ASRock AMD Radeon RX 7900 XTX Creator 24GB

python koboldcpp.py models/mistral-small.q6_K_L.gguf \
--threads 12 \
--flashattention \
--benchmark

[code]
Processing Prompt [BLAS] (3996 / 3996 tokens)
Generating (100 / 100 tokens)
[11:09:55] CtxLimit:4096/4096, Amt:100/100, Init:0.08s, Process:10.52s (379.92T/s), Generate:5.31s (18.82T/s), Total:15.83s
Benchmark Completed - v1.95.1.yr0-ROCm Results:
======
Flags: NoAVX2=False Threads=12 HighPriority=False Cublas_Args=['normal', 'mmq'] Tensor_Split=None BlasThreads=12 BlasBatchSize=512 FlashAttention=True KvCache=0
Timestamp: 2025-07-03 09:09:55.763880+00:00
Backend: koboldcpp_hipblas.so
Layers: 59
Model: mistral-small.q6_K_L
MaxCtx: 4096
GenAmount: 100
-----
ProcessingTime: 10.518s
ProcessingSpeed: 379.92T/s
GenerationTime: 5.313s
GenerationSpeed: 18.82T/s
TotalTime: 15.831s
Output: 1 1 1 1
[/code]

python koboldcpp.py ./models/gemma-3-27b-it-Q5_K_M.gguf \
--threads 12 \
--flashattention \
--benchmark

[code]
Processing Prompt [BLAS] (3996 / 3996 tokens)
Generating (100 / 100 tokens)
[15:18:01] CtxLimit:4096/4096, Amt:100/100, Init:0.16s, Process:8.49s (470.45T/s), Generate:5.65s (17.70T/s), Total:14.14s
Benchmark Completed - v1.95.1.yr0-ROCm Results:
======
Flags: NoAVX2=False Threads=12 HighPriority=False Cublas_Args=['normal', 'mmq'] Tensor_Split=None BlasThreads=12 BlasBatchSize=512 FlashAttention=True KvCache=0
Timestamp: 2025-07-03 13:18:01.862027+00:00
Backend: koboldcpp_hipblas.so
Layers: 65
Model: gemma-3-27b-it-Q5_K_M
MaxCtx: 4096
GenAmount: 100
-----
ProcessingTime: 8.494s
ProcessingSpeed: 470.45T/s
GenerationTime: 5.651s
GenerationSpeed: 17.70T/s
TotalTime: 14.145s
Output: 1 1 1 1
[/code]