Nyamdavaa Amar
commited on
Commit
·
2a9343e
1
Parent(s):
ec79aec
Add approximate measure of memories.
Browse files- README.md +1 -3
- adaptive_schedule.py +0 -11
- description1.md +5 -0
README.md
CHANGED
@@ -13,6 +13,4 @@ license: apache-2.0
|
|
13 |
|
14 |
# Pipeline Parallellism with Controllable Memory
|
15 |
|
16 |
-
Check out our paper at [Arxiv](https://arxiv.org/abs/2405.15362).
|
17 |
-
|
18 |
-
Bubble Rate here is calculated as (1 - longest stage time/(F+B+W)/m).
|
|
|
13 |
|
14 |
# Pipeline Parallellism with Controllable Memory
|
15 |
|
16 |
+
Check out our paper at [Arxiv](https://arxiv.org/abs/2405.15362).
|
|
|
|
adaptive_schedule.py
CHANGED
@@ -376,17 +376,6 @@ def squeeze_without_change_order(schedules, m):
|
|
376 |
identifier_cnt[i][identifier] += 1
|
377 |
identifier_index[_cnt * p + i][identifier] = index
|
378 |
stage_index[i] = index + 1
|
379 |
-
while True:
|
380 |
-
if(len(squeezed[0]) == 1):
|
381 |
-
break
|
382 |
-
allempty = True
|
383 |
-
for x in squeezed:
|
384 |
-
if x[-1] != ' ':
|
385 |
-
allempty = False
|
386 |
-
if allempty == False:
|
387 |
-
break
|
388 |
-
for x in squeezed:
|
389 |
-
del x[-1]
|
390 |
return squeezed
|
391 |
|
392 |
|
|
|
376 |
identifier_cnt[i][identifier] += 1
|
377 |
identifier_index[_cnt * p + i][identifier] = index
|
378 |
stage_index[i] = index + 1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
379 |
return squeezed
|
380 |
|
381 |
|
description1.md
CHANGED
@@ -6,4 +6,9 @@ From our findings, we need approximately 1/3 memory under ideal conditions (F, B
|
|
6 |
|
7 |
Check out our paper at [Arxiv](https://arxiv.org/abs/2405.15362).
|
8 |
|
|
|
|
|
|
|
|
|
|
|
9 |
Bubble Rate here is calculated as (1 - longest stage time/(F+B+W)/m).
|
|
|
6 |
|
7 |
Check out our paper at [Arxiv](https://arxiv.org/abs/2405.15362).
|
8 |
|
9 |
+
| Comparison assuming T_F=T_B=T_W | 1F1B | V-Min | V-Half | V-ZB |
|
10 |
+
| ----------------------------------------------------- | ------- |------- | ---------- | ---- |
|
11 |
+
| Bubble Rate | (p-1)/m | ???? | ??? | 0 |
|
12 |
+
| Activation Memory <br> (Compared to 1F1B) | p | (p+4)//3 | (p+2)//2 | p |
|
13 |
+
|
14 |
Bubble Rate here is calculated as (1 - longest stage time/(F+B+W)/m).
|