optimum-rbln 0.7.5rc2__tar.gz → 0.8.0__tar.gz

{optimum_rbln-0.7.5rc2 → optimum_rbln-0.8.0}/.github/workflows/rbln_optimum_pytest.yaml

@@ -25,6 +25,7 @@ env:
   REBEL_PYPI_ENDPOINT: ${{ vars.REBEL_PYPI_INTERNAL_ENDPOINT }}
   REBEL_PYPI_USERNAME: ${{ secrets.REBEL_PYPI_USERNAME }}
   REBEL_PYPI_PASSWORD: ${{ secrets.REBEL_PYPI_PASSWORD }}
+  HF_HOME: ${{ secrets.HF_HOME }}
 
 jobs:
   pytest:

{optimum_rbln-0.7.5rc2 → optimum_rbln-0.8.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.5rc2
+Version: 0.8.0
 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-0.7.5rc2 → optimum_rbln-0.8.0}/src/optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.7.5rc2'
-__version_tuple__ = version_tuple = (0, 7, 5, 'rc2')
+__version__ = version = '0.8.0'
+__version_tuple__ = version_tuple = (0, 8, 0)

{optimum_rbln-0.7.5rc2 → optimum_rbln-0.8.0}/src/optimum/rbln/transformers/models/gemma3/configuration_gemma3.py

@@ -11,9 +11,10 @@
 # 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.
-
 from typing import Optional
 
+import rebel
+
 from ....configuration_utils import RBLNModelConfig
 from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
 from ..siglip.configuration_siglip import RBLNSiglipVisionModelConfig
@@ -39,6 +40,10 @@ class RBLNGemma3ForCausalLMConfig(RBLNDecoderOnlyModelForCausalLMConfig):
             **kwargs,
         )
 
+        npu = self.npu or rebel.get_npu_name()
+        if npu == "RBLN-CA02":
+            raise NotImplementedError("Gemma3 is currently not supported on RBLN-CA02")
+
 
 class RBLNGemma3ForConditionalGenerationConfig(RBLNModelConfig):
     submodules = ["vision_tower", "language_model"]

{optimum_rbln-0.7.5rc2 → optimum_rbln-0.8.0}/src/optimum/rbln/transformers/models/gemma3/modeling_gemma3.py

@@ -858,11 +858,6 @@ class RBLNGemma3ForCausalLM(RBLNDecoderOnlyModelForCausalLM):
             max_seq_len=rbln_config.max_seq_len,
         )
 
-        if rbln_config.attn_impl == "eager":
-            raise ValueError(
-                "Eager attention is not supported for Gemma3. Please use flash attention by setting `rbln_attn_impl='flash_attn'`. Stay tuned for future updates!"
-            )
-
         validate_attention_method(
             attn_impl=rbln_config.attn_impl,
             kvcache_partition_len=rbln_config.kvcache_partition_len,

{optimum_rbln-0.7.5rc2 → optimum_rbln-0.8.0}/src/optimum/rbln/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py

@@ -595,9 +595,8 @@ class RBLNQwen2_5_VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
                 )
                 logits.append(output.logits)
             logits = torch.cat(logits, dim=0)
-            # Decoder
+        # Decoder
         else:
-            print(input_ids[0], cache_position[0])
             inputs_embeds, position_embed = self._preprocess_decoder(input_ids, cache_position)
             output = self.decoder(
                 inputs_embeds=inputs_embeds,

{optimum_rbln-0.7.5rc2 → optimum_rbln-0.8.0}/tests/test_llm.py

@@ -11,13 +11,10 @@ from transformers import AutoConfig, AutoProcessor, AutoTokenizer
 from optimum.rbln import (
     RBLNAutoModel,
     RBLNAutoModelForCausalLM,
-    RBLNAutoModelForImageTextToText,
     RBLNAutoModelForSeq2SeqLM,
     RBLNAutoModelForVision2Seq,
     RBLNBartForConditionalGeneration,
     RBLNExaoneForCausalLM,
-    RBLNGemma3ForCausalLM,
-    RBLNGemma3ForConditionalGeneration,
     RBLNGPT2LMHeadModel,
     RBLNIdefics3ForConditionalGeneration,
     RBLNLlamaForCausalLM,
@@ -68,7 +65,7 @@ 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 a chronic illness for the past"
+    EXPECTED_OUTPUT = " I am a 30-year-old woman who has been living with lupus for over 1"
     HF_CONFIG_KWARGS = {"max_position_embeddings": 1024}
 
 
@@ -395,55 +392,5 @@ class TestDisallowedLlama_4(DisallowedTestBase.DisallowedTest):
     RBLN_CLASS_KWARGS = {"rbln_config": {"attn_impl": "flash_attn", "kvcache_partition_len": 1024}}
 
 
-class TestGemma3ForConditionalGeneration(LLMTest.TestLLM):
-    RBLN_AUTO_CLASS = RBLNAutoModelForImageTextToText
-    RBLN_CLASS = RBLNGemma3ForConditionalGeneration
-    HF_MODEL_ID = "google/gemma-3-4b-it"  # No tiny model yet.
-    PROMPT = "<bos><start_of_turn>user\n<start_of_image>Describe the image.<end_of_turn>\n<start_of_turn>model\n'"
-    RBLN_CLASS_KWARGS = {"rbln_config": {"language_model": {"use_inputs_embeds": True, "kvcache_partition_len": 4096}}}
-    EXPECTED_OUTPUT = "ִיсмотрятОднакоettingiPadaenуdenlyticsinท์नियमितinท์नियमित"
-
-    @classmethod
-    def get_tokenizer(cls):
-        if cls._tokenizer is None:
-            cls._tokenizer = AutoProcessor.from_pretrained(cls.HF_MODEL_ID)
-        return cls._tokenizer
-
-    # override
-    @classmethod
-    def setUpClass(cls):
-        config = AutoConfig.from_pretrained(cls.HF_MODEL_ID)
-        text_config = json.loads(config.text_config.to_json_string())
-        text_config["num_hidden_layers"] = 2
-        text_config["sliding_window_pattern"] = 2
-        vision_config = json.loads(config.vision_config.to_json_string())
-        vision_config["num_hidden_layers"] = 1
-        kwargs = {"text_config": text_config, "vision_config": vision_config}
-        cls.HF_CONFIG_KWARGS.update(kwargs)
-        return super().setUpClass()
-
-    def get_inputs(self):
-        tokenizer = self.get_tokenizer()
-        img_path = f"{os.path.dirname(__file__)}/../assets/rbln_logo.png"
-        image = Image.open(img_path)
-        image = image.convert("RGB")
-        inputs = tokenizer(images=[image], text=[self.PROMPT], return_tensors="pt", padding=True)
-        inputs["max_new_tokens"] = 20
-        inputs["do_sample"] = False
-        return inputs
-
-
-class TestGemma3ForCausalLM(LLMTest.TestLLM):
-    RBLN_CLASS = RBLNGemma3ForCausalLM
-    HF_MODEL_ID = "google/gemma-3-1b-it"
-    EXPECTED_OUTPUT = "1st L L L L L L L L L L L L L L L L L L"
-    HF_CONFIG_KWARGS = {
-        "num_hidden_layers": 2,
-        "sliding_window_pattern": 2,
-        "max_position_embeddings": 1024,
-        "trust_remote_code": True,
-    }
-
-
 if __name__ == "__main__":
     unittest.main()