optimum-rbln 0.8.1a2__tar.gz → 0.8.1a3__tar.gz

{optimum_rbln-0.8.1a2 → optimum_rbln-0.8.1a3}/PKG-INFO

@@ -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 → optimum_rbln-0.8.1a3}/src/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-0.8.1a2 → optimum_rbln-0.8.1a3}/src/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-0.8.1a2 → optimum_rbln-0.8.1a3}/src/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-0.8.1a2 → optimum_rbln-0.8.1a3}/src/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-0.8.1a2 → optimum_rbln-0.8.1a3}/src/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 → optimum_rbln-0.8.1a3}/tests/test_llm.py

@@ -67,7 +67,7 @@ class LLMTest:
 class TestQwen2Model(LLMTest.TestLLM):
     RBLN_CLASS = RBLNQwen2ForCausalLM
     HF_MODEL_ID = "Qwen/Qwen2-0.5B-Instruct"
-    EXPECTED_OUTPUT = " I am a 30-year-old woman who has been living with lupus for over 1"
+    EXPECTED_OUTPUT = " I am a 20 year old girl from the United States. I have been studying English for"
     HF_CONFIG_KWARGS = {"max_position_embeddings": 1024}
 
 
@@ -108,7 +108,7 @@ class TestLlamaForCausalLM_Flash(LLMTest.TestLLM):
 class TestLlamaForCausalLM_Multibatch(TestLlamaForCausalLM):
     PROMPT = ["Who are you?", "What is the capital of France?", "What is the capital of Germany?"]
     EXPECTED_OUTPUT = [
-        "reress makefable R���� noethetsshss rechoolso�",
+        "reress makefable R���� noethetss0oss invetetet",
         "resget makeget makeichget makeichualichual#choolchool accngngngng",
         "resget makeget makeichget makeichualichual#choolchool accngngngng",
     ]