Skip to content
Navigation Menu
Sign in
Products
Platform
GitHub
Repository hosting, code review, and collaboration
Copilot
AI coding agent for tasks, projects, and automated work
Build & Ship
Actions
CI/CD and automation
Code Quality
Quality analysis, fixes, and enforcement
Codespaces
Cloud development environments
Secure
Code Security
Vulnerability detection and remediation
Secret Protection
Secret scanning and leak prevention
All Features
Developers
Learn
Documentation
Training
Certifications
Build
Build with Copilot
Customize Copilot
Integrations
Explore
Trending repos
GitHub Sponsors
Developer articles
Changelog
Changelog
Your source for the latest features, improvements, and fixes across GitHub products, all in one place.
Businesses
By Stage
Enterprises
Team
Startups
Nonprofits
By Industry
Healthcare
Financial services
Manufacturing
Government
See all industries
Evaluate GitHub
Why GitHub
Customer Stories
Business insights
Trust Center
Add-ons
GitHub Advanced Security
Enterprise-grade code security
Copilot for business
Enterprise-grade AI features
Premium Support
Enterprise-grade 24/7 support
Resources
Learn & Connect
Documentation
Resource center
Blog
Events & webinars
Community Forum
Customer Stories
Integrations
Support
Customer support
Training
GitHub Status
Company
About
Partners
Changelog
GitHub Universe
GitHub Universe
The event for developers, by developers — talks, workshops, and what’s next for GitHub.
Pricing
Search
/
Sign in
Sign up
Appearance settings
You signed in with another tab or window.
Reload
to refresh your session.
You signed out in another tab or window.
Reload
to refresh your session.
You switched accounts on another tab or window.
Reload
to refresh your session.
Dismiss alert
{{ message }}
ServeurpersoCom
omnivoice.cpp
Repository navigation
Code
Issues
13
(13)
Pull requests
8
(8)
Actions
Projects
Security and quality
Insights
More
items
Files
Expand file tree
master
Breadcrumbs
omnivoice.cpp
/
quantize.sh
Copy path
Blame
More file actions
Blame
More file actions
Latest commit
History
History
History
executable file
·
37 lines (33 loc) · 1.37 KB
master
Breadcrumbs
omnivoice.cpp
/
quantize.sh
Copy path
Top
File metadata and controls
Code
Blame
executable file
·
37 lines (33 loc) · 1.37 KB
Raw
Copy raw file
Download raw file
Open symbols panel
Edit and raw actions
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
#!/bin/bash
# Derive lighter GGUFs from the F32 source-of-truth produced by convert.py.
# omnivoice-base-F32.gguf -> BF16, Q8_0, Q4_K_M
# omnivoice-tokenizer-F32.gguf -> BF16, Q8_0, Q4_K_M
#
# Three variants cover the useful precision range : BF16 for max precision
# on CUDA, Q8_0 as the balanced default, Q4_K_M as the smallest variant
# that still sounds correct. Q5_K_M / Q6_K were tested and dropped : their
# size sits between Q4_K_M and Q8_0 with negligible perceptual gain.
#
# Quantization policy is centralized in tools/quantize.cpp should_quantize :
# RVQ codebooks (quantizer.quantizers.*) and the fc / fc2 linear projections
# wrapping them stay at F32 in every variant. Nearest-neighbor lookup is
# sensitive to per-row quantization noise ; even BF16 mantissa truncation
# drifts codes enough to break voice cloning. Conv weights stay at source
# dtype and are cast to F16 at load time by gf_load_conv_f16 (ARM im2col
# strict). Same policy as acestep.cpp keeping VAE-critical paths intact.
set -eu
Q="./build/quantize"
quantize() {
local src="$1" type="$2"
local out="${src/-F32.gguf/-${type}.gguf}"
if [ -f "$out" ]; then
echo "[Skip] $out"
else
$Q "$src" "$out" "$type"
fi
}
for src in models/omnivoice-base-F32.gguf models/omnivoice-tokenizer-F32.gguf; do
quantize "$src" BF16
quantize "$src" Q8_0
quantize "$src" Q4_K_M
done
You can’t perform that action at this time.