Chunhung Wang#
DCGPU Customer Software Engineer with a strong focus on supporting Tier-1 customers in the APAC region. Experienced in AMD ROCm ecosystem and responsible for enabling and debugging across the full software stack including HIP/ROCr Runtime, kernel and driver-level integration, system monitoring tools, and performance-critical libraries such as VLLM, RCCL, CK, and hipBLASLt. Proven track record in large-scale model inference, GPU system optimization, and delivering technical solutions in collaboration with customer engineering and product teams.
Posts by Chunhung Wang
Iteratively Tuning hipBLASLt TensileLite Kernels: A Smaller Search, a Faster Kernel
Learn why a small iterative search beats a big one-shot sweep for hipBLASLt TensileLite kernels, in less time and with no risk of regression.
Reverse-Engineering hipBLASLt TensileLite Kernels: From Solution Name to a Tuning Config
Pin the pool's best kernel into a TensileLite tuning config by decoding its solution name, so an expanded re-tune can only match or beat it.
Serve Kimi-K2.5-MXFP4 on MI355X with ATOM
Serve Kimi-K2.5-MXFP4 on MI355X with ATOM and gfx950 block-scaled FP4 kernels for optimized LLM inference.
When a Faster Kernel Doesn't Speed Up Serving: Profiling FP8 KV Cache on AMD Instinct MI308X
Learn how a 34% faster FP8 KV cache kernel delivered 0% E2E speedup, and how profiling attribution exposed the hidden dtype-cast cost on MI308X.
RDC and RocProfiler Compared to DCGM for Commonly Used Metrics
Learn how CLI commands and Python code help you evaluate app performance without a profiler, with examples explaining what each metric means.
Optimizing MI300X Inter-Chiplet Communication via the RCCL Tuner API
Learn how to build a topology-aware RCCL tuner plugin for MI300X CPX/NPS4 mode and validate it with rccl-tests.
Faster Kimi-K2.5-W4A8 Decoding with EAGLE3 on AMD Instinct™ MI325X
Add EAGLE3 speculative decoding and three MoE/FMHA kernel-tuning patches to Kimi-K2.5-W4A8 inference on AMD Instinct™ MI325X with SGLang, AITER, and FlyDSL.
Building and Deploying Custom hipBLASLt Libraries on AMD Instinct GPUs
Learn how to manage hipBLASLt environments with custom source builds, RPM/DEB packaging, and version switching on AMD Instinct GPUs.
Further Accelerating Kimi-K2.5 on AMD Instinct™ MI325X: W4A8 & W8A8 Quantization with AMD Quark
Quantize Kimi-K2.5 to W4A8 and W8A8 using AMD Quark and serve on MI325X with FlyDSL and AITER for further inference acceleration.
Customizing Kernels with hipBLASLt TensileLite GEMM Tuning - Advanced User Guide
Master hipBLASLt TensileLite Tuning. Learn to build custom kernels that deliver 150%-250% faster GEMM performance on AMD Instinct™ MI300X GPUs
Accelerating Kimi-K2.5 on AMD Instinct™ MI300X: Optimizing Fused MoE with FlyDSL
Optimize Kimi-K2.5 on AMD MI300X using FlyDSL for fused MoE kernel acceleration. Achieve faster TTFT, TPOT, and throughput with our step-by-step optimization guide.
Adaptive Top-K Selection: Eliminating Performance Cliffs Across All K Values on AMD GPUs
Explore adaptive Top-K on MI300X! See how auto-selection and hardware optimizations like DPP and double buffering drive peak efficiency.
Debugging NaN Results in CK Tile GEMM: A rocgdb Detective Story
Learn GPU kernel debugging with rocgdb through a real case: tracing NaN outputs to a one-character typo in CK Tile GEMM
Day 0 Developer Guide: hipBLASLt Offline GEMM Tuning Script
Learn how to improve model performance with hipBLASLt offline tuning in our easy-to-use Day 0 tool for developers to optimize GEMM efficiency
Avoiding LDS Bank Conflicts on AMD GPUs Using CK-Tile Framework
This blog shows how CK-Tile’s XOR-based swizzle optimizes shared memory access in GEMM kernels on AMD GPUs by eliminating LDS bank conflicts