optimum-rbln 0.7.3.post2__tar.gz → 0.7.4a0__tar.gz

optimum_rbln-0.7.4a0/.github/version.yaml

@@ -0,0 +1 @@
+rebel_compiler_version: 0.7.4.dev22+gcaaa7596

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.3.post2
+Version: 0.7.4a0
 Summary: Optimum RBLN is the interface between the Hugging Face Transformers and Diffusers libraries and RBLN accelerators. It provides a set of tools enabling easy model loading and inference on single and multiple rbln device settings for different downstream tasks.
 Project-URL: Homepage, https://rebellions.ai
 Project-URL: Documentation, https://docs.rbln.ai

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.3.post2'
-__version_tuple__ = version_tuple = (0, 7, 3)
+__version__ = version = '0.7.4a0'
+__version_tuple__ = version_tuple = (0, 7, 4)

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/ops/__init__.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from .attn import (
-    register_rbln_custom_add_softmax_attention,
+    register_rbln_custom_paged_add_softmax_attention,
     register_rbln_custom_paged_attention,
     register_rbln_custom_paged_causal_attention,
 )

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/ops/attn.py

@@ -182,14 +182,14 @@ def register_rbln_custom_paged_causal_attention():
 
 
 @lru_cache
-def register_rbln_custom_add_softmax_attention():
+def register_rbln_custom_paged_add_softmax_attention():
     torch.library.define(
-        "rbln_custom_ops::add_softmax_attn_decode",
-        "(Tensor x, Tensor y, Tensor z, Tensor w, Tensor a, Tensor b, Tensor c, Tensor d) -> Tensor",
+        "rbln_custom_ops::paged_add_softmax_attn_decode",
+        "(Tensor x, Tensor y, Tensor z, Tensor w, Tensor a, Tensor b, Tensor c, Tensor d, Tensor e, int f) -> Tensor",
     )
 
-    @torch.library.impl("rbln_custom_ops::add_softmax_attn_decode", "cpu")
-    def add_softmax_attn_decode_cpu(q, k, v, mask, kcache, vcache, seq, scale):
+    @torch.library.impl("rbln_custom_ops::paged_add_softmax_attn_decode", "cpu")
+    def paged_add_softmax_attn_decode_cpu(q, k, v, mask, kcache, vcache, seq, scale, block_table, block_size):
         """Defines the computation pattern for fused attention with KV cache updates.
 
         IMPORTANT: This op serves as a pattern definition for the RBLN compiler to generate
@@ -210,12 +210,14 @@ def register_rbln_custom_add_softmax_attention():
         - vcache: [batch_size, n_heads, 1, max_seq_len, head_dim] - Value cache
         - seq: [1] - Current sequence position
         - scale: [] - Attention scale factor
+        - block_table: [batch_size, max_seq_len // block_size] - Block indices for KV cache management
+        - block_size: [] - Number of tokens per block
 
         Returns:
             Tensor: attn_output: [batch=1, n_heads, 1, 1, head_dim] - Attention output
         """
         return q
 
-    @register_fake("rbln_custom_ops::add_softmax_attn_decode")
-    def add_softmax_attn_decode_abstract(q, k, v, m, kcache, vcache, seq, partition):
+    @register_fake("rbln_custom_ops::paged_add_softmax_attn_decode")
+    def paged_add_softmax_attn_decode_abstract(q, k, v, m, kcache, vcache, seq, partition, block_table, block_size):
         return q

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/transformers/models/bart/modeling_bart.py

@@ -108,6 +108,8 @@ class RBLNBartModel(RBLNModel):
 
 
 class RBLNBartForConditionalGeneration(RBLNModelForSeq2SeqLM):
+    support_causal_attn = True
+
     @classmethod
     def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
         enc_max_seq_len = (

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py

@@ -94,7 +94,7 @@ 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)
@@ -115,6 +115,7 @@ 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"]
@@ -186,13 +187,16 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
         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
+        if cls.support_causal_attn:
+            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
+        else:
+            rbln_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")
@@ -265,11 +269,7 @@ class RBLNModelForSeq2SeqLM(RBLNModel, ABC):
                 [rbln_batch_size, 1],
                 "int32",
             ),
-            (
-                "block_tables",
-                [rbln_batch_size, 1],
-                "int16",
-            ),
+            ("block_tables", [rbln_batch_size, 1], "int16"),
         ]
         dec_input_info.extend(
             [

optimum_rbln-0.7.4a0/src/optimum/rbln/transformers/models/t5/modeling_t5.py

@@ -0,0 +1,210 @@
+# 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, Dict, Optional, Tuple, Union
+
+import torch
+from transformers import (
+    AutoModelForTextEncoding,
+    PretrainedConfig,
+    T5EncoderModel,
+    T5ForConditionalGeneration,
+)
+from transformers.modeling_outputs import BaseModelOutput
+
+from ....diffusers.modeling_diffusers import RBLNDiffusionMixin
+from ....modeling import RBLNModel
+from ....modeling_config import RBLNCompileConfig, RBLNConfig
+from ....utils.logging import get_logger
+from ....utils.runtime_utils import RBLNPytorchRuntime
+from ...models.seq2seq import RBLNModelForSeq2SeqLM
+from .t5_architecture import T5Wrapper
+
+
+logger = get_logger()
+
+if TYPE_CHECKING:
+    from transformers import AutoFeatureExtractor, AutoProcessor, AutoTokenizer, PreTrainedModel
+
+
+class RBLNRuntimeModel(RBLNPytorchRuntime):
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        attention_mask: torch.FloatTensor,
+        head_mask: torch.FloatTensor,
+        inputs_embeds: torch.FloatTensor,
+        **kwargs,
+    ):
+        return super().forward(
+            input_ids,
+            attention_mask,
+            head_mask,
+            inputs_embeds,
+            **kwargs,
+        )
+
+
+class T5EncoderWrapper(torch.nn.Module):
+    def __init__(self, model: "T5EncoderModel") -> None:
+        super().__init__()
+        self.model = model
+
+    def forward(self, *args, **kwargs):
+        kwargs.pop("return_dict", None)
+        return self.model(*args, **kwargs, return_dict=False)
+
+
+class RBLNT5EncoderModel(RBLNModel):
+    auto_model_class = AutoModelForTextEncoding
+    rbln_model_input_names = ["input_ids", "attention_mask"]
+
+    def __post_init__(self, **kwargs):
+        self.model = RBLNRuntimeModel(runtime=self.model[0])
+
+    @classmethod
+    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
+        return T5EncoderWrapper(model)
+
+    @classmethod
+    def update_rbln_config_using_pipe(cls, pipe: RBLNDiffusionMixin, rbln_config: Dict[str, Any]) -> Dict[str, Any]:
+        batch_size = rbln_config.get("batch_size", 1)
+        max_sequence_length = rbln_config.get("max_sequence_length", 256)
+        model_input_names = ["input_ids"]
+
+        rbln_config.update(
+            {
+                "batch_size": batch_size,
+                "max_seq_len": max_sequence_length,
+                "model_input_names": model_input_names,
+            }
+        )
+
+        return rbln_config
+
+    @classmethod
+    def _get_rbln_config(
+        cls,
+        preprocessors: Optional[Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"]],
+        model_config: Optional["PretrainedConfig"] = None,
+        rbln_kwargs: Dict[str, Any] = {},
+    ) -> RBLNConfig:
+        rbln_max_seq_len = rbln_kwargs.get("max_seq_len", None)
+        rbln_model_input_names = rbln_kwargs.get("model_input_names", None)
+        rbln_batch_size = rbln_kwargs.get("batch_size", None)
+
+        max_position_embeddings = getattr(model_config, "n_positions", None)
+
+        if rbln_max_seq_len is None:
+            rbln_max_seq_len = max_position_embeddings
+            if rbln_max_seq_len is None:
+                for tokenizer in preprocessors:
+                    if hasattr(tokenizer, "model_max_length"):
+                        rbln_max_seq_len = tokenizer.model_max_length
+                        break
+                if rbln_max_seq_len is None:
+                    raise ValueError("`rbln_max_seq_len` should be specified!")
+
+        if max_position_embeddings is not None and rbln_max_seq_len > max_position_embeddings:
+            raise ValueError("`rbln_max_seq_len` should be less or equal than max_position_embeddings!")
+
+        signature_params = inspect.signature(cls.get_hf_class().forward).parameters.keys()
+
+        if rbln_model_input_names is None:
+            for tokenizer in preprocessors:
+                if hasattr(tokenizer, "model_input_names"):
+                    rbln_model_input_names = [name for name in signature_params if name in tokenizer.model_input_names]
+
+                    invalid_params = set(rbln_model_input_names) - set(signature_params)
+                    if invalid_params:
+                        raise ValueError(f"Invalid model input names: {invalid_params}")
+                    break
+            if rbln_model_input_names is None and hasattr(cls, "rbln_model_input_names"):
+                rbln_model_input_names = cls.rbln_model_input_names
+            elif rbln_model_input_names is None and hasattr(cls, "rbln_model_input_names") is False:
+                raise ValueError(
+                    "Specify the model input names obtained by the tokenizer via `rbln_model_input_names`, "
+                    f"and be sure to make the order of the inputs same as T5EncoderModel forward() arguments like ({list(signature_params)})"
+                )
+        else:
+            invalid_params = set(rbln_model_input_names) - set(signature_params)
+            if invalid_params:
+                raise ValueError(f"Invalid model input names: {invalid_params}")
+            rbln_model_input_names = [name for name in signature_params if name in rbln_model_input_names]
+
+        if rbln_batch_size is None:
+            rbln_batch_size = 1
+
+        input_info = [
+            (model_input_name, [rbln_batch_size, rbln_max_seq_len], "int64")
+            for model_input_name in rbln_model_input_names
+        ]
+
+        rbln_compile_config = RBLNCompileConfig(input_info=input_info)
+
+        rbln_config = RBLNConfig(
+            rbln_cls=cls.__name__,
+            compile_cfgs=[rbln_compile_config],
+            rbln_kwargs=rbln_kwargs,
+        )
+
+        rbln_config.model_cfg.update({"max_seq_len": rbln_max_seq_len})
+        return rbln_config
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:
+        encoder_outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        if not return_dict:
+            return (encoder_outputs,)
+        else:
+            return BaseModelOutput(last_hidden_state=encoder_outputs)
+
+
+class RBLNT5ForConditionalGeneration(RBLNModelForSeq2SeqLM):
+    support_causal_attn = False
+
+    @classmethod
+    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
+        enc_max_seq_len = rbln_config.model_cfg["enc_max_seq_len"]
+        dec_max_seq_len = rbln_config.model_cfg["dec_max_seq_len"]
+
+        return T5Wrapper(model, enc_max_seq_len=enc_max_seq_len, dec_max_seq_len=dec_max_seq_len)
+
+    def __getattr__(self, __name: str) -> Any:
+        def redirect(func):
+            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
+
+        val = getattr(T5ForConditionalGeneration, __name)
+
+        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
+            return redirect(val)
+
+        return val

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/transformers/models/t5/t5_architecture.py

@@ -18,7 +18,7 @@ import torch
 from torch import nn
 from transformers.utils import logging
 
-from ....ops import register_rbln_custom_add_softmax_attention
+from ....ops import register_rbln_custom_paged_add_softmax_attention
 from ..seq2seq.seq2seq_architecture import (
     Seq2SeqDecoder,
     Seq2SeqDecoderLayer,
@@ -55,7 +55,7 @@ class T5EncoderWrapper(Seq2SeqEncoderWrapper):
 
 class T5DecoderWrapper(Seq2SeqDecoderWrapper):
     def __post_init__(self, model, dec_max_seq_len: int = None):
-        register_rbln_custom_add_softmax_attention()
+        register_rbln_custom_paged_add_softmax_attention()
         self.num_layers = self.config.num_layers
         self.conditional_generation = self.convert_to_rbln_conditional_generation(model, dec_max_seq_len)
 
@@ -77,6 +77,7 @@ class T5DecoderWrapper(Seq2SeqDecoderWrapper):
         attention_mask,
         encoder_attention_mask,
         cache_position,
+        block_tables,
         cross_kv_cache,
         *self_kv_cache,
     ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor]]:
@@ -95,6 +96,7 @@ class T5DecoderWrapper(Seq2SeqDecoderWrapper):
             self_past_key_values=self_past_key_values,
             cross_past_key_values=cross_past_key_values,
             cache_position=cache_position,
+            block_tables=block_tables,
         )
 
         return lm_logits
@@ -162,7 +164,7 @@ class T5LayerSelfAttention(Seq2SeqSelfAttention):
         self.out_proj = self._original_mod.o
         self.num_heads = self._original_mod.n_heads
         self.head_dim = self._original_mod.key_value_proj_dim
-        self.attn_decode = torch.ops.rbln_custom_ops.add_softmax_attn_decode
+        self.attn_decode = torch.ops.rbln_custom_ops.paged_add_softmax_attn_decode
 
     def projection(self, hidden_states) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
         query_states = self.q_proj(hidden_states)
@@ -176,6 +178,7 @@ class T5LayerSelfAttention(Seq2SeqSelfAttention):
         past_key_value: Tuple[torch.Tensor],
         attention_mask: torch.Tensor,
         cache_position: torch.Tensor,
+        block_tables: torch.Tensor,
         **kwargs,
     ) -> Tuple[torch.Tensor, Tuple[torch.Tensor]]:
         bsz, tgt_len, _ = hidden_states.size()
@@ -185,6 +188,7 @@ class T5LayerSelfAttention(Seq2SeqSelfAttention):
         key_states = self._shape(key_states, -1, bsz)
         value_states = self._shape(value_states, -1, bsz)
 
+        block_size = past_key_value[0].shape[-2]
         attn_output = self.attn_decode(
             query_states,
             key_states,
@@ -196,6 +200,8 @@ class T5LayerSelfAttention(Seq2SeqSelfAttention):
             past_key_value[1].view(bsz, self.num_heads, 1, -1, self.head_dim),
             cache_position,
             torch.tensor(1.0, dtype=torch.float32),  # scale
+            block_tables,
+            block_size,
         )
 
         attn_output = attn_output.view(bsz, self.num_heads, -1, self.head_dim).transpose(1, 2)

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/transformers/models/whisper/modeling_whisper.py

@@ -61,6 +61,16 @@ class RBLNRuntimeEncoder(RBLNPytorchRuntime):
 class RBLNRuntimeDecoder(RBLNPytorchRuntime):
     mandatory_members = ["main_input_name"]
 
+    def __init__(
+        self,
+        runtime: rebel.Runtime,
+        batch_size: int,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(runtime, **kwargs)
+        self.batch_size = batch_size
+        self.default_block_tables = torch.arange(0, self.batch_size, dtype=torch.int16).view(self.batch_size, 1)
+
     def forward(
         self,
         decoder_input_ids: torch.Tensor = None,
@@ -76,6 +86,7 @@ class RBLNRuntimeDecoder(RBLNPytorchRuntime):
             decoder_input_ids=decoder_input_ids,
             decoder_attention_mask=decoder_attention_mask,
             cache_position=cache_position,
+            block_tables=self.default_block_tables,
         )
 
         if isinstance(outputs, torch.Tensor):
@@ -237,6 +248,7 @@ class RBLNWhisperForConditionalGeneration(RBLNModel, RBLNWhisperGenerationMixin)
             ("decoder_input_ids", [rbln_batch_size, 1], "int64"),
             ("decoder_attention_mask", [rbln_batch_size, rbln_dec_max_seq_len], "int64"),
             ("cache_position", [], "int32"),
+            ("block_tables", [rbln_batch_size, 1], "int16"),
         ]
         dec_input_info.extend(
             [

{optimum_rbln-0.7.3.post2 → optimum_rbln-0.7.4a0}/src/optimum/rbln/transformers/models/whisper/whisper_architecture.py

@@ -25,7 +25,7 @@ from transformers.modeling_outputs import (
 )
 from transformers.utils import logging
 
-from ....ops import register_rbln_custom_add_softmax_attention, register_rbln_custom_cache_update
+from ....ops import register_rbln_custom_cache_update, register_rbln_custom_paged_add_softmax_attention
 
 
 logger = logging.get_logger(__name__)
@@ -34,7 +34,7 @@ logger = logging.get_logger(__name__)
 class WhisperWrapper:
     def __init__(self, model, rbln_token_timestamps):
         register_rbln_custom_cache_update()
-        register_rbln_custom_add_softmax_attention()
+        register_rbln_custom_paged_add_softmax_attention()
         self.encoder = WhisperEncoderWrapper(model)
         self.decoder = WhisperDecoderWrapper(model, output_attentions=rbln_token_timestamps)
 
@@ -108,6 +108,7 @@ class WhisperDecoderWrapper(torch.nn.Module):
         decoder_input_ids: torch.Tensor,
         decoder_attention_mask: torch.Tensor,
         cache_position: torch.Tensor,
+        block_tables: torch.Tensor,
         cross_kv_cache: torch.Tensor,
         *self_kv_cache: torch.Tensor,
     ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
@@ -125,6 +126,7 @@ class WhisperDecoderWrapper(torch.nn.Module):
             cache_position=cache_position,
             self_past_key_values=self_past_key_values,
             cross_past_key_values=cross_past_key_values,
+            block_tables=block_tables,
         )
 
         lm_logits = self.proj_out(sequence_output)
@@ -154,6 +156,7 @@ class WhisperDecoder(nn.Module):
         self_past_key_values: Optional[torch.Tensor] = None,
         cross_past_key_values: Optional[torch.Tensor] = None,
         cache_position: Optional[torch.Tensor] = None,
+        block_tables: Optional[torch.Tensor] = None,
     ):
         input_shape = input_ids.size()
         input_ids = input_ids.view(-1, input_shape[-1])
@@ -177,6 +180,7 @@ class WhisperDecoder(nn.Module):
                 self_past_key_value=self_past_key_value,
                 cross_past_key_value=cross_past_key_value,
                 cache_position=cache_position,
+                block_tables=block_tables,
             )
             cross_attentions += (cross_attn_weights,)
 
@@ -205,6 +209,7 @@ class WhisperDecoderLayer(nn.Module):
         self_past_key_value: Optional[Tuple[torch.Tensor]] = None,
         cross_past_key_value: Optional[Tuple[torch.Tensor]] = None,
         cache_position: Optional[torch.Tensor] = None,
+        block_tables: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         # Self Attention Block
         residual = hidden_states
@@ -214,6 +219,7 @@ class WhisperDecoderLayer(nn.Module):
             past_key_value=self_past_key_value,
             attention_mask=attention_mask,
             cache_position=cache_position,
+            block_tables=block_tables,
         )
         hidden_states = residual + hidden_states
 
@@ -263,6 +269,7 @@ class WhisperSelfAttention(WhisperAttention):
         past_key_value: Optional[Tuple[torch.Tensor]] = None,
         attention_mask: Optional[torch.Tensor] = None,
         cache_position: Optional[torch.Tensor] = None,
+        block_tables: Optional[torch.Tensor] = None,
     ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
         bsz, tgt_len, _ = hidden_states.size()
         query_states = self._shape(self.q_proj(hidden_states), tgt_len, bsz)
@@ -270,8 +277,9 @@ class WhisperSelfAttention(WhisperAttention):
 
         key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
         value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+        block_size = past_key_value[0].shape[-2]
 
-        attn_output = torch.ops.rbln_custom_ops.add_softmax_attn_decode(
+        attn_output = torch.ops.rbln_custom_ops.paged_add_softmax_attn_decode(
             query_states,
             key_states,
             value_states,
@@ -280,6 +288,8 @@ class WhisperSelfAttention(WhisperAttention):
             past_key_value[1].view(bsz, self.num_heads, 1, -1, self.head_dim),
             cache_position.expand(bsz, 1),
             torch.tensor(1.0, dtype=torch.float32),  # scale
+            block_tables,
+            block_size,
         )
 
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)

optimum_rbln-0.7.3.post2/.github/version.yaml

@@ -1 +0,0 @@
-rebel_compiler_version: 0.7.3