File size: 5,301 Bytes
1227045
 
d5f4a97
1227045
d5f4a97
1227045
 
 
d5f4a97
1227045
d5f4a97
1227045
d5f4a97
1227045
d5f4a97
 
 
1227045
 
 
 
 
 
 
 
 
 
 
d5f4a97
1227045
d5f4a97
1227045
 
 
d5f4a97
1227045
d5f4a97
1227045
d5f4a97
1227045
d5f4a97
 
 
1227045
 
 
 
 
 
 
 
 
 
 
d5f4a97
1227045
d5f4a97
1227045
 
 
d5f4a97
1227045
d5f4a97
1227045
d5f4a97
1227045
d5f4a97
 
 
1227045
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
% Generated from audited outcomes; edit the analysis source, not this table.
\begingroup\footnotesize

\setlength{\tabcolsep}{3pt}
\setlength{\LTcapwidth}{\linewidth}
\begin{longtable}{@{}llrrrr@{}}
\caption{Iteration costs for Qwen-7B/GSM8K and Mistral-7B/longer-prompt code. Counts are within margin/above margin/memory failure/unresolved. Warm-up means updates 1--2; steady means 3--20. Seconds are per-run means, then mean [minimum, maximum] across completed seeds. Sync is the weight-update interval, not isolated communication time.}\label{stab:timing-conditions}\\
\toprule
Model & Condition & \multicolumn{1}{c}{Outcomes} & \multicolumn{1}{c}{Warm-up (s)} & \multicolumn{1}{c}{Steady (s)} & \multicolumn{1}{c}{Sync (s)} \\
\midrule\endfirsthead
\caption[]{\emph{Continued.}}\\
\toprule
Model & Condition & \multicolumn{1}{c}{Outcomes} & \multicolumn{1}{c}{Warm-up (s)} & \multicolumn{1}{c}{Steady (s)} & \multicolumn{1}{c}{Sync (s)} \\
\midrule\endhead
\midrule\multicolumn{6}{r}{\emph{Continued on next page.}}\\
\endfoot
\bottomrule\endlastfoot
Qwen 7B & Anchor & 3/0/0/0 & 87.8 [74.2, 95.3] & 76.1 [61.1, 85.0] & 40.2 [37.0, 44.0] \\
Qwen 7B & Offload off & 3/0/0/0 & 89.3 [69.3, 128.8] & 82.6 [62.4, 121.6] & 58.9 [36.8, 100.4] \\
Qwen 7B & Reservation .70 & 3/0/0/0 & 93.8 [89.0, 99.2] & 84.3 [81.1, 88.3] & 42.3 [40.4, 46.0] \\
Qwen 7B & State resident & 3/0/0/0 & 75.0 [71.9, 77.0] & 67.8 [64.7, 69.6] & 39.4 [37.0, 43.8] \\
Mistral 7B & Anchor & 3/0/0/0 & 108.9 [104.2, 117.6] & 101.4 [97.5, 108.9] & 40.6 [39.1, 41.5] \\
Mistral 7B & Offload off & 3/0/0/0 & 90.0 [82.9, 102.2] & 79.2 [71.9, 90.6] & 37.0 [33.2, 41.4] \\
Mistral 7B & Reservation .70 & 3/0/0/0 & 115.3 [101.5, 142.8] & 107.7 [94.1, 133.8] & 50.6 [37.1, 76.8] \\
Mistral 7B & State resident & 3/0/0/0 & 104.0 [90.8, 129.4] & 95.1 [82.8, 118.5] & 48.4 [35.6, 71.9] \\
\end{longtable}\endgroup

\begingroup\footnotesize

\setlength{\tabcolsep}{3pt}
\setlength{\LTcapwidth}{\linewidth}
\begin{longtable}{@{}llrrrr@{}}
\caption{Observed work and host memory for the same iteration-cost runs. Token rates use prompt plus response tokens divided by summed step duration (updates 3--20). Response length is the per-run mean, in tokens. Anonymous host memory is the peak charged to the job's batch cgroup, in GiB; file cache is reported separately in the artifact. Values are mean [minimum, maximum].}\label{stab:timing-work}\\
\toprule
Model & Condition & \multicolumn{1}{c}{$n$} & \multicolumn{1}{c}{Tokens/s} & \multicolumn{1}{c}{Response tokens} & \multicolumn{1}{c}{Host anon (GiB)} \\
\midrule\endfirsthead
\caption[]{\emph{Continued.}}\\
\toprule
Model & Condition & \multicolumn{1}{c}{$n$} & \multicolumn{1}{c}{Tokens/s} & \multicolumn{1}{c}{Response tokens} & \multicolumn{1}{c}{Host anon (GiB)} \\
\midrule\endhead
\midrule\multicolumn{6}{r}{\emph{Continued on next page.}}\\
\endfoot
\bottomrule\endlastfoot
Qwen 7B & Anchor & 3/3 & 311.5 [273.1, 380.9] & 257.9 [255.7, 259.5] & 176.8 [176.5, 177.0] \\
Qwen 7B & Offload off & 3/3 & 308.4 [191.0, 372.5] & 258.2 [257.5, 258.5] & 176.6 [176.2, 176.9] \\
Qwen 7B & Reservation .70 & 3/3 & 276.4 [263.5, 287.2] & 259.2 [258.7, 259.5] & 176.3 [175.2, 176.8] \\
Qwen 7B & State resident & 3/3 & 341.2 [331.6, 356.6] & 256.6 [255.9, 257.3] & 176.3 [175.8, 176.7] \\
Mistral 7B & Anchor & 3/3 & 834.6 [773.6, 865.5] & 547.7 [545.2, 550.0] & 168.2 [168.1, 168.3] \\
Mistral 7B & Offload off & 3/3 & 1075.1 [931.7, 1172.9] & 546.0 [543.8, 547.6] & 167.9 [167.9, 167.9] \\
Mistral 7B & Reservation .70 & 3/3 & 804.9 [632.8, 895.7] & 548.3 [546.0, 551.6] & 167.8 [167.5, 168.2] \\
Mistral 7B & State resident & 3/3 & 912.0 [713.6, 1019.6] & 547.9 [544.9, 550.3] & 167.9 [167.8, 168.0] \\
\end{longtable}\endgroup

\begingroup\footnotesize

\setlength{\tabcolsep}{3pt}
\setlength{\LTcapwidth}{\linewidth}
\begin{longtable}{@{}llrrrr@{}}
\caption{Same-device timing contrasts against the offload-on, reservation-0.45, release-on anchor. Time differences are alternative minus anchor, in seconds; ratios are alternative/anchor. Only jointly completed pairs with valid work records enter numeric comparisons. Values are mean [minimum, maximum] across pairs.}\label{stab:timing-pairs}\\
\toprule
Model & Alternative & \multicolumn{1}{c}{$n$} & \multicolumn{1}{c}{Time difference} & \multicolumn{1}{c}{Time ratio} & \multicolumn{1}{c}{Token-rate ratio} \\
\midrule\endfirsthead
\caption[]{\emph{Continued.}}\\
\toprule
Model & Alternative & \multicolumn{1}{c}{$n$} & \multicolumn{1}{c}{Time difference} & \multicolumn{1}{c}{Time ratio} & \multicolumn{1}{c}{Token-rate ratio} \\
\midrule\endhead
\midrule\multicolumn{6}{r}{\emph{Continued on next page.}}\\
\endfoot
\bottomrule\endlastfoot
Qwen 7B & Offload off & 3/3 & 6.5 [-21.0, 60.4] & 1.17 [0.75, 1.99] & 1.05 [0.50, 1.33] \\
Qwen 7B & State resident & 3/3 & -8.3 [-16.0, 3.5] & 0.91 [0.81, 1.06] & 1.12 [0.94, 1.23] \\
Qwen 7B & Reservation .70 & 3/3 & 8.2 [-1.6, 27.2] & 1.14 [0.98, 1.44] & 0.91 [0.69, 1.02] \\
Mistral 7B & Offload off & 3/3 & -22.2 [-25.6, -18.3] & 0.78 [0.74, 0.83] & 1.29 [1.20, 1.36] \\
Mistral 7B & State resident & 3/3 & -6.3 [-14.9, 9.6] & 0.93 [0.85, 1.09] & 1.09 [0.92, 1.18] \\
Mistral 7B & Reservation .70 & 3/3 & 6.3 [-3.7, 24.9] & 1.06 [0.96, 1.23] & 0.96 [0.82, 1.04] \\
\end{longtable}\endgroup