optimum-rbln 0.7.3.post2__py3-none-any.whl → 0.7.4__py3-none-any.whl

optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py

@@ -0,0 +1,214 @@
+import math
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+
+from ..decoderonly.decoderonly_architecture import (
+    DecoderOnlyWrapper,
+    apply_rotary_pos_emb,
+)
+
+
+class Qwen2_5_VisionTransformerWrapper(nn.Module):
+    def __init__(self, model: torch.nn.Module):
+        super().__init__()
+        self._original_mod = model
+        self.fullatt_block_indexes = model.fullatt_block_indexes
+        self.merger = model.merger
+        window_seq_len = (model.window_size // model.patch_size) ** 2
+        self.blocks = self.wrap_vision_blocks(model.blocks, window_seq_len)
+
+    def wrap_vision_blocks(self, blocks: torch.nn.ModuleList, window_seq_len: int):
+        wrapped_blocks = []
+        for i, block in enumerate(blocks):
+            is_full_attn = True if i in self.fullatt_block_indexes else False
+            wrapped_blocks.append(Qwen2_5_VLVisionBlock(block, is_full_attn, window_seq_len))
+        return nn.ModuleList(wrapped_blocks)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        full_attn_masks: torch.Tensor,
+        window_attn_masks: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+    ):
+        full_attn_masks = (1 - full_attn_masks) * torch.finfo(torch.float32).min
+        window_attn_masks = (1 - window_attn_masks) * torch.finfo(torch.float32).min
+
+        for i, block in enumerate(self.blocks):
+            attn_masks = full_attn_masks if i in self.fullatt_block_indexes else window_attn_masks
+            hidden_states = block(hidden_states, attn_masks, [cos, sin])
+
+        hidden_states = self.merger(hidden_states)
+
+        return hidden_states
+
+
+class Qwen2_5_VLVisionBlock(torch.nn.Module):
+    def __init__(self, model: torch.nn.Module, is_full_attn: bool, window_seq_len: int):
+        super().__init__()
+        self._origin_model = model
+        self.norm1 = model.norm1
+        self.norm2 = model.norm2
+
+        if is_full_attn:
+            self.attn = Qwen2_5_VLVisionFullAttention(model.attn)
+        else:
+            self.attn = Qwen2_5_VLVisionWindowAttention(model.attn, window_seq_len)
+        self.mlp = model.mlp
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attn_masks: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+    ) -> torch.Tensor:
+        hidden_states = hidden_states + self.attn(
+            self.norm1(hidden_states),
+            attn_masks,
+            position_embeddings,
+        )
+        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
+        return hidden_states
+
+
+class Qwen2_5_VLVisionFullAttention(nn.Module):
+    def __init__(self, model: nn.Module) -> None:
+        super().__init__()
+        self._origin_model = model
+        self.num_heads = model.num_heads
+        self.head_dim = model.head_dim
+        self.qkv = model.qkv
+        self.proj = model.proj
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attn_masks: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+    ) -> torch.Tensor:
+        seq_length = hidden_states.shape[0]
+        hidden_states = hidden_states.unsqueeze(0)
+        q, k, v = (
+            self.qkv(hidden_states).reshape(1, seq_length, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4).unbind(0)
+        )
+
+        cos, sin = position_embeddings
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        attn_weights = torch.matmul(q, k.transpose(2, 3)) / math.sqrt(self.head_dim)
+        attn_weights = attn_weights + attn_masks
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32)
+        attn_output = torch.matmul(attn_weights, v)
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(1, seq_length, -1)
+        attn_output = self.proj(attn_output).squeeze(0)
+
+        return attn_output
+
+
+class Qwen2_5_VLVisionWindowAttention(nn.Module):
+    def __init__(self, model: nn.Module, window_seq_len: int) -> None:
+        super().__init__()
+        self._origin_model = model
+        self.num_heads = model.num_heads
+        self.head_dim = model.head_dim
+        self.qkv = model.qkv
+        self.proj = model.proj
+        self.window_seq_len = window_seq_len
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attn_masks: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+    ) -> torch.Tensor:
+        seq_length = hidden_states.shape[0]
+        num_windows = seq_length // self.window_seq_len
+
+        window_hidden_states = []
+        for i in range(0, seq_length, self.window_seq_len):
+            window_hidden_states.append(hidden_states[i : i + self.window_seq_len])
+        hidden_states = torch.stack(window_hidden_states)
+
+        q, k, v = (
+            self.qkv(hidden_states)
+            .reshape(num_windows, self.window_seq_len, 3, self.num_heads, -1)
+            .permute(2, 0, 3, 1, 4)
+            .unbind(0)
+        )
+        cos, sin = position_embeddings
+        cos = cos.reshape(num_windows, 1, seq_length // num_windows, -1)
+        sin = sin.reshape(num_windows, 1, seq_length // num_windows, -1)
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        attn_weights = torch.matmul(q, k.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        attn_weights = attn_weights + attn_masks
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32)
+        attn_output = torch.matmul(attn_weights, v)
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(1, seq_length, -1)
+        attn_output = self.proj(attn_output).squeeze(0)
+
+        return attn_output
+
+
+class Qwen2_5_VL_LanguageModelWrapper(DecoderOnlyWrapper):
+    def forward(self, *args):
+        if self.phase == "decode":
+            if self.use_attention_mask:
+                (
+                    input_ids_or_inputs_embeds,
+                    cache_position,
+                    attention_mask,
+                    block_tables,
+                    position_emb,
+                    *past_key_values,
+                ) = args
+            else:
+                (
+                    input_ids_or_inputs_embeds,
+                    cache_position,
+                    block_tables,
+                    position_emb,
+                    *past_key_values,
+                ) = args
+                attention_mask = None
+            query_position = None
+        elif self.phase == "prefill":
+            if self.use_attention_mask:
+                (
+                    input_ids_or_inputs_embeds,
+                    cache_position,
+                    attention_mask,
+                    query_position,
+                    block_tables,
+                    position_emb,
+                    *past_key_values,
+                ) = args
+            else:
+                (
+                    input_ids_or_inputs_embeds,
+                    cache_position,
+                    query_position,
+                    block_tables,
+                    position_emb,
+                    *past_key_values,
+                ) = args
+                attention_mask = None
+
+        else:
+            raise ValueError(f"Unknown phase: {self.phase}")
+
+        return self.forward_common(
+            input_ids_or_inputs_embeds,
+            cache_position,
+            attention_mask,
+            query_position,
+            block_tables,
+            position_emb,
+            *past_key_values,
+        )

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

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .configuration_seq2seq2 import RBLNModelForSeq2SeqLMConfig
 from .modeling_seq2seq import RBLNModelForSeq2SeqLM

optimum/rbln/transformers/models/seq2seq/configuration_seq2seq2.py

@@ -0,0 +1,66 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import rebel
+
+from ....configuration_utils import RBLNModelConfig
+from ....utils.logging import get_logger
+
+
+logger = get_logger()
+
+
+class RBLNModelForSeq2SeqLMConfig(RBLNModelConfig):
+    def __init__(
+        self,
+        batch_size: Optional[int] = None,
+        enc_max_seq_len: Optional[int] = None,
+        dec_max_seq_len: Optional[int] = None,
+        use_attention_mask: Optional[bool] = None,
+        pad_token_id: Optional[int] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            batch_size (Optional[int]): The batch size for inference. Defaults to 1.
+            enc_max_seq_len (Optional[int]): Maximum sequence length for the encoder.
+            dec_max_seq_len (Optional[int]): Maximum sequence length for the decoder.
+            use_attention_mask (Optional[bool]): Whether to use attention masks during inference.
+                This is automatically set to True for RBLN-CA02 devices.
+            pad_token_id (Optional[int]): The ID of the padding token in the vocabulary.
+            **kwargs: Additional arguments passed to the parent RBLNModelConfig.
+
+        Raises:
+            ValueError: If batch_size is not a positive integer.
+        """
+        super().__init__(**kwargs)
+        self.batch_size = batch_size or 1
+        if not isinstance(self.batch_size, int) or self.batch_size < 0:
+            raise ValueError(f"batch_size must be a positive integer, got {self.batch_size}")
+
+        self.enc_max_seq_len = enc_max_seq_len
+        self.dec_max_seq_len = dec_max_seq_len
+
+        self.use_attention_mask = use_attention_mask
+        npu = self.npu or rebel.get_npu_name()
+        if npu == "RBLN-CA02":
+            if self.use_attention_mask is False:
+                logger.warning("Attention mask should be used with RBLN-CA02. Setting use_attention_mask to True.")
+            self.use_attention_mask = True
+        else:
+            self.use_attention_mask = self.use_attention_mask or False
+
+        self.pad_token_id = pad_token_id

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

@@ -22,10 +22,11 @@ from rebel.compile_context import CompileContext
 from transformers import AutoModelForSeq2SeqLM, PretrainedConfig, PreTrainedModel
 from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 
+from ....configuration_utils import RBLNCompileConfig
 from ....modeling import RBLNModel
-from ....modeling_config import RBLNCompileConfig, RBLNConfig
 from ....utils.logging import get_logger
 from ....utils.runtime_utils import RBLNPytorchRuntime
+from .configuration_seq2seq2 import RBLNModelForSeq2SeqLMConfig
 
 
 logger = get_logger(__name__)
@@ -38,8 +39,8 @@ class RBLNRuntimeEncoder(RBLNPytorchRuntime):
     mandatory_members = ["main_input_name"]
 
     def forward(self, *args: List[torch.Tensor], **kwargs: Dict[str, torch.Tensor]):
-        _ = super().forward(*args, **kwargs)
-        return BaseModelOutput(last_hidden_state=torch.tensor([1.0]))
+        output = super().forward(*args, **kwargs)
+        return BaseModelOutput(last_hidden_state=output)
 
 
 class RBLNRuntimeDecoder(RBLNPytorchRuntime):
@@ -94,7 +95,7 @@ class RBLNRuntimeDecoder(RBLNPytorchRuntime):
             decoder_attention_mask if self.use_attention_mask else None,
             attention_mask,
             cache_position,
-            block_tables,
+            block_tables=block_tables,
         )
 
         return Seq2SeqLMOutput(logits=lm_logits)
@@ -115,11 +116,12 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
 
     main_input_name = "input_ids"
     auto_model_class = AutoModelForSeq2SeqLM
+    support_causal_attn = None
 
     def __post_init__(self, **kwargs):
-        batch_size = self.rbln_config.model_cfg["batch_size"]
-        dec_max_seq_len = self.rbln_config.model_cfg["dec_max_seq_len"]
-        self.use_attention_mask = self.rbln_config.model_cfg.get("use_attention_mask", None)
+        batch_size = self.rbln_config.batch_size
+        dec_max_seq_len = self.rbln_config.dec_max_seq_len
+        self.use_attention_mask = self.rbln_config.use_attention_mask
 
         self.encoder = RBLNRuntimeEncoder(
             runtime=self.model[0],
@@ -135,7 +137,7 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
 
     @classmethod
     @torch.inference_mode()
-    def get_compiled_model(cls, model: PreTrainedModel, rbln_config: RBLNConfig):
+    def get_compiled_model(cls, model: PreTrainedModel, rbln_config: RBLNModelForSeq2SeqLMConfig):
         wrapped_model = cls.wrap_model_if_needed(model, rbln_config)
 
         enc_compile_config = rbln_config.compile_cfgs[0]
@@ -176,23 +178,15 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
         return {"encoder": compiled_encoder, "decoder": compiled_decoder}
 
     @classmethod
-    def _get_rbln_config(
+    def _update_rbln_config(
         cls,
         preprocessors: Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"],
-        model_config: "PretrainedConfig",
-        rbln_kwargs: Dict[str, Any] = {},
-    ) -> RBLNConfig:
-        rbln_enc_max_seq_len = rbln_kwargs.get("enc_max_seq_len", None)
-        rbln_dec_max_seq_len = rbln_kwargs.get("dec_max_seq_len", None)
-        rbln_batch_size = rbln_kwargs.get("batch_size", None)
-        rbln_batch_size = 1 if rbln_batch_size is None else rbln_batch_size
-        rbln_use_attention_mask = rbln_kwargs.get("use_attention_mask", None)
-
-        if rbln_use_attention_mask is None:
-            rbln_use_attention_mask = False
-            rbln_npu = rbln_kwargs.get("npu", None) or rebel.get_npu_name()
-            if rbln_npu == "RBLN-CA02":
-                rbln_use_attention_mask = True
+        model: Optional["PreTrainedModel"] = None,
+        model_config: Optional["PretrainedConfig"] = None,
+        rbln_config: Optional[RBLNModelForSeq2SeqLMConfig] = None,
+    ) -> RBLNModelForSeq2SeqLMConfig:
+        if not cls.support_causal_attn:
+            rbln_config.use_attention_mask = True
 
         n_layer = getattr(model_config, "decoder_layers", None) or getattr(model_config, "num_layers")
         n_head = getattr(model_config, "decoder_attention_heads", None) or getattr(model_config, "num_heads")
@@ -206,84 +200,85 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
             model_config, "max_position_embeddings", None
         )
 
-        rbln_pad_token_id = getattr(model_config, "pad_token_id", None)
-        if rbln_pad_token_id is None:
-            rbln_pad_token_id = getattr(model_config, "bos_token_id", None)
-            if rbln_pad_token_id is None:
-                rbln_pad_token_id = getattr(model_config, "eos_token_id", None)
-                if rbln_pad_token_id is None:
-                    rbln_pad_token_id = -1
-
-        if rbln_enc_max_seq_len is None:
-            rbln_enc_max_seq_len = max_position_embeddings
-            if rbln_enc_max_seq_len is None:
-                for tokenizer in preprocessors:
-                    if hasattr(tokenizer, "model_max_length"):
-                        rbln_enc_max_seq_len = tokenizer.model_max_length
-                        break
-                if rbln_enc_max_seq_len is None:
-                    raise ValueError("`rbln_enc_max_seq_len` should be specified!")
-        if max_position_embeddings is not None and rbln_enc_max_seq_len > max_position_embeddings:
-            raise ValueError("`rbln_enc_max_seq_len` should be less or equal than max_position_embeddings!")
-
-        if rbln_dec_max_seq_len is None:
-            rbln_dec_max_seq_len = max_position_embeddings
-            if rbln_dec_max_seq_len is None:
-                for tokenizer in preprocessors:
-                    if hasattr(tokenizer, "model_max_length"):
-                        rbln_dec_max_seq_len = tokenizer.model_max_length
-                        break
-                if rbln_dec_max_seq_len is None:
-                    raise ValueError("`rbln_dec_max_seq_len` should be specified!")
-
-        if max_position_embeddings is not None and rbln_dec_max_seq_len > max_position_embeddings:
-            raise ValueError("`rbln_dec_max_seq_len` should be less or equal than max_position_embeddings!")
+        pad_token_id = getattr(model_config, "pad_token_id", None)
+        pad_token_id = pad_token_id or getattr(model_config, "bos_token_id", None)
+        pad_token_id = pad_token_id or getattr(model_config, "eos_token_id", None)
+        pad_token_id = pad_token_id or -1
+        rbln_config.pad_token_id = pad_token_id
+
+        if rbln_config.enc_max_seq_len is None:
+            enc_max_seq_len = max_position_embeddings
+            for tokenizer in preprocessors:
+                if hasattr(tokenizer, "model_max_length"):
+                    enc_max_seq_len = enc_max_seq_len or tokenizer.model_max_length
+                    break
+
+            if enc_max_seq_len is None:
+                raise ValueError("`enc_max_seq_len` should be specified!")
+            rbln_config.enc_max_seq_len = enc_max_seq_len
+
+        if max_position_embeddings is not None and rbln_config.enc_max_seq_len > max_position_embeddings:
+            raise ValueError("`enc_max_seq_len` should be less or equal than max_position_embeddings!")
+
+        if rbln_config.dec_max_seq_len is None:
+            dec_max_seq_len = max_position_embeddings
+            for tokenizer in preprocessors:
+                if hasattr(tokenizer, "model_max_length"):
+                    dec_max_seq_len = dec_max_seq_len or tokenizer.model_max_length
+                    break
+
+            if dec_max_seq_len is None:
+                raise ValueError("`dec_max_seq_len` should be specified!")
+            rbln_config.dec_max_seq_len = dec_max_seq_len
+
+        if max_position_embeddings is not None and rbln_config.dec_max_seq_len > max_position_embeddings:
+            raise ValueError("`dec_max_seq_len` should be less or equal than max_position_embeddings!")
 
         # model input info
         enc_input_info = [
-            ("input_ids", [1, rbln_enc_max_seq_len], "int64"),
-            ("attention_mask", [1, rbln_enc_max_seq_len], "float32"),
-            (
-                "cross_key_value_states",
-                [
-                    n_layer * 2,
-                    rbln_batch_size,
-                    n_head,
-                    rbln_enc_max_seq_len,
-                    d_kv,
-                ],
-                "float32",
-            ),
+            ("input_ids", [1, rbln_config.enc_max_seq_len], "int64"),
+            ("attention_mask", [1, rbln_config.enc_max_seq_len], "float32"),
             ("block_tables", [1], "int16"),
         ]
+        enc_input_info.extend(
+            [
+                (
+                    f"cross_key_value_states_{i}",
+                    [
+                        rbln_config.batch_size,
+                        n_head,
+                        rbln_config.enc_max_seq_len,
+                        d_kv,
+                    ],
+                    "float32",
+                )
+                for i in range(n_layer * 2)
+            ]
+        )
 
         dec_input_info = [
-            ("input_ids", [rbln_batch_size, 1], "int64"),
-            ("encoder_attention_mask", [rbln_batch_size, rbln_enc_max_seq_len], "float32"),
+            ("input_ids", [rbln_config.batch_size, 1], "int64"),
+            ("encoder_attention_mask", [rbln_config.batch_size, rbln_config.enc_max_seq_len], "float32"),
             (
                 "cache_position",
-                [rbln_batch_size, 1],
+                [rbln_config.batch_size, 1],
                 "int32",
             ),
-            (
-                "block_tables",
-                [rbln_batch_size, 1],
-                "int16",
-            ),
+            ("block_tables", [rbln_config.batch_size, 1], "int16"),
         ]
         dec_input_info.extend(
             [
                 (
-                    "cross_key_value_states",
+                    f"cross_key_value_states_{i}",
                     [
-                        n_layer * 2,
-                        rbln_batch_size,
+                        rbln_config.batch_size,
                         n_head,
-                        rbln_enc_max_seq_len,
+                        rbln_config.enc_max_seq_len,
                         d_kv,
                     ],
                     "float32",
                 )
+                for i in range(n_layer * 2)
             ]
         )
         dec_input_info.extend(
@@ -291,9 +286,9 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
                 (
                     f"self_key_value_states_{i}",
                     [
-                        rbln_batch_size,
+                        rbln_config.batch_size,
                         n_head,
-                        rbln_dec_max_seq_len,
+                        rbln_config.dec_max_seq_len,
                         d_kv,
                     ],
                     "float32",
@@ -302,46 +297,38 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
             ]
         )
 
-        if rbln_use_attention_mask:
-            dec_input_info.insert(1, ("attention_mask", [rbln_batch_size, rbln_dec_max_seq_len], "float32"))
+        if rbln_config.use_attention_mask:
+            dec_input_info.insert(
+                1, ("attention_mask", [rbln_config.batch_size, rbln_config.dec_max_seq_len], "float32")
+            )
 
         enc_compile_config = RBLNCompileConfig(compiled_model_name="encoder", input_info=enc_input_info)
         dec_compile_config = RBLNCompileConfig(compiled_model_name="decoder", input_info=dec_input_info)
 
-        rbln_config = RBLNConfig(
-            rbln_cls=cls.__name__,
-            compile_cfgs=[enc_compile_config, dec_compile_config],
-            rbln_kwargs=rbln_kwargs,
-        )
-
-        rbln_config.model_cfg.update(
-            {
-                "enc_max_seq_len": rbln_enc_max_seq_len,
-                "dec_max_seq_len": rbln_dec_max_seq_len,
-                "batch_size": rbln_batch_size,
-                "pad_token_id": rbln_pad_token_id,
-                "use_attention_mask": rbln_use_attention_mask,
-            }
-        )
-
+        rbln_config.set_compile_cfgs([enc_compile_config, dec_compile_config])
         return rbln_config
 
     @classmethod
     def _create_runtimes(
         cls,
         compiled_models: List[rebel.RBLNCompiledModel],
-        rbln_device_map: Dict[str, int],
-        activate_profiler: Optional[bool] = None,
+        rbln_config: RBLNModelForSeq2SeqLMConfig,
     ) -> List[rebel.Runtime]:
-        if any(model_name not in rbln_device_map for model_name in ["encoder", "decoder"]):
+        if any(model_name not in rbln_config.device_map for model_name in ["encoder", "decoder"]):
             cls._raise_missing_compiled_file_error(["encoder", "decoder"])
 
         return [
-            compiled_models[0].create_runtime(
-                tensor_type="pt", device=rbln_device_map["encoder"], activate_profiler=activate_profiler
+            rebel.Runtime(
+                compiled_models[0],
+                tensor_type="pt",
+                device=rbln_config.device_map["encoder"],
+                activate_profiler=rbln_config.activate_profiler,
             ),
-            compiled_models[1].create_runtime(
-                tensor_type="pt", device=rbln_device_map["decoder"], activate_profiler=activate_profiler
+            rebel.Runtime(
+                compiled_models[1],
+                tensor_type="pt",
+                device=rbln_config.device_map["decoder"],
+                activate_profiler=rbln_config.activate_profiler,
             ),
         ]
 
@@ -363,7 +350,7 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
     ):
         cur_seq_len = input_ids.shape[-1]
         cache_position = cur_seq_len - 1
-        max_seq_len = self.rbln_config.model_cfg["dec_max_seq_len"]
+        max_seq_len = self.rbln_config.dec_max_seq_len
         decoder_batch_size = input_ids.shape[0]
         input_ids = input_ids[:, cur_seq_len - 1 : cur_seq_len].contiguous()
         decoder_attention_mask = torch.zeros(decoder_batch_size, max_seq_len, dtype=torch.float32)
@@ -383,7 +370,7 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
         **kwargs,
     ) -> Tuple[torch.FloatTensor]:
         # common decoder
-        cache_position = torch.full((self.rbln_config.model_cfg["batch_size"], 1), cache_position, dtype=torch.int32)
+        cache_position = torch.full((self.rbln_config.batch_size, 1), cache_position, dtype=torch.int32)
         logits = self.decoder(decoder_input_ids=decoder_input_ids, cache_position=cache_position, **kwargs).logits
 
         return Seq2SeqLMOutput(
@@ -417,11 +404,11 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
         batch_size, input_len = inputs_tensor.shape
         inputs_tensor = torch.nn.functional.pad(
             inputs_tensor,
-            (0, self.rbln_config.model_cfg["enc_max_seq_len"] - input_len),
-            value=self.rbln_config.model_cfg["pad_token_id"],
+            (0, self.rbln_config.enc_max_seq_len - input_len),
+            value=self.rbln_config.pad_token_id,
         )
         model_kwargs["attention_mask"] = torch.nn.functional.pad(
-            model_kwargs["attention_mask"], (0, self.rbln_config.model_cfg["enc_max_seq_len"] - input_len)
+            model_kwargs["attention_mask"], (0, self.rbln_config.enc_max_seq_len - input_len)
         )
 
         # 3. make sure that encoder returns `ModelOutput`