optimum-rbln 0.8.0__tar.gz → 0.8.1a0__tar.gz

{optimum_rbln-0.8.0 → optimum_rbln-0.8.1a0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.8.0
+Version: 0.8.1a0
 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-0.8.0 → optimum_rbln-0.8.1a0}/src/optimum/rbln/__version__.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.8.0'
-__version_tuple__ = version_tuple = (0, 8, 0)
+__version__ = version = '0.8.1a0'
+__version_tuple__ = version_tuple = (0, 8, 1, 'a0')

{optimum_rbln-0.8.0 → optimum_rbln-0.8.1a0}/src/optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py

@@ -709,7 +709,9 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
                 compiled_models[f"decoder_batch_{batch_size}"].get_alloc_per_node_by_key().items()
             ):
                 alloc_memory_by_key[key] += sum(memory_per_node)
-        alloc_memory_by_key.pop("PortRecur")  # kv-cache
+
+        alloc_memory_by_key.pop("PortRecur", None)  # Old compiler's kv-cache Key
+        alloc_memory_by_key.pop("DramTensor", None)  # kv-cache
         kernel_size = alloc_memory_by_key.pop("Kernel")  # model weight
 
         # Get the maximum number of blocks that can be allocated

{optimum_rbln-0.8.0 → optimum_rbln-0.8.1a0}/tests/test_llm.py

@@ -14,11 +14,13 @@ from optimum.rbln import (
     RBLNAutoModelForSeq2SeqLM,
     RBLNAutoModelForVision2Seq,
     RBLNBartForConditionalGeneration,
+    RBLNBlip2ForConditionalGeneration,
     RBLNExaoneForCausalLM,
     RBLNGPT2LMHeadModel,
     RBLNIdefics3ForConditionalGeneration,
     RBLNLlamaForCausalLM,
     RBLNLlavaNextForConditionalGeneration,
+    RBLNOPTForCausalLM,
     RBLNPhiForCausalLM,
     RBLNQwen2_5_VLForConditionalGeneration,
     RBLNQwen2ForCausalLM,
@@ -69,6 +71,13 @@ class TestQwen2Model(LLMTest.TestLLM):
     HF_CONFIG_KWARGS = {"max_position_embeddings": 1024}
 
 
+class TestOptModel(LLMTest.TestLLM):
+    RBLN_CLASS = RBLNOPTForCausalLM
+    HF_MODEL_ID = "facebook/opt-2.7b"
+    EXPECTED_OUTPUT = "\nI'm a guy who likes to play video games.\nI'm a guy who likes to"
+    HF_CONFIG_KWARGS = {"max_position_embeddings": 2048}
+
+
 class TestLlamaForCausalLM(LLMTest.TestLLM):
     RBLN_CLASS = RBLNLlamaForCausalLM
     HF_MODEL_ID = "afmck/testing-llama-tiny"
@@ -291,6 +300,51 @@ class TestLlavaNextForConditionalGeneration(LLMTest.TestLLM):
         )
 
 
+class TestBlip2ForConditionalGeneration(LLMTest.TestLLM):
+    RBLN_AUTO_CLASS = RBLNAutoModelForVision2Seq
+    RBLN_CLASS = RBLNBlip2ForConditionalGeneration
+    HF_MODEL_ID = "Salesforce/blip2-opt-2.7b"  # No tiny model yet.
+    PROMPT = "Question: Describe this image? Answer:"
+    RBLN_CLASS_KWARGS = {"rbln_config": {"language_model": {"use_inputs_embeds": True}}}
+    EXPECTED_OUTPUT = "::::::::::::::::::::"
+
+    @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"] = 1
+        kwargs = {"text_config": text_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)
+        inputs = tokenizer(images=image, text=self.PROMPT, return_tensors="pt", padding=True)
+        inputs["max_new_tokens"] = 20
+        inputs["do_sample"] = False
+        return inputs
+
+    def _inner_test_save_load(self, tmpdir):
+        super()._inner_test_save_load(tmpdir)
+        # Test loading from nested config
+        _ = self.RBLN_CLASS.from_pretrained(
+            tmpdir,
+            export=False,
+            rbln_config={"language_model": {"create_runtimes": False}},
+            **self.HF_CONFIG_KWARGS,
+        )
+
+
 class TestIdefics3ForConditionalGeneration(LLMTest.TestLLM):
     RBLN_AUTO_CLASS = RBLNAutoModelForVision2Seq
     RBLN_CLASS = RBLNIdefics3ForConditionalGeneration