DeepSeek-V4-Flash benchmarks
Hardware
- GPU 4x MI50 32G
- CPU 2x 4189 QVM7 (Xeon Platinum 8360Y)
- RAM 16x16 2666
GPU Overclocking
oc max
export MEM_MAX=1180
export GPU_MAX=1725
export TDP_MAX=300
export TDC_MAX=330
bartowski/DeepSeek-V4-Flash-GGUF
export MODEL=bartowski/DeepSeek-V4-Flash-GGUF
export HIP_VISIBLE_DEVICES="1,0,3,2"
export CTXB=2048
export GENB=256
export COMP_THREADS=32
./llama-bench \
--hf-repo $MODEL \
--split-mode layer --flash-attn 1 \
--n-prompt $CTXB --ubatch-size $CTXB --batch-size $CTXB \
--n-gen $GENB \
--n-depth 0,4096,8192,16384,32768,65536 \
--threads $COMP_THREADS \
--main-gpu 0 --tensor-split "25;25;25;25" \
--override-tensor \
'^(FROM ROCm0|blk.0.ffn_down_exps.weight|blk.0.ffn_up_exps.weight|blk.1.ffn_down_exps.weight|blk.1.ffn_up_exps.weight|blk.2.ffn_down_exps.weight|blk.2.ffn_up_exps.weight|blk.3.ffn_down_exps.weight|blk.3.ffn_up_exps.weight)$=CPU;'\
'^(FROM ROCm1|blk.11.ffn_down_exps.weight|blk.11.ffn_up_exps.weight|blk.12.ffn_down_exps.weight|blk.12.ffn_up_exps.weight|blk.13.ffn_down_exps.weight|blk.13.ffn_up_exps.weight|blk.14.ffn_down_exps.weight)$=CPU;'\
'^(FROM ROCm2|blk.22.ffn_down_exps.weight|blk.22.ffn_up_exps.weight|blk.23.ffn_down_exps.weight|blk.23.ffn_up_exps.weight|blk.24.ffn_down_exps.weight|blk.24.ffn_up_exps.weight|blk.25.ffn_down_exps.weight)$=CPU;'\
'^(FROM ROCm3|blk.33.ffn_down_exps.weight|blk.33.ffn_up_exps.weight|blk.34.ffn_down_exps.weight|blk.34.ffn_up_exps.weight)$=CPU;'
| model | size | threads | n_ubatch | fa | test | ggml b10092-rocm-7.2.4 | mxxm b10095-rocm-7.2.4 |
|---|---|---|---|---|---|---|---|
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | pp2048 | 149.50 ± 0.27 | 156.60 ± 19.40 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | tg256 | 15.66 ± 0.21 | 15.46 ± 0.27 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | pp2048 @ d4096 | 131.20 ± 0.66 | 155.22 ± 1.02 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | tg256 @ d4096 | 14.75 ± 0.34 | 14.95 ± 0.11 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | pp2048 @ d8192 | 122.50 ± 0.45 | 143.81 ± 1.96 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | tg256 @ d8192 | 14.60 ± 0.08 | 14.67 ± 0.10 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | pp2048 @ d16384 | 108.32 ± 0.29 | 125.56 ± 0.39 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | tg256 @ d16384 | 14.05 ± 0.09 | 14.10 ± 0.08 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | pp2048 @ d32768 | 88.58 ± 0.27 | 98.48 ± 0.53 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | tg256 @ d32768 | 13.45 ± 0.07 | 13.36 ± 0.12 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | pp2048 @ d65536 | 64.47 ± 0.14 | 70.35 ± 0.11 |
| deepseek4 ?B MXFP4 MoE | 145.29 GiB | 32 | 2048 | 1 | tg256 @ d65536 | 11.85 ± 0.19 | 11.96 ± 0.15 |
unsloth/DeepSeek-V4-Flash-GGUF:UD-Q8_K_XL
MODEL=unsloth/DeepSeek-V4-Flash-GGUF:UD-Q8_K_XL
export HIP_VISIBLE_DEVICES="1,0,3,2"
export CTXB=2048
export GENB=256
export COMP_THREADS=32
./llama-bench \
--hf-repo $MODEL \
--split-mode layer --flash-attn 1 \
--n-prompt $CTXB --ubatch-size $CTXB --batch-size $CTXB \
--n-gen $GENB \
--n-depth 0,4096,8192,16384,32768,65536 \
--threads $COMP_THREADS \
--main-gpu 0 --tensor-split "25;25;25;25" \
--override-tensor \
'^(FROM ROCm0|blk.0.ffn_down_exps.weight|blk.0.ffn_up_exps.weight|blk.1.ffn_down_exps.weight|blk.1.ffn_up_exps.weight|blk.2.ffn_down_exps.weight|blk.2.ffn_up_exps.weight|blk.3.ffn_down_exps.weight|blk.3.ffn_up_exps.weight|blk.4.ffn_down_exps.weight)$=CPU;'\
'^(FROM ROCm1|blk.11.ffn_down_exps.weight|blk.11.ffn_up_exps.weight|blk.12.ffn_down_exps.weight|blk.12.ffn_up_exps.weight|blk.13.ffn_down_exps.weight|blk.13.ffn_up_exps.weight|blk.14.ffn_down_exps.weight|blk.14.ffn_up_exps.weight)$=CPU;'\
'^(FROM ROCm2|blk.22.ffn_down_exps.weight|blk.22.ffn_up_exps.weight|blk.23.ffn_down_exps.weight|blk.23.ffn_up_exps.weight|blk.24.ffn_down_exps.weight|blk.24.ffn_up_exps.weight|blk.25.ffn_down_exps.weight|blk.25.ffn_up_exps.weight)$=CPU;'\
'^(FROM ROCm3|blk.33.ffn_down_exps.weight|blk.33.ffn_up_exps.weight|blk.34.ffn_down_exps.weight|blk.34.ffn_up_exps.weight|blk.35.ffn_down_exps.weight|blk.35.ffn_up_exps.weight)$=CPU;'
| model | size | threads | n_ubatch | fa | test | t/s |
|---|---|---|---|---|---|---|
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | pp2048 | 119.87 ± 0.81 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | tg256 | 12.79 ± 0.10 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | pp2048 @ d4096 | 111.02 ± 0.14 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | tg256 @ d4096 | 12.33 ± 0.08 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | pp2048 @ d8192 | 104.49 ± 0.15 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | tg256 @ d8192 | 12.04 ± 0.05 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | pp2048 @ d16384 | 93.96 ± 0.31 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | tg256 @ d16384 | 11.65 ± 0.07 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | pp2048 @ d32768 | 78.14 ± 0.48 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | tg256 @ d32768 | 10.97 ± 0.07 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | pp2048 @ d65536 | 58.62 ± 0.41 |
| deepseek4 ?B MXFP4 MoE | 150.75 GiB | 32 | 2048 | 1 | tg256 @ d65536 | 10.01 ± 0.05 |