optimum-rbln 0.8.2a3__py3-none-any.whl → 0.8.2a5__py3-none-any.whl

optimum/rbln/transformers/models/pegasus/modeling_pegasus.py

@@ -0,0 +1,69 @@
+# 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.
+
+import inspect
+from typing import TYPE_CHECKING, Any, Callable
+
+from transformers import PegasusForConditionalGeneration, PreTrainedModel
+
+from ....utils.logging import get_logger
+from ...modeling_generic import RBLNTransformerEncoderForFeatureExtraction
+from ...models.seq2seq import RBLNModelForSeq2SeqLM
+from .configuration_pegasus import RBLNPegasusForConditionalGenerationConfig
+from .pegasus_architecture import PegasusWrapper
+
+
+logger = get_logger()
+
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+
+
+class RBLNPegasusModel(RBLNTransformerEncoderForFeatureExtraction):
+    """
+    RBLN optimized PEGASUS model for feature extraction tasks.
+
+    This class provides hardware-accelerated inference for PEGASUS encoder models
+    on RBLN devices, optimized for feature extraction use cases.
+    """
+
+
+class RBLNPegasusForConditionalGeneration(RBLNModelForSeq2SeqLM):
+    """
+    RBLN optimized PEGASUS model for conditional text generation tasks.
+
+    This class provides hardware-accelerated inference for PEGASUS models
+    on RBLN devices, supporting sequence-to-sequence generation tasks
+    such as summarization, translation, and text generation.
+    """
+
+    support_causal_attn = True
+
+    @classmethod
+    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: RBLNPegasusForConditionalGenerationConfig):
+        return PegasusWrapper(
+            model, enc_max_seq_len=rbln_config.enc_max_seq_len, use_attention_mask=rbln_config.use_attention_mask
+        )
+
+    def __getattr__(self, __name: str) -> Any:
+        def redirect(func):
+            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
+
+        val = getattr(PegasusForConditionalGeneration, __name)
+
+        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
+            return redirect(val)
+
+        return val

optimum/rbln/transformers/models/pegasus/pegasus_architecture.py

@@ -0,0 +1,163 @@
+# 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 Tuple
+
+import torch
+from torch import nn
+from transformers.modeling_attn_mask_utils import (
+    _prepare_4d_attention_mask,
+)
+from transformers.utils import logging
+
+from ..seq2seq.seq2seq_architecture import (
+    Seq2SeqCrossAttention,
+    Seq2SeqDecoder,
+    Seq2SeqDecoderLayer,
+    Seq2SeqDecoderWrapper,
+    Seq2SeqEncoderWrapper,
+    Seq2SeqForConditionalGeneration,
+    Seq2SeqSelfAttention,
+)
+
+
+logger = logging.get_logger(__name__)
+
+
+class PegasusWrapper:
+    def __init__(self, model: nn.Module, enc_max_seq_len: int, use_attention_mask: bool):
+        self.encoder = Seq2SeqEncoderWrapper(model, enc_max_seq_len)
+        self.decoder = PegasusDecoderWrapper(model, use_attention_mask=use_attention_mask)
+
+
+class PegasusDecoderWrapper(Seq2SeqDecoderWrapper):
+    def convert_to_rbln_conditional_generation(self, model: nn.Module):
+        new_layers = []
+        for layer in model.get_decoder().layers:
+            self_attn = PegasusSelfAttention(layer.self_attn, use_attention_mask=self.use_attention_mask)
+            cross_attn = PegasusCrossAttention(layer.encoder_attn)
+            new_layers.append(PegasusDecoderLayer(layer, self_attn, cross_attn))
+
+        decoder_model = PegasusDecoder(model.get_decoder(), new_layers)
+        new_model = PegasusForConditionalGeneration(model, decoder_model)
+
+        return new_model
+
+
+class PegasusForConditionalGeneration(Seq2SeqForConditionalGeneration):
+    pass
+
+
+class PegasusDecoder(Seq2SeqDecoder):
+    has_pos_emb = True
+
+    def __post_init__(self):
+        self.embed_positions = self._original_mod.embed_positions
+        self.embed_scale = getattr(self._original_mod, "embed_scale", None)
+        self.final_layer_norm = getattr(self._original_mod, "layer_norm", None)
+
+    def prepare_attn_mask(self, attention_mask, encoder_attention_mask, **kwargs):
+        if attention_mask is not None:
+            attention_mask = attention_mask[:, None, None, :]
+        encoder_attention_mask = _prepare_4d_attention_mask(encoder_attention_mask, torch.float32, tgt_len=1)
+
+        return attention_mask, encoder_attention_mask
+
+    def apply_position_embedding(self, inputs_embeds, cache_position):
+        hidden_all = []
+        for i in range(inputs_embeds.shape[0]):
+            positions_idx = cache_position[i]
+            position_weight = self.embed_positions.weight
+            position = position_weight[positions_idx]
+            batch_hidden = position + inputs_embeds[i]
+            hidden_all.append(batch_hidden)
+        hidden_states = torch.stack(hidden_all, dim=0)
+
+        return hidden_states
+
+    def get_embedding(self):
+        if self.embed_scale is not None:
+            return lambda x: self.embed_tokens(x) * self.embed_scale
+        else:
+            return self.embed_tokens
+
+
+class PegasusLayerFF(nn.Module):
+    def __init__(self, decoder_layer):
+        super().__init__()
+        self.fc1 = decoder_layer.fc1
+        self.fc2 = decoder_layer.fc2
+        self.activation_fn = decoder_layer.activation_fn
+        self.layer_norm = decoder_layer.final_layer_norm
+
+    def forward(self, hidden_states):
+        # Residual Connection
+        residual = hidden_states
+        hidden_states = self.layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class PegasusDecoderLayer(Seq2SeqDecoderLayer):
+    def __post_init__(self):
+        self.self_attn_layer_norm = self._original_mod.self_attn_layer_norm
+        self.encoder_attn = self._original_mod.encoder_attn
+        self.encoder_attn_layer_norm = self._original_mod.encoder_attn_layer_norm
+        self.ff_layer = PegasusLayerFF(self._original_mod)
+
+    def pre_self_attn_layer_norm(self, hidden_states):
+        return self.self_attn_layer_norm(hidden_states)
+
+    def post_self_attn_layer_norm(self, hidden_states):
+        return hidden_states
+
+    def pre_cross_attn_layer_norm(self, hidden_states):
+        return self.encoder_attn_layer_norm(hidden_states)
+
+    def post_cross_attn_layer_norm(self, hidden_states):
+        return hidden_states
+
+
+class PegasusSelfAttention(Seq2SeqSelfAttention):
+    def __post_init__(self, use_attention_mask: bool = True):
+        self.q_proj = self._original_mod.q_proj
+        self.k_proj = self._original_mod.k_proj
+        self.v_proj = self._original_mod.v_proj
+        self.out_proj = self._original_mod.out_proj
+        self.num_heads = self._original_mod.num_heads
+        self.head_dim = self._original_mod.embed_dim // self._original_mod.num_heads
+        self.scaling = self.head_dim**-0.5
+        if use_attention_mask:
+            self.attn_decode = torch.ops.rbln_custom_ops.paged_attn_decode
+        else:
+            self.attn_decode = torch.ops.rbln_custom_ops.paged_causal_attn_decode
+
+    def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        query_states = self.q_proj(hidden_states) * self.scaling
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        return query_states, key_states, value_states
+
+
+class PegasusCrossAttention(Seq2SeqCrossAttention):
+    def __post_init__(self):
+        self.q_proj = self._original_mod.q_proj
+        self.k_proj = self._original_mod.k_proj
+        self.v_proj = self._original_mod.v_proj
+        self.out_proj = self._original_mod.out_proj
+        self.num_heads = self._original_mod.num_heads
+        self.head_dim = self._original_mod.embed_dim // self._original_mod.num_heads
+        self.embed_dim = self._original_mod.embed_dim

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

@@ -12,5 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configuration_phi import RBLNPhiForCausalLMConfig
-from .modeling_phi import RBLNPhiForCausalLM
+from .configuration_phi import RBLNPhiForCausalLMConfig, RBLNPhiModelConfig
+from .modeling_phi import RBLNPhiForCausalLM, RBLNPhiModel

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

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelConfig, RBLNDecoderOnlyModelForCausalLMConfig
 
 
 class RBLNPhiForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
@@ -40,3 +40,11 @@ class RBLNPhiForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
     )
     ```
     """
+
+
+class RBLNPhiModelConfig(RBLNDecoderOnlyModelConfig):
+    """
+    Configuration class for RBLN Phi models.
+
+    This class is an alias of RBLNDecoderOnlyModelConfig.
+    """

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

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from ....utils import logging
-from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM
+from ...models.decoderonly import RBLNDecoderOnlyModel, RBLNDecoderOnlyModelForCausalLM
 from .phi_architecture import PhiWrapper
 
 
@@ -81,3 +81,12 @@ class RBLNPhiForCausalLM(RBLNDecoderOnlyModelForCausalLM):
     """
 
     _decoder_wrapper_cls = PhiWrapper
+
+
+class RBLNPhiModel(RBLNDecoderOnlyModel):
+    """
+    The Phi Model transformer without a language modeling head.
+    This model inherits from [`RBLNDecoderOnlyModel`]. Check the superclass documentation for the generic methods the library implements for all its models.
+    """
+
+    _decoder_wrapper_cls = PhiWrapper

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

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import TYPE_CHECKING, Optional, Tuple
+from typing import TYPE_CHECKING, Optional, Tuple, Union
 
 import torch
 from transformers import PhiForCausalLM
@@ -27,7 +27,7 @@ from ..decoderonly.decoderonly_architecture import (
 
 
 if TYPE_CHECKING:
-    from transformers import PhiForCausalLM
+    from transformers import PhiForCausalLM, PhiModel
 
 
 class PhiWrapper(DecoderOnlyWrapper):
@@ -40,11 +40,11 @@ class PhiWrapper(DecoderOnlyWrapper):
     def get_rbln_model_class(self):
         return PhiModel
 
-    def get_model_layer(self, causal_lm: "PhiForCausalLM"):
-        return causal_lm.model
+    def get_model_layer(self, model: Union["PhiForCausalLM", "PhiModel"]):
+        return model.model if self.is_causal_lm else model
 
-    def get_decoder_layers(self, causal_lm: "PhiForCausalLM"):
-        return causal_lm.model.layers
+    def get_decoder_layers(self, model: Union["PhiForCausalLM", "PhiModel"]):
+        return model.model.layers if self.is_causal_lm else model.layers
 
 
 class PhiAttention(DecoderOnlyAttention):

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

@@ -12,5 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .configuration_qwen2 import RBLNQwen2ForCausalLMConfig
-from .modeling_qwen2 import RBLNQwen2ForCausalLM
+from .configuration_qwen2 import RBLNQwen2ForCausalLMConfig, RBLNQwen2ModelConfig
+from .modeling_qwen2 import RBLNQwen2ForCausalLM, RBLNQwen2Model

optimum/rbln/transformers/models/qwen2/configuration_qwen2.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelConfig, RBLNDecoderOnlyModelForCausalLMConfig
 
 
 class RBLNQwen2ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
@@ -40,3 +40,11 @@ class RBLNQwen2ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
     )
     ```
     """
+
+
+class RBLNQwen2ModelConfig(RBLNDecoderOnlyModelConfig):
+    """
+    Configuration class for RBLN Qwen2 models.
+
+    This class is an alias of RBLNDecoderOnlyModelConfig.
+    """

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

@@ -15,7 +15,11 @@
 from transformers import PretrainedConfig
 
 from ....utils import logging
-from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM, RBLNDecoderOnlyModelForCausalLMConfig
+from ...models.decoderonly import (
+    RBLNDecoderOnlyModel,
+    RBLNDecoderOnlyModelForCausalLM,
+    RBLNDecoderOnlyModelForCausalLMConfig,
+)
 from .qwen2_architecture import QWEN2Wrapper
 
 
@@ -95,3 +99,25 @@ class RBLNQwen2ForCausalLM(RBLNDecoderOnlyModelForCausalLM):
         rbln_config.sliding_window = model_config.sliding_window
         rbln_config.sliding_window_layers = list(range(model_config.num_hidden_layers))
         return rbln_config
+
+
+class RBLNQwen2Model(RBLNDecoderOnlyModel):
+    """
+    The Qwen2 Model transformer without a language modeling head.
+    This model inherits from [`RBLNDecoderOnlyModel`]. Check the superclass documentation for the generic methods the library implements for all its models.
+    """
+
+    _decoder_wrapper_cls = QWEN2Wrapper
+
+    @classmethod
+    def _update_sliding_window_config(
+        cls, model_config: PretrainedConfig, rbln_config: RBLNDecoderOnlyModelForCausalLMConfig
+    ):
+        # https://github.com/huggingface/transformers/issues/35896
+        # There seems to be a bug in transformers(v4.52.4). Therefore, similar to when attn_implementation is eager,
+        # we set all layers to use sliding window in this version. This should be updated once the bug is fixed.
+
+        rbln_config.cache_impl = "sliding_window"
+        rbln_config.sliding_window = model_config.sliding_window
+        rbln_config.sliding_window_layers = list(range(model_config.num_hidden_layers))
+        return rbln_config

optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -34,7 +34,7 @@ from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
 from ....configuration_utils import RBLNCompileConfig
 from ....modeling import RBLNModel
 from ....utils.logging import get_logger
-from ..decoderonly.modeling_decoderonly import RBLNDecoderOnlyModelForCausalLM, RBLNDecoderOnlyOutput
+from ..decoderonly.modeling_decoderonly import RBLNDecoderOnlyForCausalLMOutput, RBLNDecoderOnlyModelForCausalLM
 from .configuration_qwen2_5_vl import (
     RBLNQwen2_5_VisionTransformerPretrainedModelConfig,
     RBLNQwen2_5_VLForConditionalGenerationConfig,
@@ -595,7 +595,7 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
         generate_idx: Optional[torch.Tensor] = None,
         return_dict: Optional[bool] = None,
         **kwargs,
-    ) -> RBLNDecoderOnlyOutput:
+    ) -> RBLNDecoderOnlyForCausalLMOutput:
         # Prefill
         if cache_position is None:
             inputs_embeds, position_embed, rope_deltas = self._preprocess_prefill(
@@ -637,7 +637,7 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
         if not return_dict:
             return logits, generate_idx
         else:
-            return RBLNDecoderOnlyOutput(
+            return RBLNDecoderOnlyForCausalLMOutput(
                 logits=logits,
                 generate_idx=generate_idx,
             )

optimum/rbln/transformers/models/qwen3/configuration_qwen3.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelConfig, RBLNDecoderOnlyModelForCausalLMConfig
 
 
 class RBLNQwen3ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
@@ -43,7 +43,7 @@ class RBLNQwen3ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
     """
 
 
-class RBLNQwen3ModelConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+class RBLNQwen3ModelConfig(RBLNDecoderOnlyModelConfig):
     """
     Configuration class for RBLN Qwen3 models.