optimum-rbln 0.1.12__py3-none-any.whl → 0.1.15__py3-none-any.whl

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

@@ -21,386 +21,102 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
-import math
-from typing import Dict, Optional, Tuple
+from typing import TYPE_CHECKING, Optional, Tuple
 
 import torch
-import torch.nn as nn
-from transformers.modeling_outputs import (
-    BaseModelOutputWithPast,
-)
+from transformers import PhiForCausalLM
 
-from ...cache_utils import RebelDynamicCache
-from ..decoderonly import (
+from ..decoderonly.decoderonly_architecture import (
+    DecoderOnlyAttention,
+    DecoderOnlyForCausalLM,
+    DecoderOnlyLayer,
+    DecoderOnlyModel,
     DecoderOnlyWrapper,
-    DynamicNTKScalingRotaryEmbedding,
-    LinearScalingRotaryEmbedding,
-    RotaryEmbedding,
-    apply_rotary_pos_emb,
-    slice_and_unsqueeze_cos_sin,
+    apply_rotary_pos_emb_partial,
 )
 
 
-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(
-            {
-                "wrapper": PhiModel.forward,
-                "model": PhiDecoderLayer.forward,
-                "decoder_layer": PhiAttention.forward,
-            }
-        )
-        return forward_dict
-
+if TYPE_CHECKING:
+    from transformers import PhiForCausalLM
 
-class PhiAttention:
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_value: Optional[RebelDynamicCache] = None,
-        batch_index: Optional[int] = None,
-        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()
 
+class PhiWrapper(DecoderOnlyWrapper):
+    def convert_to_rbln_causal_lm(self, causal_lm: "PhiForCausalLM"):
+        new_layers = []
+        for layer in causal_lm.model.layers:
+            if self.attn_impl == "eager":
+                new_self_attn = PhiAttention(layer.self_attn)
+            elif self.attn_impl == "flash_attn":
+                raise NotImplementedError(f"flash attn for {self.__class__} is not implemented yet.")
+            else:
+                raise NotImplementedError(f"Unknwon attn : {self.attn_impl}")
+            new_layer = PhiLayer(layer, new_self_attn)
+            new_layers.append(new_layer)
+        new_model = PhiModel(causal_lm.model, new_layers)
+        new_causal_lm = DecoderOnlyForCausalLM(causal_lm, new_model)
+        return new_causal_lm
+
+
+class PhiAttention(DecoderOnlyAttention):
+    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.o_proj = self._original_mod.dense
+        self.qk_layernorm = self._original_mod.qk_layernorm
+        self.rotary_ndims = self._original_mod.rotary_ndims
+        self.num_key_value_heads = self.num_heads
+
+    def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         query_states = self.q_proj(hidden_states)
         key_states = self.k_proj(hidden_states)
         value_states = self.v_proj(hidden_states)
 
         if self.qk_layernorm:
-            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)
-
-        # Partial rotary embedding
-        query_rot, query_pass = (
-            query_states[..., : rotary_pos_emb.dim],
-            query_states[..., rotary_pos_emb.dim :],
-        )
-        key_rot, key_pass = (
-            key_states[..., : rotary_pos_emb.dim],
-            key_states[..., rotary_pos_emb.dim :],
-        )
-
-        # [batch_size, seq_length, num_heads, head_dim // config.partial_rotary_factor]
-        query_rot, key_rot = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
-
-        # [batch_size, seq_length, num_heads, head_dim]
-        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 = []
-
-            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,
-                )
-
-                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)
+            query_states = self._original_mod.q_layernorm(query_states)
+            key_states = self._original_mod.k_layernorm(key_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(
-                    1, q_len, self.num_heads * self.head_dim
-                )
+        return query_states, key_states, value_states
 
-                all_key_states.append(key_state)
-                all_value_states.append(value_state)
-                all_attn_output.append(attn_output)
+    def apply_rotary_pos_embed(self, query_states, key_states, cos, sin):
+        return apply_rotary_pos_emb_partial(query_states, key_states, cos, sin, ndim=self.rotary_ndims)
 
-            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)
-        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,
-            )
+class PhiLayer(DecoderOnlyLayer):
+    def get_post_attention_layernorm(self):
+        raise NotImplementedError
 
-            key_states, value_states = past_key_value.update(
-                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
-            )
-
-        attn_output = self.dense(attn_output)
-
-        if not output_attentions:
-            attn_weights = None
-
-        return attn_output, attn_weights, key_states, value_states
-
-
-class PhiDecoderLayer:
     def forward(
         self,
         hidden_states: torch.Tensor,
-        layer_idx: int,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[RebelDynamicCache] = None,
-        output_attentions: Optional[bool] = None,
-        use_cache: Optional[bool] = None,
-        batch_ids: Optional[torch.LongTensor] = None,
+        attention_mask: torch.Tensor,
+        current_steps: torch.LongTensor,
+        batch_position: torch.Tensor,
+        past_key_values: Tuple[Tuple[torch.Tensor]],
         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]]
-    ]:
-        """
-        Args:
-            hidden_states (`torch.FloatTensor`):
-                input to the layer of shape `(batch, seq_len, embed_dim)`
-            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
-                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
-            position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
-                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
-                `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
-            output_attentions (`bool`, *optional*):
-                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
-                returned tensors for more detail.
-            use_cache (`bool`, *optional*):
-                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
-                (see `past_key_values`).
-            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
-        """
-
+    ):
         residual = hidden_states
 
-        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.get_pre_attention_layernorm()(hidden_states)
 
-        # Self Attention
-        attn_outputs, self_attn_weights, key_states, value_states = forward_dict[
-            "decoder_layer"
-        ](
-            self.self_attn,
+        attn_outputs, present_key_values = self.self_attn(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_value=past_key_value,
-            output_attentions=output_attentions,
-            batch_index=batch_ids,
-            use_cache=use_cache,
+            current_steps=current_steps,
+            batch_position=batch_position,
+            past_key_values=past_key_values,
             cos=cos,
             sin=sin,
-            rotary_pos_emb=rotary_pos_emb,
-            **kwargs,
         )
-        past_key_value.assign(key_states, value_states, layer_idx)
 
-        attn_outputs = self.resid_dropout(attn_outputs)
+        feed_forward_hidden_states = self._original_mod.mlp(hidden_states)
 
-        feed_forward_hidden_states = self.resid_dropout(self.mlp(hidden_states))
         hidden_states = attn_outputs + feed_forward_hidden_states + residual
-        outputs = (hidden_states,)
-
-        if output_attentions:
-            outputs += (self_attn_weights,)
-
-        if use_cache:
-            outputs += (past_key_value,)
-
-        return outputs
-
 
-class PhiModel:
-    def forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[RebelDynamicCache] = None,
-        batch_ids: Optional[torch.LongTensor] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = True,
-        output_attentions: Optional[bool] = False,
-        output_hidden_states: Optional[bool] = False,
-        forward_dict: Optional[Dict[str, classmethod]] = None,
-        rotary_pos_emb=None,
-    ) -> BaseModelOutputWithPast:
-        # retrieve input_ids and inputs_embeds
-        if (input_ids is None) ^ (inputs_embeds is not None):
-            raise ValueError(
-                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
-            )
-
-        # embed positions
-        if inputs_embeds is None:
-            inputs_embeds = self.embed_tokens(input_ids)
-
-        hidden_states = inputs_embeds
-        attention_mask = (1 - attention_mask) * torch.finfo(torch.float16).min
-
-        # get cos,sin vector
-        cos, sin = rotary_pos_emb(inputs_embeds, attention_mask.shape[-1])
-        cos, sin = slice_and_unsqueeze_cos_sin(cos, sin, position_ids)
-
-        # decoder layers
-        all_hidden_states = () if output_hidden_states else None
-        all_self_attns = () if output_attentions else None
-
-        for layer_idx, decoder_layer in enumerate(self.layers):
-            if output_hidden_states:
-                all_hidden_states += (hidden_states,)
-            layer_outputs = forward_dict["model"](
-                decoder_layer,
-                hidden_states,
-                layer_idx,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                past_key_value=past_key_values,
-                output_attentions=output_attentions,
-                use_cache=use_cache,
-                batch_ids=batch_ids,
-                cos=cos,
-                sin=sin,
-                rotary_pos_emb=rotary_pos_emb,
-                forward_dict=forward_dict,
-            )
+        return hidden_states, present_key_values
 
-            hidden_states = layer_outputs[0]
 
-            updated_cache = layer_outputs[2 if output_attentions else 1]
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
-        hidden_states = self.final_layernorm(hidden_states)
-
-        # add hidden states from the last decoder layer
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        # convert RebelDynamicCache to legacy Tuple[Tuple[torch.Tensor]]
-        next_cache = updated_cache.to_legacy_cache()
-
-        return BaseModelOutputWithPast(
-            last_hidden_state=hidden_states,
-            past_key_values=next_cache,
-            hidden_states=all_hidden_states,
-            attentions=all_self_attns,
-        )
+class PhiModel(DecoderOnlyModel):
+    def get_last_layernorm(self):
+        return self._original_mod.final_layernorm

optimum/rbln/transformers/models/qwen2/modeling_qwen2.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 Qwen2ForCausalLM
-
-from ..decoderonly import RBLNDecoderOnlyModelForCausalLM
+from ....utils import logging
+from ...models.decoderonly import RBLNDecoderOnlyModelForCausalLM
 from .qwen2_architecture import QWEN2Wrapper
 
 
-if TYPE_CHECKING:
-    from transformers import PreTrainedModel
-
-    from ....modeling_config import RBLNConfig
-
-logger = logging.getLogger(__name__)
+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 [`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:
@@ -50,18 +40,4 @@ class RBLNQwen2ForCausalLM(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 QWEN2Wrapper(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(Qwen2ForCausalLM, __name)
-
-        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
-            return redirect(val)
-
-        return val
+    _decoder_wrapper_cls = QWEN2Wrapper

optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py

@@ -31,7 +31,7 @@ import torch  # noqa: F401
 from transformers import AutoModelForSeq2SeqLM, GenerationConfig, PretrainedConfig, PreTrainedModel
 from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 
-from ....modeling_base import RBLNModel
+from ....modeling import RBLNModel
 from ....modeling_config import DEFAULT_COMPILED_MODEL_NAME, RBLNCompileConfig, RBLNConfig
 from ....utils.runtime_utils import RBLNPytorchRuntime
 
@@ -346,51 +346,6 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
 
         return Seq2SeqLMOutput(logits=lm_logits)
 
-    def vllm_forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        cache_position: Union[List[torch.Tensor], torch.Tensor] = None,  # vllm keyword argument
-        batch_idx: Optional[torch.LongTensor] = None,
-        enc_lengths: List[int] = None,  # vllm return current attention_mask length
-        **kwargs,
-    ) -> Tuple[torch.FloatTensor]:
-        # When using vllm, need the output of the encoder (ex. vocab_size + 100) and use that value act as start_token_id in decoder (ex. vocab_size + 99)
-        # encoder
-        if batch_idx is not None:
-            enc_attention_mask = torch.zeros(1, self.rbln_config.model_cfg["enc_max_seq_len"], dtype=torch.float32)
-            enc_attention_mask[0][: enc_lengths[batch_idx] + 1] = 1
-            padding_need = self.rbln_config.model_cfg["enc_max_seq_len"] - input_ids.shape[-1]
-            input_ids = torch.nn.functional.pad(input_ids, (0, padding_need))
-            _ = self.encoder(input_ids, enc_attention_mask, batch_idx=batch_idx.to(torch.int32))
-            logits = torch.zeros(1, 1, self.config.vocab_size + 100)
-            logits[0][0][-1] = 1
-        # decoder
-        else:
-            input_ids[input_ids == (self.config.vocab_size + 99)] = self.config.decoder_start_token_id
-            cache_position[cache_position != 0] = cache_position[cache_position != 0] - 2
-
-            enc_attention_mask = torch.zeros(
-                self.rbln_config.model_cfg["batch_size"],
-                self.rbln_config.model_cfg["enc_max_seq_len"],
-                dtype=torch.float32,
-            )
-            dec_attention_mask = torch.zeros(
-                self.rbln_config.model_cfg["batch_size"],
-                self.rbln_config.model_cfg["dec_max_seq_len"],
-                dtype=torch.float32,
-            )
-            for batch_idx in range(self.rbln_config.model_cfg["batch_size"]):
-                enc_attention_mask[batch_idx, : enc_lengths[batch_idx] + 1] = 1
-
-            logits = self._forward_decoder(
-                attention_mask=enc_attention_mask,
-                decoder_input_ids=input_ids,
-                decoder_attention_mask=dec_attention_mask,
-                cache_position=cache_position,
-            ).logits
-
-        return Seq2SeqLMOutput(logits=logits)
-
     def _prepare_encoder_decoder_kwargs_for_generation(
         self,
         inputs_tensor: torch.Tensor,

optimum/rbln/transformers/models/t5/__init__.py

@@ -21,5 +21,5 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
-from .modeling_t5 import RBLNT5ForConditionalGeneration
+from .modeling_t5 import RBLNT5EncoderModel, RBLNT5ForConditionalGeneration
 from .t5_architecture import T5DecoderWrapper, T5EncoderWrapper