeNU8qmW9vyXMgHBw0d7kQcuMAA5YoHPU,795
 optimum/rbln/diffusers/pipelines/cosmos/configuration_cosmos_guardrail.py,sha256=NsEaRYgiddopL9OE_9CsZqzbikkzoxUoRQaRML-Jqrw,3709
 optimum/rbln/diffusers/pipelines/cosmos/cosmos_guardrail.py,sha256=Hmklj-sGNtUst6sTLLDXyQzt7zFoNIHytZpPQ-V1CPw,16676
-optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_text2world.py,sha256=OtPi0Qw6afNp6r6e5Flbjbn5SsVeM611qEstCNEdWQ0,3899
-optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_video2world.py,sha256=dToWLGp9qyFZaccnoFZ42AnK8rKqnB54svui3mIaRNo,3906
+optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_text2world.py,sha256=jP3xIhVrfZ0odORWvqhef9rAN7KiqJ4KdV-vaKdxe_Y,5640
+optimum/rbln/diffusers/pipelines/cosmos/pipeline_cosmos_video2world.py,sha256=U6yK1_77ubCkKZ0XFyabYLSDATQTJqlBuvRQJNolnd0,5647
 optimum/rbln/diffusers/pipelines/kandinsky2_2/__init__.py,sha256=I4YQq2HfA3xONbWsdJ870IEJPyLWeCDDG-UCJsu9YO8,1035
 optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py,sha256=2vgZZt0JrZlbHbUlBHrT0zKvYhuX1a4vwy3mxTPHisM,1335
 optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py,sha256=QqxR325Or00E77XDvR3MotltnghC-42bky1ckU1DW9U,8128
@@ -186,7 +186,7 @@ optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=hRvA37
 optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py,sha256=i_UUWhKoFjJ5CCpgeWicqABM23TxMEKPQ354LoZ6iUU,7445
 optimum/rbln/transformers/models/qwen2_vl/__init__.py,sha256=O3t6zKda92CnZDzEnz_dcisMOQ71-OOJxElXzKCH5e0,849
 optimum/rbln/transformers/models/qwen2_vl/configuration_qwen2_vl.py,sha256=OGIlUHWNymBTOxnwit-1gm2Gpl8bbGV0i076Sa4RgCw,4718
-optimum/rbln/transformers/models/qwen2_vl/modeling_qwen2_vl.py,sha256=OKjhwWe0UDczmauCNQA838BF3n1BIz8c7oM5gaBVUz8,20286
+optimum/rbln/transformers/models/qwen2_vl/modeling_qwen2_vl.py,sha256=10NR0LPJFICpPUrINpyMDWVx19HN8sXPpeAiVhE_k2k,20130
 optimum/rbln/transformers/models/qwen2_vl/qwen2_vl_architecture.py,sha256=EZlCuSRTIpSAGEjtDi4SY1V9RRdtgg76ie5jqec1UuI,4833
 optimum/rbln/transformers/models/qwen3/__init__.py,sha256=tI4KwvXpD35dUUaa8aLUXpWoU9gJGcmKXeywOlH14ZE,746
 optimum/rbln/transformers/models/qwen3/configuration_qwen3.py,sha256=BFRPggnH4VlsXlOa19C6KAID-bPgQ8ooQ29dvogh5zk,2102
@@ -200,7 +200,7 @@ optimum/rbln/transformers/models/roberta/configuration_roberta.py,sha256=6KhO-xB
 optimum/rbln/transformers/models/roberta/modeling_roberta.py,sha256=74Pswb5JJNtctvrQHlo2zYocKZN0npWhjAaKMUDVBUU,1535
 optimum/rbln/transformers/models/seq2seq/__init__.py,sha256=HiSyWFcKeZ8okfo-s-_Mf_upyvAoZwraUIJyGNLNurY,714
 optimum/rbln/transformers/models/seq2seq/configuration_seq2seq.py,sha256=Che6vOuehOD6hwCQ0NbGbh9eSI2AKl4ldiavKKfzHQo,3123
-optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=8qRESu6Go6smZJCnZroML1zFirY0Vo4CGJm9yuo8tpA,18303
+optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=G7Rkx4paSDlSE2pa1dtXTOGN9L9uy7ovVZjSrS8c5Kk,18295
 optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=4MupGjhe_DpBLl43VVYY72WWiVdAr0mRPp8cHalQh2w,20028
 optimum/rbln/transformers/models/siglip/__init__.py,sha256=X1Fc1GUnJ2EIxFx45nbeoW-T2t0OyP3W73C0HD8Vowo,712
 optimum/rbln/transformers/models/siglip/configuration_siglip.py,sha256=m1h8iDx_X9VmHdJi0sc1a2KsAO3OnpMb4cd9jW2Ic-U,3031
@@ -242,7 +242,7 @@ optimum/rbln/utils/model_utils.py,sha256=4k5879Kh75m3x_vS4-qOGfqsOiAvc2kdNFFfvsF
 optimum/rbln/utils/runtime_utils.py,sha256=R6uXDbeJP03-FWdd4vthNe2D4aCra5n12E3WB1ifiGM,7933
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=60NGLFvnhjP1DJg1opdb-FVQDsthcLCwWjW_1WQaasU,5280
-optimum_rbln-0.8.4a3.dist-info/METADATA,sha256=oCoMIFXq285yfSiKRSadH1GOD7aSHzZ1eGdXs2BLLqA,5299
-optimum_rbln-0.8.4a3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.8.4a3.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.8.4a3.dist-info/RECORD,,
+optimum_rbln-0.8.4a5.dist-info/METADATA,sha256=tNYTi5jcClIr6oyXbiAJ_69C485ec2nGx23ob_fTdwI,5299
+optimum_rbln-0.8.4a5.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.8.4a5.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.8.4a5.dist-info/RECORD,,