Update benchmarks/models/sansa/README.md
Browse files
benchmarks/models/sansa/README.md
CHANGED
|
@@ -1,8 +1,10 @@
|
|
| 1 |
## SANSA
|
| 2 |
|
| 3 |
-
Training SANSA on Yambda-500M with Listens, on Yambda-5B with Likes and on Yambda-5B with Listens is infeasible
|
|
|
|
| 4 |
|
| 5 |
-
For example, given the base weights density of \\(
|
|
|
|
| 6 |
|
| 7 |
One may observe that the good solution is to try to reduce the weights density. However given the results on Yambda-50M with Listens, further reducing density would collapse the model’s capacity, it becomes obvious that it is just practically futile.
|
| 8 |
|
|
|
|
| 1 |
## SANSA
|
| 2 |
|
| 3 |
+
Training SANSA on Yambda-500M with Listens, on Yambda-5B with Likes and on Yambda-5B with Listens is infeasible
|
| 4 |
+
due to the memory explosion during sparse matrix multiplication.
|
| 5 |
|
| 6 |
+
For example, given the base weights density of \\(d=5 \cdot 10^{-5}\\), the sparse matrix multiplication for Yambda-500M with Listens requires nearly
|
| 7 |
+
\\(N_{events} \cdot |I|^2 \cdot d = 0.5 \cdot 10^9 \cdot (3 \cdot 10^6) ^ 2 \cdot 5 \cdot 10^{-5} = 2.3 \cdot 10^17\\) operations.
|
| 8 |
|
| 9 |
One may observe that the good solution is to try to reduce the weights density. However given the results on Yambda-50M with Listens, further reducing density would collapse the model’s capacity, it becomes obvious that it is just practically futile.
|
| 10 |
|