Model card: speed from the release build
Browse files
README.md
CHANGED
|
@@ -43,12 +43,12 @@ Global-MMLU-Lite lands inside its ±2.7-point interval of the card's figure.
|
|
| 43 |
|
| 44 |
| | 1M-token window | 8 x 128K windows |
|
| 45 |
|---|---|---|
|
| 46 |
-
| Decode at 128K context, plain / with the drafter |
|
| 47 |
-
| Decode at 1M context, plain / with the drafter |
|
| 48 |
-
| Reading a cold 64K prompt | |
|
| 49 |
-
| Eight streams at once | |
|
| 50 |
|
| 51 |
-
A cold 1M-token prompt takes
|
| 52 |
|
| 53 |
## Use
|
| 54 |
|
|
|
|
| 43 |
|
| 44 |
| | 1M-token window | 8 x 128K windows |
|
| 45 |
|---|---|---|
|
| 46 |
+
| Decode at 128K context, plain / with the drafter | 25.9 / 37.9 tok/s | 23.3 / 36.9 tok/s (120K) |
|
| 47 |
+
| Decode at 1M context, plain / with the drafter | 13.3 / 21.1 tok/s | |
|
| 48 |
+
| Reading a cold 64K prompt | | 774 tok/s (84 s) |
|
| 49 |
+
| Eight streams at once | | 89-90 tok/s in total |
|
| 50 |
|
| 51 |
+
A cold 1M-token prompt takes about 46 minutes to read; follow-up turns resume from the prompt cache in seconds.
|
| 52 |
|
| 53 |
## Use
|
| 54 |
|