Chunk 2.0

In this chunk, the assistant successfully resolved the critical NaN crash during decode by using `--nsa-decode-backend trtllm` and `--nsa-prefill-backend trtllm`, which produced coherent output on the SM120 GPUs. Baseline benchmarks with 64 concurrent requests showed ~225 output tok/s and ~516 total tok/s. The assistant then tuned the server by increasing `--mem-fraction-static` to 0.92 and enabling CUDA graphs, which captured successfully without OOM, though throughput remained similar (~210–247 output tok/s). Various MoE runner backends (`flashinfer_cutlass`, `flashinfer_cutedsl`) were tested with comparable results, and the `flashinfer_trtllm` path was found to be SM100-only. The user raised the possibility of expert parallelism to improve PCIe-bound performance. The assistant analyzed memory constraints (453GB of MoE experts vs. 96GB/GPU) and concluded full replication was impossible; standard EP8 offered no advantage due to the small hidden size and similar communication volume. The user then questioned whether cross-GPU latency from the Proxmox VM environment was a bottleneck. Investigation confirmed the system is a KVM/QEMU VM with no direct GPU peer-to-peer support (NS status), forcing all cross-GPU transfers through host memory. A bandwidth test showed ~32 GB/s for large transfers but only ~1 GB/s for small messages typical of all-reduce, indicating that virtualization overhead significantly contributes to the latency-limited throughput. The main achievements include deploying GLM-5-NVFP4 with a working configuration, establishing baseline throughput metrics, and identifying virtualization-induced P2P limitations as a key performance constraint. Themes center on overcoming SM120-specific issues (NSA backend selection), tuning for throughput on PCIe-only Blackwell GPUs, and diagnosing infrastructure bottlenecks in a virtualized environment.

Deploying GLM-5-NVFP4 on Blackwell GPUs: From NaN Crashes to Virtualization Bottlenecks 3524 words

Message Articles

Subagent Sessions