fix: real calibrated per-projection input_scale (the 1.0 placeholder underflows fp8 block scales and causes long-context repetition)
Browse files
model-input-scales.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 4840704
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b812b4d7df15dedb06e74defa6380fecba61d45b8c117fb42eda1443430d2a4a
|
| 3 |
size 4840704
|