turboderp-org/exllamav3
1
0
Fork 0
mirror of https://github.com/turboderp-org/exllamav3.git synced 2026-04-20 14:29:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

convert.py: Fix overflow when mixing bitrates for expert-heavy models

Browse Source
This commit is contained in:
turboderp
2026-03-15 00:29:37 +01:00
parent cd94bf8f8f
commit 5f54aa5f57
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
exllamav3/conversion/allocation.py
View File

@@ -134,10 +134,12 @@ def allocate_linear(
numel = l.weights_numel()
budget = int(bpw * numel) + surplus_bits + 1
bpw = budget / numel
bpw = max(int(math.floor(bpw)), 1)
used_budget = bpw * numel
base_bpw = int(math.floor(bpw))
new_bpw = int(math.floor(budget / numel))
new_bpw = max(new_bpw, 1)
new_bpw = min(new_bpw, base_bpw + 2, 8)
used_budget = new_bpw * numel
strategy = {l.key: bpw}
strategy = {l.key: new_bpw}
surplus = budget - used_budget
return strategy, surplus