Radeon Pro W7900 48GB llama.cpp b9600: shisa-v2.1-unphi4-14b Q4_K_M decode
50.8 tok/s decode (tg128, batch-1, ctx 0.5k, FA on); prefill 1721 t/s
Fields
id
benchmarks/llamabench-w7900-shisa14b-q4km-tg128Metric
decode_tok_s
Value
50.8
Unit
tok/s decode (tg128, batch-1, ctx 0.5k, FA on); prefill 1721 t/s
Model
shisa-v2.1-unphi4-14b Q4_K_M
Methodology
llamabench.ai community row, reporter @lhl: llama.cpp b9600, ROCm, FA on, kv f16, 0.5k ctx, no speculative decoding. n=1.
Measured by
third_party
notes
Adds a mid-size (14B) data point to the existing 7B Q4_0 scoreboard rows (tg128 ~121-127 t/s).
Notes
Adds a mid-size (14B) data point to the existing 7B Q4_0 scoreboard rows (tg128 ~121-127 t/s).