unparallelled commited on
Commit
3849342
·
verified ·
1 Parent(s): 0429f3c

Upload logs/t64_master.log with huggingface_hub

Browse files
Files changed (1) hide show
  1. logs/t64_master.log +11 -0
logs/t64_master.log ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [04:53:57] === 680M 64k extension, 8x H100, cp=8 dp=1 ===
2
+ [04:53:57] parent : p1final (95/90/82 at x64 -- best model we have)
3
+ [04:53:57] YaRN : SCALE=16 x ORIG=4096 = 65536 == CP_T INVARIANT HOLDS
4
+ [04:53:57] virtual range 0..4095 (full), 16 tokens/virtual unit
5
+ [04:53:57] LR : peak 1.8e-4 (6e-4 x 0.3), warmup 100 / stable 900 / DECAY 200 -> 6e-5
6
+ [04:53:57] data : Data_Loader/Data/mix64 b5=50 b4=25 b3=15 b1=10, usable 1220 steps, running 1200
7
+ [04:53:57] PROBE cp=8 (8192 tok/rank, grad_accum 8)
8
+ [04:55:15] PROBE OK cp=8: peak mem: 50855MB of 81559 MiB, dt dt: 9998.16ms
9
+ [04:55:15] TRAINING 1200 steps
10
+ [06:42:22] done, run 2, loss: 2.798417
11
+ [06:42:22] === 64K TRAINING COMPLETE ===