RaynarDM commited on
Commit
458cf03
·
verified ·
1 Parent(s): 5e10383

Add MoE DFlash2 A/B: loses harder than dense (0.89x code), contradicting the prediction

Browse files
Files changed (1) hide show
  1. benchmarks.csv +10 -1
benchmarks.csv CHANGED
@@ -29,4 +29,13 @@ dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"prose prompt, 256 tok, no spec
29
  dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"prose prompt, 256 tok, draft-dflash n-max 7",generation,6.13,tok/s,2
30
  dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"code prompt, draft-dflash n-max 7",draft_acceptance,0.76,ratio,2
31
  dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"math prompt, draft-dflash n-max 7",draft_acceptance,0.72,ratio,2
32
- dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"prose prompt, draft-dflash n-max 7",draft_acceptance,0.35,ratio,2
 
 
 
 
 
 
 
 
 
 
29
  dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"prose prompt, 256 tok, draft-dflash n-max 7",generation,6.13,tok/s,2
30
  dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"code prompt, draft-dflash n-max 7",draft_acceptance,0.76,ratio,2
31
  dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"math prompt, draft-dflash n-max 7",draft_acceptance,0.72,ratio,2
32
+ dflash2,Qwen3.8-27B,dense,IQ3_S,11.20,llama.cpp,"prose prompt, draft-dflash n-max 7",draft_acceptance,0.35,ratio,2
33
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"code prompt 256 tok no spec",generation,55.28,tok/s,2
34
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"code prompt 256 tok draft-dflash n-max 7",generation,49.26,tok/s,2
35
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"math prompt 256 tok no spec",generation,52.09,tok/s,2
36
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"math prompt 256 tok draft-dflash n-max 7",generation,48.50,tok/s,2
37
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"prose prompt 256 tok no spec",generation,51.02,tok/s,2
38
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"prose prompt 256 tok draft-dflash n-max 7",generation,30.79,tok/s,2
39
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"code prompt draft-dflash n-max 7",draft_acceptance,0.60,ratio,2
40
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"math prompt draft-dflash n-max 7",draft_acceptance,0.59,ratio,2
41
+ dflash2,Qwen3.6-35B-A3B,moe_3b_active,IQ3_S,12.74,llama.cpp,"prose prompt draft-dflash n-max 7",draft_acceptance,0.28,ratio,2