optimum-rbln 0.7.2rc1__py3-none-any.whl → 0.7.3__py3-none-any.whl

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_cache_update
+from ....ops import register_rbln_custom_add_softmax_attention, register_rbln_custom_cache_update
 
 
 logger = logging.get_logger(__name__)
@@ -34,6 +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()
         self.encoder = WhisperEncoderWrapper(model)
         self.decoder = WhisperDecoderWrapper(model, output_attentions=rbln_token_timestamps)
 
@@ -77,9 +78,9 @@ class WhisperEncoderWrapper(torch.nn.Module):
         # 3. update cross_attention's past_key_value to the device-dram for optimization.
         bidx = torch.tensor(0, dtype=torch.int16)
         axis = torch.tensor(1, dtype=torch.int16)
-        cross_key_values = torch.ops.rbln_custom_ops.rbln_cache_update(cross_key_values, cross_kv, bidx, axis)
+        enc_output = torch.ops.rbln_custom_ops.rbln_cache_update(cross_key_values, cross_kv, bidx, axis)
 
-        return cross_key_values
+        return enc_output
 
 
 class WhisperDecoderWrapper(torch.nn.Module):
@@ -118,7 +119,7 @@ class WhisperDecoderWrapper(torch.nn.Module):
             cross_past_key_values = cross_past_key_values + ((cross_kv_cache[i], cross_kv_cache[i + 1]),)
 
         # Decode
-        sequence_output, self_present_key_values, cross_attentions = self.decoder(
+        sequence_output, cross_attentions = self.decoder(
             input_ids=decoder_input_ids,
             attention_mask=decoder_attention_mask,
             cache_position=cache_position,
@@ -127,9 +128,7 @@ class WhisperDecoderWrapper(torch.nn.Module):
         )
 
         lm_logits = self.proj_out(sequence_output)
-
         outputs = (lm_logits,)
-        outputs += self_present_key_values
 
         if self.output_attentions:
             # deocder's cross attention is used for token_timestamps
@@ -167,26 +166,23 @@ class WhisperDecoder(nn.Module):
         # prepare casual_attn_mask
         attention_mask = _prepare_4d_causal_attention_mask(attention_mask, input_shape, inputs_embeds, cache_position)
 
-        self_present_key_values = ()
         cross_attentions = ()
         # iterate decoder_layer
         for self_past_key_value, cross_past_key_value, decoder_layer in zip(
             self_past_key_values, cross_past_key_values, self.layers
         ):
-            layer_outputs = decoder_layer(
+            hidden_states, cross_attn_weights = decoder_layer(
                 hidden_states,
                 attention_mask=attention_mask,
                 self_past_key_value=self_past_key_value,
                 cross_past_key_value=cross_past_key_value,
                 cache_position=cache_position,
             )
-            hidden_states = layer_outputs[0]
-            self_present_key_values += layer_outputs[1]
-            cross_attentions += (layer_outputs[2],)
+            cross_attentions += (cross_attn_weights,)
 
         hidden_states = self.layer_norm(hidden_states)
 
-        return hidden_states, self_present_key_values, cross_attentions
+        return hidden_states, cross_attentions
 
 
 class WhisperDecoderLayer(nn.Module):
@@ -213,7 +209,7 @@ class WhisperDecoderLayer(nn.Module):
         # Self Attention Block
         residual = hidden_states
         hidden_states = self.self_attn_layer_norm(hidden_states)
-        hidden_states, _, self_present_key_value = self.self_attn(
+        hidden_states = self.self_attn(
             hidden_states=hidden_states,
             past_key_value=self_past_key_value,
             attention_mask=attention_mask,
@@ -224,7 +220,7 @@ class WhisperDecoderLayer(nn.Module):
         # Cross-Attention Block
         residual = hidden_states
         hidden_states = self.encoder_attn_layer_norm(hidden_states)
-        hidden_states, cross_attn_weights, cross_present_key_value = self.encoder_attn(
+        hidden_states, cross_attn_weights = self.encoder_attn(
             hidden_states=hidden_states,
             past_key_value=cross_past_key_value,
         )
@@ -237,7 +233,7 @@ class WhisperDecoderLayer(nn.Module):
         hidden_states = self.fc2(hidden_states)
         hidden_states = residual + hidden_states
 
-        return hidden_states, self_present_key_value, cross_attn_weights
+        return hidden_states, cross_attn_weights
 
 
 class WhisperAttention(nn.Module):
@@ -258,19 +254,8 @@ class WhisperAttention(nn.Module):
 
 
 class WhisperSelfAttention(WhisperAttention):
-    def rbln_cache_update(
-        self,
-        past_key_value: torch.Tensor,
-        key_states: torch.Tensor,
-        value_states: torch.Tensor,
-        cache_position: torch.Tensor,
-    ):
-        s_idx = torch.tensor(cache_position, dtype=torch.int16)
-        axis = torch.tensor(2, dtype=torch.int16)
-
-        key_states = torch.ops.rbln_custom_ops.rbln_cache_update(past_key_value[0], key_states, s_idx, axis)
-        value_states = torch.ops.rbln_custom_ops.rbln_cache_update(past_key_value[1], value_states, s_idx, axis)
-        return key_states, value_states
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int) -> torch.Tensor:
+        return tensor.view(bsz, seq_len, 1, self.num_heads, self.head_dim).transpose(1, 3)
 
     def forward(
         self,
@@ -285,22 +270,27 @@ 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)
-        key_states, value_states = self.rbln_cache_update(past_key_value, key_states, value_states, cache_position)
 
-        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
-        attn_weights = attn_weights + attention_mask
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+        attn_output = torch.ops.rbln_custom_ops.add_softmax_attn_decode(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask.unsqueeze(2),
+            past_key_value[0].view(bsz, self.num_heads, 1, -1, self.head_dim),
+            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
+        )
 
-        attn_output = torch.matmul(attn_weights, value_states)
         attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
         attn_output = attn_output.transpose(1, 2)
         attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
         attn_output = self.out_proj(attn_output)
 
-        return attn_output, attn_weights, (key_states, value_states)
+        return attn_output
 
 
-class WhisperCrossAttention(WhisperSelfAttention):
+class WhisperCrossAttention(WhisperAttention):
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -322,4 +312,4 @@ class WhisperCrossAttention(WhisperSelfAttention):
         attn_output = attn_output.reshape(batch_size, query_len, self.embed_dim)
         attn_output = self.out_proj(attn_output)
 
-        return attn_output, attn_weights, (key_states, value_states)
+        return attn_output, attn_weights

optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py

@@ -15,7 +15,6 @@
 import inspect
 from typing import TYPE_CHECKING, Optional, Union
 
-import torch
 from transformers import PretrainedConfig
 
 from ....modeling import RBLNModel
@@ -58,7 +57,7 @@ class RBLNXLMRobertaModel(RBLNModel):
         if max_position_embeddings is not None and rbln_max_seq_len > max_position_embeddings:
             raise ValueError("`rbln_enc_max_seq_len` should be less or equal than max_position_embeddings!")
 
-        signature_params = inspect.signature(cls.hf_class.forward).parameters.keys()
+        signature_params = inspect.signature(cls.get_hf_class().forward).parameters.keys()
 
         if rbln_model_input_names is None:
             for tokenizer in preprocessors:
@@ -99,15 +98,3 @@ class RBLNXLMRobertaModel(RBLNModel):
         )
         rbln_config.model_cfg.update({"max_seq_len": rbln_max_seq_len})
         return rbln_config
-
-    def forward(
-        self,
-        input_ids: "torch.Tensor",
-        attention_mask: "torch.Tensor",
-        token_type_ids: "torch.Tensor" = None,
-        **kwargs,
-    ):
-        if token_type_ids is None:
-            token_type_ids = torch.zeros_like(input=input_ids, dtype=torch.int64)
-        output = super().forward(input_ids, attention_mask, token_type_ids)
-        return output

optimum/rbln/utils/import_utils.py

@@ -28,6 +28,13 @@ class VersionCompat:
 
 
 RBLN_VERSION_COMPATS = {
+    "0.7.2": [
+        VersionCompat(
+            package_name="rebel-compiler",
+            min_version="0.7.2",
+            max_version="0.7.3",
+        ),
+    ],
     "0.2.0": [
         VersionCompat(
             package_name="rebel-compiler",

{optimum_rbln-0.7.2rc1.dist-info → optimum_rbln-0.7.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.2rc1
+Version: 0.7.3
 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-0.7.2rc1.dist-info → optimum_rbln-0.7.3.dist-info}/RECORD

@@ -1,31 +1,33 @@
-optimum/rbln/__init__.py,sha256=eHi15YM3989AcX52jka9rUmgAtlp1PHqMNwBEdOfuu8,6554
-optimum/rbln/__version__.py,sha256=TkyQ16Pgq_vGa41fqX3hjKTP9w2uZMVmlSqV9zpFgkU,414
-optimum/rbln/modeling.py,sha256=REImAAKO82CqSNABR-9E1jJEsWch9amSOwOOQhFEYLY,8283
-optimum/rbln/modeling_base.py,sha256=fQ0bI1Bb6GJquRXftmSSN9K-TXLhFltZJ6C-2w43xMg,21193
+optimum/rbln/__init__.py,sha256=ZDzXcl-oAcYJhKjJMpotjbTih9awo7HzUb6T3MUEP6Q,6894
+optimum/rbln/__version__.py,sha256=HuvheJ09dqzSRYh_uHb_Af7PEi7XiVW5Tk0cjtnAa6I,511
+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
-optimum/rbln/diffusers/__init__.py,sha256=pOyoXv3-JRzTBSwPKbgLS9H6F2K9dJdReEmpGhcLQYU,3283
-optimum/rbln/diffusers/modeling_diffusers.py,sha256=2m5gXEF7vYWzu3z7BxdDqOJ6i1aakDfWJov66h76noI,20733
+optimum/rbln/diffusers/__init__.py,sha256=Hq87CbtiCy85YmK2SB-OmUyfv77oe3j4bsTenTRnu6w,3623
+optimum/rbln/diffusers/modeling_diffusers.py,sha256=IS6Mlgexofap7f9Lefk5cKFP7ejSG_oWN3v2PX9_IDQ,20118
 optimum/rbln/diffusers/models/__init__.py,sha256=mkCvJyH1KcwrsUvYSq_bVC79oOfyqtBSFDyPS1_48wA,1478
 optimum/rbln/diffusers/models/controlnet.py,sha256=EM_HlzCdaZdnnK0oGpY2fQeigPqHhlwh4NHCzlmoumI,10512
 optimum/rbln/diffusers/models/autoencoders/__init__.py,sha256=dg17ZTUsiqTcbIaEE4fqew9uRbao0diQ21PXvRKIqKg,679
 optimum/rbln/diffusers/models/autoencoders/autoencoder_kl.py,sha256=rCbC32bJnfXtsLdVvNVVHpRAkCYy6jeCSwIZ-JSReWk,9220
 optimum/rbln/diffusers/models/autoencoders/vae.py,sha256=gB9HR7Bf7wpIXLv-Js4Pc3oyWRlqEe4cms4sI2AJicY,4380
-optimum/rbln/diffusers/models/autoencoders/vq_model.py,sha256=GunIau02_-lodYZBzd0ktJSNRT5axEFIZxSAfj2Mlyo,5974
+optimum/rbln/diffusers/models/autoencoders/vq_model.py,sha256=b36QqPbayjApKivceQVVyQxHyR1ZOZ1ffuGgdALEPTQ,6117
 optimum/rbln/diffusers/models/transformers/__init__.py,sha256=V8rSR7WzHs-i8Cwb_MNxhY2NFbwPgxu24vGtkwl-6tk,706
 optimum/rbln/diffusers/models/transformers/prior_transformer.py,sha256=VG9cQo-_eppDvQSW1q1euAGBt1socUHetN_fIN2u1iU,6169
 optimum/rbln/diffusers/models/transformers/transformer_sd3.py,sha256=n_krmMgiRxWrG--567PNpk58EG_X7x7H4gidIkRvwjo,7308
 optimum/rbln/diffusers/models/unets/__init__.py,sha256=MaICuK9CWjgzejXy8y2NDrphuEq1rkzanF8u45k6O5I,655
-optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=xHnBzFrm7aNaolxrsotbjo9GkbNiNdTleXQoeqGLlhg,15540
-optimum/rbln/diffusers/pipelines/__init__.py,sha256=DAsM4eNks3hEY-bsUKSxRKmgwUWDGDlw82gfplSOdO8,2800
+optimum/rbln/diffusers/models/unets/unet_2d_condition.py,sha256=QIjVWQQf8KBn5rU7lvipdm3gNBxZl7l6HCAj7p5FjLU,15977
+optimum/rbln/diffusers/pipelines/__init__.py,sha256=5KLZ5LrpMzBya2e_3_PvEoPwG24U8JMexfw_ygZREKc,3140
 optimum/rbln/diffusers/pipelines/controlnet/__init__.py,sha256=n1Ef22TSeax-kENi_d8K6wGGHSNEo9QkUeygELHgcao,983
 optimum/rbln/diffusers/pipelines/controlnet/multicontrolnet.py,sha256=JWKtnZYBIfgmbAo0SLFIvHBQCv2BPSFNvpcdjG4GUOY,4113
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet.py,sha256=dGdw5cwJLS4CLv6IHskk5ZCcPgS7UDuHKbfOZ8ojNUs,35187
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_img2img.py,sha256=7xCiXrH4ToCTHohVGFXqO7_f9G8HShYaHgZxoMZARkQ,33664
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl.py,sha256=Gzt2wg4dgFg0TV3Bu0cs8Xru3wVrxWUxxgciwZ-QKLE,44755
 optimum/rbln/diffusers/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py,sha256=RfwxNX_zQWFtvvFQJ5bt3qtHbdYdQV_3XLHm9WYCKOs,46084
-optimum/rbln/diffusers/pipelines/kandinsky2_2/__init__.py,sha256=YFqA76_XiMNxPwqotbHug2kd7jCbOXOu5NlxG2hbaVs,808
-optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py,sha256=9szfe1NvOr1mgDnSPZvBGq1b65RElUrqLVhuErY3Dmw,2962
-optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpaint.py,sha256=WxBbHAZSAKDSWhFerPvUlIhhWEsejW4NmhwmWX-_b54,856
+optimum/rbln/diffusers/pipelines/kandinsky2_2/__init__.py,sha256=I4YQq2HfA3xONbWsdJ870IEJPyLWeCDDG-UCJsu9YO8,1035
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py,sha256=aNFGOjth8tDvPrjYLbRWrkHr6p-8AFgcQx1Qay1fw70,904
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py,sha256=BVXOpdrezWVTCibpuAMu9KkD5oEQUY00cSqm6dFbTnk,7020
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py,sha256=fEs-WgJqWs5zvuCkKb7MuZokH9Mi6q-0DOEKxzfWxzo,932
+optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpaint.py,sha256=Ad2ZYCXaMiYpB0mz-8X1CGhILxrVbt7rRIXt6IPwYBM,932
 optimum/rbln/diffusers/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py,sha256=Mf7tzrXetwCgt7LuXfkX-CX1hltLgNZdwF9bHxAbDJM,874
 optimum/rbln/diffusers/pipelines/stable_diffusion/__init__.py,sha256=gz6CbP4T6w8XH3PIGRIJXTmKFsChJIkwcAEAsiR5Ydg,830
 optimum/rbln/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py,sha256=DgRLzO9HxtgE1jICmHoHaqeVXM4Ih-5uo2JqNMAPMcc,876
@@ -39,38 +41,38 @@ 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=MbAHevg59fXQOFgrnsFFZ97s8-YrgvYCYML_sgKEEfM,816
-optimum/rbln/ops/attn.py,sha256=QYvSMg4ps_PenHwpLVhuYRoOAFvHIo19nY0ZEdj4nTE,9700
-optimum/rbln/ops/flash_attn.py,sha256=Zn5nkouY3kk6MBivQpPjgGh4oepjpi8F3tnTrmrNfpg,2304
-optimum/rbln/ops/kv_cache_update.py,sha256=9W4WCO1Dtfy0u5i978JJRa7uLbqrfR2lHuoPynb07fw,3143
+optimum/rbln/ops/__init__.py,sha256=TxOmsN0u3PmyK4Sb89qbiC4rePOlkvUT7Lm6wVoTnY0,941
+optimum/rbln/ops/attn.py,sha256=3EqU63Oj4zI4rLbkRycorsscXeD-IpKzt9N1MhkMa5o,10374
+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
 optimum/rbln/transformers/modeling_alias.py,sha256=yx7FnZQWAnrWzivaO5hI7T6i-fyLzt2tMIXG2oDNbPo,1657
-optimum/rbln/transformers/modeling_generic.py,sha256=SD7XjpjnCn-ejNAUWgkaaHV6Fv2Y6K-hbXEXXb9W_H4,18177
+optimum/rbln/transformers/modeling_generic.py,sha256=aaZWsqVDCRvH03q-Wen7DMfLr7Gy-u-I0mTw0aYqWjk,18195
 optimum/rbln/transformers/modeling_rope_utils.py,sha256=3zwkhYUyTZhxCJUSmwCc88iiY1TppRWEY9ShwUqNB2k,14293
 optimum/rbln/transformers/models/__init__.py,sha256=zGnYODR-_T65tv6jFjtC8l01LC4vjfm41bM4doCXRvY,3835
 optimum/rbln/transformers/models/auto/__init__.py,sha256=GvGbb3ZpMv-h6euXeZ42jSizoOfrL2O1uvpAnfKxYEo,1034
 optimum/rbln/transformers/models/auto/auto_factory.py,sha256=IK9jFrJ3EEzYQa9_aKpcp2TO68M5YGkA-HcfBVpA2QU,7027
 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=dTkgMpNkyh4vT_mZU5tQ5bvH_lRZfRjaJ1gIHvJkmgs,5479
-optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=ADRbE-5N3xJ60AzzjJ4BZs_THmB71qs4XTr9iFqsEqE,5667
+optimum/rbln/transformers/models/bart/bart_architecture.py,sha256=Oo-Cdne7igKEex8wwP-gztKJHgs5GLHQjK1oc3IZIDE,5801
+optimum/rbln/transformers/models/bart/modeling_bart.py,sha256=6IpWXlBCd02v66KF77oEWfrv8-FnPBYjjjL_8KZL3Ow,5835
 optimum/rbln/transformers/models/bert/__init__.py,sha256=YVV7k_laU6yJBawZrgjIWjRmIF-Y4oQQHqyf8lsraQs,691
-optimum/rbln/transformers/models/bert/modeling_bert.py,sha256=-nv-sgmHkyHQIoQvF8-lXOJiL4eaa1pq8MpdN4uRi9M,4668
+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
 optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=NiSm7bHs4SReHDUr53BBWSX0Y8bkKOeUSpsBDrp8YDw,6628
 optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=pDogsdpJKKB5rqnVFrRjwfhUvOSV-jZ3oARMsqSvOOQ,665
-optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=eT1fbKDL92yGBXtUKA_JibD4kiRPdf3tAFJHP5nlfH4,36646
-optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=2OO8MEgFgcl1VPrQXxqkvmRJJEuFdexwu8XqbHDbR6Y,27609
+optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=m93-qKN7NMw3i0XDmFmttmRIRK4np_fWtLFlBb2RFgU,41351
+optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=uGdPGcFrWm2gAwFLjfBiALwFsl49VGCReVi4NUfOPxM,38898
 optimum/rbln/transformers/models/dpt/__init__.py,sha256=gP1tkR3XMNlHq1GT87ugIVvb2o_1eAUg1JaniXjy1Lw,651
 optimum/rbln/transformers/models/dpt/modeling_dpt.py,sha256=ZsS2SOiqcA4azULB-WFEMQZbgIoOyVUKqVKqrw_tWzA,3430
 optimum/rbln/transformers/models/exaone/__init__.py,sha256=zYH_5tVa8-juEdsOIky7I33WSC3Zuhoq1upI0OHYeVw,859
-optimum/rbln/transformers/models/exaone/exaone_architecture.py,sha256=thzWLVz3eUcst4IPiOavta5QeXZw7JQwwfdIzQ_x6Ns,3029
+optimum/rbln/transformers/models/exaone/exaone_architecture.py,sha256=ZM5vvz8KBipOiMVi8vqfvejkDSknW69xh4GrvJix-g0,3350
 optimum/rbln/transformers/models/exaone/modeling_exaone.py,sha256=WjyH8PmsMljSea7kJn_Cq1FJ96OXwXAoU7hv2Q8zUnI,1747
 optimum/rbln/transformers/models/gemma/__init__.py,sha256=7qUrekuBwCI9a6_Fq6j--FzCirRtUDz3ApY17mQS5Y4,648
-optimum/rbln/transformers/models/gemma/gemma_architecture.py,sha256=TvPWphXWG1snpnKakxQWUQZuST2GGJhCrRdUN1FLSIk,1965
+optimum/rbln/transformers/models/gemma/gemma_architecture.py,sha256=bmCx405FVcffhgrQ53qMMZDbSlPxWOjucMHbvq19Gnw,2286
 optimum/rbln/transformers/models/gemma/modeling_gemma.py,sha256=-U3w3cEOv3ps1S8aL7uOq6Kq2siCPZz7Z8MXhDQgQqo,1530
 optimum/rbln/transformers/models/gpt2/__init__.py,sha256=UwwPPYVTB9ywDWy314L2bNL0i7wfkQFA71qjgXicEPg,646
-optimum/rbln/transformers/models/gpt2/gpt2_architecture.py,sha256=sa4rU6mtLwD9q97EnBPwfPvhN7ZvG44xFtIs4inGL7o,2866
+optimum/rbln/transformers/models/gpt2/gpt2_architecture.py,sha256=1IxqHmB-GlH2Dv2Yk4z0rMxL9CpxMGHhSu_x8_4cxvs,3008
 optimum/rbln/transformers/models/gpt2/modeling_gpt2.py,sha256=qBDanUk_O-HtOIVCA4IE3FYyCsnL9xIDK00vft-0caw,1490
 optimum/rbln/transformers/models/llama/__init__.py,sha256=jo_j_eIrHYGNEhR5lb6g3r5sO0ewe3fm2TWO8mLrT58,648
 optimum/rbln/transformers/models/llama/llama_architecture.py,sha256=S7MCPfyjG5eUqgaS-QNBB0ApUD6wnb5fR0RHq7k7-pA,728
@@ -78,43 +80,43 @@ optimum/rbln/transformers/models/llama/modeling_llama.py,sha256=Z3iony7icoFhRQ11
 optimum/rbln/transformers/models/llava_next/__init__.py,sha256=VLieyWm-UgvuNxw9B38wrL1Jsa09NBDX_ebABmdpTbs,670
 optimum/rbln/transformers/models/llava_next/modeling_llava_next.py,sha256=w_plsUOzxnhkQBhQeUqW9aJqGCvCvLtsx0XNKYjOprU,26203
 optimum/rbln/transformers/models/midm/__init__.py,sha256=UJSaErsF-z6dZERIS143WTaygffZyzEGqoQ2ZPDiM-c,855
-optimum/rbln/transformers/models/midm/midm_architecture.py,sha256=mueRmMGX6UplZb0C0RFdUOa9lsNH8YJHV6rYrDLOdlQ,5302
+optimum/rbln/transformers/models/midm/midm_architecture.py,sha256=357iviqQkzI0s_lU_teH1sVOChNRDUABe3GA0HuhZZY,5444
 optimum/rbln/transformers/models/midm/modeling_midm.py,sha256=GG25BozEZriAL-OPFGpzOjyDtSFB-NfeiLJTDAqxe20,1734
 optimum/rbln/transformers/models/mistral/__init__.py,sha256=jpGdNtRLoV7WmuYpRGVXR27BTC8RIi_nhmvYlxuhqRc,652
 optimum/rbln/transformers/models/mistral/mistral_architecture.py,sha256=_aU8TE_tdvfo0K7QpgTlz_d0qwk4O82dl9268lPL16E,733
 optimum/rbln/transformers/models/mistral/modeling_mistral.py,sha256=7nrddoBIHf8S12LZWBUpotnvG3gND11vMQda9yYXJ-s,1560
 optimum/rbln/transformers/models/phi/__init__.py,sha256=mZLt1M7BbYEvSon5UlkniMUPa15SfjZFdw0kMSAF3VA,644
 optimum/rbln/transformers/models/phi/modeling_phi.py,sha256=j-6Pqd5rR2JE8I1pnKFlCi4nW5Dv3wZjoPWxohissoo,1516
-optimum/rbln/transformers/models/phi/phi_architecture.py,sha256=v1MVf9YGxSV1jJ2w81IVvoz3vxjangZJH5rOZX02pPw,3996
+optimum/rbln/transformers/models/phi/phi_architecture.py,sha256=TueyqmjPXWmOPOxBm4dIFyd0X3iV1jgw0U6c26iCAPk,4090
 optimum/rbln/transformers/models/qwen2/__init__.py,sha256=RAMWc21W_2I6DH9xBjeNxPECmAcTrbKhSIefq3Lass0,648
 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=HG_-8ufRWIls67imU1547V0bk9FUWC0haOBL7eyRV6k,16365
-optimum/rbln/transformers/models/seq2seq/seq2seq_architecture.py,sha256=_TL4-vpjM9lfRnQUXRFm3mtVdz_h5B23k01uc_XnW5I,18376
+optimum/rbln/transformers/models/seq2seq/modeling_seq2seq.py,sha256=9Pf9Y86ABDfhwIenlZqYfgqjbyFmtKBiPnbCD7zxw4M,18017
+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=MFs-3yYviV1QqSpsTB2GarTEs9wGH5AYofksLQLMBXg,8043
-optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=kkjErS42mW2jv5O_xL7BaKobvvqy7BGmYOowKyHakvI,7189
+optimum/rbln/transformers/models/t5/modeling_t5.py,sha256=8PAhPlYT1dmpcWM7hUMmZV9lPd4d75CuMuFen1pzr3Q,8088
+optimum/rbln/transformers/models/t5/t5_architecture.py,sha256=AArCQhZRETVM583wlIRzMFOSYq7t2nzxaAeyhZxyxKk,9508
 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=eP3UgkwCRaaFjc5Jc4ZEiWxr3-L7oJx9KzpJ7eFkwUs,13158
+optimum/rbln/transformers/models/whisper/whisper_architecture.py,sha256=Yn6yFpmw6IQbWlnpIMAdEUsNF6huXgaKzGMUZbhSLdo,12572
 optimum/rbln/transformers/models/xlm_roberta/__init__.py,sha256=fC7iNcdxBZ_6eOF2snStmf8r2M3c8O_-XcXnQEaHQCE,653
-optimum/rbln/transformers/models/xlm_roberta/modeling_xlm_roberta.py,sha256=lKSeL3RUwIyfuca2jZ6SFV4N59EJS4UD59JMUfh3BiA,4767
+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
 optimum/rbln/transformers/utils/rbln_quantization.py,sha256=gwBVHf97sQgPNmGa0wq87E8mPyrtXYhMnO4X4sKp3c8,7639
 optimum/rbln/utils/__init__.py,sha256=ieDBT2VFTt2E0M4v_POLBpuGW9LxSydpb_DuPd6PQqc,712
 optimum/rbln/utils/decorator_utils.py,sha256=xu-TrsNi33SRC2a7DBsyoo6-pEQxWKZPZSmM9QlDe2Y,3745
 optimum/rbln/utils/hub.py,sha256=bNmOJGEO9Jfux4Cg8Xli-898I4mxk20KuwQOhP0Zs1U,4198
-optimum/rbln/utils/import_utils.py,sha256=ec-tISKIjUPHIfjzj6p-w78NVejWVBohb59f7J-HJvQ,4059
+optimum/rbln/utils/import_utils.py,sha256=n4HcvZPzFW2ytl45qJ4ZQYlrRSoOb0-nnqhyT2_JA8M,4224
 optimum/rbln/utils/logging.py,sha256=VKKBmlQSdg6iZCGmAXaWYiW67K84jyp1QJhLQSSjPPE,3453
 optimum/rbln/utils/model_utils.py,sha256=DfD_Z2qvZHqcddXqnzTM1AN8khanj3-DXK2lJvVxDvs,1278
 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-0.7.2rc1.dist-info/METADATA,sha256=s8trW-wJbj_RzACjxvDgwVIwy4DxeTkdY7MB3QLjmhc,5301
-optimum_rbln-0.7.2rc1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.7.2rc1.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.7.2rc1.dist-info/RECORD,,
+optimum_rbln-0.7.3.dist-info/METADATA,sha256=XUq3lQDHCViSK9l_txn0h0psqls1nDN4ZFFJGGpifZ0,5298
+optimum_rbln-0.7.3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.7.3.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.7.3.dist-info/RECORD,,