Chunk 6.0

In this chunk, the assistant significantly improved the inference throughput for GLM-5-NVFP4 on 8x RTX PRO 6000 GPUs by addressing several key bottlenecks. The primary fix was enabling FlashInfer CUTLASS MoE autotune for SM120 by patching `model_runner.py` to include `flashinfer_cutlass` in the autotune list, and raising `--max-running-requests` from 64 to 1024. Combined with `--disable-cuda-graph` and `--disable-radix-cache`, total token throughput jumped from ~880 tok/s to ~1,950 at 256 concurrency, ~2,800 at 512, and ~3,740 at 1024 concurrency, with peak output reaching nearly 4,000 tok/s. However, GPU power draw remained around 250W out of a 600W TDP, indicating the hardware was still underutilized. The assistant then investigated the root cause of the low power and identified that FlashInfer's allreduce fusion was disabled on SM120 because the underlying TRT-LLM communication kernels only supported SM90/SM100. Attempts to patch the flashinfer kernel to add SM120 support (by modifying the architecture gate and version check) allowed the server to start, but the fusion performed poorly, dropping throughput to 236 tok/s and power to 125W, suggesting synchronization issues on SM120. The assistant reverted those changes and instead explored NCCL tuning and `--num-continuous-decode-steps 4`, neither of which yielded meaningful gains. The chunk concludes with the assistant trying the `flashinfer_trtllm` MoE backend as an alternative path to better GPU utilization. The overarching theme is the challenge of adapting a large MoE inference stack designed for datacenter Blackwell (SM100) to consumer Blackwell (SM120) hardware, which has different constraints like smaller shared memory and missing kernel support for advanced features like allreduce fusion. The assistant's approach of forking and patching both sglang and flashinfer code demonstrates a willingness to modify upstream dependencies to unlock performance, but the allreduce fusion attempt highlights the risk of using architecture-specific synchronization primitives without thorough validation. The next steps will likely involve either finding a working allreduce fusion approach for SM120 or accepting the PCIe bottleneck and focusing on other optimizations like TP4+PP2 or deeper MoE kernel tuning.

The Blackwell Divide: How a 4× Throughput Breakthrough Revealed the Hidden Gap Between Consumer and Datacenter GPU Inference 2403 words

Message Articles

Subagent Sessions