optimum-rbln 0.7.3.post1__tar.gz → 0.7.4__tar.gz

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/ISSUE_TEMPLATE/model_request.md

@@ -8,7 +8,7 @@ assignees: ''
 
 ### Model Information
 - Model name: 
-- Model link (Hugging Face/Paper/GitHub): 
+- Model link (HuggingFace/Paper/GitHub): 
 - Task type (e.g., Text Generation, Image Classification): 
 
 ### Model Details

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/scripts/auto_code_review.py

@@ -22,8 +22,7 @@ from github import Github
 
 model_name = os.environ["GOOGLE_MODEL_ID"]
 genai.configure(api_key=os.environ["GOOGLE_API_KEY"])
-max_context_token = 100000
-force_review = False
+max_context_token = 500000
 
 
 def get_pr_diff():
@@ -37,7 +36,7 @@ def get_pr_diff():
 
 
 def get_prompt(diff, pr):
-    system_prompt = """You are an experienced software engineer specializing in code reviews for deep learning libraries. Your task is to review code changes and related pull request (PR) information for `optimum-rbln`, a Python library that optimizes Hugging Face models for execution on RBLN NPUs.
+    system_prompt = """You are an experienced software engineer specializing in code reviews for deep learning libraries. Your task is to review code changes and related pull request (PR) information for `optimum-rbln`, a Python library that optimizes HuggingFace models for execution on RBLN NPUs.
 
 Focus on providing actionable and constructive feedback. Don't make generalized suggestions."""
 
@@ -58,19 +57,6 @@ Review the following code changes(GIT DIFF) along with the pull request (PR) det
     return system_prompt, prompt
 
 
-def translate_review(review):
-    model = genai.GenerativeModel(
-        model_name,
-        system_instruction="You are a professional translator specializing in technical and software-related content. Keep the technical words in English, but understand the whole sentence and rephrase it in Korean.",
-    )
-    prompt = f"""Translate the following English text into Korean, maintaining technical accuracy and clarity. Include ONLY the translation, NO OTHER EXPLANATIONS or RESPONSES as a chatbot. :
-
-{review}"""
-    response = model.generate_content(prompt)
-
-    return response.text
-
-
 def review_code(system_prompt, prompt):
     model = genai.GenerativeModel(model_name, system_instruction=system_prompt)
     response = model.generate_content(prompt)
@@ -125,7 +111,7 @@ def main():
     system_prompt, prompt = get_prompt(diff, pr)
     model = genai.GenerativeModel(model_name=model_name, system_instruction=system_prompt)
     num_tokens = model.count_tokens(prompt).total_tokens
-    if num_tokens > max_context_token and not force_review:
+    if num_tokens > max_context_token:
         msg = f"Diff ({len(diff)}) exceeds maximum allowed tokens ({max_context_token}) > ({num_tokens}). Skipping review."
         print(msg)
         pr.create_issue_comment(msg)
@@ -133,14 +119,10 @@ def main():
 
     # Get Auto review
     review = review_code(system_prompt, prompt)
-    translation = translate_review(review)
 
     # Post comment on PR
-    pr.create_issue_comment(f"""# Auto Code Review
-
-- [참고] Auto Code Review를 invoke하려면, commit message의 시작을 [autoreview]로 시작하거나, "/autoreview" 를 comment로 작성한 후,
-해당 commit의 github action에서 code review를 re-run 하시면 됩니다.
-\n\n{review}\n\n{translation}""")
+    pr.create_issue_comment(f"""# Auto Code Review by {model_name}
+\n\n{review}""")
 
 
 if __name__ == "__main__":

optimum_rbln-0.7.4/.github/version.yaml

@@ -0,0 +1 @@
+rebel_compiler_version: 0.7.4.dev61+gb562a7f0

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/workflows/auto_code_review.yml

@@ -14,7 +14,7 @@ env:
 
 jobs:
   auto-review:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-latest-rbln
     steps:
       - name: Checkout repository
         uses: actions/checkout@v3

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/workflows/check_code_quality.yml

@@ -25,7 +25,7 @@ jobs:
       fail-fast: false
       matrix:
         python-version: [3.9]
-        os: [ubuntu-latest]
+        os: [ubuntu-latest-rbln]
 
     runs-on: ${{ matrix.os }}
     steps:

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/workflows/deploy-on-tag.yaml

@@ -28,7 +28,7 @@ jobs:
               workflow_id: 'rebel_dispatch_model_generation_for_vllm.yaml',
               ref: 'dev',
               inputs: {
-                optimum_rbln_version: ${{ github.ref_name }},
+                optimum_rbln_version: "${{ github.ref_name }}",
               }
             })
             console.log(result)

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/workflows/pr-title-check.yaml

@@ -20,7 +20,7 @@ permissions:
 jobs:
   title-checker:
     name: Check PR title
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-latest-rbln
     steps:
       - uses: amannn/action-semantic-pull-request@v5
         id: lint_pr_title

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/workflows/pr_checklist_validator.yml

@@ -10,7 +10,7 @@ on:
 
 jobs:
   validate-pr-checklist:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-latest-rbln
     steps:
       - name: Checkout repository
         uses: actions/checkout@v2

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/.github/workflows/rbln_optimum_pytest.yaml

@@ -43,7 +43,9 @@ jobs:
         if: ${{ inputs.commit_message == '' }}
         run: |
           COMMIT_MESSAGE=$(git log -1 --pretty=%B)
-          echo "message=$COMMIT_MESSAGE" >> $GITHUB_OUTPUT
+          echo "message<<EOF" >> $GITHUB_OUTPUT
+          echo "$COMMIT_MESSAGE" >> $GITHUB_OUTPUT
+          echo "EOF" >> $GITHUB_OUTPUT
 
       - name: Setup uv
         uses: astral-sh/setup-uv@v3
@@ -62,7 +64,13 @@ jobs:
         run: |
           PYPI_URL=$(echo ${{ env.REBEL_PYPI_ENDPOINT }} | sed "s/\/\//\0${{ env.REBEL_PYPI_USERNAME}}:${{ env.REBEL_PYPI_PASSWORD}}@/")
           uv pip install --extra-index-url $PYPI_URL rebel-compiler==${{ inputs.rebel_compiler_version }}
-      
+
+      - name: Run pytest (config)
+        env:
+          OPTIMUM_RBLN_TEST_LEVEL: ${{ inputs.test_level }}
+        run: |
+          uv run --no-sync pytest tests/test_config.py
+
       - name: Run pytest (transformers)
         env:
           OPTIMUM_RBLN_TEST_LEVEL: ${{ inputs.test_level }}

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/PKG-INFO

@@ -1,7 +1,7 @@
 Metadata-Version: 2.4
 Name: optimum-rbln
-Version: 0.7.3.post1
-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.
+Version: 0.7.4
+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
 Project-URL: Repository, https://github.com/rebellions-sw/optimum-rbln
@@ -25,10 +25,10 @@ Requires-Python: <3.13,>=3.9
 Requires-Dist: accelerate>=1.0.1
 Requires-Dist: diffusers<=0.31.0
 Requires-Dist: packaging>=24.1
-Requires-Dist: torch<=2.5.1
-Requires-Dist: torchaudio<=2.5.1
-Requires-Dist: torchvision<=0.20.1
-Requires-Dist: transformers==4.48.3
+Requires-Dist: torch==2.6.0
+Requires-Dist: torchaudio<=2.6.0
+Requires-Dist: torchvision<=0.21.0
+Requires-Dist: transformers==4.50.3
 Description-Content-Type: text/markdown
 
 

optimum_rbln-0.7.4/examples/image-to-text/run_idefics3.py

@@ -0,0 +1,67 @@
+import os
+import typing
+
+import fire
+from datasets import load_dataset
+from transformers import AutoProcessor
+
+from optimum.rbln import RBLNIdefics3ForConditionalGeneration
+
+
+def main(
+    model_id: str = "HuggingFaceM4/Idefics3-8B-Llama3",
+    batch_size: int = 1,
+    from_transformers: bool = False,
+    prompt: typing.Optional[str] = None,
+    max_seq_len: typing.Optional[int] = None,
+    tensor_parallel_size: typing.Optional[int] = 4,
+):
+    processor = AutoProcessor.from_pretrained(model_id)
+
+    if from_transformers:
+        model = RBLNIdefics3ForConditionalGeneration.from_pretrained(
+            model_id,
+            export=True,
+            rbln_config={
+                "text_model": {
+                    "attn_impl": "flash_attn",
+                    "max_seq_len": max_seq_len,
+                    "use_inputs_embeds": True,
+                    "tensor_parallel_size": tensor_parallel_size,
+                    "batch_size": batch_size,
+                }
+            },
+        )
+        model.save_pretrained(os.path.basename(model_id))
+    else:
+        model = RBLNIdefics3ForConditionalGeneration.from_pretrained(
+            os.path.basename(model_id),
+            export=False,
+        )
+
+    ds = load_dataset("HuggingFaceM4/the_cauldron", "ai2d", split="train")
+    samples = ds.select(range(batch_size))
+    images = []
+    prompts = []
+
+    for sample in samples:
+        img = sample["images"]
+        images.append(img)
+
+        message = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "Describe this image."}]}]
+        prompt = processor.apply_chat_template(message, add_generation_prompt=True)
+        prompts.append(prompt)
+
+    inputs = processor(text=prompts, images=images, return_tensors="pt", padding=True)
+    inputs = dict(inputs)
+    # Generate
+
+    generated_ids = model.generate(**inputs, max_new_tokens=500)
+    generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+    for i, text in enumerate(generated_texts):
+        print(f"Sample {i + 1} generate:\n{text}\n")
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/examples/kandinsky2_2/run_kandinsky2_2_img2img.py

@@ -20,6 +20,7 @@ def main(
         prior_pipe = RBLNKandinskyV22PriorPipeline.from_pretrained(
             model_id=prior_model_id,
             export=True,
+            rbln_config={"prior": {"batch_size": 2}},
         )
         prior_pipe.save_pretrained(os.path.basename(prior_model_id))
 
@@ -28,6 +29,7 @@ def main(
             export=True,
             rbln_img_height=768,
             rbln_img_width=768,
+            rbln_config={"unet": {"batch_size": 2}},
         )
         pipe.save_pretrained(os.path.basename(inpaint_model_id))
     else:

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/examples/question-answering/run_question_answering.py

@@ -30,7 +30,7 @@ def main(
         model=model,
         tokenizer=model_id,
         padding="max_length",
-        max_seq_len=model.rbln_config.model_cfg["max_seq_len"],
+        max_seq_len=model.rbln_config.max_seq_len,
     )
     question, text = (
         ["What is Rebellions?"] * batch_size,

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/examples/stable-diffusion/run_stable_diffusion.py

@@ -16,6 +16,11 @@ def main(
         pipe = RBLNStableDiffusionPipeline.from_pretrained(
             model_id=model_id,
             export=True,
+            rbln_config={
+                "unet": {
+                    "batch_size": 2,
+                }
+            },
         )
         pipe.save_pretrained(os.path.basename(model_id))
     else:

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/examples/stable-diffusion/run_stable_diffusion_controlnet.py

@@ -38,6 +38,14 @@ def main(
             controlnet=controlnet,
             rbln_img_width=img_width,
             rbln_img_height=img_height,
+            rbln_config={
+                "unet": {
+                    "batch_size": 2,
+                },
+                "controlnet": {
+                    "batch_size": 2,
+                },
+            },
             export=True,
             scheduler=UniPCMultistepScheduler.from_pretrained(diffusion_model_id, subfolder="scheduler"),
         )

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/examples/stable-diffusion/run_stable_diffusion_img2img.py

@@ -29,6 +29,11 @@ def main(
             export=True,
             rbln_img_width=img_width,
             rbln_img_height=img_height,
+            rbln_config={
+                "unet": {
+                    "batch_size": 2,
+                },
+            },
         )
         pipe.save_pretrained(os.path.basename(model_id))
     else:

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/examples/stable-diffusion/run_stable_diffusion_img2img_controlnet.py

@@ -40,6 +40,14 @@ def main(
             controlnet=controlnet,
             rbln_img_width=img_width,
             rbln_img_height=img_height,
+            rbln_config={
+                "unet": {
+                    "batch_size": 2,
+                },
+                "controlnet": {
+                    "batch_size": 2,
+                },
+            },
             export=True,
             scheduler=UniPCMultistepScheduler.from_pretrained(diffusion_model_id, subfolder="scheduler"),
         )

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/examples/stable-diffusion/run_stable_diffusion_multicontrolnet.py

@@ -75,6 +75,7 @@ def main(
         images,
         negative_prompt=negative_prompt,
         num_inference_steps=20,
+        guidance_scale=0.0,
         controlnet_conditioning_scale=[1.0, 0.8],
         generator=torch.Generator(device="cpu").manual_seed(42),
     ).images[0]

optimum_rbln-0.7.4/examples/time-series-forecasting/run_time_series_forecasting.py

@@ -0,0 +1,43 @@
+import os
+
+import fire
+import torch
+from huggingface_hub import hf_hub_download
+
+from optimum.rbln import RBLNTimeSeriesTransformerForPrediction
+
+
+def main(
+    model_id: str = "huggingface/time-series-transformer-tourism-monthly",
+    batch_size: int = 1,
+    num_parallel_samples: int = 100,
+    from_transformers: bool = False,
+):
+    if from_transformers:
+        model = RBLNTimeSeriesTransformerForPrediction.from_pretrained(
+            model_id, export=True, rbln_batch_size=batch_size, num_parallel_samples=num_parallel_samples
+        )
+        model.save_pretrained(os.path.basename(model_id))
+    else:
+        model = RBLNTimeSeriesTransformerForPrediction.from_pretrained(
+            os.path.basename(model_id),
+            export=False,
+        )
+
+    dataset = hf_hub_download(
+        repo_id="hf-internal-testing/tourism-monthly-batch", filename="val-batch.pt", repo_type="dataset"
+    )
+    data = torch.load(dataset, weights_only=True)
+
+    batched_data = {}
+    for k, v in data.items():
+        batched_data[k] = v[:batch_size]
+
+    rbln_outputs = model.generate(**batched_data)
+    mean_prediction = rbln_outputs.sequences.mean(dim=1)
+
+    print(mean_prediction)
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

{optimum_rbln-0.7.3.post1 → optimum_rbln-0.7.4}/pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "optimum-rbln"
 description = """
-Optimum RBLN is the interface between the Hugging Face Transformers and Diffusers libraries and RBLN accelerators.
+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.
 """
 authors = [
@@ -28,11 +28,11 @@ classifiers = [
 ]
 keywords = ["transformers", "diffusers", "inference", "rbln", "atom", "rebel"]
 dependencies = [
-    "torch<=2.5.1",
-    "torchaudio<=2.5.1",
-    "torchvision<=0.20.1",
+    "torch==2.6.0",
+    "torchaudio<=2.6.0",
+    "torchvision<=0.21.0",
     "accelerate>=1.0.1",
-    "transformers==4.48.3",
+    "transformers==4.50.3",
     "diffusers<=0.31.0",
     "packaging>=24.1",
 ]
@@ -53,6 +53,8 @@ tests = [
     "sacremoses>=0.1.1",
     "safetensors>=0.4.2",
     "protobuf>=5.27.2",
+    "soundfile>=0.13.1",
+    "librosa>=0.11.0",
 ]
 quality = [
     "ruff>=0.3.3",