DeepLink
Official@deeplink-org
Offers specialized optimization and precision alignment for domestic hardware backends including Ascend, CAMB, and MACA architectures.
Agent Skills by DeepLink
Showing 28 vetted skills indexed across 2 GitHub repositories.
DeepLink-orgtemplate-skill
Provides a reusable template for authoring hardware benchmark evaluation skills.
DeepLink-orghygon-science-weather
Runs weather forecasting model inference and RMSE/ACC evaluation on Hygon DCU.
DeepLink-orghygon-audio-asr
Benchmarks SenseVoice speech recognition inference performance on Hygon DCU hardware.
DeepLink-orghygon-audio-au
Evaluates ECAPA-TDNN language identification inference accuracy and latency on Hygon DCU.
DeepLink-orghygon-nlp-operator
Benchmarks GEMM, Conv2d, longtail, and Transformer Block operators on Hygon DCU hardware.
DeepLink-orgmetax-nlp-operator
Benchmarks GEMM, Conv2d, long-tail operators, and Transformer blocks on MetaX GPUs.
DeepLink-orgiluvatar-nlp-operator
Benchmarks GEMM, Conv2d, longtail, and Transformer Block operators on Iluvatar BI-V150 GPUs.
DeepLink-orgascend-audio-au
Benchmarks language identification model inference accuracy and latency on Ascend NPU.
DeepLink-orgascend-audio-asr
Benchmarks SenseVoice speech recognition inference on Ascend NPU with CER and WER metrics.
DeepLink-orgascend-nlp-operator
Runs operator accuracy and latency benchmarks on Huawei Ascend 910B NPUs.
DeepLink-orgcambricon-nlp-operator
Benchmarks operator accuracy and latency on Cambricon MLU hardware.
DeepLink-orgnvidia-audio-asr
Evaluates SenseVoice speech recognition inference performance on NVIDIA GPUs with CER and WER metrics.
DeepLink-orgnvidia-audio-au
Benchmarks language identification model inference accuracy and latency on NVIDIA GPUs.
DeepLink-orgnvidia-nlp-operator
Benchmarks GEMM, Conv2d, long-tail, and Transformer Block operators on NVIDIA GPUs.
DeepLink-orgnvidia-nlp-inference
Benchmarks SGLang-served HuggingFace text models on NVIDIA GPUs for throughput and latency.
DeepLink-orgnvidia-nlp-training-qwen3
Benchmarks Qwen3-8B pretraining throughput on NVIDIA GPUs using NeMo and Megatron-LM.
DeepLink-orgnvidia-nlp-finetune-mixtral
Runs Mixtral-8x7B finetuning benchmarks on NVIDIA GPUs and collects throughput metrics.
DeepLink-orgnvidia-nlp-finetune-llama-7b
Benchmarks Alpaca-LoRA LLaMA-7B fine-tuning throughput on NVIDIA GPUs.
DeepLink-orgnvidia-mm-t2v
Benchmarks Open-Sora v2 text-to-video inference performance on a single NVIDIA GPU.
DeepLink-orgnvidia-mm-t2i
Benchmarks Stable Diffusion text-to-image inference on NVIDIA GPUs using TensorRT.
DeepLink-orgnvidia-cv-pretrain
Benchmarks image classification model training performance on NVIDIA GPUs using mmpretrain.
DeepLink-orgnvidia-cv-segmentation
Benchmarks semantic segmentation training performance on NVIDIA GPUs using mmsegmentation.
DeepLink-orgnvidia-cv-detection
Benchmarks object detection model training performance on NVIDIA GPUs using mmdetection.
DeepLink-orgppu-audio-au
Evaluates speech understanding model inference accuracy and latency on PPU accelerators.
Frequently Asked Questions About DeepLink
FAQPage SchemaWhat hardware backends are supported for model adaptation?▼
DeepLink supports model adaptation and precision alignment specifically for Ascend, CAMB, MACA, and PPU hardware backends. These capabilities ensure that models function correctly when deployed on domestic compute infrastructure by addressing backend-specific execution requirements.
How can I resolve precision mismatches in my model deployment?▼
You can resolve precision mismatches by utilizing the precision-align capability to diagnose discrepancies between lmdeploy, dlinfer, and reference implementations. This process identifies numerical errors occurring during execution on Ascend, CAMB, or MACA backends to ensure output parity.
What is the primary focus of the graph-mode-internals skill?▼
The graph-mode-internals skill focuses on explaining the capture and replay mechanisms required for efficient execution on Ascend, Camb, MACA, and PPU backends. It provides technical insight into how computational graphs are managed across these specific hardware environments.