optimum-rbln 0.2.1a1__tar.gz → 0.2.1a2__tar.gz

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/.github/workflows/rbln_trigger_on_pr.yaml

@@ -9,7 +9,7 @@ env:
   REBEL_PYPI_ENDPOINT: ${{ vars.REBEL_PYPI_INTERNAL_ENDPOINT }}
   REBEL_PYPI_USERNAME: ${{ secrets.REBEL_PYPI_USERNAME }}
   REBEL_PYPI_PASSWORD: ${{ secrets.REBEL_PYPI_PASSWORD }}
-  REBEL_COMPILER_VERSION: 0.7.1.dev63+ge6c4fb68
+  REBEL_COMPILER_VERSION: 0.7.2.dev151+g19f099fd
 
 jobs:
   check-rebel-compiler-version:

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.2.1a1
+Version: 0.2.1a2
 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
@@ -28,7 +28,7 @@ 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==4.45.2
+Requires-Dist: transformers==4.48.3
 Description-Content-Type: text/markdown
 
 

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/pyproject.toml

@@ -32,7 +32,7 @@ dependencies = [
     "torchaudio<=2.5.1",
     "torchvision<=0.20.1",
     "accelerate>=1.0.1",
-    "transformers==4.45.2",
+    "transformers==4.48.3",
     "diffusers<=0.31.0",
     "packaging>=24.1",
 ]

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/__init__.py

@@ -71,6 +71,7 @@ _import_structure = {
         "RBLNRobertaForSequenceClassification",
         "RBLNRobertaForMaskedLM",
         "RBLNViTForImageClassification",
+        "RBLNBertForMaskedLM",
     ],
     "diffusers": [
         "RBLNStableDiffusionPipeline",
@@ -141,6 +142,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForVision2Seq,
         RBLNBartForConditionalGeneration,
         RBLNBartModel,
+        RBLNBertForMaskedLM,
         RBLNBertForQuestionAnswering,
         RBLNBertModel,
         RBLNCLIPTextModel,

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/__version__.py

@@ -12,5 +12,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.2.1a1'
+__version__ = version = '0.2.1a2'
 __version_tuple__ = version_tuple = (0, 2, 1)

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/ops/attn.py

@@ -152,16 +152,16 @@ def register_rbln_custom_attention_add_softmax():
         """
         return (
             q,
-            torch.empty(1, *kcache.shape[1:], device=kcache.device),
-            torch.empty(1, *vcache.shape[1:], device=vcache.device),
+            torch.empty(*kcache.shape, device=kcache.device),
+            torch.empty(*vcache.shape, device=vcache.device),
         )
 
     @register_fake("rbln_custom_ops::attn_decode_add_softmax")
     def attn_decode_add_softmax_abstract(q, k, v, m, kcache, vcache, seq, partition):
         return (
             q,
-            torch.empty(1, *kcache.shape[1:], device=kcache.device),
-            torch.empty(1, *vcache.shape[1:], device=vcache.device),
+            torch.empty(*kcache.shape, device=kcache.device),
+            torch.empty(*vcache.shape, device=vcache.device),
         )
 
     torch.library.define(

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/transformers/__init__.py

@@ -35,6 +35,7 @@ _import_structure = {
         "RBLNBartForConditionalGeneration",
         "RBLNBartModel",
         "RBLNBertModel",
+        "RBLNBertForMaskedLM",
         "RBLNBertForQuestionAnswering",
         "RBLNCLIPTextModel",
         "RBLNCLIPTextModelWithProjection",
@@ -92,6 +93,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForVision2Seq,
         RBLNBartForConditionalGeneration,
         RBLNBartModel,
+        RBLNBertForMaskedLM,
         RBLNBertForQuestionAnswering,
         RBLNBertModel,
         RBLNCLIPTextModel,

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/transformers/models/__init__.py

@@ -33,7 +33,7 @@ _import_structure = {
         "RBLNAutoModelForVision2Seq",
     ],
     "bart": ["RBLNBartForConditionalGeneration", "RBLNBartModel"],
-    "bert": ["RBLNBertModel", "RBLNBertForQuestionAnswering"],
+    "bert": ["RBLNBertModel", "RBLNBertForQuestionAnswering", "RBLNBertForMaskedLM"],
     "clip": ["RBLNCLIPTextModel", "RBLNCLIPTextModelWithProjection", "RBLNCLIPVisionModel"],
     "dpt": ["RBLNDPTForDepthEstimation"],
     "exaone": ["RBLNExaoneForCausalLM"],
@@ -67,7 +67,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForVision2Seq,
     )
     from .bart import RBLNBartForConditionalGeneration, RBLNBartModel
-    from .bert import RBLNBertForQuestionAnswering, RBLNBertModel
+    from .bert import RBLNBertForMaskedLM, RBLNBertForQuestionAnswering, RBLNBertModel
     from .clip import RBLNCLIPTextModel, RBLNCLIPTextModelWithProjection, RBLNCLIPVisionModel
     from .dpt import RBLNDPTForDepthEstimation
     from .exaone import RBLNExaoneForCausalLM

optimum_rbln-0.2.1a2/src/optimum/rbln/transformers/models/bert/__init__.py

@@ -0,0 +1,15 @@
+# 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 .modeling_bert import RBLNBertForMaskedLM, RBLNBertForQuestionAnswering, RBLNBertModel

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/transformers/models/bert/modeling_bert.py

@@ -20,7 +20,7 @@ from transformers import PretrainedConfig
 
 from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
-from ...modeling_generic import RBLNModelForQuestionAnswering
+from ...modeling_generic import RBLNModelForMaskedLM, RBLNModelForQuestionAnswering
 
 
 logger = logging.getLogger(__name__)
@@ -100,5 +100,9 @@ class RBLNBertModel(RBLNModel):
         return rbln_config
 
 
+class RBLNBertForMaskedLM(RBLNModelForMaskedLM):
+    rbln_model_input_names = ["input_ids", "attention_mask", "token_type_ids"]
+
+
 class RBLNBertForQuestionAnswering(RBLNModelForQuestionAnswering):
     rbln_model_input_names = ["input_ids", "attention_mask", "token_type_ids"]

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py

@@ -544,15 +544,19 @@ class DecoderOnlyAttention(nn.Module):
         super().__init__()
         self._original_mod = self_attn
         self.layer_idx = self_attn.layer_idx
-        self.num_heads = self._original_mod.num_heads
+        self.num_heads = getattr(self._original_mod, "num_heads", None) or getattr(
+            self._original_mod.config, "num_attention_heads"
+        )
         self.head_dim = self._original_mod.head_dim
         self._phase = "prefill"
         self.scale = torch.tensor(self.get_attn_scale())
 
         if hasattr(self._original_mod, "num_key_value_heads"):
             self.num_key_value_heads = self._original_mod.num_key_value_heads
+        elif hasattr(self._original_mod, "config") and hasattr(self._original_mod.config, "num_key_value_heads"):
+            self.num_key_value_heads = self._original_mod.config.num_key_value_heads
         else:
-            self.num_key_value_heads = self._original_mod.num_heads
+            self.num_key_value_heads = self.num_heads
 
         self.attention = self.get_attention()
         self.__post_init__()

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py

@@ -420,7 +420,7 @@ class Seq2SeqSelfAttention(nn.Module):
         pass
 
     def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int) -> torch.Tensor:
-        return tensor.view(bsz, 1, seq_len, 1, self.num_heads, self.head_dim).transpose(2, 4)
+        return tensor.view(bsz, seq_len, 1, self.num_heads, self.head_dim).transpose(1, 3)
 
     def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         """Projects input hidden states into query, key, and value representations.
@@ -450,38 +450,21 @@ class Seq2SeqSelfAttention(nn.Module):
         key_states = self._shape(key_states, -1, bsz)
         value_states = self._shape(value_states, -1, bsz)
 
-        all_key_states = []
-        all_value_states = []
-        all_attn_output = []
-        for b_idx in range(bsz):
-            query_state = query_states[b_idx]
-            key_state = key_states[b_idx]
-            value_state = value_states[b_idx]
-            attn_mask = attention_mask[b_idx].unsqueeze(0).unsqueeze(2)
-            past_key_state = past_key_value[0].view(bsz, self.num_heads, 1, -1, self.head_dim)
-            past_value_state = past_key_value[1].view(bsz, self.num_heads, 1, -1, self.head_dim)
-
-            attn_output, key_state, value_state = self.attn_decode(
-                query_state,
-                key_state,
-                value_state,
-                attn_mask,
-                past_key_state,
-                past_value_state,
-                cache_position[b_idx][0],
-                torch.tensor(1.0, dtype=torch.float32),  # scale
-            )
-
-            attn_output = attn_output.view(1, self.num_heads, -1, self.head_dim).transpose(1, 2)
-            attn_output = attn_output.reshape(1, -1, self.num_heads * self.head_dim)
-
-            all_key_states.append(key_state.squeeze(2))
-            all_value_states.append(value_state.squeeze(2))
-            all_attn_output.append(attn_output)
+        attn_output, key_states, value_states = self.attn_decode(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask.unsqueeze(
+                2
+            ),  # Unsqueeze group axis since CustomKernel expects it for group query attention
+            past_key_value[0].view(bsz, self.num_heads, 1, -1, self.head_dim),
+            past_key_value[1].view(bsz, self.num_heads, 1, -1, self.head_dim),
+            cache_position.squeeze(1),
+            torch.tensor(1.0, dtype=torch.float32),  # scale
+        )
 
-        key_states = torch.cat(all_key_states, dim=0)
-        value_states = torch.cat(all_value_states, dim=0)
-        attn_output = torch.cat(all_attn_output, dim=0)
+        attn_output = attn_output.view(bsz, self.num_heads, -1, self.head_dim).transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, -1, self.num_heads * self.head_dim)
 
         attn_output = self.out_proj(attn_output)
         present_key_value = (key_states, value_states)

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/src/optimum/rbln/transformers/models/t5/t5_architecture.py

@@ -147,6 +147,11 @@ class T5CrossAttention(nn.Module):
     def __init__(self, attn):
         super().__init__()
         self.attn = attn
+        self.q = attn.q
+        self.o = attn.o
+        self.n_heads = attn.n_heads
+        self.key_value_proj_dim = attn.key_value_proj_dim
+        self.inner_dim = attn.inner_dim
 
     def forward(
         self,
@@ -155,9 +160,27 @@ class T5CrossAttention(nn.Module):
         attention_mask: torch.Tensor = None,
         key_value_states: torch.Tensor = None,
     ):
-        return self.attn(
-            hidden_states=hidden_states,
-            past_key_value=past_key_value,
-            position_bias=attention_mask,
-            key_value_states=key_value_states,
-        )
+        batch_size = hidden_states.shape[0]
+
+        query_states = self.q(hidden_states)
+        query_states = query_states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
+
+        # reuse k,v, cross_attentions
+        key_states = past_key_value[0]
+        value_states = past_key_value[1]
+
+        # compute scores, equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9
+        scores = torch.matmul(query_states, key_states.transpose(3, 2))
+        scores += attention_mask
+
+        # (batch_size, n_heads, seq_length, key_length)
+        attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(batch_size, -1, self.inner_dim)
+        attn_output = self.o(attn_output)
+
+        outputs = (attn_output, past_key_value)
+
+        return outputs

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/tests/test_llm.py

@@ -47,7 +47,7 @@ class LLMTest:
 
         def get_inputs(self):
             inputs = self.tokenizer(self.PROMPT, return_tensors="pt")
-            inputs["max_new_tokens"] = 16
+            inputs["max_new_tokens"] = 20
             inputs["do_sample"] = False
             return inputs
 
@@ -62,7 +62,7 @@ class LLMTest:
 class TestQwen2Model(LLMTest.TestLLM):
     RBLN_CLASS = RBLNQwen2ForCausalLM
     HF_MODEL_ID = "Qwen/Qwen2-0.5B-Instruct"
-    EXPECTED_OUTPUT = " I am a 30-year-old woman who has been living with lupus"
+    EXPECTED_OUTPUT = " I am a 30-year-old woman who has been living with lupus for over 1"
     HF_CONFIG_KWARGS = {"max_position_embeddings": 1024}
 
 
@@ -70,7 +70,7 @@ class TestLlamaForCausalLM(LLMTest.TestLLM):
     RBLN_CLASS = RBLNLlamaForCausalLM
     HF_MODEL_ID = "afmck/testing-llama-tiny"
     TEST_LEVEL = TestLevel.ESSENTIAL
-    EXPECTED_OUTPUT = "reress makefable R���� noethetssh"
+    EXPECTED_OUTPUT = "reress makefable R���� noethetsshss rechoolso�"
     HF_CONFIG_KWARGS = {"num_hidden_layers": 1, "max_position_embeddings": 1024}
 
     def get_inputs(self):
@@ -83,7 +83,7 @@ class TestLlamaForCausalLM_Flash(LLMTest.TestLLM):
     RBLN_CLASS = RBLNLlamaForCausalLM
     HF_MODEL_ID = "afmck/testing-llama-tiny"
     TEST_LEVEL = TestLevel.ESSENTIAL
-    EXPECTED_OUTPUT = "reress makefable R���� noethetssh"
+    EXPECTED_OUTPUT = "reress makefable R���� noethetsshss rechoolso�"
     HF_CONFIG_KWARGS = {"num_hidden_layers": 1, "max_position_embeddings": 8192}
     RBLN_CLASS_KWARGS = {"rbln_config": {"attn_impl": "flash_attn", "kvcache_partition_len": 4096}}
 
@@ -96,7 +96,9 @@ class TestLlamaForCausalLM_Flash(LLMTest.TestLLM):
 class TestGPT2LMHeadModel(LLMTest.TestLLM):
     RBLN_CLASS = RBLNGPT2LMHeadModel
     # TEST_LEVEL = TestLevel.FULL
-    EXPECTED_OUTPUT = " What kind kind kind kind kind kind kind kind kind kind kind kind kind kind kind"
+    EXPECTED_OUTPUT = (
+        " What kind kind kind kind kind kind kind kind kind kind kind kind kind kind kind kind kind kind kind"
+    )
     HF_MODEL_ID = "openai-community/gpt2"
     HF_CONFIG_KWARGS = {"n_layer": 1, "max_position_embeddings": 1024}
 
@@ -106,7 +108,7 @@ class TestPhiForCausalLM(LLMTest.TestLLM):
 
     # HF_MODEL_ID = "hf-internal-testing/tiny-random-PhiForCausalLM"
     HF_MODEL_ID = "microsoft/phi-2"
-    EXPECTED_OUTPUT = "\nAnswer: Theorettebrates']['<<<urlskolegate"
+    EXPECTED_OUTPUT = "\nAnswer: Theorettebrates']['<<<urlskolegateezzingrill"
     HF_CONFIG_KWARGS = {"num_hidden_layers": 1, "max_position_embeddings": 1024, "trust_remote_code": True}
 
 
@@ -114,7 +116,7 @@ class TestExaoneForCausalLM(LLMTest.TestLLM):
     RBLN_CLASS = RBLNExaoneForCausalLM
     # HF_MODEL_ID = "katuni4ka/tiny-random-exaone"
     HF_MODEL_ID = "LGAI-EXAONE/EXAONE-3.5-2.4B-Instruct"
-    EXPECTED_OUTPUT = "????????????????"
+    EXPECTED_OUTPUT = "????????????????????"
     HF_CONFIG_KWARGS = {"num_hidden_layers": 1, "max_position_embeddings": 1024, "trust_remote_code": True}
 
 
@@ -134,7 +136,7 @@ class TestT5Model(LLMTest.TestLLM):
         inputs = self.tokenizer(
             self.PROMPT, padding="max_length", max_length=512, truncation=True, return_tensors="pt"
         )
-        inputs["max_new_tokens"] = 16
+        inputs["max_new_tokens"] = 20
         inputs["do_sample"] = False
         inputs["num_beams"] = 1
         return inputs
@@ -158,16 +160,14 @@ class TestBartModel(LLMTest.TestLLM):
     }
     RBLN_CLASS_KWARGS = {"rbln_config": {"enc_max_seq_len": 512, "dec_max_seq_len": 512}}
     PROMPT = "summarize: studies have shown that owning a dog is good for you"
-    EXPECTED_OUTPUT = (
-        "InsteadInsteadInsteadHoweverHoweverHoweverAlthoughAlthoughAlthoughWhileWhileWhileAlthoughAlthough"
-    )
+    EXPECTED_OUTPUT = "InsteadInsteadInsteadHoweverHoweverHoweverAlthoughAlthoughAlthoughWhileWhileWhileAlthoughAlthoughHoweverHoweverManyMany"
     TEST_LEVEL = TestLevel.ESSENTIAL
 
     def get_inputs(self):
         inputs = self.tokenizer(
             self.PROMPT, padding="max_length", max_length=512, truncation=True, return_tensors="pt"
         )
-        inputs["max_new_tokens"] = 16
+        inputs["max_new_tokens"] = 20
         inputs["do_sample"] = False
         inputs["num_beams"] = 1
         return inputs

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/tests/test_transformers.py

@@ -7,6 +7,7 @@ from transformers import T5EncoderModel
 
 from optimum.rbln import (
     RBLNASTForAudioClassification,
+    RBLNBertForMaskedLM,
     RBLNBertForQuestionAnswering,
     RBLNCLIPTextModel,
     RBLNDPTForDepthEstimation,
@@ -23,6 +24,7 @@ from optimum.rbln.transformers.models.auto.modeling_auto import (
     RBLNAutoModelForCTC,
     RBLNAutoModelForDepthEstimation,
     RBLNAutoModelForImageClassification,
+    RBLNAutoModelForMaskedLM,
     RBLNAutoModelForQuestionAnswering,
     RBLNAutoModelForSequenceClassification,
     RBLNAutoModelForSpeechSeq2Seq,
@@ -98,6 +100,17 @@ class TestBertModel(BaseTest.TestModel):
     }
 
 
+class TestBertForMaskedLM(BaseTest.TestModel):
+    RBLN_AUTO_CLASS = RBLNAutoModelForMaskedLM
+    RBLN_CLASS = RBLNBertForMaskedLM
+    HF_MODEL_ID = "hf-internal-testing/tiny-random-BertForMaskedLM"
+    GENERATION_KWARGS = {
+        "input_ids": RANDOM_INPUT_IDS,
+        "attention_mask": RANDOM_ATTN_MASK,
+        "token_type_ids": RANDOM_TOKEN_TYPE_IDS,
+    }
+
+
 class TestDPTModel(BaseTest.TestModel):
     RBLN_AUTO_CLASS = RBLNAutoModelForDepthEstimation
     RBLN_CLASS = RBLNDPTForDepthEstimation

{optimum_rbln-0.2.1a1 → optimum_rbln-0.2.1a2}/uv.lock

@@ -996,7 +996,7 @@ requires-dist = [
     { name = "torchaudio", marker = "sys_platform == 'darwin'", specifier = "<=2.5.1" },
     { name = "torchvision", marker = "sys_platform != 'darwin'", specifier = "<=0.20.1", index = "https://download.pytorch.org/whl/cpu" },
     { name = "torchvision", marker = "sys_platform == 'darwin'", specifier = "<=0.20.1" },
-    { name = "transformers", specifier = "==4.45.2" },
+    { name = "transformers", specifier = "==4.48.3" },
 ]
 
 [package.metadata.requires-dev]
@@ -1637,55 +1637,23 @@ wheels = [
 
 [[package]]
 name = "tokenizers"
-version = "0.20.3"
+version = "0.21.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "huggingface-hub", marker = "sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/da/25/b1681c1c30ea3ea6e584ae3fffd552430b12faa599b558c4c4783f56d7ff/tokenizers-0.20.3.tar.gz", hash = "sha256:2278b34c5d0dd78e087e1ca7f9b1dcbf129d80211afa645f214bd6e051037539", size = 340513 }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/e3/e8/0e9f81a09ab79f409eabfd99391ca519e315496694671bebca24c3e90448/tokenizers-0.20.3-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f128d5da1202b78fa0a10d8d938610472487da01b57098d48f7e944384362514", size = 2892896 },
-    { url = "https://files.pythonhosted.org/packages/b0/72/15fdbc149e05005e99431ecd471807db2241983deafe1e704020f608f40e/tokenizers-0.20.3-cp310-cp310-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:79c4121a2e9433ad7ef0769b9ca1f7dd7fa4c0cd501763d0a030afcbc6384481", size = 2802785 },
-    { url = "https://files.pythonhosted.org/packages/26/44/1f8aea48f9bb117d966b7272484671b33a509f6217a8e8544d79442c90db/tokenizers-0.20.3-cp310-cp310-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:b7850fde24197fe5cd6556e2fdba53a6d3bae67c531ea33a3d7c420b90904141", size = 3086060 },
-    { url = "https://files.pythonhosted.org/packages/2e/83/82ba40da99870b3a0b801cffaf4f099f088a84c7e07d32cc6ca751ce08e6/tokenizers-0.20.3-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:b357970c095dc134978a68c67d845a1e3803ab7c4fbb39195bde914e7e13cf8b", size = 3096760 },
-    { url = "https://files.pythonhosted.org/packages/f3/46/7a025404201d937f86548928616c0a164308aa3998e546efdf798bf5ee9c/tokenizers-0.20.3-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:a333d878c4970b72d6c07848b90c05f6b045cf9273fc2bc04a27211721ad6118", size = 3380165 },
-    { url = "https://files.pythonhosted.org/packages/aa/49/15fae66ac62e49255eeedbb7f4127564b2c3f3aef2009913f525732d1a08/tokenizers-0.20.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1fd9fee817f655a8f50049f685e224828abfadd436b8ff67979fc1d054b435f1", size = 2994038 },
-    { url = "https://files.pythonhosted.org/packages/f4/64/693afc9ba2393c2eed85c02bacb44762f06a29f0d1a5591fa5b40b39c0a2/tokenizers-0.20.3-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:9e7816808b402129393a435ea2a509679b41246175d6e5e9f25b8692bfaa272b", size = 8977285 },
-    { url = "https://files.pythonhosted.org/packages/be/7e/6126c18694310fe07970717929e889898767c41fbdd95b9078e8aec0f9ef/tokenizers-0.20.3-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:ba96367db9d8a730d3a1d5996b4b7babb846c3994b8ef14008cd8660f55db59d", size = 9294890 },
-    { url = "https://files.pythonhosted.org/packages/46/54/033b5b2ba0c3ae01e026c6f7ced147d41a2fa1c573d00a66cb97f6d7f9b3/tokenizers-0.20.3-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ef820880d5e4e8484e2fa54ff8d297bb32519eaa7815694dc835ace9130a3eea", size = 2892476 },
-    { url = "https://files.pythonhosted.org/packages/e6/b0/cc369fb3297d61f3311cab523d16d48c869dc2f0ba32985dbf03ff811041/tokenizers-0.20.3-cp311-cp311-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:67ef4dcb8841a4988cd00dd288fb95dfc8e22ed021f01f37348fd51c2b055ba9", size = 2802775 },
-    { url = "https://files.pythonhosted.org/packages/1a/74/62ad983e8ea6a63e04ed9c5be0b605056bf8aac2f0125f9b5e0b3e2b89fa/tokenizers-0.20.3-cp311-cp311-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ff1ef8bd47a02b0dc191688ccb4da53600df5d4c9a05a4b68e1e3de4823e78eb", size = 3086138 },
-    { url = "https://files.pythonhosted.org/packages/6b/ac/4637ba619db25094998523f9e6f5b456e1db1f8faa770a3d925d436db0c3/tokenizers-0.20.3-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:444d188186eab3148baf0615b522461b41b1f0cd58cd57b862ec94b6ac9780f1", size = 3098076 },
-    { url = "https://files.pythonhosted.org/packages/58/ce/9793f2dc2ce529369807c9c74e42722b05034af411d60f5730b720388c7d/tokenizers-0.20.3-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:37c04c032c1442740b2c2d925f1857885c07619224a533123ac7ea71ca5713da", size = 3379650 },
-    { url = "https://files.pythonhosted.org/packages/50/f6/2841de926bc4118af996eaf0bdf0ea5b012245044766ffc0347e6c968e63/tokenizers-0.20.3-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:453c7769d22231960ee0e883d1005c93c68015025a5e4ae56275406d94a3c907", size = 2994005 },
-    { url = "https://files.pythonhosted.org/packages/a3/b2/00915c4fed08e9505d37cf6eaab45b12b4bff8f6719d459abcb9ead86a4b/tokenizers-0.20.3-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:4bb31f7b2847e439766aaa9cc7bccf7ac7088052deccdb2275c952d96f691c6a", size = 8977488 },
-    { url = "https://files.pythonhosted.org/packages/e9/ac/1c069e7808181ff57bcf2d39e9b6fbee9133a55410e6ebdaa89f67c32e83/tokenizers-0.20.3-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:843729bf0f991b29655a069a2ff58a4c24375a553c70955e15e37a90dd4e045c", size = 9294935 },
-    { url = "https://files.pythonhosted.org/packages/27/37/d108df55daf4f0fcf1f58554692ff71687c273d870a34693066f0847be96/tokenizers-0.20.3-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:abe4e08c7d0cd6154c795deb5bf81d2122f36daf075e0c12a8b050d824ef0a64", size = 2898389 },
-    { url = "https://files.pythonhosted.org/packages/b2/27/32f29da16d28f59472fa7fb38e7782069748c7e9ab9854522db20341624c/tokenizers-0.20.3-cp312-cp312-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:ca94fc1b73b3883c98f0c88c77700b13d55b49f1071dfd57df2b06f3ff7afd64", size = 2795866 },
-    { url = "https://files.pythonhosted.org/packages/29/4e/8a9a3c89e128c4a40f247b501c10279d2d7ade685953407c4d94c8c0f7a7/tokenizers-0.20.3-cp312-cp312-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ef279c7e239f95c8bdd6ff319d9870f30f0d24915b04895f55b1adcf96d6c60d", size = 3085446 },
-    { url = "https://files.pythonhosted.org/packages/b4/3b/a2a7962c496ebcd95860ca99e423254f760f382cd4bd376f8895783afaf5/tokenizers-0.20.3-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:16384073973f6ccbde9852157a4fdfe632bb65208139c9d0c0bd0176a71fd67f", size = 3094378 },
-    { url = "https://files.pythonhosted.org/packages/1f/f4/a8a33f0192a1629a3bd0afcad17d4d221bbf9276da4b95d226364208d5eb/tokenizers-0.20.3-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:312d522caeb8a1a42ebdec87118d99b22667782b67898a76c963c058a7e41d4f", size = 3385755 },
-    { url = "https://files.pythonhosted.org/packages/9e/65/c83cb3545a65a9eaa2e13b22c93d5e00bd7624b354a44adbdc93d5d9bd91/tokenizers-0.20.3-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f2b7cb962564785a83dafbba0144ecb7f579f1d57d8c406cdaa7f32fe32f18ad", size = 2997679 },
-    { url = "https://files.pythonhosted.org/packages/55/e9/a80d4e592307688a67c7c59ab77e03687b6a8bd92eb5db763a2c80f93f57/tokenizers-0.20.3-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:124c5882ebb88dadae1fc788a582299fcd3a8bd84fc3e260b9918cf28b8751f5", size = 8989296 },
-    { url = "https://files.pythonhosted.org/packages/90/af/60c957af8d2244321124e893828f1a4817cde1a2d08d09d423b73f19bd2f/tokenizers-0.20.3-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:2b6e54e71f84c4202111a489879005cb14b92616a87417f6c102c833af961ea2", size = 9303621 },
-    { url = "https://files.pythonhosted.org/packages/b0/39/073836c1d73e63268b1c67a682a8ba23e2688a43e737166be45ab8243701/tokenizers-0.20.3-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7eb2fb1c432f5746b22f8a7f09fc18c4156cb0031c77f53cb19379d82d43297a", size = 2893676 },
-    { url = "https://files.pythonhosted.org/packages/c1/d9/b9ff819c3df4bc73ad93629804f7b85321a78bc2da4f54fb774a90e995c6/tokenizers-0.20.3-cp39-cp39-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:bfa8d029bb156181b006643309d6b673615a24e4ed24cf03aa191d599b996f51", size = 2804173 },
-    { url = "https://files.pythonhosted.org/packages/3e/d5/6b2b519ba2d9a6d3435f22918f0ad5850c40cf5357f6d989e6d68ef40fb9/tokenizers-0.20.3-cp39-cp39-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:6f90549622de3bf476ad9f1dd6f3f952ec3ed6ab8615ae88ef060d0c5bfad55d", size = 3086866 },
-    { url = "https://files.pythonhosted.org/packages/01/e1/d96e90ef872dd9b3a4b7a78874411f1c48476019f95a87a2cfd54c470a57/tokenizers-0.20.3-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a1d469c74eebf5c43fd61cd9b030e271d17198edd7bd45392e03a3c091d7d6d4", size = 3099004 },
-    { url = "https://files.pythonhosted.org/packages/0c/6a/a94248dc5915907e18d55c9739cd018f5aeb4146f198622f45f9748dcb9f/tokenizers-0.20.3-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:bee8f53b2594749f4460d53253bae55d718f04e9b633efa0f5df8938bd98e4f0", size = 3381574 },
-    { url = "https://files.pythonhosted.org/packages/29/9e/c95f8821d6bc93eba7c5db95e6299c009db523d1c646da8563b42ad892c4/tokenizers-0.20.3-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:938441babf3e5720e4459e306ef2809fb267680df9d1ff2873458b22aef60248", size = 2994953 },
-    { url = "https://files.pythonhosted.org/packages/95/ff/01fdcf9a77776730baf63a9f66adf75c3aa4bdb1bdc77c7d1a3e03b2a25e/tokenizers-0.20.3-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:7310ab23d7b0caebecc0e8be11a1146f320f5f07284000f6ea54793e83de1b75", size = 8977698 },
-    { url = "https://files.pythonhosted.org/packages/ef/2d/8b823741c64e9726b82076fa09f6d66285b61bd2c77e109871415b1ed9e2/tokenizers-0.20.3-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:16121eb030a2b13094cfec936b0c12e8b4063c5f839591ea7d0212336d8f9921", size = 9295649 },
-    { url = "https://files.pythonhosted.org/packages/8e/c1/6af62ef61316f33ecf785bbb2bee4292f34ea62b491d4480ad9b09acf6b6/tokenizers-0.20.3-pp310-pypy310_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:39270a7050deaf50f7caff4c532c01b3c48f6608d42b3eacdebdc6795478c8df", size = 2897936 },
-    { url = "https://files.pythonhosted.org/packages/9a/0b/c076b2ff3ee6dc70c805181fbe325668b89cfee856f8dfa24cc9aa293c84/tokenizers-0.20.3-pp310-pypy310_pp73-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:e005466632b1c5d2d2120f6de8aa768cc9d36cd1ab7d51d0c27a114c91a1e6ee", size = 3082688 },
-    { url = "https://files.pythonhosted.org/packages/0a/60/56510124933136c2e90879e1c81603cfa753ae5a87830e3ef95056b20d8f/tokenizers-0.20.3-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a07962340b36189b6c8feda552ea1bfeee6cf067ff922a1d7760662c2ee229e5", size = 2998924 },
-    { url = "https://files.pythonhosted.org/packages/68/60/4107b618b7b9155cb34ad2e0fc90946b7e71f041b642122fb6314f660688/tokenizers-0.20.3-pp310-pypy310_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:55046ad3dd5f2b3c67501fcc8c9cbe3e901d8355f08a3b745e9b57894855f85b", size = 8989514 },
-    { url = "https://files.pythonhosted.org/packages/e8/bd/48475818e614b73316baf37ac1e4e51b578bbdf58651812d7e55f43b88d8/tokenizers-0.20.3-pp310-pypy310_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:efcf0eb939988b627558aaf2b9dc3e56d759cad2e0cfa04fcab378e4b48fc4fd", size = 9303476 },
-    { url = "https://files.pythonhosted.org/packages/ce/32/37ff2ced2c169c2e7586fcd51314f59d02c60fd2eeafea527c2f9d1bb512/tokenizers-0.20.3-pp39-pypy39_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a292392f24ab9abac5cfa8197e5a6208f2e43723420217e1ceba0b4ec77816ac", size = 2897613 },
-    { url = "https://files.pythonhosted.org/packages/79/e4/fdd7ad2aedaa4a3f148aa28670bf0b0856211a3fec3e6554ed6ceec9a928/tokenizers-0.20.3-pp39-pypy39_pp73-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:8dcd91f4e60f62b20d83a87a84fe062035a1e3ff49a8c2bbdeb2d441c8e311f4", size = 3085434 },
-    { url = "https://files.pythonhosted.org/packages/e0/b8/479ab7349faf1da001b861ea521055ad18a34a9b1053079e0c9b5c476f50/tokenizers-0.20.3-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:900991a2b8ee35961b1095db7e265342e0e42a84c1a594823d5ee9f8fb791958", size = 2998651 },
-    { url = "https://files.pythonhosted.org/packages/6b/7f/3a1d5ded5f841764d67aa4c6e2e4b40d9dac5fbd2df135bccc58284a6917/tokenizers-0.20.3-pp39-pypy39_pp73-musllinux_1_1_aarch64.whl", hash = "sha256:5a8d8261ca2133d4f98aa9627c748189502b3787537ba3d7e2beb4f7cfc5d627", size = 8989010 },
-    { url = "https://files.pythonhosted.org/packages/2b/a7/e0b5d5fea8cb69afdbab3c0e0cc3a02b5dd888ce0f933312f7c0ca6b017e/tokenizers-0.20.3-pp39-pypy39_pp73-musllinux_1_1_x86_64.whl", hash = "sha256:c4fd4d71e6deb6ddf99d8d0eab87d1d16f635898906e631914a9bae8ae9f2cfb", size = 9303287 },
+sdist = { url = "https://files.pythonhosted.org/packages/20/41/c2be10975ca37f6ec40d7abd7e98a5213bb04f284b869c1a24e6504fd94d/tokenizers-0.21.0.tar.gz", hash = "sha256:ee0894bf311b75b0c03079f33859ae4b2334d675d4e93f5a4132e1eae2834fe4", size = 343021 }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f7/14/83429177c19364df27d22bc096d4c2e431e0ba43e56c525434f1f9b0fd00/tokenizers-0.21.0-cp39-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6b177fb54c4702ef611de0c069d9169f0004233890e0c4c5bd5508ae05abf193", size = 2903304 },
+    { url = "https://files.pythonhosted.org/packages/7e/db/3433eab42347e0dc5452d8fcc8da03f638c9accffefe5a7c78146666964a/tokenizers-0.21.0-cp39-abi3-manylinux_2_17_armv7l.manylinux2014_armv7l.whl", hash = "sha256:6b43779a269f4629bebb114e19c3fca0223296ae9fea8bb9a7a6c6fb0657ff8e", size = 2804378 },
+    { url = "https://files.pythonhosted.org/packages/57/8b/7da5e6f89736c2ade02816b4733983fca1c226b0c42980b1ae9dc8fcf5cc/tokenizers-0.21.0-cp39-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:9aeb255802be90acfd363626753fda0064a8df06031012fe7d52fd9a905eb00e", size = 3095488 },
+    { url = "https://files.pythonhosted.org/packages/4d/f6/5ed6711093dc2c04a4e03f6461798b12669bc5a17c8be7cce1240e0b5ce8/tokenizers-0.21.0-cp39-abi3-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d8b09dbeb7a8d73ee204a70f94fc06ea0f17dcf0844f16102b9f414f0b7463ba", size = 3121410 },
+    { url = "https://files.pythonhosted.org/packages/81/42/07600892d48950c5e80505b81411044a2d969368cdc0d929b1c847bf6697/tokenizers-0.21.0-cp39-abi3-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:400832c0904f77ce87c40f1a8a27493071282f785724ae62144324f171377273", size = 3388821 },
+    { url = "https://files.pythonhosted.org/packages/22/06/69d7ce374747edaf1695a4f61b83570d91cc8bbfc51ccfecf76f56ab4aac/tokenizers-0.21.0-cp39-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e84ca973b3a96894d1707e189c14a774b701596d579ffc7e69debfc036a61a04", size = 3008868 },
+    { url = "https://files.pythonhosted.org/packages/c8/69/54a0aee4d576045b49a0eb8bffdc495634309c823bf886042e6f46b80058/tokenizers-0.21.0-cp39-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:eb7202d231b273c34ec67767378cd04c767e967fda12d4a9e36208a34e2f137e", size = 8975831 },
+    { url = "https://files.pythonhosted.org/packages/f7/f3/b776061e4f3ebf2905ba1a25d90380aafd10c02d406437a8ba22d1724d76/tokenizers-0.21.0-cp39-abi3-musllinux_1_2_armv7l.whl", hash = "sha256:089d56db6782a73a27fd8abf3ba21779f5b85d4a9f35e3b493c7bbcbbf0d539b", size = 8920746 },
+    { url = "https://files.pythonhosted.org/packages/d8/ee/ce83d5ec8b6844ad4c3ecfe3333d58ecc1adc61f0878b323a15355bcab24/tokenizers-0.21.0-cp39-abi3-musllinux_1_2_i686.whl", hash = "sha256:c87ca3dc48b9b1222d984b6b7490355a6fdb411a2d810f6f05977258400ddb74", size = 9161814 },
+    { url = "https://files.pythonhosted.org/packages/18/07/3e88e65c0ed28fa93aa0c4d264988428eef3df2764c3126dc83e243cb36f/tokenizers-0.21.0-cp39-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:4145505a973116f91bc3ac45988a92e618a6f83eb458f49ea0790df94ee243ff", size = 9357138 },
 ]
 
 [[package]]
@@ -1869,7 +1837,7 @@ wheels = [
 
 [[package]]
 name = "transformers"
-version = "4.45.2"
+version = "4.48.3"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "filelock", marker = "sys_platform == 'linux'" },
@@ -1884,9 +1852,9 @@ dependencies = [
     { name = "tokenizers", marker = "sys_platform == 'linux'" },
     { name = "tqdm", marker = "sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/4b/4c/3862b2dd6cdf83b187897bd351da0f7fb74d0df642b03c6f5d06353a3ca0/transformers-4.45.2.tar.gz", hash = "sha256:72bc390f6b203892561f05f86bbfaa0e234aab8e927a83e62b9d92ea7e3ae101", size = 8478357 }
+sdist = { url = "https://files.pythonhosted.org/packages/e3/82/cebeb7af5e64440f1638f18c4ed0f89156d0eeaa6290d98da8ca93ac3872/transformers-4.48.3.tar.gz", hash = "sha256:a5e8f1e9a6430aa78215836be70cecd3f872d99eeda300f41ad6cc841724afdb", size = 8373458 }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/f9/9d/030cc1b3e88172967e22ee1d012e0d5e0384eb70d2a098d1669d549aea29/transformers-4.45.2-py3-none-any.whl", hash = "sha256:c551b33660cfc815bae1f9f097ecfd1e65be623f13c6ee0dda372bd881460210", size = 9881312 },
+    { url = "https://files.pythonhosted.org/packages/b6/1a/efeecb8d83705f2f4beac98d46f2148c95ecd7babfb31b5c0f1e7017e83d/transformers-4.48.3-py3-none-any.whl", hash = "sha256:78697f990f5ef350c23b46bf86d5081ce96b49479ab180b2de7687267de8fd36", size = 9669412 },
 ]
 
 [[package]]

optimum_rbln-0.2.1a1/src/optimum/rbln/transformers/models/bert/__init__.py

@@ -1,15 +0,0 @@
-# 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 .modeling_bert import RBLNBertForQuestionAnswering, RBLNBertModel