Thanks for your excellent work!
I noticed some inconsistencies between the DFlash(16) numbers reported in your paper (Table 2) and those reported in the original DFlash paper (Table 1).
For example, for Q3-4B at Temperature=0 on MT-Bench:
- Your paper reports: Speedup = 2.69×, τ = 5.29
- DFlash paper reports: Speedup = 2.85×, τ = 4.35
Could the discrepancy in τ be due to a difference in how average acceptance length is computed — for instance, averaging over per-sample acceptance lengths vs. averaging over all individual drafting steps?
Or do you have any other insight into where this gap might come from?
Thank you!
Thanks for your excellent work!
I noticed some inconsistencies between the DFlash(16) numbers reported in your paper (Table 2) and those reported in the original DFlash paper (Table 1).
For example, for Q3-4B at Temperature=0 on MT-Bench:
Could the discrepancy in τ be due to a difference in how average acceptance length is computed — for instance, averaging over per-sample acceptance lengths vs. averaging over all individual drafting steps?
Or do you have any other insight into where this gap might come from?
Thank you!