llama.cpp Vulkan on RTX 4090: Llama-2-7B Q4_0 FA-off tg128 decode tok/s (scoreboard #10879)
187.97 tok/s (tg128, per-request, batch 1)
Fields
id
benchmarks/llamacpp-vulkan-rtx-4090-llama27b-q40-nofa-tg128Metric
decode_tok_s
Value
187.97
Unit
tok/s (tg128, per-request, batch 1)
Model
llama-2-7b.Q4_0.gguf (TheBloke, 3.56 GiB, 6.74B params)
Methodology
COMMUNITY-SCOREBOARD (llama.cpp discussion #10879, maintainer-curated). llama-bench -m llama-2-7b.Q4_0.gguf -ngl 99 -fa 0. RTX 4090. Build commit 4ae88d0. llama-bench default repeats, mean. Single device. Per-request batch-1, NOT aggregate.
Measured by
third_party
Reproducible
yes
notes
Maintainer-curated scoreboard row from ggml-org/llama.cpp discussion #10879, non-FA table (llama-bench -fa 0). Build commit 4ae88d0. Scoreboard comment column says coopmat2 (Vulkan cooperative-matrix-2 path).
Notes
Maintainer-curated scoreboard row from ggml-org/llama.cpp discussion #10879, non-FA table (llama-bench -fa 0). Build commit 4ae88d0. Scoreboard comment column says coopmat2 (Vulkan cooperative-matrix-2 path).