I work on efficient LLM inference and GPU kernel optimization, with a focus on MoE and distributed inference.
Selected contributions:
| Project / Area | Status | Contributions |
|---|---|---|
| LMDeploy · GLM-5.3-Flash | 🟣 | #4968 Model support, #5027 Inference optimizations, #5045 FP8 MLA cache |
| LMDeploy · Kimi-K2.6 | 🟣 | #4846 PyTorch backend support |
| LMDeploy · Quantization | 🟣 | #4897 TurboMind W4A16 backend for AWQ linear |
| LMDeploy · Communication | 🟣 | #4915 Symmetric-memory LM-head all-gather |
| LMDeploy · KPool | 🟣 | #5067 Limit JIT specialization with bounded scoring |
| LMDeploy · MoE dispatch | 🟣 | #5071 Skip dummy/padded tokens in DeepEP low-latency dispatch. |
| LMDeploy · MoE kernels | 🟢 | #4947 CuTeDSL W4A16 MoE backend, #4956 FP8 MoE decode optimizations |
| LMDeploy · KDA | 🟢 | #5050 Unified KDA prefill with optional FlashKDA checkpoints, #5073 FlashKDA Docker integration |
| DeepGEMM · Grouped GEMM | 🟢 | #350 Shape-aware alignment for SM90 small-M grouped GEMM |
| Triton_Grouped_GEMM | 🟣 | Optimized BF16/FP8 grouped GEMM kernels in Triton. M-grouped / K-grouped · TMA support |
| DLBlas · Triton kernels | 🟣 | #140 MatterSim optimizations, #66 Permute / unpermute, #60 GRPO loss backward |
🟣 Merged · 🟢 Open · Status checked on 2026-10-29.
Complete contributions: