Skip to content
View WEIXILIN0307's full-sized avatar

Block or report WEIXILIN0307

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
WEIXILIN0307/README.md

Hi, I'm WEIXI LIN

I'm a master's student in Computer Technology at Northwestern Polytechnical University (NWPU). I work on efficient LLM inference and training systems, with a focus on runtime performance, GPU/NPU kernels, and KV cache management.

I'm looking for internships in LLM inference optimization and AI systems, remote or on-site.

Email · GitHub

Open-source contributions

Merged

  • verl-omni: Added CPU placement for native reward models (#694) and migrated the SD3.5 OCR recipe to named rewards (#695).
  • vLLM-Omni: Added remote SeedVR2 video restoration for ComfyUI (#8505).

In progress

  • verl-omni: Streaming named-reward scoring (#730) and two-rank native PickScore FSDP replicas (#855, draft).

  • Megatron-LM: Activation offloading for GDN recurrence (#7908), microbatch regression coverage (#8243), and context-parallel scheduling metrics (#8431).

  • SGLang-Omni: PersonaPlex streaming and startup improvements, including incremental history and session warmup.

  • vLLM: Multimodal prefix-cache boundary handling for hybrid Mamba models (#56818).

  • Mooncake: Encoder–Prefill–Decode disaggregation and reusable agent state (#2836).

View my upstream pull requests

Selected projects

  • MiniCPM-o on Ascend: Optimized multimodal inference across two NPUs, including CPU/NPU overlap, CANN operators, graph execution, and W8A8 quantization.
  • Efficient VLM inference: Optimized Qwen3-VL-2B with KV cache reuse, projection fusion, Triton kernels, and CUDA Graphs.
  • MetaX C500 kernels: Built and tuned fused quantization kernels with TileLang, profiling, and correctness tests.

Interests and tools

  • LLM inference, multimodal serving, KV and prefix caching
  • Distributed training, activation offloading, and runtime scheduling
  • Quantization, GPU/NPU kernels, profiling, and performance analysis

Python · PyTorch · CUDA · Triton · TileLang · CANN / AscendC · vLLM · llama.cpp

Competitions and awards

  • MiniCPM × Ascend Inference Optimization Challenge: 5th on the public leaderboard
  • IEEE AICAS 2026 Grand Challenge: 7th place
  • CCF Open Source Innovation Competition: Third Prize, Mooncake KVCache
  • Huawei MindSpore Model Development Challenge: 3rd place / Silver, S1-MoE
  • MetaX Open Source Talent Camp: Best Engineering Practice Award

Education

  • Northwestern Polytechnical University: Master's in Computer Technology, 2025–present
  • Shaanxi Normal University: Bachelor's in Software Engineering, 2021–2025

Pinned Loading

  1. Megatron-LM Megatron-LM Public

    Forked from NVIDIA/Megatron-LM

    Ongoing research training transformer models at scale

    Python

  2. sglang-omni sglang-omni Public

    Forked from sgl-project/sglang-omni

    High Performance Serving Runtime For Interactive Super Intelligence System

    Python

  3. verl verl Public

    Forked from verl-project/verl

    verl/HybridFlow: A Flexible and Efficient RL Post-Training Framework

    Python

  4. verl-omni verl-omni Public

    Forked from verl-project/verl-omni

    Multimodal RL training framework for diffusion & omni models

    Python

  5. vllm vllm Public

    Forked from vllm-project/vllm

    A high-throughput and memory-efficient inference and serving engine for LLMs

    Python

  6. vllm-omni vllm-omni Public

    Forked from vllm-project/vllm-omni

    A framework for efficient model inference with omni-modality models

    Python