Strix Halo retained-PM4 llama-server Qwen3.8-Flash-Next MTP decode
43.9 tok/s (per-request, parallel 1)
Fields
id
benchmarks/kyuz0-8060s-strixllama-qwen38-flash-next-mtp-decodeMetric
decode_tok_s
Value
43.9
Unit
tok/s (per-request, parallel 1)
Model
Qwen3.8-Flash-Next (shared MTP head)
Methodology
kyuz0 docs/building.md, retained-PM4 Strix Halo build named rocm-10.0-strix-llama. llama-server from halo-box/strix-llama.cpp with a custom ROCm runtime built from pwilkin/rocm-systems:ilintar-experiments. Flags: -dev ROCm0 -ngl all -fa on -fit off --load-mode none --lazy-mode on-direct -ctk f16 -ctv f16 -c 262144 -b 16384 -ub 16384 --parallel 1 --spec-type draft-mtp,ngram-mod --spec-draft-n-max 3. pp2048 at depth 0, speculative decoding on (draft MTP plus n-gram, draft-n-max 3). The page does not state the llama.cpp commit or the host RAM size. Per-request because --parallel 1.
Measured by
third_party
Reproducible
yes
notes
Decode with the shared MTP head and n-gram draft, draft-n-max 3. Roughly double the 19.5-20.6 tok/s baseline decode of the same model without speculation on the same class of box.
Notes
Decode with the shared MTP head and n-gram draft, draft-n-max 3. Roughly double the 19.5-20.6 tok/s baseline decode of the same model without speculation on the same class of box.