GLM 5.3 Flash on TP4 DGX Sparks (switchless)

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

In case anyone is interested:

I’m quite proud of this one and open to any feedback:

  • Upstream FP8 GLM 5.3 Flash
  • Switchless direct QSFP DAC setup (no switch needed)
  • GB10 performance-tuned, near parity speed to NVFP4/EXL3
  • Optimized for agent parallelisation and coding tasks
  • Recipe w 256k context; tested up to 500k

Any feedback would be welcome.

2

So what is different with your setup compared to this one:

3

Different focus!! FujitsuPolycom have been a huge driver and have basically “solved” che switchless setup for my sparks, but that’s not where I have been focusing! My repo is focusing only on GLM 5.3 Flash, my goal was to make it FAST without using any quantised version of the model. I wanted fast, full precision, upstream FP8 - and I have optimised a lot of the numbers for performance! For a local upstream model of GLM 5.3 Flash, I think it now runs quite fast!

Sorry for the X’ish style punchcard, but the numbers here are very nice for an FP8 and I think it will answer your question:

bf3de099-5f66-4d8e-a6e9-47e4c4463e25
bf3de099-5f66-4d8e-a6e9-47e4c4463e251536×1024 429 KB

4

I was simply wondering if you had come up with a different approach to switchless.

I’m glad his patched NCCL is gaining tracktion.