model card: prefill sweep numbers for the fused tiled prologue build
Browse files
README.md
CHANGED
|
@@ -58,7 +58,7 @@ MXFP4 plus the rotation tensors, so a dequantize-and-rotate reference is a few l
|
|
| 58 |
|---|---|---|---|
|
| 59 |
| GSM8K 500q, greedy, served path | 97.4-98.0% | **97.4-97.6%** | 97.8% |
|
| 60 |
| decode step @ctx 25 / 8k / 32k | 24.19 / 25.74 / 26.70 ms | **24.53 / 25.89 / 26.80 ms** | 28.6 ms @25 |
|
| 61 |
-
| prefill, 8k / 32k
|
| 62 |
| BetterBench combined decode, single pass | 226 t/s | 203 t/s | 186 t/s |
|
| 63 |
| in-serve kernel check vs exact dequant (real inputs) | -- | rel 0.0012-0.0021 | -- |
|
| 64 |
|
|
|
|
| 58 |
|---|---|---|---|
|
| 59 |
| GSM8K 500q, greedy, served path | 97.4-98.0% | **97.4-97.6%** | 97.8% |
|
| 60 |
| decode step @ctx 25 / 8k / 32k | 24.19 / 25.74 / 26.70 ms | **24.53 / 25.89 / 26.80 ms** | 28.6 ms @25 |
|
| 61 |
+
| prefill sweep, 2k / 8k / 16k / 32k / 64k (BetterBench PP t/s) | 3782 / 3700 / 3725 / 3621 / 3450 | **4770 / 4827 / 4649 / 4495 / 4273** | -- |
|
| 62 |
| BetterBench combined decode, single pass | 226 t/s | 203 t/s | 186 t/s |
|
| 63 |
| in-serve kernel check vs exact dequant (real inputs) | -- | rel 0.0012-0.0021 | -- |
|
| 64 |
|