optimum-rbln 0.7.4a0__tar.gz → 0.7.4a1__tar.gz

optimum_rbln-0.7.4a1/.github/version.yaml

@@ -0,0 +1 @@
+rebel_compiler_version: 0.7.4.dev61+gb562a7f0

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.4a0
+Version: 0.7.4a1
 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.
 Project-URL: Homepage, https://rebellions.ai
 Project-URL: Documentation, https://docs.rbln.ai
@@ -25,10 +25,10 @@ Requires-Python: <3.13,>=3.9
 Requires-Dist: accelerate>=1.0.1
 Requires-Dist: diffusers<=0.31.0
 Requires-Dist: packaging>=24.1
-Requires-Dist: torch<=2.5.1
-Requires-Dist: torchaudio<=2.5.1
-Requires-Dist: torchvision<=0.20.1
-Requires-Dist: transformers==4.48.3
+Requires-Dist: torch==2.6.0
+Requires-Dist: torchaudio<=2.6.0
+Requires-Dist: torchvision<=0.21.0
+Requires-Dist: transformers==4.50.3
 Description-Content-Type: text/markdown
 
 

optimum_rbln-0.7.4a1/examples/time-series-forecasting/run_time_series_forecasting.py

@@ -0,0 +1,43 @@
+import os
+
+import fire
+import torch
+from huggingface_hub import hf_hub_download
+
+from optimum.rbln import RBLNTimeSeriesTransformerForPrediction
+
+
+def main(
+    model_id: str = "huggingface/time-series-transformer-tourism-monthly",
+    batch_size: int = 1,
+    num_parallel_samples: int = 100,
+    from_transformers: bool = False,
+):
+    if from_transformers:
+        model = RBLNTimeSeriesTransformerForPrediction.from_pretrained(
+            model_id, export=True, rbln_batch_size=batch_size, num_parallel_samples=num_parallel_samples
+        )
+        model.save_pretrained(os.path.basename(model_id))
+    else:
+        model = RBLNTimeSeriesTransformerForPrediction.from_pretrained(
+            os.path.basename(model_id),
+            export=False,
+        )
+
+    dataset = hf_hub_download(
+        repo_id="hf-internal-testing/tourism-monthly-batch", filename="val-batch.pt", repo_type="dataset"
+    )
+    data = torch.load(dataset, weights_only=True)
+
+    batched_data = {}
+    for k, v in data.items():
+        batched_data[k] = v[:batch_size]
+
+    rbln_outputs = model.generate(**batched_data)
+    mean_prediction = rbln_outputs.sequences.mean(dim=1)
+
+    print(mean_prediction)
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/pyproject.toml

@@ -28,11 +28,11 @@ classifiers = [
 ]
 keywords = ["transformers", "diffusers", "inference", "rbln", "atom", "rebel"]
 dependencies = [
-    "torch<=2.5.1",
-    "torchaudio<=2.5.1",
-    "torchvision<=0.20.1",
+    "torch==2.6.0",
+    "torchaudio<=2.6.0",
+    "torchvision<=0.21.0",
     "accelerate>=1.0.1",
-    "transformers==4.48.3",
+    "transformers==4.50.3",
     "diffusers<=0.31.0",
     "packaging>=24.1",
 ]
@@ -53,6 +53,8 @@ tests = [
     "sacremoses>=0.1.1",
     "safetensors>=0.4.2",
     "protobuf>=5.27.2",
+    "soundfile>=0.13.1",
+    "librosa>=0.11.0",
 ]
 quality = [
     "ruff>=0.3.3",

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/__init__.py

@@ -73,6 +73,7 @@ _import_structure = {
         "RBLNRobertaForMaskedLM",
         "RBLNViTForImageClassification",
         "RBLNBertForMaskedLM",
+        "RBLNTimeSeriesTransformerForPrediction",
     ],
     "diffusers": [
         "RBLNAutoencoderKL",
@@ -184,6 +185,7 @@ if TYPE_CHECKING:
         RBLNRobertaForSequenceClassification,
         RBLNT5EncoderModel,
         RBLNT5ForConditionalGeneration,
+        RBLNTimeSeriesTransformerForPrediction,
         RBLNViTForImageClassification,
         RBLNWav2Vec2ForCTC,
         RBLNWhisperForConditionalGeneration,

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.4a0'
+__version__ = version = '0.7.4a1'
 __version_tuple__ = version_tuple = (0, 7, 4)

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/ops/__init__.py

@@ -19,3 +19,4 @@ from .attn import (
 )
 from .flash_attn import register_rbln_custom_paged_flash_attention, register_rbln_custom_paged_flash_causal_attention
 from .kv_cache_update import register_rbln_custom_cache_update
+from .linear import linear

optimum_rbln-0.7.4a1/src/optimum/rbln/ops/linear.py

@@ -0,0 +1,25 @@
+# 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 typing import Optional
+
+import torch
+from torch import Tensor
+
+
+@torch.library.custom_op("rbln_custom_ops::linear", mutates_args=())
+def linear(input: Tensor, weight: Tensor, bias: Optional[Tensor] = None) -> Tensor:
+    output_shape = list(input.shape[:-1])
+    output_shape += [weight.shape[0]]
+    return torch.empty(size=output_shape, dtype=input.dtype, device=input.device, requires_grad=input.requires_grad)

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/transformers/__init__.py

@@ -52,6 +52,7 @@ _import_structure = {
         "RBLNPhiForCausalLM",
         "RBLNT5EncoderModel",
         "RBLNT5ForConditionalGeneration",
+        "RBLNTimeSeriesTransformerForPrediction",
         "RBLNLlavaNextForConditionalGeneration",
         "RBLNMidmLMHeadModel",
         "RBLNXLMRobertaModel",
@@ -113,6 +114,7 @@ if TYPE_CHECKING:
         RBLNQwen2ForCausalLM,
         RBLNT5EncoderModel,
         RBLNT5ForConditionalGeneration,
+        RBLNTimeSeriesTransformerForPrediction,
         RBLNWav2Vec2ForCTC,
         RBLNWhisperForConditionalGeneration,
         RBLNXLMRobertaModel,

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/transformers/models/__init__.py

@@ -50,6 +50,7 @@ _import_structure = {
     "mistral": ["RBLNMistralForCausalLM"],
     "phi": ["RBLNPhiForCausalLM"],
     "qwen2": ["RBLNQwen2ForCausalLM"],
+    "time_series_transformers": ["RBLNTimeSeriesTransformerForPrediction"],
     "t5": ["RBLNT5EncoderModel", "RBLNT5ForConditionalGeneration"],
     "wav2vec2": ["RBLNWav2Vec2ForCTC"],
     "whisper": ["RBLNWhisperForConditionalGeneration"],
@@ -90,6 +91,7 @@ if TYPE_CHECKING:
     from .phi import RBLNPhiForCausalLM
     from .qwen2 import RBLNQwen2ForCausalLM
     from .t5 import RBLNT5EncoderModel, RBLNT5ForConditionalGeneration
+    from .time_series_transformers import RBLNTimeSeriesTransformerForPrediction
     from .wav2vec2 import RBLNWav2Vec2ForCTC
     from .whisper import RBLNWhisperForConditionalGeneration
     from .xlm_roberta import RBLNXLMRobertaModel

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/transformers/models/bart/modeling_bart.py

@@ -94,12 +94,11 @@ class RBLNBartModel(RBLNModel):
             for model_input_name in rbln_model_input_names
         ]
 
-        enc_compile_config = RBLNCompileConfig(input_info=input_info, compiled_model_name="encoder")
-        dec_compile_config = RBLNCompileConfig(input_info=input_info, compiled_model_name="decoder")
+        rbln_compile_config = RBLNCompileConfig(input_info=input_info)
 
         rbln_config = RBLNConfig(
             rbln_cls=cls.__name__,
-            compile_cfgs=[enc_compile_config, dec_compile_config],
+            compile_cfgs=[rbln_compile_config],
             rbln_kwargs=rbln_kwargs,
         )
 

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py

@@ -222,8 +222,6 @@ class RBLNRuntimeModel(RBLNPytorchRuntime):
 
             attention_mask = self.dec_attn_mask
 
-            attention_mask = self.dec_attn_mask
-
         logits = super().forward(
             inputs,
             cache_position,
@@ -547,22 +545,27 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
 
         @QuantizationManager.with_quantization_env
         def compile_model(*args, **kwargs):
-            wrapped_model.phase = "prefill"
-            compiled_prefill = RBLNModel.compile(
-                wrapped_model,
-                prefill_compile_config,
-                example_inputs=prefill_example_inputs,
-                compile_context=context,
-            )
+            try:
+                original_linear = torch.nn.functional.linear
+                torch.nn.functional.linear = torch.ops.rbln_custom_ops.linear
+                wrapped_model.phase = "prefill"
+                compiled_prefill = RBLNModel.compile(
+                    wrapped_model,
+                    prefill_compile_config,
+                    example_inputs=prefill_example_inputs,
+                    compile_context=context,
+                )
 
-            wrapped_model.phase = "decode"
-            compiled_decoder = RBLNModel.compile(
-                wrapped_model,
-                dec_compile_config,
-                example_inputs=dec_example_inputs,
-                compile_context=context,
-            )
-            return {"prefill": compiled_prefill, "decoder": compiled_decoder}
+                wrapped_model.phase = "decode"
+                compiled_decoder = RBLNModel.compile(
+                    wrapped_model,
+                    dec_compile_config,
+                    example_inputs=dec_example_inputs,
+                    compile_context=context,
+                )
+                return {"prefill": compiled_prefill, "decoder": compiled_decoder}
+            finally:
+                torch.nn.functional.linear = original_linear
 
         return compile_model(quantize_config=quantize_config)
 

{optimum_rbln-0.7.4a0 → optimum_rbln-0.7.4a1}/src/optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py

@@ -38,8 +38,8 @@ class RBLNRuntimeEncoder(RBLNPytorchRuntime):
     mandatory_members = ["main_input_name"]
 
     def forward(self, *args: List[torch.Tensor], **kwargs: Dict[str, torch.Tensor]):
-        _ = super().forward(*args, **kwargs)
-        return BaseModelOutput(last_hidden_state=torch.tensor([1.0]))
+        output = super().forward(*args, **kwargs)
+        return BaseModelOutput(last_hidden_state=output)
 
 
 class RBLNRuntimeDecoder(RBLNPytorchRuntime):

optimum_rbln-0.7.4a1/src/optimum/rbln/transformers/models/time_series_transformers/__init__.py

@@ -0,0 +1,24 @@
+# Copyright 2024 Rebellions Inc.
+
+# 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.
+
+# Portions of this software are licensed under the Apache License,
+# Version 2.0. See the NOTICE file distributed with this work for
+# additional information regarding copyright ownership.
+
+# All other portions of this software, including proprietary code,
+# are the intellectual property of Rebellions Inc. and may not be
+# copied, modified, or distributed without prior written permission
+# from Rebellions Inc.
+
+from .modeling_time_series_transformers import RBLNTimeSeriesTransformerForPrediction