optimum-rbln 0.8.1a2__py3-none-any.whl → 0.8.1a3__py3-none-any.whl

optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.8.1a2'
-__version_tuple__ = version_tuple = (0, 8, 1, 'a2')
+__version__ = version = '0.8.1a3'
+__version_tuple__ = version_tuple = (0, 8, 1, 'a3')

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

@@ -177,8 +177,8 @@ class RBLNRuntimeModel(RBLNPytorchRuntime):
                 )
             elif block_tables is None and local_block_tables is None:
                 return False
-        else:
-            return True
+
+        return True
 
     def forward(
         self,

optimum/rbln/transformers/models/gemma3/modeling_gemma3.py

@@ -559,7 +559,7 @@ class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
         (
             inputs,
             cache_position,
-            chunked_attention_mask,
+            padded_attention_mask,
             out_buffers,
             position_ids,
             position_embed,
@@ -571,7 +571,7 @@ class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
         )
         if not is_external_block_tables:
             local_block_tables = torch.tensor([batch_idx], dtype=torch.int16)
-            self.dec_attn_mask[batch_idx : batch_idx + 1] = chunked_attention_mask[:1]
+            self.dec_attn_mask[batch_idx : batch_idx + 1] = padded_attention_mask[:1]
 
         if self.rbln_config.use_attention_mask and self.rbln_config.use_position_ids:
             chunked_attention_mask = torch.zeros(1, self.rbln_config.max_seq_len, dtype=torch.float32)
@@ -587,18 +587,10 @@ class RBLNGemma3RuntimeModel(RBLNRuntimeModel):
                 else None
             )
 
-            # Not used in Gemma3 yet.
             if self.rbln_config.use_attention_mask:
                 if self.rbln_config.use_position_ids:
-                    chunked_attention_mask[0, step : step + self.rbln_config.prefill_chunk_size] = self.dec_attn_mask[
-                        batch_idx, step : step + self.rbln_config.prefill_chunk_size
-                    ]
-                else:
-                    # Update attention mask to ensure proper causal behavior
-                    if step >= self.rbln_config.prefill_chunk_size:
-                        chunked_attention_mask[:, :, :, step - self.rbln_config.prefill_chunk_size : step] = 1
-                    chunked_attention_mask[:, :, :, step : step + self.rbln_config.prefill_chunk_size] = (
-                        self.causal_mask
+                    chunked_attention_mask[0, step : step + self.rbln_config.prefill_chunk_size] = (
+                        padded_attention_mask[0, step : step + self.rbln_config.prefill_chunk_size]
                     )
 
             # Define query position

optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -28,7 +28,6 @@ from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
     Qwen2_5_VisionPatchEmbed,
     Qwen2_5_VisionRotaryEmbedding,
     Qwen2_5_VisionTransformerPretrainedModel,
-    Qwen2_5_VLModel,
     Qwen2_5_VLRotaryEmbedding,
 )
 
@@ -391,14 +390,6 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
     def can_generate(self):
         return True
 
-    @classmethod
-    def get_pytorch_model(cls, *args, **kwargs):
-        model = super().get_pytorch_model(*args, **kwargs)
-        model.model.lm_head = model.lm_head
-        model.lm_head = None
-        del model.lm_head
-        return model
-
     @classmethod
     def update_kwargs(cls, kwargs):
         kwargs.update(
@@ -540,8 +531,7 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
             vision_tokens = input_id[0][vision_start_indices + 1]
             image_nums = (vision_tokens == image_token_id).sum()
             video_nums = (vision_tokens == video_token_id).sum()
-            position_ids, rope_deltas = Qwen2_5_VLModel.get_rope_index(
-                self,
+            position_ids, rope_deltas = self.get_rope_index(
                 input_id,
                 image_grid_thw[image_idx : image_idx + image_nums] if image_grid_thw is not None else None,
                 video_grid_thw[video_idx : video_idx + video_nums] if video_grid_thw is not None else None,

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

@@ -3,14 +3,8 @@ from typing import Tuple
 
 import torch
 import torch.nn as nn
-from transformers import PreTrainedModel
 
 from ..decoderonly.decoderonly_architecture import (
-    DecoderOnlyAttention,
-    DecoderOnlyFlashAttention,
-    DecoderOnlyForCausalLM,
-    DecoderOnlyLayer,
-    DecoderOnlyModel,
     DecoderOnlyWrapper,
     apply_rotary_pos_emb,
 )
@@ -203,40 +197,3 @@ class Qwen2_5_VL_LanguageModelWrapper(DecoderOnlyWrapper):
             past_key_values,
             position_embeds,
         )
-
-    def convert_to_rbln_causal_lm(self, causal_lm: PreTrainedModel, max_seq_len: int):
-        new_layers = []
-
-        for layer in causal_lm.model.language_model.layers:
-            if self.attn_impl == "eager":
-                new_self_attn = DecoderOnlyAttention(
-                    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(
-                    layer.self_attn,
-                    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}")
-
-            new_layer = DecoderOnlyLayer(layer, new_self_attn)
-            new_layers.append(new_layer)
-
-        new_model = DecoderOnlyModel(
-            causal_lm.model.language_model,
-            new_layers,
-            partition_len=self.kvcache_partition_len,
-            max_seq_len=max_seq_len,
-            kvcache_block_size=self.kvcache_block_size,
-            use_learned_pos_emb=self.use_learned_pos_emb,
-            sliding_window_layers=self.sliding_window_layers,
-        )
-        new_causal_lm = DecoderOnlyForCausalLM(causal_lm.model, new_model)
-        return new_causal_lm

{optimum_rbln-0.8.1a2.dist-info → optimum_rbln-0.8.1a3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.1a2
+Version: 0.8.1a3
 Summary: Optimum RBLN is the interface between the HuggingFace 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-0.8.1a2.dist-info → optimum_rbln-0.8.1a3.dist-info}/RECORD

@@ -1,5 +1,5 @@
 optimum/rbln/__init__.py,sha256=qJJTumXhoFnawXGpeGJbAm4J4A9FFwD1SQ2MqcKDXoM,14436
-optimum/rbln/__version__.py,sha256=shHuZmlC1NNfixc5X-chXIonN-Rp-nB4KxJDhJNxa3k,519
+optimum/rbln/__version__.py,sha256=bE7uZ_Vr3gK5nEF_YfozhLcF3_Q9HvGMYkpifHovJxI,519
 optimum/rbln/configuration_utils.py,sha256=o5oer7fBdE-MHLGNXoP35FjmuQbMmjEIDv0QE_k3kpo,32336
 optimum/rbln/modeling.py,sha256=ZlJ_tOCWiFjDIlwJ_B_HOCO0kBduWrBAbW9VSEVIAFg,12088
 optimum/rbln/modeling_base.py,sha256=5fUb1FaxfjApzJIkT8-SrPhuygGo_1Uc0i7UedawOeE,23393
@@ -88,7 +88,7 @@ optimum/rbln/transformers/models/clip/modeling_clip.py,sha256=0u1JTlO47qoH_-qxWG
 optimum/rbln/transformers/models/decoderonly/__init__.py,sha256=vQYZDDdoddwA7yKc5zzrq2Zs9sax-0p8rNF_aYfF4bk,1006
 optimum/rbln/transformers/models/decoderonly/configuration_decoderonly.py,sha256=cakn8RGo8gS3nmXdEqOfC2xUBOMGInROgLEbCOoLFR0,13398
 optimum/rbln/transformers/models/decoderonly/decoderonly_architecture.py,sha256=YAn8J_lIq4IS-HM_gbi5Qov8_osxhWtBr5z_28QRbGM,49667
-optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=cVk7UTXIc5HryADMSJdMN1ENW4ZZM_S5I7RuCk6srEI,53273
+optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py,sha256=J8eJqg56vPcYnnSP3EYU6X1e5BvdRdPyRcWWlcHzK1c,53256
 optimum/rbln/transformers/models/distilbert/__init__.py,sha256=zXL78SOEORTnUN_wrdoaDaYpntG8lcFHvPobM6jC0CI,841
 optimum/rbln/transformers/models/distilbert/configuration_distilbert.py,sha256=qfxCk15hW47i1oO8dCo-xntSbKyW-WOu30h28rIw6eA,766
 optimum/rbln/transformers/models/distilbert/modeling_distilbert.py,sha256=_Ubhbvrhi7jBC5uS9ITstIAE5VJVwAuDwvQ_Hrr6Ny4,797
@@ -106,7 +106,7 @@ optimum/rbln/transformers/models/gemma/modeling_gemma.py,sha256=Ojvum34EhDHWfMB4
 optimum/rbln/transformers/models/gemma3/__init__.py,sha256=6rugk3615SEt4lh7gduo_J9VyGiSReuEIvL0Uno0eaI,790
 optimum/rbln/transformers/models/gemma3/configuration_gemma3.py,sha256=eupMGTHJGJNNrAZ3GE6M6GQBAQzBb7KFJvalyDmbM-A,3063
 optimum/rbln/transformers/models/gemma3/gemma3_architecture.py,sha256=sgFQQbvEr15tb2Sxk_tgcgQFcjhKGbNSW6fm2u7-Vck,8609
-optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=BjICf2xmiUKGkraoGmkTRMIpgqjHrBCiJxc4XgeQw9I,39483
+optimum/rbln/transformers/models/gemma3/modeling_gemma3.py,sha256=9c6-Qz4EGGbSnKwoz2zH5r6W7sVfjb-m5Z-dnQkAOXU,38992
 optimum/rbln/transformers/models/gpt2/__init__.py,sha256=socBMIBZSiLbrVN12rQ4nL9gFeT0axMgz6SWaCaD4Ac,704
 optimum/rbln/transformers/models/gpt2/configuration_gpt2.py,sha256=9sS6-EGapmow3rG9ViejK9qwrqy_X86VBxQ7u9x0Yqk,923
 optimum/rbln/transformers/models/gpt2/gpt2_architecture.py,sha256=pnGgixjgjW7HULbs5211cC2guw_4e4-MlS69vdCRMMg,3206
@@ -143,8 +143,8 @@ optimum/rbln/transformers/models/qwen2/modeling_qwen2.py,sha256=OKd7SXQLLtzPVolr
 optimum/rbln/transformers/models/qwen2/qwen2_architecture.py,sha256=XlNAMYAcDLohnSAhIFGKOPuCB5XLgzYs5ABWdeQSaZs,720
 optimum/rbln/transformers/models/qwen2_5_vl/__init__.py,sha256=rAW3DKQUzGL6EMwa5r1iLu94yhpiZpk6zfoD7TtYXrc,865
 optimum/rbln/transformers/models/qwen2_5_vl/configuration_qwen2_5_vl.py,sha256=U3ngIfkA58itqQZqTf-gbISMPoV7ipDttI7V2uwK_18,4155
-optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=x14mGDbCGVFNTw3yXoRJ70W6U5wAlIe_lxqk190P3z4,26636
-optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py,sha256=G7zAMC2FSqz1dYisBUmBJF5hkyueaOrJSI2gD3nLvCQ,9068
+optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py,sha256=Q4U-avMkby-CunNXEERqvRZx9duC5i-6UmfF1376ciU,26336
+optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py,sha256=PAQz__9o_f5phlozhhXAB8JErBlS1jc4FYZkZkSYJuI,7312
 optimum/rbln/transformers/models/resnet/__init__.py,sha256=0QqtEQF1IMYgEmmfXMGarCDS8kJB5tzODfwTEzDVZRg,837
 optimum/rbln/transformers/models/resnet/configuration_resnet.py,sha256=KQd887jgNOl_Am3b407P2OvKtzkkeBS1cEhCfiN0tJg,769
 optimum/rbln/transformers/models/resnet/modeling_resnet.py,sha256=E8vg3Rw_KsHt6vaOg0ungZD7sXe0T4OMP0X8NFG1EXI,816
@@ -191,7 +191,7 @@ optimum/rbln/utils/model_utils.py,sha256=4k5879Kh75m3x_vS4-qOGfqsOiAvc2kdNFFfvsF
 optimum/rbln/utils/runtime_utils.py,sha256=LoKNK3AQNV_BSScstIZWjICkJf265MnUgy360BOocVI,5454
 optimum/rbln/utils/save_utils.py,sha256=hG5uOtYmecSXZuGTvCXsTM-SiyZpr5q3InUGCCq_jzQ,3619
 optimum/rbln/utils/submodule.py,sha256=w5mgPgncI740gVKMu3S-69DGNdUSI0bTZxegQGcZ98Y,5011
-optimum_rbln-0.8.1a2.dist-info/METADATA,sha256=qiUSuqA0Rledv8xJhhmMT4DgtLU-wPUuvDXnyGv78CY,5299
-optimum_rbln-0.8.1a2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-optimum_rbln-0.8.1a2.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-optimum_rbln-0.8.1a2.dist-info/RECORD,,
+optimum_rbln-0.8.1a3.dist-info/METADATA,sha256=e2Q0Hat0Lk5pWpTSk_kbikGUsOezSiz7nM-01GXFU8w,5299
+optimum_rbln-0.8.1a3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+optimum_rbln-0.8.1a3.dist-info/licenses/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+optimum_rbln-0.8.1a3.dist-info/RECORD,,