Views
No views yet
This is a DFlash draft model, not a standalone language model.It must be used together with a compatible google/gemma-4-31B-it GGUF target model.
z-lab/gemma-4-31B-it-DFlashgoogle/gemma-4-31B-itgemma-4-31B-it-UD-Q4_K_XL1llama-server \
2 -m gemma-4-31B-it-UD-Q4_K_XL.gguf \
3 -md gemma-4-31B-it-DFlash-Q4_K_M.gguf \
4 --spec-type draft-dflash \
5 --spec-draft-n-max 41llama-cli \
2 -m gemma-4-31B-it-UD-Q4_K_XL.gguf \
3 -md gemma-4-31B-it-DFlash-Q4_K_M.gguf \
4 --spec-type draft-dflash \
5 --spec-draft-n-max 4n_max = 4 gave the best aggregate wall-clock result.n_max = 2 gave the highest overall acceptance rate and performed better on several high-entropy natural-language tasks.n_max = 4–5 performed well for predictable outputs such as code completion, JSON, and repeated patterns.n_max = 5 was slower overall in this mixed benchmark, despite improving several structured tasks.--spec-draft-n-max 4n_max = 2 may be preferable.convert_hf_to_gguf.py.--target-model-dir.gemma-4-31B-it-UD-Q4_K_XLgemma-4-31B-it-DFlash-Q4_K_Mn_max | Draft acceptance | Predicted tokens | Draft tokens | Accepted tokens | Wall time |
|---|---|---|---|---|---|
| 2 | 74.5% | 1,990 | 1,588 | 1,183 | 42.37 s |
| 3 | 64.0% | 1,990 | 2,032 | 1,300 | 41.36 s |
| 4 | 60.3% | 1,999 | 2,332 | 1,405 | 40.39 s |
| 5 | 53.3% | 1,999 | 2,710 | 1,444 | 43.02 s |
n_max = 21code_python pred= 192 draft= 145 acc= 118 rate=0.814 tok/s=53.7
2code_cpp pred= 192 draft= 144 acc= 118 rate=0.819 tok/s=54.7
3explain_concept pred= 192 draft= 192 acc= 95 rate=0.495 tok/s=41.5
4summarize pred= 46 draft= 42 acc= 25 rate=0.595 tok/s=45.1
5qa_factual pred= 166 draft= 142 acc= 94 rate=0.662 tok/s=48.6
6translation pred= 16 draft= 18 acc= 7 rate=0.389 tok/s=36.5
7creative_short pred= 34 draft= 36 acc= 16 rate=0.444 tok/s=39.1
8stepwise_math pred= 192 draft= 148 acc= 117 rate=0.790 tok/s=53.5
9json_output pred= 192 draft= 132 acc= 125 rate=0.947 tok/s=60.2
10long_reasoning pred= 192 draft= 144 acc= 119 rate=0.826 tok/s=54.6
11repeat_pattern pred= 192 draft= 127 acc= 127 rate=1.000 tok/s=60.9
12code_completion pred= 192 draft= 136 acc= 123 rate=0.904 tok/s=58.0
13long_code_review pred= 192 draft= 182 acc= 99 rate=0.544 tok/s=42.3
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1990
18 draft tokens: 1588
19 accepted tokens: 1183
20 acceptance rate: 0.7450
21 total wall time: 42.37 sn_max = 31code_python pred= 192 draft= 180 acc= 130 rate=0.722 tok/s=56.5
2code_cpp pred= 192 draft= 176 acc= 132 rate=0.750 tok/s=58.7
3explain_concept pred= 192 draft= 269 acc= 101 rate=0.376 tok/s=38.7
4summarize pred= 46 draft= 54 acc= 27 rate=0.500 tok/s=45.9
5qa_factual pred= 166 draft= 198 acc= 100 rate=0.505 tok/s=45.4
6translation pred= 16 draft= 27 acc= 7 rate=0.259 tok/s=31.7
7creative_short pred= 34 draft= 54 acc= 17 rate=0.315 tok/s=33.9
8stepwise_math pred= 192 draft= 183 acc= 130 rate=0.710 tok/s=56.7
9json_output pred= 192 draft= 152 acc= 140 rate=0.921 tok/s=67.3
10long_reasoning pred= 192 draft= 176 acc= 132 rate=0.750 tok/s=58.3
11repeat_pattern pred= 192 draft= 144 acc= 143 rate=0.993 tok/s=71.8
12code_completion pred= 192 draft= 159 acc= 137 rate=0.862 tok/s=63.5
13long_code_review pred= 192 draft= 260 acc= 104 rate=0.400 tok/s=38.7
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1990
18 draft tokens: 2032
19 accepted tokens: 1300
20 acceptance rate: 0.6398
21 total wall time: 41.36 sn_max = 41code_python pred= 192 draft= 186 acc= 144 rate=0.774 tok/s=64.4
2code_cpp pred= 192 draft= 203 acc= 139 rate=0.685 tok/s=59.3
3explain_concept pred= 192 draft= 319 acc= 111 rate=0.348 tok/s=38.6
4summarize pred= 47 draft= 60 acc= 33 rate=0.550 tok/s=49.8
5qa_factual pred= 165 draft= 236 acc= 107 rate=0.453 tok/s=44.7
6translation pred= 16 draft= 36 acc= 7 rate=0.194 tok/s=28.2
7creative_short pred= 43 draft= 92 acc= 21 rate=0.228 tok/s=29.9
8stepwise_math pred= 192 draft= 204 acc= 140 rate=0.686 tok/s=60.0
9json_output pred= 192 draft= 168 acc= 149 rate=0.887 tok/s=71.6
10long_reasoning pred= 192 draft= 204 acc= 140 rate=0.686 tok/s=59.6
11repeat_pattern pred= 192 draft= 153 acc= 152 rate=0.994 tok/s=77.9
12code_completion pred= 192 draft= 180 acc= 146 rate=0.811 tok/s=66.9
13long_code_review pred= 192 draft= 291 acc= 116 rate=0.399 tok/s=40.1
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1999
18 draft tokens: 2332
19 accepted tokens: 1405
20 acceptance rate: 0.6025
21 total wall time: 40.39 sn_max = 51code_python pred= 192 draft= 201 acc= 150 rate=0.746 tok/s=64.0
2code_cpp pred= 192 draft= 236 acc= 143 rate=0.606 tok/s=55.9
3explain_concept pred= 192 draft= 402 acc= 110 rate=0.274 tok/s=33.1
4summarize pred= 47 draft= 60 acc= 36 rate=0.600 tok/s=54.0
5qa_factual pred= 165 draft= 280 acc= 110 rate=0.393 tok/s=40.9
6translation pred= 16 draft= 45 acc= 7 rate=0.156 tok/s=24.6
7creative_short pred= 43 draft= 120 acc= 20 rate=0.167 tok/s=24.8
8stepwise_math pred= 192 draft= 222 acc= 146 rate=0.658 tok/s=59.5
9json_output pred= 192 draft= 187 acc= 153 rate=0.818 tok/s=69.6
10long_reasoning pred= 192 draft= 221 acc= 146 rate=0.661 tok/s=59.4
11repeat_pattern pred= 192 draft= 160 acc= 159 rate=0.994 tok/s=82.8
12code_completion pred= 192 draft= 182 acc= 154 rate=0.846 tok/s=71.7
13long_code_review pred= 192 draft= 394 acc= 110 rate=0.279 tok/s=32.7
14
15Aggregate:
16 requests: 13
17 predicted tokens: 1999
18 draft tokens: 2710
19 accepted tokens: 1444
20 acceptance rate: 0.5328
21 total wall time: 43.02 sn_max = 2 achieved the highest aggregate draft acceptance rate.n_max = 4 achieved the shortest total wall time in this mixed benchmark.n_max = 5 improved highly predictable tasks such as repeated patterns and code completion, but increased total wall time.n_max increased.google/gemma-4-31B-it GGUF target model is required. The target GGUF may use a different quantization from the draft model, but both models must share compatible tokenizer and model architecture assumptions.google/gemma-4-31B-itz-lab/gemma-4-31B-it-DFlash and google/gemma-4-31B-it repositories identify their applicable licensing terms. Users should review the upstream model cards and licenses before redistribution or commercial use.