optimum-rbln 0.1.11__py3-none-any.whl → 0.1.13__py3-none-any.whl

optimum/rbln/transformers/models/midm/midm_architecture.py

@@ -58,23 +58,12 @@ class MidmLMHeadModelWrapper(torch.nn.Module):
         self.model = model.transformer
         self.lm_head = model.lm_head
         self.config = model.config
-        self.head_dim = self.config.n_embd // self.config.n_head
-        self.max_position_embeddings = (
-            self.config.max_position_embeddings if max_seq_len > self.config.max_position_embeddings else max_seq_len
-        )
         self.max_seq_len = max_seq_len
-        self.rotary_dim = int(
-            model.config.hidden_size // model.config.num_attention_heads * model.config.rotary_percentage
-        )
-        self.rotary_emb = self._init_rope()
 
-    def _init_rope(self):
-        """Initializes the Rotary Position Embeddings."""
-        rotary_emb = RotaryEmbedding(
-            self.rotary_dim,
-            max_position_embeddings=self.max_position_embeddings,
-        )
-        return rotary_emb
+        self.config.partial_rotary_factor = model.config.rotary_percentage
+        self.config.head_dim = self.config.n_embd // self.config.n_head
+        self.config.rope_theta = 10000
+        self.rotary_emb = RotaryEmbedding(config=self.config, max_seq_len_cached=max_seq_len)
 
     def forward(
         self,

optimum/rbln/transformers/models/midm/modeling_midm.py

@@ -21,11 +21,7 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
-import inspect
-import logging
-from typing import TYPE_CHECKING, Any, Callable
-
-from ....modeling_config import RBLNConfig
+from ....utils import logging
 from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM
 from .hf_hub_cached.modeling_midm import MidmLMHeadModel
 from .midm_architecture import (
@@ -33,11 +29,7 @@ from .midm_architecture import (
 )
 
 
-logger = logging.getLogger(__name__)
-if TYPE_CHECKING:
-    from transformers import (
-        PreTrainedModel,
-    )
+logger = logging.get_logger(__name__)
 
 
 class RBLNMidmLMHeadModel(RBLNDecoderOnlyModelForCausalLM):
@@ -54,22 +46,10 @@ class RBLNMidmLMHeadModel(RBLNDecoderOnlyModelForCausalLM):
 
     """
 
-    @classmethod
-    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
-        rbln_max_seq_len = rbln_config.model_cfg["max_seq_len"]
-        return MidmLMHeadModelWrapper(model, rbln_max_seq_len).eval()
-
-    def __getattr__(self, __name: str) -> Any:
-        """This is the key method to implement RBLN-Midm.
+    _decoder_wrapper_cls = MidmLMHeadModelWrapper
+    _original_cls = MidmLMHeadModel
 
-        Returns:
-            Any: Midm's corresponding method
-        """
-
-        def redirect(func):
-            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
-
-        val = getattr(MidmLMHeadModel, __name)
-        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
-            return redirect(val)
-        return val
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        kwargs.setdefault("trust_remote_code", True)
+        return super().from_pretrained(*args, **kwargs)

optimum/rbln/transformers/models/mistral/modeling_mistral.py

@@ -21,29 +21,18 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
-import inspect
-import logging
-from typing import TYPE_CHECKING, Any, Callable
-
-from transformers import MistralForCausalLM
-
+from ....utils import logging
 from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM
 from .mistral_architecture import MistralForCausalLMWrapper
 
 
-if TYPE_CHECKING:
-    from transformers import PreTrainedModel
-
-    from ....modeling_config import RBLNConfig
-
-
-logger = logging.getLogger(__name__)
+logger = logging.get_logger(__name__)
 
 
 class RBLNMistralForCausalLM(RBLNDecoderOnlyModelForCausalLM):
     """
     The Llama Model transformer with a language modeling head (linear layer) on top.
-    This model inherits from [`RBLNMultiModel`]. Check the superclass documentation for the generic methods the library implements for all its models.
+    This model inherits from [`RBLNDecoderOnlyModelForCausalLM`]. Check the superclass documentation for the generic methods the library implements for all its models.
 
     A class to convert and run pre-trained transformers based LlamaForCausalLM model on RBLN devices.
     It implements the methods to convert a pre-trained transformers LlamaForCausalLM model into a RBLN transformer model by:
@@ -51,18 +40,4 @@ class RBLNMistralForCausalLM(RBLNDecoderOnlyModelForCausalLM):
     - compiling the resulting graph using the RBLN compiler.
     """
 
-    @classmethod
-    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
-        rbln_max_seq_len = rbln_config.model_cfg["max_seq_len"]
-        return MistralForCausalLMWrapper(model, rbln_max_seq_len).eval()
-
-    def __getattr__(self, __name: str) -> Any:
-        def redirect(func):
-            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
-
-        val = getattr(MistralForCausalLM, __name)
-
-        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
-            return redirect(val)
-
-        return val
+    _decoder_wrapper_cls = MistralForCausalLMWrapper

optimum/rbln/transformers/models/phi/modeling_phi.py

@@ -21,28 +21,18 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
-import inspect
-import logging
-from typing import TYPE_CHECKING, Any, Callable
-
-from transformers import PhiForCausalLM
-
-from ..decoderonly import RBLNDecoderOnlyModelForCausalLM
+from ....utils import logging
+from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM
 from .phi_architecture import PhiWrapper
 
 
-if TYPE_CHECKING:
-    from transformers import PreTrainedModel
-
-    from ....modeling_config import RBLNConfig
-
-logger = logging.getLogger(__name__)
+logger = logging.get_logger(__name__)
 
 
 class RBLNPhiForCausalLM(RBLNDecoderOnlyModelForCausalLM):
     """
     The Phi Model transformer with a language modeling head (linear layer) on top.
-    This model inherits from [`RBLNMultiModel`]. Check the superclass documentation for the generic methods the library implements for all its models.
+    This model inherits from [`RBLNDecoderOnlyModelForCausalLM`]. Check the superclass documentation for the generic methods the library implements for all its models.
 
     A class to convert and run pre-trained transformers based PhiForCausalLM model on RBLN devices.
     It implements the methods to convert a pre-trained transformers PhiForCausalLM model into a RBLN transformer model by:
@@ -50,20 +40,4 @@ class RBLNPhiForCausalLM(RBLNDecoderOnlyModelForCausalLM):
     - compiling the resulting graph using the RBLN compiler.
     """
 
-    @classmethod
-    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
-        rbln_max_seq_len = rbln_config.model_cfg["max_seq_len"]
-        return PhiWrapper(model, rbln_max_seq_len).eval()
-
-    def __getattr__(self, __name: str) -> Any:
-        def redirect(func):
-            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
-
-        val = getattr(PhiForCausalLM, __name)
-
-        if isinstance(val, Callable) and "self" in set(
-            inspect.signature(val).parameters
-        ):
-            return redirect(val)
-
-        return val
+    _decoder_wrapper_cls = PhiWrapper

optimum/rbln/transformers/models/phi/phi_architecture.py

@@ -33,46 +33,12 @@ from transformers.modeling_outputs import (
 from ...cache_utils import RebelDynamicCache
 from ..decoderonly import (
     DecoderOnlyWrapper,
-    DynamicNTKScalingRotaryEmbedding,
-    LinearScalingRotaryEmbedding,
-    RotaryEmbedding,
     apply_rotary_pos_emb,
     slice_and_unsqueeze_cos_sin,
 )
 
 
 class PhiWrapper(DecoderOnlyWrapper):
-    def _init_rope(self):
-        if self.rope_scaling is None:
-            rotary_emb = RotaryEmbedding(
-                int(self.config.partial_rotary_factor * self.head_dim),
-                max_position_embeddings=self.max_position_embeddings,
-                base=self.config.rope_theta,
-            )
-        else:
-            scaling_type = self.rope_scaling["type"]
-            scaling_factor = self.rope_scaling["factor"]
-            if scaling_type == "linear":
-                rotary_emb = LinearScalingRotaryEmbedding(
-                    int(self.config.partial_rotary_factor * self.head_dim),
-                    max_position_embeddings=self.max_position_embeddings,
-                    scaling_factor=scaling_factor,
-                    base=self.config.rope_theta,
-                    max_seq_len=self.max_seq_len,
-                )
-            elif scaling_type == "dynamic":
-                rotary_emb = DynamicNTKScalingRotaryEmbedding(
-                    int(self.config.partial_rotary_factor * self.head_dim),
-                    max_position_embeddings=self.max_position_embeddings,
-                    scaling_factor=scaling_factor,
-                    base=self.config.rope_theta,
-                    max_seq_len=self.max_seq_len,
-                )
-            else:
-                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
-
-        return rotary_emb
-
     def get_forward_dict(self):
         forward_dict = {}
         forward_dict.update(
@@ -86,6 +52,41 @@ class PhiWrapper(DecoderOnlyWrapper):
 
 
 class PhiAttention:
+    def _attn(self, query_state, key_state, value_state, attn_mask, past_key_value, batch_idx=0, is_prefill=False):
+        # reshape for removing repeat_kv (batch=1 , num_head, 1, q_len=1, head_dim)
+        key_state = key_state.unsqueeze(2)
+        value_state = value_state.unsqueeze(2)
+        attn_mask = attn_mask.unsqueeze(2)
+
+        query_state = query_state.view(
+            1,
+            self.num_key_value_heads,
+            self.num_heads // self.num_key_value_heads,
+            -1,
+            self.head_dim,
+        )
+
+        key_state, value_state = past_key_value.update(key_state, value_state, self.layer_idx, batch_idx, is_prefill)
+
+        # Queries and keys upcast to fp32 is required by Phi-2 to avoid overflow
+        attn_weights = torch.matmul(
+            query_state.to(torch.float32),
+            key_state.to(torch.float32).transpose(3, 4),
+        ) / math.sqrt(self.head_dim)
+        attn_weights = attn_weights + attn_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_state.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_state)
+
+        # reshape for removing repeat_kv
+        attn_output = attn_output.view(1, self.num_heads, -1, self.head_dim)
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(1, -1, self.num_heads * self.head_dim)
+
+        return attn_output, key_state, value_state
+
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -95,7 +96,6 @@ class PhiAttention:
         output_attentions: bool = False,
         cos: Optional[torch.Tensor] = None,
         sin: Optional[torch.Tensor] = None,
-        rotary_pos_emb=None,
         **kwargs,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
         bsz, q_len, _ = hidden_states.size()
@@ -108,24 +108,18 @@ class PhiAttention:
             query_states = self.q_layernorm(query_states)
             key_states = self.k_layernorm(key_states)
 
-        query_states = query_states.view(
-            bsz, q_len, self.num_heads, self.head_dim
-        ).transpose(1, 2)
-        key_states = key_states.view(
-            bsz, q_len, self.num_key_value_heads, self.head_dim
-        ).transpose(1, 2)
-        value_states = value_states.view(
-            bsz, q_len, self.num_key_value_heads, self.head_dim
-        ).transpose(1, 2)
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
         # Partial rotary embedding
         query_rot, query_pass = (
-            query_states[..., : rotary_pos_emb.dim],
-            query_states[..., rotary_pos_emb.dim :],
+            query_states[..., : self.rotary_ndims],
+            query_states[..., self.rotary_ndims :],
         )
         key_rot, key_pass = (
-            key_states[..., : rotary_pos_emb.dim],
-            key_states[..., rotary_pos_emb.dim :],
+            key_states[..., : self.rotary_ndims],
+            key_states[..., self.rotary_ndims :],
         )
 
         # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
@@ -135,113 +129,38 @@ class PhiAttention:
         query_states = torch.cat((query_rot, query_pass), dim=-1)
         key_states = torch.cat((key_rot, key_pass), dim=-1)
 
-        # Decoder
-        if (batch_index is None or batch_index == -1) and bsz > 1:
-            all_key_states = []
-            all_value_states = []
-            all_attn_output = []
-
+        # Decoder (bsz > 1)
+        if bsz > 1:
+            iterate_results = {"key_states": [], "value_states": [], "attn_output": []}
             for b in range(bsz):
-                query_state = query_states[b].unsqueeze(0)
-                attn_mask = attention_mask[b].unsqueeze(0)
-                key_state = key_states[b].unsqueeze(0)
-                value_state = value_states[b].unsqueeze(0)
-
-                # reshape for removing repeat_kv (batch=1 , num_head, 1, q_len=1, head_dim)
-                key_state = key_state.unsqueeze(2)
-                value_state = value_state.unsqueeze(2)
-                attn_mask = attn_mask.unsqueeze(2)
-
-                query_state = query_state.view(
-                    1,
-                    self.num_key_value_heads,
-                    self.num_heads // self.num_key_value_heads,
-                    q_len,
-                    self.head_dim,
+                attn_output, key_state, value_state = PhiAttention._attn(
+                    self,
+                    query_states[b].unsqueeze(0),
+                    key_states[b].unsqueeze(0),
+                    value_states[b].unsqueeze(0),
+                    attention_mask[b].unsqueeze(0),
+                    past_key_value,
+                    batch_idx=b,
+                    is_prefill=False,
                 )
-
-                key_state, value_state = past_key_value.update(
-                    key_state,
-                    value_state,
-                    self.layer_idx,
-                    b,
-                )
-
-                # Queries and keys upcast to fp32 is required by Phi-2 to avoid overflow
-                attn_weights = torch.matmul(
-                    query_state.to(torch.float32),
-                    key_state.to(torch.float32).transpose(3, 4),
-                ) / math.sqrt(self.head_dim)
-                attn_weights = attn_weights + attn_mask
-
-                # upcast attention to fp32
-                attn_weights = nn.functional.softmax(
-                    attn_weights, dim=-1, dtype=torch.float32
-                ).to(query_states.dtype)
-                attn_weights = nn.functional.dropout(
-                    attn_weights, p=self.attention_dropout, training=self.training
-                )
-                attn_output = torch.matmul(attn_weights, value_state)
-
-                # reshape for removing repeat_kv
-                attn_output = attn_output.view(1, self.num_heads, q_len, self.head_dim)
-                attn_output = attn_output.transpose(1, 2).contiguous()
-                attn_output = attn_output.reshape(
-                    1, q_len, self.num_heads * self.head_dim
-                )
-
-                all_key_states.append(key_state)
-                all_value_states.append(value_state)
-                all_attn_output.append(attn_output)
-
-            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)
+                iterate_results["key_states"].append(key_state)
+                iterate_results["value_states"].append(value_state)
+                iterate_results["attn_output"].append(attn_output)
+
+            key_states = torch.cat(iterate_results["key_states"], dim=0)
+            value_states = torch.cat(iterate_results["value_states"], dim=0)
+            attn_output = torch.cat(iterate_results["attn_output"], dim=0)
+        # Prefill & Decoder (bsz == 1)
         else:
-            if batch_index is None or batch_index == -1:
-                batch_index = 0
-
-            # reshape for removing repeat_kv
-            key_states = key_states.unsqueeze(2)
-            value_states = value_states.unsqueeze(2)
-            attention_mask = attention_mask.unsqueeze(2)
-            query_states = query_states.view(
-                1,
-                self.num_key_value_heads,
-                self.num_heads // self.num_key_value_heads,
-                q_len,
-                self.head_dim,
-            )
-
-            key_states, value_states = past_key_value.update(
+            attn_output, key_states, value_states = PhiAttention._attn(
+                self,
+                query_states,
                 key_states,
                 value_states,
-                self.layer_idx,
-                batch_index,
-                read_first_step=True,
-            )
-
-            # Queries and keys upcast to fp32 is required by Phi-2 to avoid overflow
-            attn_weights = torch.matmul(
-                query_states.to(torch.float32),
-                key_states.to(torch.float32).transpose(3, 4),
-            ) / math.sqrt(self.head_dim)
-            attn_weights = attn_weights + attention_mask
-
-            # upcast attention to fp32
-            attn_weights = torch.nn.functional.softmax(
-                attn_weights, dim=-1, dtype=torch.float32
-            ).to(value_states.dtype)
-            attn_weights = torch.nn.functional.dropout(
-                attn_weights, p=self.attention_dropout, training=self.training
-            )
-            attn_output = torch.matmul(attn_weights, value_states)
-
-            # reshape for removing repeat_kv
-            attn_output = attn_output.view(1, self.num_heads, q_len, self.head_dim)
-            attn_output = attn_output.transpose(1, 2).contiguous()
-            attn_output = attn_output.reshape(
-                bsz, q_len, self.num_heads * self.head_dim
+                attention_mask,
+                past_key_value,
+                batch_idx=batch_index,
+                is_prefill=True,
             )
 
         attn_output = self.dense(attn_output)
@@ -265,12 +184,9 @@ class PhiDecoderLayer:
         batch_ids: Optional[torch.LongTensor] = None,
         cos: Optional[torch.Tensor] = None,
         sin: Optional[torch.Tensor] = None,
-        rotary_pos_emb=None,
         forward_dict: Optional[Dict[str, classmethod]] = None,
         **kwargs,
-    ) -> Tuple[
-        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
-    ]:
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
         """
         Args:
             hidden_states (`torch.FloatTensor`):
@@ -294,9 +210,7 @@ class PhiDecoderLayer:
         hidden_states = self.input_layernorm(hidden_states)
 
         # Self Attention
-        attn_outputs, self_attn_weights, key_states, value_states = forward_dict[
-            "decoder_layer"
-        ](
+        attn_outputs, self_attn_weights, key_states, value_states = forward_dict["decoder_layer"](
             self.self_attn,
             hidden_states=hidden_states,
             attention_mask=attention_mask,
@@ -307,7 +221,6 @@ class PhiDecoderLayer:
             use_cache=use_cache,
             cos=cos,
             sin=sin,
-            rotary_pos_emb=rotary_pos_emb,
             **kwargs,
         )
         past_key_value.assign(key_states, value_states, layer_idx)
@@ -339,6 +252,8 @@ class PhiModel:
         use_cache: Optional[bool] = True,
         output_attentions: Optional[bool] = False,
         output_hidden_states: Optional[bool] = False,
+        cache_pos_for_partitions: Optional[torch.Tensor] = None,
+        kvcache_partition_size: Optional[torch.Tensor] = None,
         forward_dict: Optional[Dict[str, classmethod]] = None,
         rotary_pos_emb=None,
     ) -> BaseModelOutputWithPast:
@@ -378,7 +293,8 @@ class PhiModel:
                 batch_ids=batch_ids,
                 cos=cos,
                 sin=sin,
-                rotary_pos_emb=rotary_pos_emb,
+                cache_pos_for_partitions=cache_pos_for_partitions,
+                kvcache_partition_size=kvcache_partition_size,
                 forward_dict=forward_dict,
             )
 

optimum/rbln/transformers/models/qwen2/__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_qwen2 import RBLNQwen2ForCausalLM

optimum/rbln/transformers/models/qwen2/modeling_qwen2.py

@@ -0,0 +1,43 @@
+# 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 ....utils import logging
+from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM
+from .qwen2_architecture import QWEN2Wrapper
+
+
+logger = logging.get_logger(__name__)
+
+
+class RBLNQwen2ForCausalLM(RBLNDecoderOnlyModelForCausalLM):
+    """
+    The Llama Model transformer with a language modeling head (linear layer) on top.
+    This model inherits from [`RBLNDecoderOnlyModelForCausalLM`]. Check the superclass documentation for the generic methods the library implements for all its models.
+
+    A class to convert and run pre-trained transformers based LlamaForCausalLM model on RBLN devices.
+    It implements the methods to convert a pre-trained transformers LlamaForCausalLM model into a RBLN transformer model by:
+    - transferring the checkpoint weights of the original into an optimized RBLN graph,
+    - compiling the resulting graph using the RBLN compiler.
+    """
+
+    _decoder_wrapper_cls = QWEN2Wrapper

optimum/rbln/transformers/models/qwen2/qwen2_architecture.py

@@ -0,0 +1,29 @@
+# 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 ..decoderonly.decoderonly_architecture import DecoderOnlyWrapper
+
+
+class QWEN2Wrapper(DecoderOnlyWrapper):
+    pass

optimum/rbln/transformers/models/seq2seq/__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_seq2seq import RBLNModelForSeq2SeqLM