fix: qkv_proj per-rank scale-grid repair (ranks 1..3 of all 48 layers)\n\nPre-fix converter dequantized fused qkv with a flat scale grid, corrupting\nranks 1..NR-1 (rank 0 intact). Repaired from XiaomiMiMo/MiMo-V2.5 fp8 source\nwith the fixed per-rank dequant + identical int8 requant. Verified:\nverify_mimo_qkv 5/5 layers PASS all ranks; ppl prose 9.97 / code 2.07\n(code matches the 12.40/2.07 reference axis); greedy 'The capital of France\nis' -> 'Paris.' See peng findings.md §46.
Browse files- out-00001.safetensors +1 -1
out-00001.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 12635280000
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e5c62b9dd039d0023449ff75bdff33e60b88c35dbabe5b21c02c1f03d975be22
|
| 3 |
size 12635280000
|