optimum-rbln 0.8.2a1__tar.gz → 0.8.2a2__tar.gz

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.2a1
+Version: 0.8.2a2
 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-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/__init__.py

@@ -110,6 +110,10 @@ _import_structure = {
         "RBLNQwen2_5_VisionTransformerPretrainedModelConfig",
         "RBLNQwen2_5_VLForConditionalGeneration",
         "RBLNQwen2_5_VLForConditionalGenerationConfig",
+        "RBLNQwen3ForCausalLM",
+        "RBLNQwen3ForCausalLMConfig",
+        "RBLNQwen3Model",
+        "RBLNQwen3ModelConfig",
         "RBLNResNetForImageClassification",
         "RBLNResNetForImageClassificationConfig",
         "RBLNRobertaForMaskedLM",
@@ -357,6 +361,10 @@ if TYPE_CHECKING:
         RBLNQwen2_5_VLForConditionalGenerationConfig,
         RBLNQwen2ForCausalLM,
         RBLNQwen2ForCausalLMConfig,
+        RBLNQwen3ForCausalLM,
+        RBLNQwen3ForCausalLMConfig,
+        RBLNQwen3Model,
+        RBLNQwen3ModelConfig,
         RBLNResNetForImageClassification,
         RBLNResNetForImageClassificationConfig,
         RBLNRobertaForMaskedLM,

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.8.2a1'
-__version_tuple__ = version_tuple = (0, 8, 2, 'a1')
+__version__ = version = '0.8.2a2'
+__version_tuple__ = version_tuple = (0, 8, 2, 'a2')

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/configuration_utils.py

@@ -147,7 +147,7 @@ class RBLNCompileConfig:
         return asdict(self)
 
 
-RUNTIME_KEYWORDS = ["create_runtimes", "optimize_host_memory", "device", "device_map", "activate_profiler"]
+RUNTIME_KEYWORDS = ["create_runtimes", "optimize_host_memory", "device", "device_map", "activate_profiler", "timeout"]
 CONFIG_MAPPING: Dict[str, Type["RBLNModelConfig"]] = {}
 
 
@@ -481,6 +481,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
         "device",
         "device_map",
         "activate_profiler",
+        "timeout",
     ]
     submodules: List[str] = []
     subclass_non_save_attributes = []
@@ -561,6 +562,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
         activate_profiler: Optional[bool] = None,
         npu: Optional[str] = None,
         tensor_parallel_size: Optional[int] = None,
+        timeout: Optional[int] = None,
         optimum_rbln_version: Optional[str] = None,
         _compile_cfgs: List[RBLNCompileConfig] = [],
         **kwargs: Dict[str, Any],
@@ -577,6 +579,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
             activate_profiler (Optional[bool]): Whether to activate the profiler for performance analysis.
             npu (Optional[str]): The NPU device name to use for compilation.
             tensor_parallel_size (Optional[int]): Size for tensor parallelism to distribute the model across devices.
+            timeout (Optional[int]): The timeout for the runtime in seconds. If it isn't provided, it will be set to 60 by default.
             optimum_rbln_version (Optional[str]): The optimum-rbln version used for this configuration.
             _compile_cfgs (List[RBLNCompileConfig]): List of compilation configurations for the model.
             **kwargs: Additional keyword arguments.
@@ -599,6 +602,7 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
         self._runtime_options["device"] = device
         self._runtime_options["device_map"] = device_map
         self._runtime_options["activate_profiler"] = activate_profiler
+        self._runtime_options["timeout"] = timeout
 
         # Automatically pass npu, tensor_parallel_size to compile_cfgs
         self.npu = npu
@@ -838,3 +842,14 @@ class RBLNModelConfig(RBLNSerializableConfigProtocol):
     @activate_profiler.setter
     def activate_profiler(self, activate_profiler: bool):
         self._runtime_options["activate_profiler"] = activate_profiler
+
+    @property
+    def timeout(self):
+        context = ContextRblnConfig.get_current_context()["timeout"]
+        if context is not None:
+            return context
+        return self._runtime_options["timeout"]
+
+    @timeout.setter
+    def timeout(self, timeout: int):
+        self._runtime_options["timeout"] = timeout

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/configurations/models/configuration_transformer_cosmos.py

@@ -52,6 +52,9 @@ class RBLNCosmosTransformer3DModelConfig(RBLNModelConfig):
         Raises:
             ValueError: If batch_size is not a positive integer.
         """
+        if kwargs.get("timeout") is None:
+            kwargs["timeout"] = 80
+
         super().__init__(**kwargs)
         self.batch_size = batch_size or 1
         self.num_frames = num_frames or 121

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/modeling_diffusers.py

@@ -230,6 +230,7 @@ class RBLNDiffusionMixin:
             create_runtimes=rbln_config.create_runtimes,
             optimize_host_mem=rbln_config.optimize_host_memory,
             activate_profiler=rbln_config.activate_profiler,
+            timeout=rbln_config.timeout,
         ):
             model = super().from_pretrained(pretrained_model_name_or_path=model_id, **kwargs)
 

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/models/autoencoders/autoencoder_kl.py

@@ -209,6 +209,7 @@ class RBLNAutoencoderKL(RBLNModel):
                 tensor_type="pt",
                 device=device_val,
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model, device_val in zip(compiled_models, device_vals)
         ]

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/models/autoencoders/autoencoder_kl_cosmos.py

@@ -200,6 +200,7 @@ class RBLNAutoencoderKLCosmos(RBLNModel):
                 tensor_type="pt",
                 device=device_val,
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model, device_val in zip(compiled_models, device_vals)
         ]

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/models/autoencoders/vq_model.py

@@ -165,6 +165,7 @@ class RBLNVQModel(RBLNModel):
                 tensor_type="pt",
                 device=device_val,
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model, device_val in zip(compiled_models, device_vals)
         ]

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/models/transformers/transformer_cosmos.py

@@ -279,7 +279,7 @@ class RBLNCosmosTransformer3DModel(RBLNModel):
                 tensor_type="pt",
                 device=rbln_config.device_map[DEFAULT_COMPILED_MODEL_NAME],
                 activate_profiler=rbln_config.activate_profiler,
-                timeout=120,
+                timeout=rbln_config.timeout,
             )
             for compiled_model in compiled_models
         ]

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/pipelines/cosmos/configuration_cosmos_guardrail.py

@@ -15,7 +15,7 @@
 from typing import Any, Dict, Optional, Tuple
 
 from ....configuration_utils import RBLNAutoConfig, RBLNModelConfig
-from ....transformers import RBLNSiglipVisionModelConfig
+from ....transformers import RBLNLlamaForCausalLMConfig, RBLNSiglipVisionModelConfig
 
 
 class RBLNVideoSafetyModelConfig(RBLNModelConfig):
@@ -75,7 +75,15 @@ class RBLNCosmosSafetyCheckerConfig(RBLNModelConfig):
         if height is not None and width is not None:
             image_size = (height, width)
 
-        self.aegis = self.init_submodule_config(RBLNModelConfig, aegis)
+        tensor_parallel_size = kwargs.get("tensor_parallel_size")
+
+        self.aegis = self.init_submodule_config(
+            RBLNLlamaForCausalLMConfig,
+            aegis,
+            batch_size=batch_size,
+            tensor_parallel_size=tensor_parallel_size,
+        )
+
         self.siglip_encoder = self.init_submodule_config(
             RBLNSiglipVisionModelConfig,
             siglip_encoder,

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/diffusers/pipelines/cosmos/cosmos_guardrail.py

@@ -127,25 +127,13 @@ class RBLNSigLIPEncoder(SigLIPEncoder):
 
             # We don't use RBLNSiglipModel, but we need to override get_image_features to return pooler_output
             self.model = RBLNSiglipVisionModel.from_pretrained(
-                self.checkpoint_dir,
-                rbln_device=rbln_config.siglip_encoder.device,
-                rbln_create_runtimes=rbln_config.siglip_encoder.create_runtimes,
-                rbln_activate_profiler=rbln_config.siglip_encoder.activate_profiler,
-                rbln_optimize_host_memory=rbln_config.siglip_encoder.optimize_host_memory,
+                self.checkpoint_dir, rbln_config=rbln_config.siglip_encoder
             )
         else:
             super().__init__(model_name, checkpoint_id)
             model = self.model
             del self.model
-            self.model = RBLNSiglipVisionModel.from_model(
-                model,
-                rbln_device=rbln_config.siglip_encoder.device,
-                rbln_image_size=rbln_config.siglip_encoder.image_size,
-                rbln_npu=rbln_config.siglip_encoder.npu,
-                rbln_create_runtimes=rbln_config.siglip_encoder.create_runtimes,
-                rbln_activate_profiler=rbln_config.siglip_encoder.activate_profiler,
-                rbln_optimize_host_memory=rbln_config.siglip_encoder.optimize_host_memory,
-            )
+            self.model = RBLNSiglipVisionModel.from_model(model, rbln_config=rbln_config.siglip_encoder)
         self.rbln_config = rbln_config
 
         # Override get_image_features to return pooler_output
@@ -336,28 +324,14 @@ class RBLNAegis(Aegis):
             torch.nn.Module.__init__(self)
             cache_dir = pathlib.Path(checkpoint_id) / "aegis"
             self.tokenizer = AutoTokenizer.from_pretrained(cache_dir)
-            self.model = RBLNAutoModelForCausalLM.from_pretrained(
-                cache_dir,
-                rbln_device=rbln_config.aegis.device,
-                rbln_create_runtimes=rbln_config.aegis.create_runtimes,
-                rbln_activate_profiler=rbln_config.aegis.activate_profiler,
-                rbln_optimize_host_memory=rbln_config.aegis.optimize_host_memory,
-            )
+            self.model = RBLNAutoModelForCausalLM.from_pretrained(cache_dir, rbln_config=rbln_config.aegis)
 
         else:
             super().__init__(checkpoint_id, base_model_id, aegis_adapter)
             model = self.model.merge_and_unload()  # peft merge
             del self.model
 
-            self.model = RBLNAutoModelForCausalLM.from_model(
-                model,
-                rbln_tensor_parallel_size=4,
-                rbln_device=rbln_config.aegis.device,
-                rbln_create_runtimes=rbln_config.aegis.create_runtimes,
-                rbln_npu=rbln_config.aegis.npu,
-                rbln_activate_profiler=rbln_config.aegis.activate_profiler,
-                rbln_optimize_host_memory=rbln_config.aegis.optimize_host_memory,
-            )
+            self.model = RBLNAutoModelForCausalLM.from_model(model, rbln_config=rbln_config.aegis)
 
         self.rbln_config = rbln_config
         self.dtype = torch.bfloat16

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/modeling.py

@@ -238,6 +238,7 @@ class RBLNModel(RBLNBaseModel):
                 tensor_type="pt",
                 device=rbln_config.device_map[DEFAULT_COMPILED_MODEL_NAME],
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             )
             for compiled_model in compiled_models
         ]

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/transformers/__init__.py

@@ -98,6 +98,10 @@ _import_structure = {
         "RBLNQwen2_5_VLForConditionalGenerationConfig",
         "RBLNQwen2ForCausalLM",
         "RBLNQwen2ForCausalLMConfig",
+        "RBLNQwen3ForCausalLM",
+        "RBLNQwen3ForCausalLMConfig",
+        "RBLNQwen3Model",
+        "RBLNQwen3ModelConfig",
         "RBLNResNetForImageClassification",
         "RBLNResNetForImageClassificationConfig",
         "RBLNRobertaForMaskedLM",
@@ -204,6 +208,10 @@ if TYPE_CHECKING:
         RBLNQwen2_5_VLForConditionalGenerationConfig,
         RBLNQwen2ForCausalLM,
         RBLNQwen2ForCausalLMConfig,
+        RBLNQwen3ForCausalLM,
+        RBLNQwen3ForCausalLMConfig,
+        RBLNQwen3Model,
+        RBLNQwen3ModelConfig,
         RBLNResNetForImageClassification,
         RBLNResNetForImageClassificationConfig,
         RBLNRobertaForMaskedLM,

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/transformers/models/__init__.py

@@ -113,6 +113,7 @@ _import_structure = {
     "mistral": ["RBLNMistralForCausalLM", "RBLNMistralForCausalLMConfig"],
     "phi": ["RBLNPhiForCausalLM", "RBLNPhiForCausalLMConfig"],
     "qwen2": ["RBLNQwen2ForCausalLM", "RBLNQwen2ForCausalLMConfig"],
+    "qwen3": ["RBLNQwen3ForCausalLM", "RBLNQwen3ForCausalLMConfig", "RBLNQwen3Model", "RBLNQwen3ModelConfig"],
     "resnet": ["RBLNResNetForImageClassification", "RBLNResNetForImageClassificationConfig"],
     "roberta": [
         "RBLNRobertaForMaskedLM",
@@ -241,6 +242,7 @@ if TYPE_CHECKING:
         RBLNQwen2_5_VLForConditionalGeneration,
         RBLNQwen2_5_VLForConditionalGenerationConfig,
     )
+    from .qwen3 import RBLNQwen3ForCausalLM, RBLNQwen3ForCausalLMConfig, RBLNQwen3Model, RBLNQwen3ModelConfig
     from .resnet import RBLNResNetForImageClassification, RBLNResNetForImageClassificationConfig
     from .roberta import (
         RBLNRobertaForMaskedLM,

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py

@@ -1085,6 +1085,7 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
                 tensor_type="pt",
                 device=rbln_config.device_map["prefill"],
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             ),
             *[
                 rebel.Runtime(
@@ -1092,6 +1093,7 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
                     tensor_type="pt",
                     device=rbln_config.device_map[f"decoder_batch_{batch_size}"],
                     activate_profiler=rbln_config.activate_profiler,
+                    timeout=rbln_config.timeout,
                 )
                 for i, batch_size in enumerate(rbln_config.decoder_batch_sizes)
             ],
@@ -1190,6 +1192,11 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
         if cache_position is None:
             logits = []
             inputs = inputs_embeds if inputs_embeds is not None else input_ids
+            # for only use forward
+            if generate_idx is None:
+                generate_idx = attention_mask.sum(dim=-1, keepdim=True).int()
+            if padded_cache_lengths is None:
+                padded_cache_lengths = torch.zeros_like(generate_idx)
             batch_size = inputs.shape[0]
             for b_idx in range(batch_size):
                 cache_position = torch.arange(0, generate_idx[b_idx].item(), dtype=torch.int32).unsqueeze(0)

{optimum_rbln-0.8.2a1 → optimum_rbln-0.8.2a2}/src/optimum/rbln/transformers/models/gemma3/modeling_gemma3.py

@@ -884,12 +884,14 @@ class RBLNGemma3ForCausalLM(RBLNDecoderOnlyModelForCausalLM):
                 tensor_type="pt",
                 device=rbln_config.device_map["prefill"],
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             ),
             rebel.Runtime(
                 compiled_models[1],
                 tensor_type="pt",
                 device=rbln_config.device_map["image_prefill"],
                 activate_profiler=rbln_config.activate_profiler,
+                timeout=rbln_config.timeout,
             ),
             *[
                 rebel.Runtime(
@@ -897,6 +899,7 @@ class RBLNGemma3ForCausalLM(RBLNDecoderOnlyModelForCausalLM):
                     tensor_type="pt",
                     device=rbln_config.device_map[f"decoder_batch_{batch_size}"],
                     activate_profiler=rbln_config.activate_profiler,
+                    timeout=rbln_config.timeout,
                 )
                 for i, batch_size in enumerate(rbln_config.decoder_batch_sizes)
             ],

optimum_rbln-0.8.2a2/src/optimum/rbln/transformers/models/qwen3/__init__.py

@@ -0,0 +1,16 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_qwen3 import RBLNQwen3ForCausalLMConfig, RBLNQwen3ModelConfig
+from .modeling_qwen3 import RBLNQwen3ForCausalLM, RBLNQwen3Model

optimum_rbln-0.8.2a2/src/optimum/rbln/transformers/models/qwen3/configuration_qwen3.py

@@ -0,0 +1,71 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNQwen3ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    """
+    Configuration class for RBLN Qwen3 models.
+
+    This class is an alias of RBLNDecoderOnlyModelForCausalLMConfig.
+
+    Example usage:
+    ```python
+    from optimum.rbln import RBLNQwen3ForCausalLM, RBLNQwen3ForCausalLMConfig
+
+    # Create a configuration object
+    config = RBLNQwen3ForCausalLMConfig(
+        batch_size=1,
+        max_seq_len=40960,
+        tensor_parallel_size=4,
+        kvcache_partition_len=16384
+    )
+
+    # Use the configuration with from_pretrained
+    model = RBLNQwen3ForCausalLM.from_pretrained(
+        "Qwen/Qwen3-4B",
+        export=True,
+        rbln_config=config
+    )
+    ```
+    """
+
+
+class RBLNQwen3ModelConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    """
+    Configuration class for RBLN Qwen3 models.
+
+    This class is an alias of RBLNDecoderOnlyModelForCausalLMConfig.
+
+    Example usage:
+    ```python
+    from optimum.rbln import RBLNQwen3Model, RBLNQwen3ModelConfig
+
+    # Create a configuration object
+    config = RBLNQwen3ModelConfig(
+        batch_size=1,
+        max_seq_len=40960,
+        tensor_parallel_size=4,
+        kvcache_partition_len=16384
+    )
+
+    # Use the configuration with from_pretrained
+    model = RBLNQwen3Model.from_pretrained(
+        "Qwen/Qwen3-Embedding-4B",
+        export=True,
+        rbln_config=config
+    )
+    ```
+    """