Strix Halo (gorgon-halo) DwarfStar ROCm 10.0 GLM-5.3-Flash Q2 64k-depth decode

Protocol deterministic-prime-extend-completions-v1, runner local-llm-speed v1, measured at the OpenAI-completions API endpoint. Mean of 2 repeats after 1 warmup. Concurrency 1, so the figure is per-request, not aggregate. Depth is the primed KV prefix; at depth 0 the whole prompt is computed (actual cached depth 0, recomputed history 0). Prefill is effective follow-up prompt processing: uncached tokens divided by client-observed time to first output. Decode is client-observed generation throughput over the tokens after the first, so TTFT is excluded. KV cache type default, speculative decoding off. Source: local-llm-benchmarks.dev/data/results.json. Run run-ebfd0d8410eb2fd8, completed 2026-09-25. Engine DwarfStar 8588ab028357104b2d4aab32d249948ab9b5d41f5bd3bcf27f924907200f20eb /usr/local/bin/ds4-server, runtime ROCm 10.0. Container docker.io/kyuz0/strix-halo-ds4-toolbox:rocm-10.0. Weights: local DwarfStar GGUF of zai-org/GLM-5.3-Flash file GLM-5.3-Flash-Q2.gguf, quant Q2. Prompt 2048 tokens, 256 generated, primed depth 65536.

Decode drops from 11.92 tok/s at depth 0 to this figure at a 65536 primed prefix.