optimum-rbln 0.7.5a1__py3-none-any.whl → 0.7.5rc1__py3-none-any.whl

optimum/rbln/__init__.py

@@ -38,6 +38,7 @@ _import_structure = {
         "RBLNAutoModelForCTC",
         "RBLNAutoModelForDepthEstimation",
         "RBLNAutoModelForImageClassification",
+        "RBLNAutoModelForImageTextToText",
         "RBLNAutoModelForMaskedLM",
         "RBLNAutoModelForQuestionAnswering",
         "RBLNAutoModelForSeq2SeqLM",
@@ -78,6 +79,10 @@ _import_structure = {
         "RBLNExaoneForCausalLMConfig",
         "RBLNGemmaForCausalLM",
         "RBLNGemmaForCausalLMConfig",
+        "RBLNGemma3ForCausalLM",
+        "RBLNGemma3ForCausalLMConfig",
+        "RBLNGemma3ForConditionalGeneration",
+        "RBLNGemma3ForConditionalGenerationConfig",
         "RBLNGPT2LMHeadModel",
         "RBLNGPT2LMHeadModelConfig",
         "RBLNIdefics3VisionTransformer",
@@ -259,6 +264,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForCTC,
         RBLNAutoModelForDepthEstimation,
         RBLNAutoModelForImageClassification,
+        RBLNAutoModelForImageTextToText,
         RBLNAutoModelForMaskedLM,
         RBLNAutoModelForQuestionAnswering,
         RBLNAutoModelForSeq2SeqLM,
@@ -297,6 +303,10 @@ if TYPE_CHECKING:
         RBLNDPTForDepthEstimationConfig,
         RBLNExaoneForCausalLM,
         RBLNExaoneForCausalLMConfig,
+        RBLNGemma3ForCausalLM,
+        RBLNGemma3ForCausalLMConfig,
+        RBLNGemma3ForConditionalGeneration,
+        RBLNGemma3ForConditionalGenerationConfig,
         RBLNGemmaForCausalLM,
         RBLNGemmaForCausalLMConfig,
         RBLNGPT2LMHeadModel,

optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.5a1'
-__version_tuple__ = version_tuple = (0, 7, 5, 'a1')
+__version__ = version = '0.7.5rc1'
+__version_tuple__ = version_tuple = (0, 7, 5, 'rc1')

optimum/rbln/diffusers/models/transformers/prior_transformer.py

@@ -58,8 +58,7 @@ class _PriorTransformer(torch.nn.Module):
 class RBLNPriorTransformer(RBLNModel):
     hf_library_name = "diffusers"
     auto_model_class = PriorTransformer
-    output_class = PriorTransformerOutput
-    output_key = "predicted_image_embedding"
+    _output_class = PriorTransformerOutput
 
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)

optimum/rbln/diffusers/models/transformers/transformer_sd3.py

@@ -61,8 +61,7 @@ class SD3Transformer2DModelWrapper(torch.nn.Module):
 class RBLNSD3Transformer2DModel(RBLNModel):
     hf_library_name = "diffusers"
     auto_model_class = SD3Transformer2DModel
-    output_class = Transformer2DModelOutput
-    output_key = "sample"
+    _output_class = Transformer2DModelOutput
 
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)

optimum/rbln/diffusers/models/unets/unet_2d_condition.py

@@ -143,8 +143,7 @@ class RBLNUNet2DConditionModel(RBLNModel):
     hf_library_name = "diffusers"
     auto_model_class = UNet2DConditionModel
     _rbln_config_class = RBLNUNet2DConditionModelConfig
-    output_class = UNet2DConditionOutput
-    output_key = "sample"
+    _output_class = UNet2DConditionOutput
 
     def __post_init__(self, **kwargs):
         super().__post_init__(**kwargs)

optimum/rbln/modeling.py

@@ -14,7 +14,7 @@
 
 from pathlib import Path
 from tempfile import TemporaryDirectory
-from typing import TYPE_CHECKING, Dict, List, Optional, Union
+from typing import TYPE_CHECKING, Dict, List, Optional, Union, get_args, get_origin, get_type_hints
 
 import rebel
 import torch
@@ -49,8 +49,7 @@ class RBLNModel(RBLNBaseModel):
         ```
     """
 
-    output_class = None
-    output_key = "last_hidden_state"
+    _output_class = None
 
     @classmethod
     def update_kwargs(cls, kwargs):
@@ -245,16 +244,61 @@ class RBLNModel(RBLNBaseModel):
         # Format output according to task requirements
         return self._prepare_output(output, return_dict)
 
+    @classmethod
+    def get_hf_output_class(cls):
+        """
+        Dynamically gets the output class from the corresponding HuggingFace model class.
+
+        Returns:
+            type: The appropriate output class from transformers or diffusers
+        """
+        if cls._output_class:
+            return cls._output_class
+
+        hf_class = cls.get_hf_class()
+        if hf_class is None:
+            raise ValueError(f"No HuggingFace model class found for {cls.__name__}")
+
+        hints = get_type_hints(hf_class.forward) if hasattr(hf_class, "forward") else {}
+        ret = hints.get("return")
+
+        if ret is not None:
+            candidates = get_args(ret) if get_origin(ret) is Union else (ret,)
+
+            for t in candidates:
+                if t is type(None):  # Skip NoneType in Union
+                    continue
+                mod = getattr(t, "__module__", "")
+                if "transformers" in mod or "diffusers" in mod:
+                    cls._output_class = t
+                    return t
+
+        # Fallback to BaseModelOutput
+        cls._output_class = BaseModelOutput
+        return BaseModelOutput
+
     def _prepare_output(self, output, return_dict):
         """
         Prepare model output based on return_dict flag.
         This method can be overridden by subclasses to provide task-specific output handling.
         """
+        tuple_output = (output,) if not isinstance(output, (tuple, list)) else output
         if not return_dict:
-            return (output,) if not isinstance(output, (tuple, list)) else output
+            return tuple_output
         else:
-            if self.output_class is None:
-                return BaseModelOutput(last_hidden_state=output)
-
-            # Create output with the appropriate class and key
-            return self.output_class(**{self.output_key: output})
+            output_class = self.get_hf_output_class()
+            if hasattr(output_class, "loss"):
+                tuple_output = (None,) + tuple_output
+
+            # Truncate if we have too many outputs, otherwise use as is
+            if hasattr(output_class, "__annotations__"):
+                num_fields = len(output_class.__annotations__)
+                if len(tuple_output) > num_fields:
+                    tuple_output = tuple_output[:num_fields]
+                    logger.warning(
+                        f"Truncating output to {num_fields} fields for {output_class.__name__}. "
+                        f"Expected {num_fields} fields, but got {len(tuple_output)} fields."
+                        "This is unexpected. Please report this issue to the developers."
+                    )
+
+            return output_class(*tuple_output)

optimum/rbln/modeling_base.py

@@ -178,9 +178,27 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
         return str(model_path)
 
     @classmethod
-    def _load_compiled_models(cls, model_path: str):
+    def _load_compiled_models(cls, model_path: str, expected_compiled_model_names: List[str]):
         compiled_models = Path(model_path).glob("*.rbln")
-        rbln_compiled_models = {cm.stem: rebel.RBLNCompiledModel(cm) for cm in compiled_models}
+        expected_compiled_models = [
+            Path(model_path) / f"{compiled_model_name}.rbln" for compiled_model_name in expected_compiled_model_names
+        ]
+        unexpected_compiled_models = [cm for cm in compiled_models if cm not in expected_compiled_models]
+        if unexpected_compiled_models:
+            # TODO(jongho): fix after May release. raise error if unexpected compiled models are found
+            logger.warning(
+                f"Unexpected compiled models found: {[cm.name for cm in unexpected_compiled_models]}. "
+                f"Please check the model path: {model_path}"
+            )
+
+        rbln_compiled_models = {}
+        for compiled_model in expected_compiled_models:
+            if not compiled_model.exists():
+                raise FileNotFoundError(
+                    f"Expected RBLN compiled model '{compiled_model.name}' not found at '{model_path}'. "
+                    "Please ensure all models specified in `rbln_config` are present."
+                )
+            rbln_compiled_models[compiled_model.stem] = rebel.RBLNCompiledModel(compiled_model)
         return rbln_compiled_models
 
     @classmethod
@@ -271,7 +289,8 @@ class RBLNBaseModel(SubModulesMixin, PushToHubMixin, PreTrainedModel):
                     )
                     config = PretrainedConfig(**config)
 
-            rbln_compiled_models = cls._load_compiled_models(model_path_subfolder)
+            compiled_model_names = [cfg.compiled_model_name for cfg in rbln_config.compile_cfgs]
+            rbln_compiled_models = cls._load_compiled_models(model_path_subfolder, compiled_model_names)
 
             if subfolder != "":
                 model_save_dir = Path(model_path_subfolder).absolute().parent

optimum/rbln/transformers/__init__.py

@@ -34,6 +34,7 @@ _import_structure = {
         "RBLNAutoModelForCTC",
         "RBLNAutoModelForDepthEstimation",
         "RBLNAutoModelForImageClassification",
+        "RBLNAutoModelForImageTextToText",
         "RBLNAutoModelForMaskedLM",
         "RBLNAutoModelForQuestionAnswering",
         "RBLNAutoModelForSeq2SeqLM",
@@ -72,6 +73,10 @@ _import_structure = {
         "RBLNExaoneForCausalLMConfig",
         "RBLNGemmaForCausalLM",
         "RBLNGemmaForCausalLMConfig",
+        "RBLNGemma3ForCausalLM",
+        "RBLNGemma3ForCausalLMConfig",
+        "RBLNGemma3ForConditionalGeneration",
+        "RBLNGemma3ForConditionalGenerationConfig",
         "RBLNGPT2LMHeadModel",
         "RBLNGPT2LMHeadModelConfig",
         "RBLNIdefics3VisionTransformer",
@@ -148,6 +153,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForCTC,
         RBLNAutoModelForDepthEstimation,
         RBLNAutoModelForImageClassification,
+        RBLNAutoModelForImageTextToText,
         RBLNAutoModelForMaskedLM,
         RBLNAutoModelForQuestionAnswering,
         RBLNAutoModelForSeq2SeqLM,
@@ -184,6 +190,10 @@ if TYPE_CHECKING:
         RBLNDPTForDepthEstimationConfig,
         RBLNExaoneForCausalLM,
         RBLNExaoneForCausalLMConfig,
+        RBLNGemma3ForCausalLM,
+        RBLNGemma3ForCausalLMConfig,
+        RBLNGemma3ForConditionalGeneration,
+        RBLNGemma3ForConditionalGenerationConfig,
         RBLNGemmaForCausalLM,
         RBLNGemmaForCausalLMConfig,
         RBLNGPT2LMHeadModel,

optimum/rbln/transformers/modeling_generic.py

@@ -36,11 +36,7 @@ from transformers import (
 )
 from transformers.modeling_outputs import (
     BaseModelOutput,
-    DepthEstimatorOutput,
-    ImageClassifierOutput,
-    MaskedLMOutput,
     QuestionAnsweringModelOutput,
-    SequenceClassifierOutput,
 )
 
 from ..configuration_utils import RBLNCompileConfig
@@ -63,8 +59,6 @@ class _RBLNTransformerEncoder(RBLNModel):
     auto_model_class = AutoModel
     rbln_model_input_names = ["input_ids", "attention_mask", "token_type_ids"]
     rbln_dtype = "int64"
-    output_class = BaseModelOutput
-    output_key = "last_hidden_state"
 
     @classmethod
     def _update_rbln_config(
@@ -149,7 +143,6 @@ class _RBLNImageModel(RBLNModel):
     auto_model_class = AutoModel
     main_input_name = "pixel_values"
     output_class = BaseModelOutput
-    output_key = "last_hidden_state"
 
     @classmethod
     def _update_rbln_config(
@@ -223,15 +216,11 @@ class RBLNModelForQuestionAnswering(_RBLNTransformerEncoder):
 class RBLNModelForSequenceClassification(_RBLNTransformerEncoder):
     auto_model_class = AutoModelForSequenceClassification
     rbln_model_input_names = ["input_ids", "attention_mask"]
-    output_class = SequenceClassifierOutput
-    output_key = "logits"
 
 
 class RBLNModelForMaskedLM(_RBLNTransformerEncoder):
     auto_model_class = AutoModelForMaskedLM
     rbln_model_input_names = ["input_ids", "attention_mask"]
-    output_class = MaskedLMOutput
-    output_key = "logits"
 
 
 class RBLNModelForTextEncoding(_RBLNTransformerEncoder):
@@ -243,20 +232,14 @@ class RBLNTransformerEncoderForFeatureExtraction(_RBLNTransformerEncoder):
     # TODO: RBLNModel is also for feature extraction.
     auto_model_class = AutoModel
     rbln_model_input_names = ["input_ids", "attention_mask"]
-    output_class = BaseModelOutput
-    output_key = "last_hidden_state"
 
 
 class RBLNModelForImageClassification(_RBLNImageModel):
     auto_model_class = AutoModelForImageClassification
-    output_class = ImageClassifierOutput
-    output_key = "logits"
 
 
 class RBLNModelForDepthEstimation(_RBLNImageModel):
     auto_model_class = AutoModelForDepthEstimation
-    output_class = DepthEstimatorOutput
-    output_key = "predicted_depth"
 
 
 class RBLNModelForAudioClassification(RBLNModel):
@@ -273,8 +256,6 @@ class RBLNModelForAudioClassification(RBLNModel):
     """
 
     auto_model_class = AutoModelForAudioClassification
-    output_class = SequenceClassifierOutput
-    output_key = "logits"
 
     @classmethod
     def _update_rbln_config(

optimum/rbln/transformers/models/__init__.py

@@ -31,6 +31,7 @@ _import_structure = {
         "RBLNAutoModelForSequenceClassification",
         "RBLNAutoModelForSpeechSeq2Seq",
         "RBLNAutoModelForVision2Seq",
+        "RBLNAutoModelForImageTextToText",
     ],
     "bart": [
         "RBLNBartForConditionalGeneration",
@@ -80,6 +81,12 @@ _import_structure = {
     ],
     "exaone": ["RBLNExaoneForCausalLM", "RBLNExaoneForCausalLMConfig"],
     "gemma": ["RBLNGemmaForCausalLM", "RBLNGemmaForCausalLMConfig"],
+    "gemma3": [
+        "RBLNGemma3ForCausalLM",
+        "RBLNGemma3ForCausalLMConfig",
+        "RBLNGemma3ForConditionalGeneration",
+        "RBLNGemma3ForConditionalGenerationConfig",
+    ],
     "gpt2": ["RBLNGPT2LMHeadModel", "RBLNGPT2LMHeadModelConfig"],
     "idefics3": [
         "RBLNIdefics3VisionTransformer",
@@ -121,6 +128,7 @@ if TYPE_CHECKING:
         RBLNAutoModelForCTC,
         RBLNAutoModelForDepthEstimation,
         RBLNAutoModelForImageClassification,
+        RBLNAutoModelForImageTextToText,
         RBLNAutoModelForMaskedLM,
         RBLNAutoModelForQuestionAnswering,
         RBLNAutoModelForSeq2SeqLM,
@@ -170,6 +178,12 @@ if TYPE_CHECKING:
     )
     from .exaone import RBLNExaoneForCausalLM, RBLNExaoneForCausalLMConfig
     from .gemma import RBLNGemmaForCausalLM, RBLNGemmaForCausalLMConfig
+    from .gemma3 import (
+        RBLNGemma3ForCausalLM,
+        RBLNGemma3ForCausalLMConfig,
+        RBLNGemma3ForConditionalGeneration,
+        RBLNGemma3ForConditionalGenerationConfig,
+    )
     from .gpt2 import RBLNGPT2LMHeadModel, RBLNGPT2LMHeadModelConfig
     from .idefics3 import (
         RBLNIdefics3ForConditionalGeneration,

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

@@ -19,6 +19,7 @@ from .modeling_auto import (
     RBLNAutoModelForCTC,
     RBLNAutoModelForDepthEstimation,
     RBLNAutoModelForImageClassification,
+    RBLNAutoModelForImageTextToText,
     RBLNAutoModelForMaskedLM,
     RBLNAutoModelForQuestionAnswering,
     RBLNAutoModelForSeq2SeqLM,

optimum/rbln/transformers/models/auto/modeling_auto.py

@@ -23,6 +23,8 @@ from transformers.models.auto.modeling_auto import (
     MODEL_FOR_DEPTH_ESTIMATION_MAPPING_NAMES,
     MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
     MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
+    MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING,
+    MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES,
     MODEL_FOR_MASKED_LM_MAPPING,
     MODEL_FOR_MASKED_LM_MAPPING_NAMES,
     MODEL_FOR_QUESTION_ANSWERING_MAPPING,
@@ -90,6 +92,11 @@ class RBLNAutoModelForVision2Seq(_BaseAutoModelClass):
     _model_mapping_names = MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES
 
 
+class RBLNAutoModelForImageTextToText(_BaseAutoModelClass):
+    _model_mapping = MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING
+    _model_mapping_names = MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES
+
+
 class RBLNAutoModelForMaskedLM(_BaseAutoModelClass):
     _model_mapping = MODEL_FOR_MASKED_LM_MAPPING
     _model_mapping_names = MODEL_FOR_MASKED_LM_MAPPING_NAMES

optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py

@@ -157,7 +157,11 @@ class DecoderOnlyWrapper(nn.Module):
         self.config = causal_lm.config
 
         if use_rotary_emb:
-            self.rotary_emb = self.get_rotary_emb(max_seq_len=max_seq_len)
+            rotary_embs = self.get_rotary_emb(max_seq_len=max_seq_len)
+            if isinstance(rotary_embs, tuple):
+                self.rotary_emb_global, self.rotary_emb_local = rotary_embs
+            else:
+                self.rotary_emb = rotary_embs
         else:
             self.rotary_emb = None
 
@@ -195,7 +199,10 @@ class DecoderOnlyWrapper(nn.Module):
         for layer in causal_lm.model.layers:
             if self.attn_impl == "eager":
                 new_self_attn = DecoderOnlyAttention(
-                    layer.self_attn, self.use_attention_mask, kvcache_block_size=self.kvcache_block_size
+                    layer.self_attn,
+                    self.use_attention_mask,
+                    self.use_position_ids,
+                    kvcache_block_size=self.kvcache_block_size,
                 )
             elif self.attn_impl == "flash_attn":
                 new_self_attn = DecoderOnlyFlashAttention(
@@ -203,6 +210,7 @@ class DecoderOnlyWrapper(nn.Module):
                     kvcache_partition_len=self.kvcache_partition_len,
                     kvcache_block_size=self.kvcache_block_size,
                     use_attention_mask=self.use_attention_mask,
+                    use_position_ids=self.use_position_ids,
                 )
             else:
                 raise NotImplementedError(f"Unknwon attn : {self.attn_impl}")
@@ -363,6 +371,13 @@ class DecoderOnlyForCausalLM(nn.Module):
             hidden_states = hidden_states[:, query_position.to(torch.int).unsqueeze(0)]
 
         logits = self.lm_head(hidden_states)
+
+        # Apply final logit softmaxing if configured, e.g. for Gemma2
+        if getattr(self.config, "final_logit_softcapping", None) is not None:
+            logits = logits / self.config.final_logit_softcapping
+            logits = torch.tanh(logits)
+            logits = logits * self.config.final_logit_softcapping
+
         return logits
 
 
@@ -610,7 +625,7 @@ class DecoderOnlyAttention(nn.Module):
         self_attn: Original attention module from the base model
     """
 
-    def __init__(self, self_attn, use_attention_mask, kvcache_block_size):
+    def __init__(self, self_attn, use_attention_mask, use_position_ids, kvcache_block_size):
         super().__init__()
         self._original_mod = self_attn
         self.layer_idx = self_attn.layer_idx
@@ -629,6 +644,7 @@ class DecoderOnlyAttention(nn.Module):
             self.num_key_value_heads = self.num_heads
 
         self.use_attention_mask = use_attention_mask
+        self.use_position_ids = use_position_ids
         self.attention = self.get_attention()
         self.kvcache_block_size = kvcache_block_size
         self.__post_init__()
@@ -643,7 +659,9 @@ class DecoderOnlyAttention(nn.Module):
         self.attention.phase = phase
 
     def get_attention(self):
-        return AttentionOp(self.num_heads, self.head_dim, self.num_key_value_heads, self.use_attention_mask)
+        return AttentionOp(
+            self.num_heads, self.head_dim, self.num_key_value_heads, self.use_attention_mask, self.use_position_ids
+        )
 
     def __post_init__(self):
         self.q_proj = self._original_mod.q_proj
@@ -716,13 +734,16 @@ class DecoderOnlyAttention(nn.Module):
 
 
 class AttentionOp(nn.Module):
-    def __init__(self, num_heads: int, head_dim: int, num_key_value_heads: int, use_attention_mask: bool):
+    def __init__(
+        self, num_heads: int, head_dim: int, num_key_value_heads: int, use_attention_mask: bool, use_position_ids: bool
+    ):
         super().__init__()
         self.num_heads = num_heads
         self.head_dim = head_dim
         self.num_key_value_heads = num_key_value_heads
         self.phase = "prefill"
         self.use_attention_mask = use_attention_mask
+        self.use_position_ids = use_position_ids
 
     def forward(
         self,
@@ -755,7 +776,8 @@ class AttentionOp(nn.Module):
         # reshape for removing repeat_kv (batch=1 , num_head, 1, q_len=1, head_dim)
         key_state = key_state.unsqueeze(2)  # 1, 32, 1, 128, 128
         value_state = value_state.unsqueeze(2)
-        if self.use_attention_mask:
+
+        if self.use_attention_mask and not self.use_position_ids:
             attn_mask = attn_mask.unsqueeze(2)
 
         if self.phase == "decode":
@@ -772,7 +794,7 @@ class AttentionOp(nn.Module):
         )
 
         if self.phase == "decode":
-            if self.use_attention_mask:
+            if self.use_attention_mask and not self.use_position_ids:
                 attn_output = torch.ops.rbln_custom_ops.paged_attn_decode(
                     q=query_state,
                     k=key_state,
@@ -796,11 +818,11 @@ class AttentionOp(nn.Module):
                     scale=scale,
                     block_table=block_tables,
                     block_size=block_size,
-                    mask=None,
+                    mask=attn_mask if self.use_position_ids else None,
                 )
 
         else:
-            if self.use_attention_mask:
+            if self.use_attention_mask and not self.use_position_ids:
                 attn_output = torch.ops.rbln_custom_ops.paged_attn_prefill(
                     q=query_state,
                     k=key_state,
@@ -824,8 +846,8 @@ class AttentionOp(nn.Module):
                     scale=scale,
                     block_table=block_tables,
                     block_size=block_size,
-                    is_bidirectional=False,
-                    mask=None,
+                    is_bidirectional=True if self.phase == "image_prefill" else False,  # FIXME, Hard-coded for Gemma3.
+                    mask=attn_mask if self.use_position_ids else None,
                 )
 
         attn_output = attn_output.view(batch_size, self.num_heads, -1, self.head_dim)
@@ -927,10 +949,13 @@ class RotaryEmbedding(nn.Module):
 
 
 class DecoderOnlyFlashAttention(DecoderOnlyAttention):
-    def __init__(self, self_attn, kvcache_partition_len, kvcache_block_size, use_attention_mask):
+    def __init__(self, self_attn, kvcache_partition_len, kvcache_block_size, use_attention_mask, use_position_ids):
         self.kvcache_partition_size = kvcache_partition_len
         super().__init__(
-            self_attn=self_attn, use_attention_mask=use_attention_mask, kvcache_block_size=kvcache_block_size
+            self_attn=self_attn,
+            use_attention_mask=use_attention_mask,
+            use_position_ids=use_position_ids,
+            kvcache_block_size=kvcache_block_size,
         )
 
     def get_attention(self):
@@ -940,6 +965,7 @@ class DecoderOnlyFlashAttention(DecoderOnlyAttention):
             self.num_key_value_heads,
             self.kvcache_partition_size,
             self.use_attention_mask,
+            self.use_position_ids,
         )
 
     def forward(
@@ -991,12 +1017,14 @@ class FlashAttentionOp(AttentionOp):
         num_key_value_heads: int,
         kvcache_partition_len: int,
         use_attention_mask: bool,
+        use_position_ids: bool,
     ):
         super().__init__(
             num_heads=num_heads,
             head_dim=head_dim,
             num_key_value_heads=num_key_value_heads,
             use_attention_mask=use_attention_mask,
+            use_position_ids=use_position_ids,
         )
         self.kvcache_partition_size = kvcache_partition_len
 
@@ -1016,7 +1044,7 @@ class FlashAttentionOp(AttentionOp):
         # 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)
-        if self.use_attention_mask:
+        if self.use_attention_mask and not self.use_position_ids:
             attn_mask = attn_mask.unsqueeze(2)
 
         if self.phase == "decode":
@@ -1033,7 +1061,7 @@ class FlashAttentionOp(AttentionOp):
         )
 
         if self.phase == "decode":
-            if self.use_attention_mask:
+            if self.use_attention_mask and not self.use_position_ids:
                 attn_output = torch.ops.rbln_custom_ops.paged_flash_attn_decode(
                     q=query_state,
                     k=key_state,
@@ -1059,10 +1087,10 @@ class FlashAttentionOp(AttentionOp):
                     block_table=block_tables,
                     block_size=kvcache_block_size,
                     partition=self.kvcache_partition_size,
-                    mask=None,
+                    mask=attn_mask if self.use_position_ids else None,
                 )
         else:
-            if self.use_attention_mask:
+            if self.use_attention_mask and not self.use_position_ids:
                 attn_output = torch.ops.rbln_custom_ops.paged_flash_attn_prefill(
                     q=query_state,
                     k=key_state,
@@ -1088,8 +1116,8 @@ class FlashAttentionOp(AttentionOp):
                     block_table=block_tables,
                     block_size=kvcache_block_size,
                     partition=self.kvcache_partition_size,
-                    is_bidirectional=False,
-                    mask=None,
+                    is_bidirectional=True if self.phase == "image_prefill" else False,
+                    mask=attn_mask if self.use_position_ids else None,
                 )
 
         # reshape for removing repeat_kv
@@ -1098,3 +1126,70 @@ class FlashAttentionOp(AttentionOp):
         attn_output = attn_output.reshape(batch_size, -1, self.num_heads * self.head_dim)
 
         return attn_output
+
+
+class SlidingWindowAttentionOp(AttentionOp):
+    def forward(
+        self,
+        query_state: torch.Tensor,
+        key_state: torch.Tensor,
+        value_state: torch.Tensor,
+        attn_mask: torch.Tensor,
+        past_key_state: torch.Tensor,
+        past_value_state: torch.Tensor,
+        seq_position: Tuple[torch.Tensor],
+        scale: torch.Tensor,
+        block_tables: torch.Tensor,
+        block_size: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        # 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)
+
+        if self.phase == "decode":
+            batch_size = key_state.shape[0]
+        else:
+            batch_size = 1
+
+        query_state = query_state.view(
+            batch_size,
+            self.num_key_value_heads,
+            self.num_heads // self.num_key_value_heads,
+            -1,  # seq len
+            self.head_dim,
+        )
+
+        if self.phase == "decode":
+            attn_output = torch.ops.rbln_custom_ops.paged_sliding_window_attn_decode(
+                q=query_state,
+                k=key_state,
+                v=value_state,
+                kcache=past_key_state.unsqueeze(2),
+                vcache=past_value_state.unsqueeze(2),
+                cache_seq_len=seq_position[0],
+                cache_offset=seq_position[1],
+                scale=scale,
+                block_table=block_tables,
+                block_size=block_size,
+            )
+        else:
+            attn_output = torch.ops.rbln_custom_ops.paged_sliding_window_attn_prefill(
+                q=query_state,
+                k=key_state,
+                v=value_state,
+                kcache=past_key_state.unsqueeze(2),
+                vcache=past_value_state.unsqueeze(2),
+                cache_seq_len=seq_position[0],
+                cache_offset=seq_position[1],
+                scale=scale,
+                block_table=block_tables,
+                block_size=block_size,
+                is_bidirectional=True if self.phase == "image_prefill" else False,
+            )
+
+        # reshape for removing repeat_kv
+        attn_output = attn_output.view(batch_size, self.num_heads, -1, self.head_dim)
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, -1, self.num_heads * self.head_dim)
+
+        return attn_output