Chunk 38.0

This chunk marked the transition from experimental benchmarking to a hardened production deployment of the Kimi-K2.5 INT4 model. The optimal configuration (EAGLE-3 topk=1, spec_v2 overlap) was finalized after confirming it matched or beat baseline throughput at high concurrency. A crash caused by a missing attribute in the dynamic speculation disable patch was fixed, and the setup was codified into a systemd service (`sglang-kimi.service`) with auto-start on boot. Critical operational fixes included adding `--tool-call-parser kimi_k2` and `--reasoning-parser kimi_k2` to properly structure tool calls and reasoning content in the API output, and enabling the hierarchical KV cache (`--enable-hierarchical-cache --hicache-ratio 4.0`) to utilize ~358 GB of system RAM as an L2 prefix cache, significantly boosting effective memory capacity. A tangent experiment involved passing the 8 GPUs to a VM, which required managing the driver binding on the Proxmox host. The main challenge was that the proprietary NVIDIA driver in the VM failed to initialize the Blackwell GPUs. This was resolved by installing the open kernel module package (`nvidia-dkms-590-open`), which is required for SM120 support. A workflow was established for rebinding GPUs between the host nvidia driver and the VM's `vfio-pci` driver, enabling relatively painless switching between the LXC container and the VM. The user then pivoted to deploying a newer, more efficient model (`nvidia/Qwen3.5-397B-A17B-NVFP4`). This required building the latest SGLang main branch from source to support the new model architecture and `modelopt_fp4` quantization, alongside applying the necessary SM120 patches. Initial startup produced NaN outputs due to incompatible default FP4 GEMM and MoE backends on Blackwell. The issue was resolved by explicitly setting `--moe-runner-backend flashinfer_cutlass` and `--fp4-gemm-runner-backend flashinfer_cudnn`, which stabilized the model and produced correct generations.

From Optimization to Production: The Hardening and Evolution of a Multi-GPU LLM Deployment 2819 words

Message Articles

Subagent Sessions