optimum-rbln 0.7.3.post2__py3-none-any.whl → 0.7.4a0__py3-none-any.whl

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/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/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/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/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/transformers/models/t5/modeling_t5.py

@@ -13,9 +13,8 @@
 # limitations under the License.
 
 import inspect
-from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Any, Callable, Dict, Optional, Tuple, Union
 
-import rebel
 import torch
 from transformers import (
     AutoModelForTextEncoding,
@@ -23,7 +22,7 @@ from transformers import (
     T5EncoderModel,
     T5ForConditionalGeneration,
 )
-from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
+from transformers.modeling_outputs import BaseModelOutput
 
 from ....diffusers.modeling_diffusers import RBLNDiffusionMixin
 from ....modeling import RBLNModel
@@ -58,63 +57,6 @@ class RBLNRuntimeModel(RBLNPytorchRuntime):
         )
 
 
-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]))
-
-
-class RBLNRuntimeDecoder(RBLNPytorchRuntime):
-    mandatory_members = ["main_input_name"]
-
-    def __init__(
-        self,
-        runtime: rebel.Runtime,
-        batch_size: int,
-        dec_max_seq_len: int,
-        **kwargs: Any,
-    ) -> None:
-        super().__init__(runtime, **kwargs)
-        self.batch_size = batch_size
-        self.dec_max_seq_len = dec_max_seq_len
-
-    def forward(
-        self,
-        decoder_input_ids: Optional[torch.LongTensor] = None,
-        attention_mask: Optional[torch.FloatTensor] = None,
-        decoder_attention_mask: Optional[torch.BoolTensor] = None,
-        cache_position: Optional[torch.Tensor] = None,
-        **kwargs,
-    ) -> Tuple[torch.FloatTensor]:
-        batch_size = decoder_input_ids.shape[0]
-        if batch_size != self.batch_size:
-            raise RuntimeError(
-                f"Batch size mismatch: got {batch_size}, expected {self.batch_size} (compiled batch size)."
-            )
-
-        if batch_size != cache_position.shape[0]:
-            raise RuntimeError(f"Cache position size mismatch: got {cache_position.shape[0]}, expected {batch_size}.")
-
-        for b_idx in range(self.batch_size):
-            decoding_step = cache_position[b_idx].item()
-            if not (0 <= decoding_step < self.dec_max_seq_len):
-                raise ValueError(
-                    f"Decoding step {decoding_step} out of bounds for attention mask with shape {self.dec_attn_mask.shape}."
-                )
-            decoder_attention_mask[b_idx, : decoding_step + 1] = 1
-
-        lm_logits = super().forward(
-            decoder_input_ids,
-            decoder_attention_mask,
-            attention_mask,
-            cache_position,
-        )
-
-        return Seq2SeqLMOutput(logits=lm_logits)
-
-
 class T5EncoderWrapper(torch.nn.Module):
     def __init__(self, model: "T5EncoderModel") -> None:
         super().__init__()
@@ -247,20 +189,7 @@ class RBLNT5EncoderModel(RBLNModel):
 
 
 class RBLNT5ForConditionalGeneration(RBLNModelForSeq2SeqLM):
-    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.encoder = RBLNRuntimeEncoder(
-            runtime=self.model[0],
-            main_input_name="input_ids",
-        )
-        self.decoder = RBLNRuntimeDecoder(
-            runtime=self.model[1],
-            main_input_name="input_ids",
-            batch_size=batch_size,
-            dec_max_seq_len=dec_max_seq_len,
-        )
+    support_causal_attn = False
 
     @classmethod
     def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
@@ -279,139 +208,3 @@ class RBLNT5ForConditionalGeneration(RBLNModelForSeq2SeqLM):
             return redirect(val)
 
         return val
-
-    @classmethod
-    def _get_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
-
-        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")
-        d_kv = (
-            model_config.d_kv
-            if hasattr(model_config, "d_kv")
-            else model_config.d_model // model_config.encoder_attention_heads
-        )
-
-        max_position_embeddings = getattr(model_config, "n_positions", None) or getattr(
-            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!")
-
-        # 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",
-            ),
-            ("block_tables", [1], "int16"),
-        ]
-
-        dec_input_info = [
-            ("input_ids", [rbln_batch_size, 1], "int64"),
-            ("attention_mask", [rbln_batch_size, rbln_dec_max_seq_len], "float32"),
-            ("encoder_attention_mask", [rbln_batch_size, rbln_enc_max_seq_len], "float32"),
-            (
-                "cache_position",
-                [rbln_batch_size, 1],
-                "int32",
-            ),
-        ]
-        dec_input_info.extend(
-            [
-                (
-                    "cross_key_value_states",
-                    [
-                        n_layer * 2,
-                        rbln_batch_size,
-                        n_head,
-                        rbln_enc_max_seq_len,
-                        d_kv,
-                    ],
-                    "float32",
-                )
-            ]
-        )
-        dec_input_info.extend(
-            [
-                (
-                    f"self_key_value_states_{i}",
-                    [
-                        rbln_batch_size,
-                        n_head,
-                        rbln_dec_max_seq_len,
-                        d_kv,
-                    ],
-                    "float32",
-                )
-                for i in range(n_layer * 2)
-            ]
-        )
-
-        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,
-            }
-        )
-
-        return rbln_config

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/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/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.dist-info → optimum_rbln-0.7.4a0.dist-info}/METADATA

@@ -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.dist-info → optimum_rbln-0.7.4a0.dist-info}/RECORD

@@ -1,5 +1,5 @@
 optimum/rbln/__init__.py,sha256=ZDzXcl-oAcYJhKjJMpotjbTih9awo7HzUb6T3MUEP6Q,6894
-optimum/rbln/__version__.py,sha256=OJRzB6Y7xaNgH7EkerbumPEoY0Nlzs1_xYhBJvOXTzQ,517
+optimum/rbln/__version__.py,sha256=xyj1Oj5eR1yz0oBU9FRdubMKrBiNrPrrW8h8ohd1iG8,513
 optimum/rbln/modeling.py,sha256=nJsAs5zs--VVOYGFjYNpqfxYIemJIK4Lr0WEzlDLdP0,8390
 optimum/rbln/modeling_base.py,sha256=dNCL-BhrWCpuOVkZaj8-MW567Tf4lLo3p3Z3ldjWJfU,21779
 optimum/rbln/modeling_config.py,sha256=7104bxmrvKW4Q6XTruQayiIGl8GHDFmPkJ3cknMIInE,11335
@@ -41,8 +41,8 @@ optimum/rbln/diffusers/pipelines/stable_diffusion_xl/__init__.py,sha256=9iIMZYvp
 optimum/rbln/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py,sha256=OvB5bxX6HUiqJeIc3uukuEmUXYEx1pTqGNOtdG2l1m8,902
 optimum/rbln/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py,sha256=3aB1Rw-OgKytQOHwOaShbEvq_XVHPOGvsGm8pstEmKU,930
 optimum/rbln/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py,sha256=MzVP1wscaO1sUIiBIPJqG6zuGyez9VUbA42-JSIm-mk,930
-optimum/rbln/ops/__init__.py,sha256=TxOmsN0u3PmyK4Sb89qbiC4rePOlkvUT7Lm6wVoTnY0,941
-optimum/rbln/ops/attn.py,sha256=3EqU63Oj4zI4rLbkRycorsscXeD-IpKzt9N1MhkMa5o,10374
+optimum/rbln/ops/__init__.py,sha256=Wv2cJhEw8mqc6-To24bHzf4qQL8gM0Zh_2Ck77LB65g,947
+optimum/rbln/ops/attn.py,sha256=OSgPoEgCwvR7HdjbnaVkFVMBcJ5RpRWcE6OCg2lVyGk,10634
 optimum/rbln/ops/flash_attn.py,sha256=wfyiCxDGf034IngzwRU160R7_DlKYpd-uWT0BDEGFks,3408
 optimum/rbln/ops/kv_cache_update.py,sha256=pxf8kAptPaQF5xE8qItvmlFOq_sgim6ZERD7AVaOtec,3221
 optimum/rbln/transformers/__init__.py,sha256=AGo3BqVIZrsOzYsQAnnQ25HCstTPBclrXbvvUxVMlqE,4255
@@ -55,7 +55,7 @@ optimum/rbln/transformers/models/auto/auto_factory.py,sha256=IK9jFrJ3EEzYQa9_aKp
 optimum/rbln/transformers/models/auto/modeling_auto.py,sha256=Un9qoqdy3dO8JBza_bTJF_6_fRVNM9QisihSgTRFI-o,3933
 optimum/rbln/transformers/models/bart/__init__.py,sha256=32HPe0_GIO0hp9U464Iv6Jd7M-1nop9g8hA1UZMHhyw,674
 optimum/rbln/transformers/models/bart/bart_architecture.py,sha256=Oo-Cdne7igKEex8wwP-gztKJHgs5GLHQjK1oc3IZIDE,5801
-optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=iI3ubPOVvHmhLt0wEz_vkOfMyNTHVNjmnkLtbpOX760,5797
+optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=CUF5PE9TxJxtO1VpuGgeKrL_u6PdsKxstlZDthYSXgU,5829
 optimum/rbln/transformers/models/bert/__init__.py,sha256=YVV7k_laU6yJBawZrgjIWjRmIF-Y4oQQHqyf8lsraQs,691
 optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=p3utRqf3dv9_RkHwaMCa1EfXttNJkqCJUIZo3CeZ9YY,4674
 optimum/rbln/transformers/models/clip/__init__.py,sha256=H9vuBwrmFO0-CqZhXUrKF-uQL6igCqMlqrT1X_ELaAI,754
@@ -92,17 +92,17 @@ optimum/rbln/transformers/models/qwen2/__init__.py,sha256=RAMWc21W_2I6DH9xBjeNxP
 optimum/rbln/transformers/models/qwen2/modeling_qwen2.py,sha256=9-aFDvjMzPNUyGOz0qo33RE18bUFGYZ3Wt_68zb5uJY,1530
 optimum/rbln/transformers/models/qwen2/qwen2_architecture.py,sha256=XlNAMYAcDLohnSAhIFGKOPuCB5XLgzYs5ABWdeQSaZs,720
 optimum/rbln/transformers/models/seq2seq/__init__.py,sha256=EmEMV4rOYqKyruX85d0fR73-b8N6BSD6CPcbpYdBuVk,651
-optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=NPfJf9Uk_bYOae7hXGHwteGiWH0va63Z-D93RmAMENg,17611
+optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=QelhuCWEHPL2Ut7fm0gLnzTVveBAaKSNpoa9X1AmwTI,17709
 optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=tvzacIZam1sIr_1BvvZ_fDr8u5dXAiYiynFdX9tArtY,18877
 optimum/rbln/transformers/models/t5/__init__.py,sha256=1skR1RmnG62WTAP3-F5P1x-V_ReFhMyirH3u56vWwvc,675
-optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=nKRR3eH1EAu1YkKvhlqGyTrJXDRd-IWB5LOeG9jrcb4,16021
-optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=AArCQhZRETVM583wlIRzMFOSYq7t2nzxaAeyhZxyxKk,9508
+optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=-fG-h0wwsfjZ3par0QHbXKA7hbvw_lPJOIf8iXQDOfM,8082
+optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=Ups6drBbYe4wEAiBLcBIyO9wqrIQbvOPFR_ybbAgR8c,9722
 optimum/rbln/transformers/models/wav2vec2/__init__.py,sha256=YpgA0K-vyg9veh0eL_jxauosbRpb_kpGKHvvQLBspKM,649
 optimum/rbln/transformers/models/wav2vec2/modeling_wav2vec2.py,sha256=JYJmV52j6cBwim4RanVJryfKnV80V96ol0A-oR6o7cg,3856
 optimum/rbln/transformers/models/whisper/__init__.py,sha256=ktnNe5ri3ycCWZ_W_voFB9y9-vgGgxS1X9s8LBRZmWc,665
 optimum/rbln/transformers/models/whisper/generation_whisper.py,sha256=GIHTca3b1VtW81kp7BzKQ7f77c2t9OsEsbZetripgDo,4582
-optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=0nBADNxE0A1ozBbRutTBvxpo_Y1qkOycT_zronkN-ZU,15840
-optimum/rbln/transformers/models/whisper/whisper_architecture.py,sha256=Yn6yFpmw6IQbWlnpIMAdEUsNF6huXgaKzGMUZbhSLdo,12572
+optimum/rbln/transformers/models/whisper/modeling_whisper.py,sha256=U9zK49DcSdXuoK_UOsVPsyKe6EJ5CQR8QZhpgi23EUU,16275
+optimum/rbln/transformers/models/whisper/whisper_architecture.py,sha256=ArQPOgiRVu-XddEN5FXVl1OlCoGF6uY7jGoWTj3Nfe4,13005
 optimum/rbln/transformers/models/xlm_roberta/__init__.py,sha256=fC7iNcdxBZ_6eOF2snStmf8r2M3c8O_-XcXnQEaHQCE,653
 optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py,sha256=8YNLz0bc5ze-QuU8rN-QhUfGzlSUs3iMJiWTxO3o6AM,4366
 optimum/rbln/transformers/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -116,7 +116,7 @@ optimum/rbln/utils/model_utils.py,sha256=DfD_Z2qvZHqcddXqnzTM1AN8khanj3-DXK2lJvV
 optimum/rbln/utils/runtime_utils.py,sha256=5-DYniyP59nx-mrrbi7AqA77L85b4Cm5oLpaxidSyss,3699
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=oZoGrItB8WqY4i-K9WJPlLlcLohc1YGB9OHB8_XZw3A,4071
-optimum_rbln-0.7.3.post2.dist-info/METADATA,sha256=YgOp5SEpJ_VfYEohAoBhSQ20TaX1usvkRAzV7s7mS5I,5304
-optimum_rbln-0.7.3.post2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.7.3.post2.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.7.3.post2.dist-info/RECORD,,
+optimum_rbln-0.7.4a0.dist-info/METADATA,sha256=tXU0EmgjFJug_Cvmw8S9NeEZ2z9XpgamFwgMQTTCa1U,5300
+optimum_rbln-0.7.4a0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.7.4a0.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.7.4a0.dist-info/RECORD,,