kyuz0 Strix Halo 8060S gemma-4-31B-it-BF16 ROCm 7 nightlies 32k tg32
3.25 tok/s
Fields
id
benchmarks/kyuz0-8060s-gemma-4-31b-it-bf16-rocm7n-32k-tg32Metric
decode_tok_s
Value
3.25
Unit
tok/s
Model
gemma-4-31B-it-BF16
Methodology
kyuz0 amd-strix-halo-toolboxes llama-bench sweep (results.json generated 2026-05-18, Fedora 43 kernel 6.19.12-200, llama.cpp build 0253fb21f / b9187). ROCm 7 nightlies, -fa 1 -ngl 99 --mmap 0. Long-context cell: context depth 32768 tokens, generation of 32 tokens. ONE request, batch size 1, so the number is per-request, not aggregate across concurrent requests. Mean of the repeated runs; std 0.0 tok/s. Model file 57.18 GiB, 30.7 B params, quant BF16.
Measured by
third_party
Reproducible
yes
notes
Not in the earlier kyuz0 filing, which covered only the default-context pp512/tg128 cells. Same results.json, the longctx32768 and longctx65536 rows.
Notes
Not in the earlier kyuz0 filing, which covered only the default-context pp512/tg128 cells. Same results.json, the longctx32768 and longctx65536 rows.