Tag
Explore by tags
GPU kernel library for LLM inference attention, sampling, and KV-cache, built on block-sparse formats with JIT-compiled customizable templates. Reports 29-69% inter-token-latency cuts vs compiler backends; powers SGLang, vLLM, and MLC-Engine.
Applies quantization, pruning, distillation, speculative decoding and sparsity to compress and accelerate deep learning models, producing exportable checkpoints ready for deployment on TensorRT‑LLM, TensorRT, vLLM and similar inference runtimes.
Streamlines the full lifecycle of foundation models — data prep, fine-tuning (SFT/LoRA/QLoRA/GRPO), evaluation, and deployment — with ready-to-run recipes, multi-engine inference support, and cloud/CLI workflows for both laptop experiments and large-scale runs.
Cloud-native control plane that scales vLLM on Kubernetes, adding the routing, autoscaling, and fault tolerance single-instance serving lacks. Brings high-density LoRA management, an LLM gateway, distributed KV cache reuse, and SLO-aware GPU serving.
Disaggregated LLM serving architecture that splits prefill and decode into separate clusters and pools spare CPU, DRAM, and SSD into a distributed KVCache. Powers Kimi in production, handling 75% more requests under the same SLOs.
Open-source TTS that clones a voice from 3-10s of audio and synthesizes cross-lingual speech in 9 languages and 18+ Chinese dialects. Supports streaming at ~150ms latency with instruction control over emotion, speed, and accent.
Open-source HybridFlow implementation for RL post-training of LLMs. Decouples control flow from compute so PPO, GRPO, GSPO and DAPO share one dataflow; pairs FSDP/Megatron with vLLM/SGLang rollout and reports 1.5-20x throughput over prior RLHF stacks.
Full-stack AI red‑teaming platform that fingerprints AI infrastructure for known CVEs, audits MCP servers and agent skills with LLM-driven analysis, and runs cross-model jailbreak evaluations; designed for hands-on security assessment of AI deployments.
A 671B-parameter Mixture-of-Experts language model (37B activated) trained on 14.8T tokens with 128K context, FP8-first training, a Multi-Token Prediction module, and Hugging Face weights—focused on efficient MoE training and long-context use cases.
Provides an open platform of omnimodal world models, datasets, and tools to build Physical AI — joint perception, generation, and action reasoning for robots, autonomous vehicles, and smart infrastructure. Supports images, video, audio, and action-conditioned workflows.
Provides a hardware plugin that runs vLLM on Huawei Ascend NPUs by mapping vLLM execution and memory management to the Ascend runtime. Key features: support for Transformer/MoE/embedding/multimodal models, official docs, CI-backed release branches and community maintenance.
Zero-shot, single‑reference voice cloning TTS with multilingual support (ZH/EN/JA/ES/AR), fine-grained emotion and duration control, and pronunciation hooks (Pinyin/CMU/Kana); ships model weights, Web UI and production deployment recipes for local or server use.