GLM-5.3-Flash + DFlash2 on DGX Station GB300

No DFlash2 post exists yet for this model on GB300-class hardware. The published recipes (tonyd2wild on vLLM (h/t @tonyd615), beastllama on SGLang) target DGX Spark / GB10 at SM121. This is the SM100 lane: GLM-5.3-Flash NVFP4 with an incoai DFlash2 drafter on a single GB300, TP1.

The bit that decides everything on the Station: the MXFP8 drafter needs the B12X kernel stack, which is SM120/SM121-only and will not run on GB300 (SM100). So on the Station you use the BF16 incoai drafter. That is what this is.

Stack

  • Target: local-inference-lab/GLM-5.3-Flash-NVFP4 (calibrated, ~169B), fp8 KV, 1M context
  • Drafter: incoai/GLM-5.3-Flash-DFlash2 (BF16 block-diffusion, 7-token blocks)
  • Runtime: vLLM 0.1.dev20051+g487ecf187 (the glm53-flash image) plus a ported DFlash2 overlay, tonyd2wild’s GB10 overlay adapted to SM100 geometry
  • Build: ~25s overlay on the public vllm/vllm-openai:glm53-flash base. No source build, no B12X.

Numbers (single GB300, TP1, temp 0, stream:false, no-think)

config decode
no spec ~137 tok/s
DFlash2 (7-token blocks) ~305-470 tok/s

Draft acceptance runs 54-67%, rising warm, with healthy per-position decay across the 7-token block (207 down to 96). The drafter, target, and rejection sampler are wired correctly on SM100. The range reflects prompt, context, and RTT; code-style prompts land at the top.

The port’s hard part is done up front. GLM-5.3’s custom KV grouping (34 KDA/mamba + 11 MLA slot-sharing) is the DFlash2 minefield. I re-derived the exact-fit geometry for GB300 and made it a build-time check (sim_glm5_drafter_hades.py): drafter block 2048 to 4288, real page == mla_page (4,390,912 B), per-block KV cost unchanged. The build fails if it ever breaks. First boot came up clean.

Everything (overlay patches, GB300 geometry sim, stack recipe) is in a prebuilt image: ghcr.io/ebfio/glm53-flash-dflash2-gb300:latest (pin sm100-gb300-20260901).

Compose change vs your non-spec glm53-flash compose: swap the image and add

--speculative-config '{"method":"dflash","model":"/models/.../incoai--GLM-5.3-Flash-DFlash2/snapshots/<snap>","num_speculative_tokens":7}'

num_speculative_tokens must be 7 (drafter block 8 minus the target’s own token).

Caveats

  • DFlash2 drafts text only. Image and video requests pass through unspeculated.
  • Long-prefill stability past ~32K tokens on this stack is fully verified!
  • incoai/GLM-5.3-Flash-DFlash2 is CC BY-NC-ND 4.0 (non-commercial). Respect the terms.
  • Validated on one GB300 at TP1. Other SM100 configs are untested.
  • Credit to tonyd2wild’s GB10 overlay this is ported from, and to incoai for DFlash2. Happy to share more numbers or logs for anyone reproducing on a Station.

No one has GB300 money except you ebfio but goodness those numbers are crazy

We (Local Inference Lab) will be working on this soon, including the fused MoE support from b12x for sm103.

Assuming Exxact ships my Station!!

@el8 I’m a big fan of your work! My 4x RTX 6000 Pro hums along amazingly well with GLM-5.3!

Still scratching the surface on the Station.. another one is coming soon 😄

Join us on our Discord. Several users have their Stations.