릴리스 노트

릴리스 노트의 각 변경사항은 명확한 이해를 위해 영문으로 작성되어 있습니다.

2025.03.28

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo
2025.03.28.0	v1.2.92	v0.7.3	v0.7.3.post2	v0.7.3	v0.5.8

• Install Command

  pip3 install --extra-index-url https://pypi.rbln.ai/simple rebel-compiler==0.7.3 optimum-rbln==0.7.3.post2 vllm-rbln==0.7.3
  

• RBLN Compiler:

  • Features
    • Added support for Paged Attention
    • Expanded the coverage of the LayerNorm operation
  • Optimization
    • Improved memory tiling algorithm for better Neural Engine Clusters' utilization
  • BugFix
    • Fixed an issue where the RBLN_NUM_THREADS environment variable did not work with the AsyncRuntime
  • API
    • Added a context manager, the profile class, to support saving profiled data to a user-defined path for RBLN Profiler

• Optimum RBLN:

  • Added functions:
    • RBLNKandinskyV22Pipeline()
    • RBLNKandinskyV22Img2ImgPipeline()
    • RBLNKandinskyV22CombinedPipeline()
    • RBLNKandinskyV22Img2ImgCombinedPipeline()
  • BugFix:
    • Removed the ad-hoc forward statement in RBLNXLMRobertaModel() to fix "the number of inputs exceeds the expected number" error for BGE-M3 model
  • Features:
    • Updated to support Paged Attention
      • Note: Inference with the generate() API cannot be completed when no memory blocks are available for allocation due to the absence of the Paged Attention block manager. This issue can be resolved by using vLLM RBLN, which fully supports Paged Attention
  • Others:
    • The usage of class property, which had been deprecated since Python 3.11, had been removed from the codebase to align with Python's deprecation schedule

• vLLM RBLN:

  • Updated to support Paged Attention

• RBLN Model Zoo:

  • Added new models:
    • HuggingFace
      • EXAONE-3.5-32b
      • BGE-Small-en-v1.5
      • BGE-Base-en-v1.5
      • BGE-Large-en-v1.5
      • Kandinsky v2.2
        • Text2Image
        • Image2Image
  • Updated to restructure the directories for the HuggingFace models, differentiating between the diffusers and transformers models
  • Changed the default input image size from 512 to 768 for knadinsky v2.2 Inpaint examples

• RBLNServe:

  • Deprecation Notice: As announced in the RBLN SDK 2025.02.28 release, RBLNServe has been deprecated and removed in this release. Please transition to alternative solutions such as Nvidia Triton Inference Server, vLLM, or TorchServe

• Other Changes:

  • SDK Documentation:
    • Added a Quick Start page as part of a tutorial to help users understand how to use the RBLN SDK
    • Added an Error Codes page that provides explanations of error codes encountered during model compilation, including the affected compilation pass and their underlying causes
    • Added a Performance Tuning section to the Software > RBLN Compiler > Troubleshoot
    • Added tutorials using Flash Attention:
      • vLLM RBLN
        • vLLM Native API
        • OpenAI Compatible Server
      • Nvidia Triton Inference Server
        • Llama3.1-8B with Flash Attention
      • TorchServe
        • Llama3.1-8B with Flash Attention
    • Updated the tutorials that demonstrate using RBLN SDK with the Llama2-7B to now use Llama3-8B:
      • Optimum RBLN
        • Chatbot tutorial
      • vLLM RBLN
        • vLLM Native API tutorial
        • OpenAI Compatible Server tutorial
      • Nvidia Triton Inference Server
        • Llama3-8b tutorial

---

2025.02.28

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2025.02.28.0	v1.2.92	v0.7.2	v0.7.2	v0.7.2	v0.5.7	v0.5.0

• Install Command

  pip3 install --extra-index-url https://pypi.rbln.ai/simple rebel-compiler==0.7.2 optimum-rbln==0.7.2 vllm-rbln==0.7.2
  

• RBLN Compiler:

  • Features
    • Improved the attention kernel to support multiple batches
    • Added new supported operations
      • depth_to_space
      • ConsineSimilarity
      • MaxUnpool2d
      • polyval
      • mirror_pad
  • Optimization
    • Optimized prefill performance of the flash attention supported models
    • Improved the internal buffer management algorithm for faster model build time
    • Enhanced LLM attention mask transfer logic for RSD models
  • API
    • Added debugging functions for RuntimeBase class
      • get_elasped_times()
        • Calculate the average total execution time of operations in microseconds
      • get_elasped_device_times()
        • Calculate the average device-side execution time of operations in microseconds
      • get_reports()
        • Retrieve all pending reports from the runtime
      • flush_reports()
        • Fetch and discard all pending reports from the runtime
    • Added a new class method inspect() for RBLNCompiledModel class
      • Provides metadata information about the compiled model without loading it into host memory

• Optimum RBLN:

  • Added functions:
    • RBLNBertForMaskedLM()
    • RBLNKandinskyV22InpaintCombinedPipeline()
    • RBLNKandinskyV22InpaintPipeline()
    • RBLNKandinskyV22PriorPipeline()
  • BugFix:
    • Resolved incorrect version dependency warning for the RBLN Compiler
    • Fixed issue with handling large input images in Llava-Next model
    • Corrected behavior of rbln_model_input_names() in BERT models
  • Features:
    • Updated decoder-only models to call multi-batch attention kernel to simplify model and clear attention kernel writing
  • Others:
    • Algined the version number of the Optimum RBLN with the version number of the RBLN Compiler to eliminate any confusion for users
    • Updated to support the latest transformers (v4.48.3) and diffusers (v0.31.0)

• vLLM RBLN:

  • Updated to sync with vllm v0.7.1
  • Algined the version number of the vLLM RBLN with the version number of the RBLN Compiler to eliminate any confusion for users

• RBLN Model Zoo:

  • Added new models:
    • HuggingFace
      • DeepSeek-R1-Distill-Llama-8B
      • DeepSeek-R1-Distill-Llama-70B
      • DeepSeek-R1-Distill-Qwen-1.5B
      • DeepSeek-R1-Distill-Qwen-7B
      • DeepSeek-R1-Distill-Qwen-14B
      • DeepSeek-R1-Distill-Qwen-32B
      • Kandinsky v2.2
        • Inpainting
      • Ko-Reranker
      • KR-SBERT
    • PyTorch
      • LaBSE
  • Added supplementary guides for the TorchServe tutorials

• RBLNServe:

  • Deprecation Notice: Please be advised that RBLNServe will be deprecated after March 2025. We recommend transitioning to alternative solutions (Triton server, vLLM, or torchserve) before this date to ensure continued support and functionality. After the deprecation, RBLNServe will be removed in future versions. We encourage all users to plan accordingly and migrate to the recommended alternatives

• Other Changes:

  • Added new documentatin pages
    • TorchServe Support
    • Docker Support
    • Support Matrix
  • Documentation refactoring:
    • The following pages were revised to provide clearer explanations and better guidance for the RBLN SDK users:
      • OVERVIEW
      • GETTING STARTED > Installation Guide
      • GETTING STARTED > Tutorials

---

2025.02.04: Breaking Changes

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2025.02.04.0	v1.2.92	v0.7.1	v0.2.0	v0.2.0	v0.5.6	v0.5.0

Note

BREAKING CHANGES: Please update the RBLN Compiler to the latest version (v0.7.1 or higher) for compatibility with the updated RBLN Driver.

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.7.1 vllm-rbln==0.2.0
  pip3 install optimum-rbln==0.2.0
  

• RBLN Driver:

  • Enabled Low Power Management (LPM) feature
  • Added support for RSD with long sequence LLM models (>32K)
  • Supported driver package installation on RHEL and AlmaLinux
  • Improved hDMA performance

• RBLN Compiler:

  • Features
    • Updated RBLN Compiler for compatibility with the RBLN Driver
    • Added support for Flash Attention
    • Added support for the depth_to_space operation
  • Optimization
    • Improved memory tiling algorithms
    • Enhanced collective communication processing for RSD models
    • Increased the write speed of *.rbln files
  • API
    • Added a mode option to the options argument in torch.compile()
  • Functionality
    • Initial release of the RBLN Profiler
      • The RBLN Compiler v0.7.1 includes the performance profiler, which allows users to view the time spent on each step of the model inference process. For more details, refer to the RBLN Profiler.

• Optimum RBLN

  • Public release of Optimum RBLN: GitHub Repository
  • Added rbln_attn_impl and rbln_kvcache_partition_len arguments to decoder-only transformer model APIs to support Flash Attention:
    • RBLNLlamaForCausalLM
    • RBLNGemmaForCausalLM
    • RBLNMistralForCausalLM
    • RBLNQwen2ForCausalLM
    • RBLNExaoneForCausalLM

• vLLM RBLN

  • Updated to support Flash Attention

• RBLN Model Zoo

  • Added new models
    • HuggingFace
      • Llama3.1-8b
      • Llama3.1-70b
      • Llama3.2-3b
      • Llama3.3-70b
      • KONI-Llama3.1-8b
    • PyTorch
      • YOLOv10-N/S/M/B/L/X

• RBLNServe:

  • Pinned rebel-compiler version to <0.8, >=0.7

• Other Changes:

  • Documentation refactoring:
    • Added a new section for Software > RBLN Profiler
    • Consolidated model serving content into a single section under Software > Model Serving, which now includes:
      • RBLNServe
      • vLLM Support
      • Nvidia Triton Inference Server Support

---

2024.12.27

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.12.27.0	v1.1.67	v0.6.2	v0.1.15	v0.1.3	v0.5.5	v0.4.0

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.6.2 optimum-rbln==0.1.15 vllm-rbln==0.1.3
  

• RBLN Compiler:

  • Added get_total_device_alloc() methods to RBLNCompiledModel class, enabling efficient retrieval of the total device memory allocation (in bytes) used by the compiled graph across all NPUs
  • Fixed bug for split operation handling logic
  • Added exception handling for previously uncovered edge cases in the error handling logic

• Optimum RBLN:

  • Added functions:
    • RBLNStableDiffusionInpaintPipeline()
    • RBLNStableDiffusionXLInpaintPipeline()
    • RBLNStableDiffusionXLControlNetPipeline()
    • RBLNStableDiffusionXLControlNetImg2ImgPipeline()
    • RBLNStableDiffusion3Pipeline()
    • RBLNStableDiffusion3Img2ImgPipeline()
    • RBLNStableDiffusion3InpaintPipeline()
  • Removed the dependency on optimum
    • This change eliminated the automatic installation of GPU-related dependencies, resulting in a significantly faster installation process

• vLLM RBLN:

  • Updated to sync with vllm v0.6.5
  • Updated to support EXAONE v3.5 models

• RBLN Model Zoo:

  • Added new models
    • HuggingFace
      • EXAONE-3.5-2.4b
      • EXAONE-3.5-7.8b
      • Stable Diffusion v3
        • Text to image
        • Image to image
        • Inpainting
      • Stable Diffusion
        • Inpainting
      • Stable Diffusion XL
        • Inpainting
        • Text to image + controlNet
        • Image to image + controlNet
    • PyTorch
      • YOLOv5-Face
  • Improved the formatting of all model code for better readability and maintainability
  • Added new examples demonstrating how to use vLLM's native APIs with a wider range of model architectures:
    • Decoder-only (Llama3)
    • Encoder-decoder (BART)
    • Multi-modal (Llava-next)
  • Added supplementary guides for the model serving:
    • Tutorial > Advanced > LLM Serving > LLM Serving with Continous Batching: RBLN Model Zoo Link
    • Software > Model Serving > Nvidia Triton Infernece Server: RBLN Model Zoo Link

---

2024.11.27

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.11.27.0	v1.1.67	v0.6.1	v0.1.13	v0.1.2	v0.5.4	v0.4.0

Note

Deprecation Notice: Python 3.8 Support

As part of our commitment to maintaining compatibility with supported and secure versions of Python, we are officailly deprecating support for Python 3.8. This version will no longer be included in future releases, and users are encouraged to upgrade to a more recent Python version to ensure continued support and compatibility.

New Additions: Python 3.11 and Python 3.12

We are pleased to announce that Python 3.11 and Python 3.12 are now fully supported and included in our release package.

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.6.1 optimum-rbln==0.1.13 vllm-rbln==0.1.2
  

• RBLN Compiler:

  • Improved operation efficiency:
    • New fusion logic for masked softmax
    • Accelerated take operation for tensor indexing
  • Enhanced communication logic for RSD models
  • Refactored .rbln file format

• Optimum RBLN:

  • Added functions:
    • RBLNT5EncoderModel()
  • Refactored RBLNStableDiffusion pipelines to enhance functionality and flexibility:
    • LoRA support:
      • The pipeline included support for Low-Rank Adaptation (LoRA).
    • rbln_config input argument:
      • A new rbln_config argument has been introduced. This configuration is designed specifically for RBLN compilation and incldues:
        • Global settings: Parameters such as npu, device, and create_rutniems
        • Image generation settings: Prameters such as batch_size, img_height, img_width, and guidacne_scale
    • For more detailed information, please refer to the Model API documentation and the RBLN Model Zoo example
  • Added a new from_model() method to the RBLN<ModelName>ForCausalLM classes. This method enables LoRA support by accepting a HuggingFace PreTrainedModel as input, allowing the base model and LoRA adapter to be merged using the merge_and_unload() approach. For more details, please refer to the RBLN Model Zoo example
  • Enabled to support various RoPE (Rotary Position Embedding) methods
    • default
    • linear
    • dynamic
    • yarn
    • longrope
    • llama3

• vLLM RBLN:

  • Updated to sync with vllm v0.6.4
  • The --compiled_model_dir configuration has been deprecated and will be removed in future release. Users are encourged to use --model argument instead. Please refer LLM serving with Continous Batching tutorial for actual usecase

• RBLN Model Zoo:

  • Added new models
    • HuggingFace
      • Qwen2.5-7b/14b
      • Llama3-8b + LoRA
      • StableDiffusion + LoRA
      • StableDiffusionXL + LoRA
  • Dependency Updates
    • PyTorch: updated to version 2.5.1
    • TensorFlow: updated to version 2.18.0
  • Model deprecation
    • The 3DDFA model has been deprecated due to maintenance discontinuation

• RBLNServe:

  • Pinned rebel-compiler version to <0.7, >=0.6

---

2024.10.30

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.10.30.0	v1.1.67	v0.5.12	v0.1.12	v0.1.0	v0.5.3	v0.3.0

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.5.12 optimum-rbln==0.1.12 vllm-rbln==0.1.0
  

• RBLN Compiler:

  • Updated to support cosine_similarity operation
  • Enabled a runtime initialization with RBLNCompiledModel and deprecated the path argument
    • Runtime
    • AsyncRuntime
  • Added double buffering on/off option on AsyncRuntime creation with the parallel arugment.
  • Added example_info argument to compile_from_torch() to support compilation without InputInfo creation
  • Added device argument to torch.compile to specify the NPU device ID for execution
• Optimum RBLN:
  • Added functions:
    • RBLNQwen2ForCausalLM()
    • RBLNExaoneForCausalLM()
    • RBLNPhiForCausalLM()
    • RBLNViTImageClassification()
  • Updated to support latest transformers (v4.45.2)
• vLLM RBLN:
  • Updated to support Qwen2, EXAONE, and Phi-2 architectures
• RBLN Model Zoo:
  • Added new models
    • HuggingFace
      • Qwen2-7b
      • EXAONE-3.0-7.8b
      • Salamandra-7b
      • Phi-2
      • ViT-large
      • Whisper-large-v3-turbo
    • PyTorch Dynamo
      • SAM2.1_hiera_large
  • Separated the framework-specific requirements.txt files into individual requirements.txt files for each model

---

2024.09.27

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.09.27.0	v1.1.67	v0.5.10	v0.1.11	v0.0.7	v0.5.2	v0.3.0

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.5.10 optimum-rbln==0.1.11 vllm-rbln==0.0.7
  

• RBLN Driver:

  • Added runtime power management with dynamic PCIe link speed change and PCIe ASPM (Active State Power Management)
  • Improved P2P throughput
  • Enhanced stability for Rebellions Scalable Design (RSD)
• RBLN Compiler:
  • Improved internal memory management algorithm
  • Updated the runtime description to show NPU version
  • Refactored .rbln file format
• Optimum RBLN:
  • Added functions:
    • RBLNBertModel()
    • RBLNBartModel()
    • RBLNLlavaNextForConditionalGeneration()
  • Updated RBLNWhisperForConditionalGeneration to support generating token timestamps and long-form transcription
• vLLM RBLN:
  • Updated to support Llava-Next, BART, and T5 models
• RBLN Model Zoo:
  • Added new models
    • HuggingFace
      • Llava-Next
      • E5-Base-4k
      • KoBART
      • BGE-Reranker-Base/Large
    • PyTorch
      • MotionBERT Action Recognition
  • Updated Whisper models to support generating token timestamps and long-form transcription
• Others
  • Split LLM Serving tutorial into with Triton Inference Server and with Continuous Batching
  • Added vLLM API standalone example in LLM serving tutorial

---

2024.08.30

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.08.30.0	v1.0.1	v0.5.9	v0.1.9	-	v0.5.1	v0.3.0
2024.08.30.1	v1.0.5	v0.5.9	v0.1.9	v0.0.6	v0.5.1	v0.3.0

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.5.9 optimum-rbln==0.1.9 vllm-rbln==0.0.6
  

• RBLN Compiler:

  • Added model_description() method in Runtime class
  • Updated to support where and einsum operations
  • Fixed bug for strided_slice operation
• Optimum RBLN:
  • Added functions:
    • RBLNGemmaForCausalLM()
    • RBLNMistralForCausalLM()
    • RBLNDistilBertForQuestionAnswering()
• vLLM RBLN:
  • Updated to support Gemma and Mistral architectures
• RBLN Model Zoo:
  • Added new models
    • HuggingFace
      • Gemma-2B
      • Gemma-7B
      • Mistral-7B
      • DistilBERT
    • PyTorch
      • MotionBERT
    • PyTorch Dynamo
      • YOLOv3
      • YOLOv4
      • YOLOv5
      • YOLOv6
      • YOLOvX

---

2024.08.16

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.08.16.0	v1.0.1	v0.5.8	v0.1.8	-	v0.5.0	v0.3.0
2024.08.16.1	v1.0.5	v0.5.8	v0.1.8	v0.0.4	v0.5.0	v0.3.0

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.5.8 optimum-rbln==0.1.8 vllm-rbln==0.0.4
  

• RBLN Compiler:

  • Improved visualization of the compilation progress bar
  • Optimized performance for long sequence LLM models
  • Reduced DRAM memory consumption for RSD models
  • Fixed bug for PReLU handling logic
  • Initial release of C/C++ runtime libraries:
    • Refer Software > API > Language Binding > C/C++ for installation, API docs, and tutorials.
• Optimum RBLN:
  • Added functions:
    • RBLNRobertaForMaskedLM()
    • RBLNRobertaForSequenceClassification()
    • RBLNXLMRobertaModel()
    • RBLNXLMRobertaForSequenceClassification()
• vLLM RBLN:
  • Updated to support GPT2 and Mi:dm architectures
• RBLN Model Zoo:
  • Initial release to support torch.compile() in PyTorch2.0:
    • Visit Tutorial > Basic > PyTorch (Vision), Tutorial > Basic > PyTorch (NLP), and Software > API > Python API pages for more information
    • Examples can be found in RBLN Model Zoo repository
  • Added new models (HuggingFace)
    • Mi:dm-7b
    • BGE-M3
    • BGE-Reranker-v2-M3
    • SecureBERT
    • Roberta

---

2024.07.25

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.07.25.0	v1.0.1	v0.5.7	v0.1.7	-	v0.4.1	v0.3.0
2024.07.25.1	v1.0.5	v0.5.7	v0.1.7	v0.0.3	v0.4.1	v0.3.0

• Install Command

  pip3 install -i https://pypi.rbln.ai/simple rebel-compiler==0.5.7 optimum-rbln==0.1.7 vllm-rbln==0.0.3
  

• RBLN Compiler:

  • Optimized RSD performance for long sequence LLMs
• Optimum RBLN:
  • Added warning messge for dependency version compatibilities
  • Added RBLNDPTForDepthEstimation() functiuon
  • Fixed bug for memory leak in GPT models
• RBLN Model Zoo:
  • Added a new model (HuggingFace)
    • DPT-large
• Others
  • Updated LLM Serving tutorial page
    • Revised the Serving with Triton Inference Server and Continuous Batching Support with vllm-rbln sections
    • Added OpenAI Compatible API Server section

---

2024.07.10

SDK Version	Driver	Compiler	Optimum RBLN	vLLM RBLN	Model Zoo	RBLNServe
2024.07.10.0	v1.0.1	v0.5.2	v0.1.4	-	v0.4.0	v0.3.0
2024.07.10.1	v1.0.5	v0.5.2	v0.1.4	v0.0.3	v0.4.0	v0.3.0

• RBLN Driver:
  • Enhanced stability for Rebellions Scalable Design (RSD)
• RBLN Compiler:
  • Updated to support continuous batching
• Optimum RBLN:
  • Updated LlamaForCausalLM() class to support continuous batching
• vLLM RBLN
  • Initial release to support continuous batching
  • Updated the LLM Serving page to include information on continuous batching
• RBLN Model Zoo:
  • Public release of the RBLN Model Zoo:
    • https://github.com/rebellions-sw/rbln-model-zoo
  • Added a new model (PyTorch)
    • ConvTasNet
  • Miscellaneous:
    • Removed pipeline() from BERT mlm inference.py
    • Removed pipeline() from BERT qa inference.py
    • Added trust_remote_code=True to the load_dataset() method in AST & Wav2Vec.

---

2024.06.11: Breaking Changes

SDK Version	Driver	Compiler	Optimum RBLN	Model Zoo	RBLNServe
2024.05.23.0	v0.10.42	v0.3.11	v0.1.0	v0.3.6	v0.1.5
2024.06.11.0	v1.0.1	v0.5.0	v0.1.1	v0.3.10	v0.3.0

Note

BREAKING CHANGES: Please update the RBLN Compiler to the appropriate version as below for compatibility with the updated RBLN Driver. You can check your RBLN Driver version with the rbln-stat -j | grep KMD_version command.

• 0.10.42: pip install -i https://pypi.rbln.ai/simple rebel-compiler==0.3.11
• 1.0.1: pip install -i https://pypi.rbln.ai/simple rebel-compiler==0.5.0

• RBLN Driver:
  • Stable release for Rebellions Scalable Design (RSD)
• RBLN Compiler:
  • Updated RBLN Compiler to be compatible with RBLN Driver
  • Added utility APIs:
    • npu_is_available()
    • get_npu_name()
• Optimum RBLN:
  • Updated model APIs
    • Please refer to RBLN Model Zoo below
• RBLN Model Zoo:
  • Added new models (HuggingFace)
    • With Rebellions Scalable Design (RSD)
      • Llama3-8b
      • SOLAR-10.7b
      • EEVE-Korean-10.8b
    • SDXL-base-1.0
    • ControlNet
• RBLNServe:
  • Pinned rebel-compiler version to <0.6, >=0.5

---

2024.05.23: Breaking Changes

SDK Version	Driver	Compiler	Optimum RBLN	Model Zoo	RBLNServe
2024.05.23.0	v0.10.42	v0.3.11	v0.1.0	v0.3.6	v0.1.5
2024.05.23.1	v0.12.37	v0.4.0	v0.1.0	v0.3.6	v0.2.0

Note

BREAKING CHANGES: Please update the RBLN Compiler to the appropriate version as below for compatibility with the updated RBLN Driver. You can check your RBLN Driver version with the rbln-stat -j | grep KMD_version command.

• 0.10.42: pip install -i https://pypi.rbln.ai/simple rebel-compiler==0.3.11
• 0.12.37: pip install -i https://pypi.rbln.ai/simple rebel-compiler==0.4.0

• RBLN Driver:
  • Added support for Rebellions Scalable Design (RSD)
  • rbln-stat (CLI tool) update: Added new columns Name and Power for NPU version and power consumption, respectively
• RBLN Compiler:
  • Updated RBLN Compiler to be compatible with the RBLN Driver
  • Updated input arguments of python user APIs
  • Added new user APIs for concurrent processing
  • Enabled LLM compilation & inference for Rebellions Scalable Design (RSD)
  • Added a new page - Nvidia Triton Inference Server
• Optimum RBLN:
  • Initial release
  • Add new pages - HuggingFace Model Support
• RBLN Model Zoo:
  • Added new models (HuggingFace)
    • With Rebellions Scalable Design (RSD)
      • Llama2-7b
      • Llama2-13b
    • GPT2, GPT2-medium/large/xl
    • T5-small/base/large/3B
    • BART-base/large
    • BERT-base/large
    • Stable Diffusion v1.5
    • SDXL-turbo
    • Whisper-tiny/base/small/medium/large
    • Wav2Vec2
    • Audio Spectrogram Transformer
• RBLNServe:
  • Pinned rebel-compiler version to <0.5, >=0.4

---

2024.01.31: Breaking Changes

SDK Version	Driver	Compiler	Model Zoo	RBLNServe
2024.01.31.0	v0.10.42	v0.3.5	v0.2.0	v0.1.5

Note

BREAKING CHANGES: Please update the RBLN Compiler to the latest version (v0.3.5 or higher) for compatibility with the updated RBLN Driver.

• RBLN Driver:
  • Refactored device internal command processing logic for stability & scalability
• RBLN Compiler:
  • Updated RBLN Compiler to be compatible with the RBLN Driver
  • Updated device memory scheduling logic
  • Enhanced functionality for operation fusion logic
  • Updated supported OP list for both TensorFlow and Pytorch
• RBLN Model Zoo:
  • Added new models (PyTorch):
    • YOLOv4: v4/v4-csp-s-mish/v4-csp-x-mish
    • Video ResNet: r3d_18/mc3_18/r2plus1D_18
    • Video S3D: s3d
  • Changed default input size:
    • YOLOv3/4/5/6/7/8
    • deeplabv3_resnet50/resnet101/mobilenetv3_large, fcn_resnet50/101, unet
  • Restructured directories:
    • PyTorch image classification examples are moved from rbln_model_zoo/pytorch/vision/classification to rbln_model_zoo/pytorch/vision/image_classification
• RBLNServe:
  • Set rebel-compiler version pinned to <0.4, >=0.3

---

2023.10.06

SDK Version	Driver	Compiler	Model Zoo	RBLNServe
2023.10.06.0	v0.9.34	v0.2.13	v0.1.9	v0.1.4

• RBLN Compiler:
  • Updated version parsing module of runtime APIs
  • Updated runtime input size calculation logic
  • Enhanced functionality for tensor slicing operations
• RBLNServe:
  • Updated configuration for gRPC/REST protocol

---

2023.09.12

SDK Version	Driver	Compiler	Model Zoo	RBLNServe
2023.09.12.0	v0.9.34	v0.2.10	v0.1.9	v0.1.1

• RBLN Compiler:
  • Enabled print() for the rebel.Runtime module - print(module) will show basic information of the loaded model
  • Refactored compiler internal large op handling passes for scalability
  • Updated error message handling logic
  • Fixed bug in a type cast pass
• RBLN Model Zoo:
  • Updated submodule - YOLOv3
• RBLNServe:
  • Added --version command

---

2023.08.18

SDK Version	Driver	Compiler	Model Zoo	RBLNServe
2023.08.18.0	v0.9.34	v0.2.1	v0.1.8	v0.1.0

• RBLN Compiler:
  • Fixed bug for the destruction issue in rebel.Runtime
• RBLNServe:
  • Initial release
  • Added a new page - RBLNServe (Model Server)

---

2023.08.12: Breaking Changes

SDK Version	Driver	Compiler	Model Zoo
2023.08.12.0	v0.9.34	v0.2.0	v0.1.8

Note

BREAKING CHANGES: Please update the RBLN Compiler to the latest version (v0.2.0 or higher) for compatibility with the updated RBLN Driver.

• RBLN Driver:
  • Refactored host-device communication protocol for stability & scalability
• RBLN Compiler:
  • Updated RBLN Compiler to be compatible with the RBLN Driver
• Others:
  • Added a new page - Kubernetes Support

---

2023.07.31

SDK Version	Driver	Compiler	Model Zoo
2023.07.31.0	v0.8.44	v0.1.17	v0.1.8

• RBLN Compiler:
  • Enhanced functionality for normalization operations
  • Updated compiler internal scheduling logic
  • Updated error message handling logic
• RBLN Model Zoo:
  • Updated requirements.txt to use ultralytics 8.0.145
  • Applied ultralytics 8.0.145 to YOLOv8

---

2023.07.10

SDK Version	Driver	Compiler	Model Zoo
2023.07.10.0	v0.8.44	v0.1.14	v0.1.7

• RBLN Driver:
  • Enhanced stability for device reset and recovery
  • rbln-stat (CLI tool) update: status categorization of the process
• RBLN Compiler:
  • Updated input arguments for compile_from_torchscript()
  • Enhanced functionality for unary and binary operations
  • Optimized build time
• RBLN Model Zoo:
  • Added new models (PyTorch):
    • YOLOv6: v6s/v6n/v6m/v6l
    • YOLOv7: v7-tiny/v7/v7x
    • YOLOv8: v8s/v8n/v8m/v8l/v8x
  • Added new models (TF Keras Applications)
    • MobileNetV3: Small/Large
    • ConvNeXt: Tiny/Small/Base/Large/XLarge
    • RegNetX: 002/004/006/008/016/032/040/064/080/120/160/320
    • RegNetY: 002/004/006/008/016/032/040/064/080/120/160/320

---

2023.06.20

SDK Version	Driver	Compiler	Model Zoo
2023.06.20.0	v0.7.34	v0.1.8	v0.1.5

• RBLN Compiler:
  • Added a new compile function - compile_from_torchscript()
  • Enhanced functionality for matrix multiplication and pooling operations
  • Optimized device memory scheduling
• RBLN Model Zoo:
  • Added new models (PyTorch):
    • YOLOv3: v3-tiny/v3/v3-spp
    • YOLOv5: v5s/v5n/v5m/v5l/v5x

---

2023.05.26

SDK Version	Driver	Compiler	Model Zoo
2023.05.26.0	v0.7.34	v0.1.5	v0.1.4

• Initial release