Add llama.cpp ROCm benchmark results

README.md

@@ -0,0 +1,74 @@
+---
+license: mit
+tags:
+- llama.cpp
+- gguf
+- rocm
+- amd
+- rx-7800-xt
+- benchmark
+- local-llm
+pretty_name: llama.cpp ROCm Benchmarks on RX 7800 XT
+---
+
+# llama.cpp ROCm Benchmarks on AMD Radeon RX 7800 XT
+
+This dataset contains local `llama.cpp` benchmark outputs collected on an AMD Radeon RX 7800 XT Linux system.
+
+## Contents
+
+- `llama_bench_results.csv` - parsed `llama-bench` rows.
+- `llama_bench_results.jsonl` - same parsed results as JSONL.
+- `metadata.json` - host/build metadata and notes.
+- `raw/` - original benchmark text, CSV, and log artifacts.
+
+## Important note about Qwen3.6 MTP speculative run
+
+`raw/qwen36-35b-a3b-q3km-mtp-speculative-20260529-103518.txt` is included for reproducibility, but it is not a valid throughput benchmark. All attempted `llama-speculative` runs failed immediately with:
+
+```text
+--model-draft is required
+```
+
+Do not compare that file against normal `llama-bench` results.
+
+## Parsed results
+
+| Model | Test | tok/s | Source |
+|---|---:|---:|---|
+| lfm2moe 8B.A1B Q4_K - Medium | pp512 | 4844.05 ± 20.61 | llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt |
+| lfm2moe 8B.A1B Q4_K - Medium | tg128 | 167.88 ± 1.81 | llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt |
+| lfm2moe 8B.A1B Q5_K - Medium | pp512 | 4625.49 ± 33.50 | llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt |
+| lfm2moe 8B.A1B Q5_K - Medium | tg128 | 199.70 ± 1.92 | llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt |
+| lfm2moe 8B.A1B Q6_K | pp512 | 3336.65 ± 23.56 | llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt |
+| lfm2moe 8B.A1B Q6_K | tg128 | 195.71 ± 1.27 | llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt |
+| lfm2moe 8B.A1B Q8_0 | pp512 | 6002.21 ± 85.75 | llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt |
+| lfm2moe 8B.A1B Q8_0 | tg128 | 171.18 ± 1.99 | llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt |
+| qwen3 14B Q3_K - Medium | pp512 | 1004.75 ± 31.82 | llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt |
+| qwen3 14B Q3_K - Medium | tg128 | 42.93 ± 0.05 | llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt |
+| qwen3 14B Q4_K - Medium | pp512 | 1091.35 ± 23.42 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt |
+| qwen3 14B Q4_K - Medium | tg128 | 36.75 ± 0.19 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt |
+| qwen3 14B Q4_K - Medium | pp512 | 1095.16 ± 17.51 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt |
+| qwen3 14B Q4_K - Medium | tg128 | 36.83 ± 0.02 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt |
+| qwen3 14B Q4_K - Medium | pp512 | 1091.85 ± 17.52 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt |
+| qwen3 14B Q4_K - Medium | tg128 | 36.76 ± 0.02 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt |
+| qwen3 14B Q4_K - Medium | pp512 | 1086.99 ± 21.08 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt |
+| qwen3 14B Q4_K - Medium | tg128 | 36.53 ± 0.07 | llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt |
+| qwen3 14B Q5_K - Medium | pp512 | 1036.95 ± 17.77 | llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt |
+| qwen3 14B Q5_K - Medium | tg128 | 41.57 ± 0.01 | llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt |
+| qwen3 14B Q2_K - Medium | pp512 | 1064.51 ± 16.34 | llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt |
+| qwen3 14B Q2_K - Medium | tg128 | 49.66 ± 0.08 | llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt |
+| qwen3 14B Q3_K - Medium | pp512 | 1038.61 ± 18.82 | llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt |
+| qwen3 14B Q3_K - Medium | tg128 | 43.11 ± 0.04 | llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt |
+| qwen3 14B Q4_K - Medium | pp512 | 1087.87 ± 31.11 | llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt |
+| qwen3 14B Q4_K - Medium | tg128 | 38.14 ± 0.02 | llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt |
+| qwen3 14B Q5_K - Medium | pp512 | 1047.31 ± 15.11 | llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt |
+| qwen3 14B Q5_K - Medium | tg128 | 40.31 ± 0.05 | llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt |
+| qwen3 14B Q6_K | pp512 | 1101.35 ± 30.22 | llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt |
+| qwen3 14B Q6_K | tg128 | 36.60 ± 0.05 | llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt |
+| qwen35 27B Q3_K - Medium | pp512 | 549.09 ± 5.50 | llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt |
+| qwen35 27B Q3_K - Medium | tg128 | 23.39 ± 0.03 | llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt |
+
+## Reproduction context
+
+See `metadata.json` and the raw files for exact commands/output where available.

llama_bench_results.csv

@@ -0,0 +1,33 @@
+source_file,model,size,params,backend,ngl,n_batch,test,t_per_s,t_per_s_std,raw_t_per_s
+llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt,lfm2moe 8B.A1B Q4_K - Medium,4.79 GiB,8.47 B,ROCm,999,512,pp512,4844.05,20.61,4844.05 ± 20.61
+llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt,lfm2moe 8B.A1B Q4_K - Medium,4.79 GiB,8.47 B,ROCm,999,512,tg128,167.88,1.81,167.88 ± 1.81
+llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt,lfm2moe 8B.A1B Q5_K - Medium,5.61 GiB,8.47 B,ROCm,999,512,pp512,4625.49,33.50,4625.49 ± 33.50
+llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt,lfm2moe 8B.A1B Q5_K - Medium,5.61 GiB,8.47 B,ROCm,999,512,tg128,199.70,1.92,199.70 ± 1.92
+llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt,lfm2moe 8B.A1B Q6_K,6.47 GiB,8.47 B,ROCm,999,512,pp512,3336.65,23.56,3336.65 ± 23.56
+llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt,lfm2moe 8B.A1B Q6_K,6.47 GiB,8.47 B,ROCm,999,512,tg128,195.71,1.27,195.71 ± 1.27
+llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt,lfm2moe 8B.A1B Q8_0,8.38 GiB,8.47 B,ROCm,999,512,pp512,6002.21,85.75,6002.21 ± 85.75
+llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt,lfm2moe 8B.A1B Q8_0,8.38 GiB,8.47 B,ROCm,999,512,tg128,171.18,1.99,171.18 ± 1.99
+llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt,qwen3 14B Q3_K - Medium,6.81 GiB,14.77 B,ROCm,999,512,pp512,1004.75,31.82,1004.75 ± 31.82
+llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt,qwen3 14B Q3_K - Medium,6.81 GiB,14.77 B,ROCm,999,512,tg128,42.93,0.05,42.93 ± 0.05
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,pp512,1091.35,23.42,1091.35 ± 23.42
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,tg128,36.75,0.19,36.75 ± 0.19
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,pp512,1095.16,17.51,1095.16 ± 17.51
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,tg128,36.83,0.02,36.83 ± 0.02
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,pp512,1091.85,17.52,1091.85 ± 17.52
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,tg128,36.76,0.02,36.76 ± 0.02
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,pp512,1086.99,21.08,1086.99 ± 21.08
+llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt,qwen3 14B Q4_K - Medium,8.38 GiB,14.77 B,ROCm,999,512,tg128,36.53,0.07,36.53 ± 0.07
+llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt,qwen3 14B Q5_K - Medium,9.79 GiB,14.77 B,ROCm,999,512,pp512,1036.95,17.77,1036.95 ± 17.77
+llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt,qwen3 14B Q5_K - Medium,9.79 GiB,14.77 B,ROCm,999,512,tg128,41.57,0.01,41.57 ± 0.01
+llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt,qwen3 14B Q2_K - Medium,5.65 GiB,14.77 B,ROCm,999,512,pp512,1064.51,16.34,1064.51 ± 16.34
+llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt,qwen3 14B Q2_K - Medium,5.65 GiB,14.77 B,ROCm,999,512,tg128,49.66,0.08,49.66 ± 0.08
+llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt,qwen3 14B Q3_K - Medium,7.07 GiB,14.77 B,ROCm,999,512,pp512,1038.61,18.82,1038.61 ± 18.82
+llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt,qwen3 14B Q3_K - Medium,7.07 GiB,14.77 B,ROCm,999,512,tg128,43.11,0.04,43.11 ± 0.04
+llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt,qwen3 14B Q4_K - Medium,8.53 GiB,14.77 B,ROCm,999,512,pp512,1087.87,31.11,1087.87 ± 31.11
+llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt,qwen3 14B Q4_K - Medium,8.53 GiB,14.77 B,ROCm,999,512,tg128,38.14,0.02,38.14 ± 0.02
+llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt,qwen3 14B Q5_K - Medium,9.82 GiB,14.77 B,ROCm,999,512,pp512,1047.31,15.11,1047.31 ± 15.11
+llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt,qwen3 14B Q5_K - Medium,9.82 GiB,14.77 B,ROCm,999,512,tg128,40.31,0.05,40.31 ± 0.05
+llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt,qwen3 14B Q6_K,12.37 GiB,14.77 B,ROCm,999,512,pp512,1101.35,30.22,1101.35 ± 30.22
+llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt,qwen3 14B Q6_K,12.37 GiB,14.77 B,ROCm,999,512,tg128,36.60,0.05,36.60 ± 0.05
+llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt,qwen35 27B Q3_K - Medium,12.86 GiB,27.32 B,ROCm,999,512,pp512,549.09,5.50,549.09 ± 5.50
+llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt,qwen35 27B Q3_K - Medium,12.86 GiB,27.32 B,ROCm,999,512,tg128,23.39,0.03,23.39 ± 0.03

llama_bench_results.jsonl

@@ -0,0 +1,32 @@
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt", "model": "lfm2moe 8B.A1B Q4_K - Medium", "size": "4.79 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "4844.05", "t_per_s_std": "20.61", "raw_t_per_s": "4844.05 \u00b1 20.61"}
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt", "model": "lfm2moe 8B.A1B Q4_K - Medium", "size": "4.79 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "167.88", "t_per_s_std": "1.81", "raw_t_per_s": "167.88 \u00b1 1.81"}
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt", "model": "lfm2moe 8B.A1B Q5_K - Medium", "size": "5.61 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "4625.49", "t_per_s_std": "33.50", "raw_t_per_s": "4625.49 \u00b1 33.50"}
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt", "model": "lfm2moe 8B.A1B Q5_K - Medium", "size": "5.61 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "199.70", "t_per_s_std": "1.92", "raw_t_per_s": "199.70 \u00b1 1.92"}
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt", "model": "lfm2moe 8B.A1B Q6_K", "size": "6.47 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "3336.65", "t_per_s_std": "23.56", "raw_t_per_s": "3336.65 \u00b1 23.56"}
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt", "model": "lfm2moe 8B.A1B Q6_K", "size": "6.47 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "195.71", "t_per_s_std": "1.27", "raw_t_per_s": "195.71 \u00b1 1.27"}
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt", "model": "lfm2moe 8B.A1B Q8_0", "size": "8.38 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "6002.21", "t_per_s_std": "85.75", "raw_t_per_s": "6002.21 \u00b1 85.75"}
+{"source_file": "llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt", "model": "lfm2moe 8B.A1B Q8_0", "size": "8.38 GiB", "params": "8.47 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "171.18", "t_per_s_std": "1.99", "raw_t_per_s": "171.18 \u00b1 1.99"}
+{"source_file": "llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt", "model": "qwen3 14B Q3_K - Medium", "size": "6.81 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1004.75", "t_per_s_std": "31.82", "raw_t_per_s": "1004.75 \u00b1 31.82"}
+{"source_file": "llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt", "model": "qwen3 14B Q3_K - Medium", "size": "6.81 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "42.93", "t_per_s_std": "0.05", "raw_t_per_s": "42.93 \u00b1 0.05"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1091.35", "t_per_s_std": "23.42", "raw_t_per_s": "1091.35 \u00b1 23.42"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "36.75", "t_per_s_std": "0.19", "raw_t_per_s": "36.75 \u00b1 0.19"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1095.16", "t_per_s_std": "17.51", "raw_t_per_s": "1095.16 \u00b1 17.51"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "36.83", "t_per_s_std": "0.02", "raw_t_per_s": "36.83 \u00b1 0.02"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1091.85", "t_per_s_std": "17.52", "raw_t_per_s": "1091.85 \u00b1 17.52"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "36.76", "t_per_s_std": "0.02", "raw_t_per_s": "36.76 \u00b1 0.02"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1086.99", "t_per_s_std": "21.08", "raw_t_per_s": "1086.99 \u00b1 21.08"}
+{"source_file": "llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.38 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "36.53", "t_per_s_std": "0.07", "raw_t_per_s": "36.53 \u00b1 0.07"}
+{"source_file": "llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt", "model": "qwen3 14B Q5_K - Medium", "size": "9.79 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1036.95", "t_per_s_std": "17.77", "raw_t_per_s": "1036.95 \u00b1 17.77"}
+{"source_file": "llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt", "model": "qwen3 14B Q5_K - Medium", "size": "9.79 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "41.57", "t_per_s_std": "0.01", "raw_t_per_s": "41.57 \u00b1 0.01"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt", "model": "qwen3 14B Q2_K - Medium", "size": "5.65 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1064.51", "t_per_s_std": "16.34", "raw_t_per_s": "1064.51 \u00b1 16.34"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt", "model": "qwen3 14B Q2_K - Medium", "size": "5.65 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "49.66", "t_per_s_std": "0.08", "raw_t_per_s": "49.66 \u00b1 0.08"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt", "model": "qwen3 14B Q3_K - Medium", "size": "7.07 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1038.61", "t_per_s_std": "18.82", "raw_t_per_s": "1038.61 \u00b1 18.82"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt", "model": "qwen3 14B Q3_K - Medium", "size": "7.07 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "43.11", "t_per_s_std": "0.04", "raw_t_per_s": "43.11 \u00b1 0.04"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.53 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1087.87", "t_per_s_std": "31.11", "raw_t_per_s": "1087.87 \u00b1 31.11"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt", "model": "qwen3 14B Q4_K - Medium", "size": "8.53 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "38.14", "t_per_s_std": "0.02", "raw_t_per_s": "38.14 \u00b1 0.02"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt", "model": "qwen3 14B Q5_K - Medium", "size": "9.82 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1047.31", "t_per_s_std": "15.11", "raw_t_per_s": "1047.31 \u00b1 15.11"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt", "model": "qwen3 14B Q5_K - Medium", "size": "9.82 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "40.31", "t_per_s_std": "0.05", "raw_t_per_s": "40.31 \u00b1 0.05"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt", "model": "qwen3 14B Q6_K", "size": "12.37 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "1101.35", "t_per_s_std": "30.22", "raw_t_per_s": "1101.35 \u00b1 30.22"}
+{"source_file": "llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt", "model": "qwen3 14B Q6_K", "size": "12.37 GiB", "params": "14.77 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "36.60", "t_per_s_std": "0.05", "raw_t_per_s": "36.60 \u00b1 0.05"}
+{"source_file": "llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt", "model": "qwen35 27B Q3_K - Medium", "size": "12.86 GiB", "params": "27.32 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "pp512", "t_per_s": "549.09", "t_per_s_std": "5.50", "raw_t_per_s": "549.09 \u00b1 5.50"}
+{"source_file": "llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt", "model": "qwen35 27B Q3_K - Medium", "size": "12.86 GiB", "params": "27.32 B", "backend": "ROCm", "ngl": "999", "n_batch": "512", "test": "tg128", "t_per_s": "23.39", "t_per_s_std": "0.03", "raw_t_per_s": "23.39 \u00b1 0.03"}

metadata.json

@@ -0,0 +1,11 @@
+{
+  "created_at": "2026-05-29T03:32:54.044372+00:00",
+  "host_kernel": "Linux AMDxt7800 6.17.0-29-generic #29~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Mon May 11 10:30:58 UTC 2 x86_64 x86_64 x86_64 GNU/Linux",
+  "gpu": "GPU: 0\n    BDF: 0000:07:00.0\n    UUID: 0cff747e-0000-1000-80a7-3042394d6f22\n    KFD_ID: 34454\n    NODE_ID: 1\n    PARTITION_ID: 0",
+  "llama_cpp_commit": "f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe",
+  "notes": [
+    "Benchmarks run locally with llama.cpp HIP/ROCm build on AMD Radeon RX 7800 XT system.",
+    "qwen36-35b-a3b-q3km-mtp-speculative-20260529-103518.txt is a failed/invalid MTP speculative attempt: llama-speculative required --model-draft; no throughput numbers.",
+    "Raw result files are included under raw/."
+  ]
+}

raw/lfm25-8b-a1b-bench-batch-20260529-100753.log

@@ -0,0 +1,117 @@
+=== LFM2.5-8B-A1B benchmark batch ===
+2026-05-29T10:07:53+10:00
+batch_log=/home/neo/bench/results/llama/lfm25-8b-a1b-bench-batch-20260529-100753.log
+runner=/home/neo/bench/scripts/llama-bench-runner.sh
+standard_flags=-ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5
+
+=== GPU ===
+GPU: 0
+    BDF: 0000:07:00.0
+    UUID: 0cff747e-0000-1000-80a7-3042394d6f22
+    KFD_ID: 34454
+    NODE_ID: 1
+    PARTITION_ID: 0
+
+
+=== START LFM2.5-8B-A1B-Q4_K_M.gguf ===
+2026-05-29T10:07:53+10:00
+=== RUN METADATA ===
+2026-05-29T10:07:53+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q4_K_M.gguf
+sha256=3ef9879143342a46cdd00b5f6f3919b77ee693945f9d73eea99344eadd6950b6
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q4_K - Medium   |   4.79 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      4844.05 ± 20.61 |
+| lfm2moe 8B.A1B Q4_K - Medium   |   4.79 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        167.88 ± 1.81 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt
+=== OK LFM2.5-8B-A1B-Q4_K_M.gguf ===
+2026-05-29T10:08:02+10:00
+
+=== START LFM2.5-8B-A1B-Q5_K_M.gguf ===
+2026-05-29T10:08:02+10:00
+=== RUN METADATA ===
+2026-05-29T10:08:02+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q5_K_M.gguf
+sha256=925a297ddd1c00d433db11e2cc1f62b1ed4f6624240a00691a581444e98a788e
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q5_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q5_K - Medium   |   5.61 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      4625.49 ± 33.50 |
+| lfm2moe 8B.A1B Q5_K - Medium   |   5.61 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        199.70 ± 1.92 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt
+=== OK LFM2.5-8B-A1B-Q5_K_M.gguf ===
+2026-05-29T10:08:12+10:00
+
+=== START LFM2.5-8B-A1B-Q6_K.gguf ===
+2026-05-29T10:08:12+10:00
+=== RUN METADATA ===
+2026-05-29T10:08:12+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q6_K.gguf
+sha256=ec239b926ba38d421dc2623b2568b9057105a27a988a2db23a04372270d03f91
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q6_K.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q6_K            |   6.47 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      3336.65 ± 23.56 |
+| lfm2moe 8B.A1B Q6_K            |   6.47 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        195.71 ± 1.27 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt
+=== OK LFM2.5-8B-A1B-Q6_K.gguf ===
+2026-05-29T10:08:22+10:00
+
+=== START LFM2.5-8B-A1B-Q8_0.gguf ===
+2026-05-29T10:08:22+10:00
+=== RUN METADATA ===
+2026-05-29T10:08:22+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q8_0.gguf
+sha256=c183230108711aabe8b304b7b5b5ea0ac439e62392a470d92f0aff016b24d2a1
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q8_0.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q8_0            |   8.38 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      6002.21 ± 85.75 |
+| lfm2moe 8B.A1B Q8_0            |   8.38 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        171.18 ± 1.99 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt
+=== OK LFM2.5-8B-A1B-Q8_0.gguf ===
+2026-05-29T10:08:38+10:00
+
+=== BATCH DONE ===
+2026-05-29T10:08:38+10:00

raw/lfm25-8b-a1b-summary-20260529-100753.csv

@@ -0,0 +1,5 @@
+timestamp,model_file,model_display,sha256,size,params,backend,ngl,n_batch,prompt_tokens,gen_tokens,batch,ubatch,repeats,pp512_tps,tg128_tps,llama_cpp_commit,kernel,result_file
+2026-05-29T10:07-10:08+10:00,LFM2.5-8B-A1B-Q4_K_M.gguf,lfm2moe 8B.A1B Q4_K - Medium,3ef9879143342a46cdd00b5f6f3919b77ee693945f9d73eea99344eadd6950b6,4.79 GiB,8.47 B,ROCm,999,512,512,128,512,512,5,4844.05 ± 20.61,167.88 ± 1.81,f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe,6.17.0-29-generic,/home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt
+2026-05-29T10:07-10:08+10:00,LFM2.5-8B-A1B-Q5_K_M.gguf,lfm2moe 8B.A1B Q5_K - Medium,925a297ddd1c00d433db11e2cc1f62b1ed4f6624240a00691a581444e98a788e,5.61 GiB,8.47 B,ROCm,999,512,512,128,512,512,5,4625.49 ± 33.50,199.70 ± 1.92,f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe,6.17.0-29-generic,/home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt
+2026-05-29T10:07-10:08+10:00,LFM2.5-8B-A1B-Q6_K.gguf,lfm2moe 8B.A1B Q6_K,ec239b926ba38d421dc2623b2568b9057105a27a988a2db23a04372270d03f91,6.47 GiB,8.47 B,ROCm,999,512,512,128,512,512,5,3336.65 ± 23.56,195.71 ± 1.27,f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe,6.17.0-29-generic,/home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt
+2026-05-29T10:07-10:08+10:00,LFM2.5-8B-A1B-Q8_0.gguf,lfm2moe 8B.A1B Q8_0,c183230108711aabe8b304b7b5b5ea0ac439e62392a470d92f0aff016b24d2a1,8.38 GiB,8.47 B,ROCm,999,512,512,128,512,512,5,6002.21 ± 85.75,171.18 ± 1.99,f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe,6.17.0-29-generic,/home/neo/bench/results/llama/llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt

raw/llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-29T10:07:53+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q4_K_M.gguf
+sha256=3ef9879143342a46cdd00b5f6f3919b77ee693945f9d73eea99344eadd6950b6
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q4_K - Medium   |   4.79 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      4844.05 ± 20.61 |
+| lfm2moe 8B.A1B Q4_K - Medium   |   4.79 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        167.88 ± 1.81 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-29T10:08:02+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q5_K_M.gguf
+sha256=925a297ddd1c00d433db11e2cc1f62b1ed4f6624240a00691a581444e98a788e
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q5_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q5_K - Medium   |   5.61 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      4625.49 ± 33.50 |
+| lfm2moe 8B.A1B Q5_K - Medium   |   5.61 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        199.70 ± 1.92 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-29T10:08:12+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q6_K.gguf
+sha256=ec239b926ba38d421dc2623b2568b9057105a27a988a2db23a04372270d03f91
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q6_K.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q6_K            |   6.47 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      3336.65 ± 23.56 |
+| lfm2moe 8B.A1B Q6_K            |   6.47 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        195.71 ± 1.27 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-29T10:08:22+10:00
+model=/home/neo/bench/models/LFM2.5-8B-A1B-Q8_0.gguf
+sha256=c183230108711aabe8b304b7b5b5ea0ac439e62392a470d92f0aff016b24d2a1
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/LFM2.5-8B-A1B-Q8_0.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| lfm2moe 8B.A1B Q8_0            |   8.38 GiB |     8.47 B | ROCm       | 999 |     512 |           pp512 |      6002.21 ± 85.75 |
+| lfm2moe 8B.A1B Q8_0            |   8.38 GiB |     8.47 B | ROCm       | 999 |     512 |           tg128 |        171.18 ± 1.99 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:40:19+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q3_K_M.gguf
+sha256=9694f1db203d5a18bea6f19a36be01b2d2c79bf9b7f8df553a4cc5cff870b47e
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q3_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q3_K - Medium        |   6.81 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1004.75 ± 31.82 |
+| qwen3 14B Q3_K - Medium        |   6.81 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         42.93 ± 0.05 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T11:21:04+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf
+sha256=500a8806e85ee9c83f3ae08420295592451379b4f8cf2d0f41c15dffeb6b81f0
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1091.35 ± 23.42 |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         36.75 ± 0.19 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T11:24:15+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf
+sha256=500a8806e85ee9c83f3ae08420295592451379b4f8cf2d0f41c15dffeb6b81f0
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1095.16 ± 17.51 |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         36.83 ± 0.02 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T11:46:37+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf
+sha256=500a8806e85ee9c83f3ae08420295592451379b4f8cf2d0f41c15dffeb6b81f0
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1091.85 ± 17.52 |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         36.76 ± 0.02 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:41:28+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf
+sha256=5eaa0870bd81ed3b58a630a271234cfa604e43ffb3a19cd68e54a80dd9d52a66
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1086.99 ± 21.08 |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         36.53 ± 0.07 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:42:47+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q5_K_M.gguf
+sha256=305309039ea44df963c01f66d65e949f8eebca238554916df437f02295a3a413
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q5_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q5_K - Medium        |   9.79 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1036.95 ± 17.77 |
+| qwen3 14B Q5_K - Medium        |   9.79 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         41.57 ± 0.01 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:39:47+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q2_K_XL.gguf
+sha256=6abefcfb88de4157acc6109a711af65fdb5a6da2a77a1ac91bdada2ded22fd49
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q2_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q2_K - Medium        |   5.65 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1064.51 ± 16.34 |
+| qwen3 14B Q2_K - Medium        |   5.65 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         49.66 ± 0.08 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:40:54+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q3_K_XL.gguf
+sha256=46b8c8a3fda3de374f465be47176c956e44c7ef56ceddf6aa8bafddb1c73d5b6
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q3_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q3_K - Medium        |   7.07 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1038.61 ± 18.82 |
+| qwen3 14B Q3_K - Medium        |   7.07 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         43.11 ± 0.04 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:42:10+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q4_K_XL.gguf
+sha256=7f371adeaaf51dc97da927c61e6573e75e3ea33d76ec09dddb8cad00d0d5b6cb
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q4_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q4_K - Medium        |   8.53 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1087.87 ± 31.11 |
+| qwen3 14B Q4_K - Medium        |   8.53 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         38.14 ± 0.02 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:43:27+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q5_K_XL.gguf
+sha256=f62da26805a70d9963400ca6160843c82e1aaadc714da8b299d0a676676741b8
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q5_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q5_K - Medium        |   9.82 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1047.31 ± 15.11 |
+| qwen3 14B Q5_K - Medium        |   9.82 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         40.31 ± 0.05 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T19:44:09+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q6_K_XL.gguf
+sha256=83e7dcec42952a4ed6e092977c587a9ab8d528d52d0db481df6d694d5181811b
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q6_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q6_K                 |  12.37 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1101.35 ± 30.22 |
+| qwen3 14B Q6_K                 |  12.37 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         36.60 ± 0.05 |
+
+build: f12cc6d0f (9371)

raw/llama-bench-Qwen3-14B-UD-Q8_K_XL.gguf-20260528-194455.txt

@@ -0,0 +1,13 @@
+=== RUN METADATA ===
+2026-05-28T19:44:55+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q8_K_XL.gguf
+sha256=2da4750f82103624cbc6f831a5907628670c9b2ba59c4d57e59186c19860f4b3
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q8_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |

raw/llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt

@@ -0,0 +1,17 @@
+=== RUN METADATA ===
+2026-05-28T12:42:39+10:00
+model=/home/neo/bench/models/Qwen3.6-27B-Q3_K_M.gguf
+sha256=f8a9329b78446254200d5cd387c3d8c8f374a2b824696d4b550f42461b970310
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3.6-27B-Q3_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen35 27B Q3_K - Medium       |  12.86 GiB |    27.32 B | ROCm       | 999 |     512 |           pp512 |        549.09 ± 5.50 |
+| qwen35 27B Q3_K - Medium       |  12.86 GiB |    27.32 B | ROCm       | 999 |     512 |           tg128 |         23.39 ± 0.03 |
+
+build: f12cc6d0f (9371)

raw/llama-cpp-commit.txt

@@ -0,0 +1 @@
+f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe

raw/model-sha256.txt

@@ -0,0 +1,2 @@
+500a8806e85ee9c83f3ae08420295592451379b4f8cf2d0f41c15dffeb6b81f0  /home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf
+f8a9329b78446254200d5cd387c3d8c8f374a2b824696d4b550f42461b970310  /home/neo/bench/models/Qwen3.6-27B-Q3_K_M.gguf

raw/qwen3-14b-all-quants-batch-20260528-193947.log

@@ -0,0 +1,945 @@
+batch_start=2026-05-28T19:39:47+10:00
+batch_log=/home/neo/bench/results/llama/qwen3-14b-all-quants-batch-20260528-193947.log
+models=9
+Linux AMDxt7800 6.17.0-29-generic #29~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC Mon May 11 10:30:58 UTC 2 x86_64 x86_64 x86_64 GNU/Linux
+GPU: 0
+    BDF: 0000:07:00.0
+    UUID: 0cff747e-0000-1000-80a7-3042394d6f22
+    KFD_ID: 34454
+    NODE_ID: 1
+    PARTITION_ID: 0
+
+
+===== BENCH START 2026-05-28T19:39:47+10:00 models/Qwen3-14B-UD-Q2_K_XL.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:39:47+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q2_K_XL.gguf
+sha256=6abefcfb88de4157acc6109a711af65fdb5a6da2a77a1ac91bdada2ded22fd49
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q2_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q2_K - Medium        |   5.65 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1064.51 ± 16.34 |
+| qwen3 14B Q2_K - Medium        |   5.65 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         49.66 ± 0.08 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt
+===== BENCH OK 2026-05-28T19:40:08+10:00 models/Qwen3-14B-UD-Q2_K_XL.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 62 %
+        UMC_ACTIVITY: 24 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 143 W
+        GFX_VOLTAGE: 736 mV
+        SOC_VOLTAGE: 865 mV
+        MEM_VOLTAGE: 768 mV
+        THROTTLE_STATUS: THROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 1777 MHz
+            MIN_CLK: 42 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 1218 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1950 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 1285 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 39 °C
+        HOTSPOT: 49 °C
+        MEM: 48 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 94
+        MAX: 255
+        RPM: 977
+        USAGE: 36.86 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:40:19+10:00 models/Qwen3-14B-Q3_K_M.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:40:19+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q3_K_M.gguf
+sha256=9694f1db203d5a18bea6f19a36be01b2d2c79bf9b7f8df553a4cc5cff870b47e
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q3_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q3_K - Medium        |   6.81 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1004.75 ± 31.82 |
+| qwen3 14B Q3_K - Medium        |   6.81 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         42.93 ± 0.05 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt
+===== BENCH OK 2026-05-28T19:40:43+10:00 models/Qwen3-14B-Q3_K_M.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 98 %
+        UMC_ACTIVITY: 48 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 223 W
+        GFX_VOLTAGE: 943 mV
+        SOC_VOLTAGE: 929 mV
+        MEM_VOLTAGE: 804 mV
+        THROTTLE_STATUS: THROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 2421 MHz
+            MIN_CLK: 27 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 1218 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1950 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 1285 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 43 °C
+        HOTSPOT: 54 °C
+        MEM: 54 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 99
+        MAX: 255
+        RPM: 1102
+        USAGE: 38.82 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:40:54+10:00 models/Qwen3-14B-UD-Q3_K_XL.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:40:54+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q3_K_XL.gguf
+sha256=46b8c8a3fda3de374f465be47176c956e44c7ef56ceddf6aa8bafddb1c73d5b6
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q3_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q3_K - Medium        |   7.07 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1038.61 ± 18.82 |
+| qwen3 14B Q3_K - Medium        |   7.07 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         43.11 ± 0.04 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt
+===== BENCH OK 2026-05-28T19:41:18+10:00 models/Qwen3-14B-UD-Q3_K_XL.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 55 %
+        UMC_ACTIVITY: 22 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 131 W
+        GFX_VOLTAGE: 835 mV
+        SOC_VOLTAGE: 863 mV
+        MEM_VOLTAGE: 779 mV
+        THROTTLE_STATUS: THROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 1952 MHz
+            MIN_CLK: 43 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 1218 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 24 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 24 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1950 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 1285 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 46 °C
+        HOTSPOT: 56 °C
+        MEM: 58 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 99
+        MAX: 255
+        RPM: 1092
+        USAGE: 38.82 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:41:28+10:00 models/Qwen3-14B-Q4_K_M.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:41:28+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf
+sha256=5eaa0870bd81ed3b58a630a271234cfa604e43ffb3a19cd68e54a80dd9d52a66
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q4_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1086.99 ± 21.08 |
+| qwen3 14B Q4_K - Medium        |   8.38 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         36.53 ± 0.07 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt
+===== BENCH OK 2026-05-28T19:41:59+10:00 models/Qwen3-14B-Q4_K_M.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 1 %
+        UMC_ACTIVITY: 0 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 43 W
+        GFX_VOLTAGE: 695 mV
+        SOC_VOLTAGE: 949 mV
+        MEM_VOLTAGE: 799 mV
+        THROTTLE_STATUS: UNTHROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 43 MHz
+            MIN_CLK: 43 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 1218 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1950 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 1285 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 46 °C
+        HOTSPOT: 57 °C
+        MEM: 58 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 99
+        MAX: 255
+        RPM: 1103
+        USAGE: 38.82 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:42:10+10:00 models/Qwen3-14B-UD-Q4_K_XL.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:42:10+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q4_K_XL.gguf
+sha256=7f371adeaaf51dc97da927c61e6573e75e3ea33d76ec09dddb8cad00d0d5b6cb
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q4_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q4_K - Medium        |   8.53 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1087.87 ± 31.11 |
+| qwen3 14B Q4_K - Medium        |   8.53 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         38.14 ± 0.02 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt
+===== BENCH OK 2026-05-28T19:42:37+10:00 models/Qwen3-14B-UD-Q4_K_XL.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 97 %
+        UMC_ACTIVITY: 51 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 218 W
+        GFX_VOLTAGE: 985 mV
+        SOC_VOLTAGE: 927 mV
+        MEM_VOLTAGE: 804 mV
+        THROTTLE_STATUS: THROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 2514 MHz
+            MIN_CLK: 500 MHz
+            MAX_CLK: 2514 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 1218 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1950 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 1285 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 48 °C
+        HOTSPOT: 59 °C
+        MEM: 62 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 99
+        MAX: 255
+        RPM: 1101
+        USAGE: 38.82 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:42:47+10:00 models/Qwen3-14B-Q5_K_M.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:42:47+10:00
+model=/home/neo/bench/models/Qwen3-14B-Q5_K_M.gguf
+sha256=305309039ea44df963c01f66d65e949f8eebca238554916df437f02295a3a413
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-Q5_K_M.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q5_K - Medium        |   9.79 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1036.95 ± 17.77 |
+| qwen3 14B Q5_K - Medium        |   9.79 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         41.57 ± 0.01 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt
+===== BENCH OK 2026-05-28T19:43:17+10:00 models/Qwen3-14B-Q5_K_M.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 49 %
+        UMC_ACTIVITY: 28 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 113 W
+        GFX_VOLTAGE: 537 mV
+        SOC_VOLTAGE: 842 mV
+        MEM_VOLTAGE: 751 mV
+        THROTTLE_STATUS: THROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 1286 MHz
+            MIN_CLK: 10 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 96 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1000 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 600 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 48 °C
+        HOTSPOT: 56 °C
+        MEM: 60 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 99
+        MAX: 255
+        RPM: 1097
+        USAGE: 38.82 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:43:27+10:00 models/Qwen3-14B-UD-Q5_K_XL.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:43:27+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q5_K_XL.gguf
+sha256=f62da26805a70d9963400ca6160843c82e1aaadc714da8b299d0a676676741b8
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q5_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q5_K - Medium        |   9.82 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1047.31 ± 15.11 |
+| qwen3 14B Q5_K - Medium        |   9.82 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         40.31 ± 0.05 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt
+===== BENCH OK 2026-05-28T19:43:59+10:00 models/Qwen3-14B-UD-Q5_K_XL.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 61 %
+        UMC_ACTIVITY: 34 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 138 W
+        GFX_VOLTAGE: 656 mV
+        SOC_VOLTAGE: 860 mV
+        MEM_VOLTAGE: 761 mV
+        THROTTLE_STATUS: THROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 1609 MHz
+            MIN_CLK: 500 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 1218 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1950 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 1285 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 48 °C
+        HOTSPOT: 60 °C
+        MEM: 62 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 99
+        MAX: 255
+        RPM: 1104
+        USAGE: 38.82 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:44:09+10:00 models/Qwen3-14B-UD-Q6_K_XL.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:44:09+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q6_K_XL.gguf
+sha256=83e7dcec42952a4ed6e092977c587a9ab8d528d52d0db481df6d694d5181811b
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q6_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+| qwen3 14B Q6_K                 |  12.37 GiB |    14.77 B | ROCm       | 999 |     512 |           pp512 |      1101.35 ± 30.22 |
+| qwen3 14B Q6_K                 |  12.37 GiB |    14.77 B | ROCm       | 999 |     512 |           tg128 |         36.60 ± 0.05 |
+
+build: f12cc6d0f (9371)
+llama-bench result: /home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt
+===== BENCH OK 2026-05-28T19:44:45+10:00 models/Qwen3-14B-UD-Q6_K_XL.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 99 %
+        UMC_ACTIVITY: 71 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 225 W
+        GFX_VOLTAGE: 943 mV
+        SOC_VOLTAGE: 918 mV
+        MEM_VOLTAGE: 804 mV
+        THROTTLE_STATUS: THROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 2429 MHz
+            MIN_CLK: 43 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 1218 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1950 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 1285 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 48 °C
+        HOTSPOT: 59 °C
+        MEM: 64 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 99
+        MAX: 255
+        RPM: 1104
+        USAGE: 38.82 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+===== BENCH START 2026-05-28T19:44:55+10:00 models/Qwen3-14B-UD-Q8_K_XL.gguf =====
+=== RUN METADATA ===
+2026-05-28T19:44:55+10:00
+model=/home/neo/bench/models/Qwen3-14B-UD-Q8_K_XL.gguf
+sha256=2da4750f82103624cbc6f831a5907628670c9b2ba59c4d57e59186c19860f4b3
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+
+=== COMMAND ===
+/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench -m /home/neo/bench/models/Qwen3-14B-UD-Q8_K_XL.gguf -ngl 999 -p 512 -n 128 -b 512 -ub 512 -r 5 
+
+=== OUTPUT ===
+ggml_cuda_init: found 1 ROCm devices (Total VRAM: 16368 MiB):
+  Device 0: AMD Radeon RX 7800 XT, gfx1101 (0x1101), VMM: no, Wave Size: 32, VRAM: 16368 MiB
+llama_bench: error: failed to load model '/home/neo/bench/models/Qwen3-14B-UD-Q8_K_XL.gguf'
+| model                          |       size |     params | backend    | ngl | n_batch |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | --------------: | -------------------: |
+===== BENCH FAILED rc=1 2026-05-28T19:45:12+10:00 models/Qwen3-14B-UD-Q8_K_XL.gguf =====
+
+GPU: 0
+    USAGE:
+        GFX_ACTIVITY: 1 %
+        UMC_ACTIVITY: 0 %
+        MM_ACTIVITY: 0 %
+        VCN_ACTIVITY: [0 %, N/A, N/A, N/A]
+        JPEG_ACTIVITY: [N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A, N/A]
+        GFX_BUSY_INST: N/A
+        JPEG_BUSY: N/A
+        VCN_BUSY: N/A
+    POWER:
+        SOCKET_POWER: 26 W
+        GFX_VOLTAGE: 689 mV
+        SOC_VOLTAGE: 773 mV
+        MEM_VOLTAGE: 753 mV
+        THROTTLE_STATUS: UNTHROTTLED
+        POWER_MANAGEMENT: ENABLED
+    CLOCK:
+        GFX_0:
+            CLK: 43 MHz
+            MIN_CLK: 31 MHz
+            MAX_CLK: 2213 MHz
+            DEEP_SLEEP: DISABLED
+        MEM_0:
+            CLK: 772 MHz
+            MIN_CLK: 96 MHz
+            MAX_CLK: 1218 MHz
+            DEEP_SLEEP: DISABLED
+        VCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2934 MHz
+            DEEP_SLEEP: DISABLED
+        DCLK_0:
+            CLK: 25 MHz
+            MIN_CLK: 25 MHz
+            MAX_CLK: 2200 MHz
+            DEEP_SLEEP: DISABLED
+        FCLK_0:
+            CLK: 1200 MHz
+            MIN_CLK: 601 MHz
+            MAX_CLK: 2301 MHz
+            DEEP_SLEEP: DISABLED
+        SOCCLK_0:
+            CLK: 750 MHz
+            MIN_CLK: 300 MHz
+            MAX_CLK: 1500 MHz
+            DEEP_SLEEP: DISABLED
+    TEMPERATURE:
+        EDGE: 42 °C
+        HOTSPOT: 50 °C
+        MEM: 52 °C
+    PCIE:
+        WIDTH: 16
+        SPEED: 16 GT/s
+        BANDWIDTH: N/A
+        REPLAY_COUNT: N/A
+        L0_TO_RECOVERY_COUNT: N/A
+        REPLAY_ROLL_OVER_COUNT: N/A
+        NAK_SENT_COUNT: N/A
+        NAK_RECEIVED_COUNT: N/A
+        CURRENT_BANDWIDTH_SENT: N/A
+        CURRENT_BANDWIDTH_RECEIVED: N/A
+        MAX_PACKET_SIZE: N/A
+        LC_PERF_OTHER_END_RECOVERY: N/A
+    ECC:
+        TOTAL_CORRECTABLE_COUNT: 0
+        TOTAL_UNCORRECTABLE_COUNT: 0
+        TOTAL_DEFERRED_COUNT: 0
+        CACHE_CORRECTABLE_COUNT: N/A
+        CACHE_UNCORRECTABLE_COUNT: N/A
+    ECC_BLOCKS: N/A
+    FAN:
+        SPEED: 0
+        MAX: 255
+        RPM: 0
+        USAGE: 0.0 %
+    VOLTAGE_CURVE:
+        POINT_0_FREQUENCY: N/A
+        POINT_0_VOLTAGE: N/A
+
+batch_end=2026-05-28T19:45:23+10:00

raw/qwen3-14b-all-quants-summary-20260528-193947.csv

@@ -0,0 +1,10 @@
+quant,model_file,size,params,backend,ngl,batch,pp512,tg128,status,sha256,result_file
+Q3_K_M,Qwen3-14B-Q3_K_M.gguf,6.81 GiB,14.77 B,ROCm,999,512,1004.75 ± 31.82,42.93 ± 0.05,ok,9694f1db203d5a18bea6f19a36be01b2d2c79bf9b7f8df553a4cc5cff870b47e,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt
+Q4_K_M,Qwen3-14B-Q4_K_M.gguf,8.38 GiB,14.77 B,ROCm,999,512,1086.99 ± 21.08,36.53 ± 0.07,ok,5eaa0870bd81ed3b58a630a271234cfa604e43ffb3a19cd68e54a80dd9d52a66,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt
+Q5_K_M,Qwen3-14B-Q5_K_M.gguf,9.79 GiB,14.77 B,ROCm,999,512,1036.95 ± 17.77,41.57 ± 0.01,ok,305309039ea44df963c01f66d65e949f8eebca238554916df437f02295a3a413,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt
+UD-Q2_K_XL,Qwen3-14B-UD-Q2_K_XL.gguf,5.65 GiB,14.77 B,ROCm,999,512,1064.51 ± 16.34,49.66 ± 0.08,ok,6abefcfb88de4157acc6109a711af65fdb5a6da2a77a1ac91bdada2ded22fd49,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt
+UD-Q3_K_XL,Qwen3-14B-UD-Q3_K_XL.gguf,7.07 GiB,14.77 B,ROCm,999,512,1038.61 ± 18.82,43.11 ± 0.04,ok,46b8c8a3fda3de374f465be47176c956e44c7ef56ceddf6aa8bafddb1c73d5b6,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt
+UD-Q4_K_XL,Qwen3-14B-UD-Q4_K_XL.gguf,8.53 GiB,14.77 B,ROCm,999,512,1087.87 ± 31.11,38.14 ± 0.02,ok,7f371adeaaf51dc97da927c61e6573e75e3ea33d76ec09dddb8cad00d0d5b6cb,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt
+UD-Q5_K_XL,Qwen3-14B-UD-Q5_K_XL.gguf,9.82 GiB,14.77 B,ROCm,999,512,1047.31 ± 15.11,40.31 ± 0.05,ok,f62da26805a70d9963400ca6160843c82e1aaadc714da8b299d0a676676741b8,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt
+UD-Q6_K_XL,Qwen3-14B-UD-Q6_K_XL.gguf,12.37 GiB,14.77 B,ROCm,999,512,1101.35 ± 30.22,36.60 ± 0.05,ok,83e7dcec42952a4ed6e092977c587a9ab8d528d52d0db481df6d694d5181811b,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt
+UD-Q8_K_XL,Qwen3-14B-UD-Q8_K_XL.gguf,,,,,,,,failed/no result rows,2da4750f82103624cbc6f831a5907628670c9b2ba59c4d57e59186c19860f4b3,/home/neo/bench/results/llama/llama-bench-Qwen3-14B-UD-Q8_K_XL.gguf-20260528-194455.txt

raw/qwen36-35b-a3b-q3km-mtp-speculative-20260529-103518.txt

@@ -0,0 +1,32 @@
+=== Qwen3.6 MTP speculative benchmark ===
+2026-05-29T10:35:18+10:00
+model=/home/neo/bench/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q3_K_M.gguf
+sha256=8966dd0cd8c543c4228490a2a8b0e0814fc4f1e6a8e199ceed4de6754ae7b8e1
+llama_cpp_commit=f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
+kernel=6.17.0-29-generic
+gpu=GPU: 0     BDF: 0000:07:00.0     UUID: 0cff747e-0000-1000-80a7-3042394d6f22     KFD_ID: 34454     NODE_ID: 1     PARTITION_ID: 0  
+note=separate MTP path; not normal llama-bench comparable table
+
+=== ATTEMPT c8192_b512_ub512 ===
+2026-05-29T10:35:31+10:00
+COMMAND: ./build-hip/bin/llama-speculative -m /home/neo/bench/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q3_K_M.gguf -ngl auto -c 8192 -b 512 -ub 512 -fa on -np 1 --spec-type draft-mtp --spec-draft-n-max 2 -p '<prompt>' -n 128 --perf
+0.00.022.868 E main: --model-draft is required
+ATTEMPT_RC=1
+2026-05-29T10:35:31+10:00
+
+=== ATTEMPT c4096_b256_ub256 ===
+2026-05-29T10:35:31+10:00
+COMMAND: ./build-hip/bin/llama-speculative -m /home/neo/bench/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q3_K_M.gguf -ngl auto -c 4096 -b 256 -ub 256 -fa on -np 1 --spec-type draft-mtp --spec-draft-n-max 2 -p '<prompt>' -n 128 --perf
+0.00.022.077 E main: --model-draft is required
+ATTEMPT_RC=1
+2026-05-29T10:35:31+10:00
+
+=== ATTEMPT c2048_b128_ub128 ===
+2026-05-29T10:35:31+10:00
+COMMAND: ./build-hip/bin/llama-speculative -m /home/neo/bench/models/Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q3_K_M.gguf -ngl auto -c 2048 -b 128 -ub 128 -fa on -np 1 --spec-type draft-mtp --spec-draft-n-max 2 -p '<prompt>' -n 128 --perf
+0.00.019.807 E main: --model-draft is required
+ATTEMPT_RC=1
+2026-05-29T10:35:31+10:00
+
+=== RESULT_FILE ===
+/home/neo/bench/results/llama/qwen36-35b-a3b-q3km-mtp-speculative-20260529-103518.txt