feat: support Moondream Parakeet Ultra and Redux - #77

Open

localai-org-maint-bot wants to merge 31 commits into

feat/hf-ternary-vad

Commits on Sep 29, 2026

feat(convert): import HF safetensors Parakeet TDT checkpoints
mudler
committed
ae599e1
feat(convert): keep ternary weights packed and carry the VAD head
mudler
committed
b9105f7
feat(loader): read the ternary and VAD flags from the GGUF
mudler
committed
c70d419
feat(ternary): repack, int8 activation quantization and scalar reference
mudler
committed
bd1ff06
feat(ternary): run packed ternary encoder linears through a ggml op pair
mudler
committed
e54e0c5
fix(ternary): validate and repack packed weights at load
mudler
committed
bc8775e
feat(ternary): AVX-512 VNNI and AVX2 kernels with runtime dispatch
mudler
committed
4bf3e96
feat(ternary): NEON dot-product kernel
mudler
committed
f2eb381
fix(ternary): move NEON -ffp-contract=off flag inside NOT MSVC guard
mudler
committed
c8499e6
docs: ternary Redux support, measurements and converter
mudler
committed
9d151fb
bench: compare ternary kernels with ggml Q8_0 and F16 mul_mat
mudler
committed
f7c9e8c
docs: fix ternary numbers and wording after review
mudler
committed
b4164da
ternary: one output row per vector lane in the kernels
mudler
committed
7c8cc48
bench: run the scalar ternary reference with fewer repetitions
mudler
committed
93028b6
ternary: tune the x86 tile shapes
mudler
committed
6136e9b
ternary: vectorize the activation quantization on x86
mudler
committed
f228376
ternary: build the AVX-512 quantizer for avx512f only
mudler
committed
8b66c63
docs: ternary speed after the kernel rework
mudler
committed
8b18b08

Commits on Sep 30, 2026

feat(vad): VAD head and a vad-probe tool
mudler
committed
2a370e6
fix(vad): read head weights through the backend, asymmetric tests
mudler
committed
ff2b86b
feat(vad): cut long audio at pauses
mudler
committed
dcc7666
fix(vad_segmenter_test): correct test expectations and add rule tests
mudler
committed
fed1141
feat(vad): transcribe long audio in VAD-cut segments
mudler
committed
ec0d983
test(vad): self-contained long-clip checks, harden slicing and CLI flags
mudler
committed
31a0058
scripts: long-form clip builder and VAD WER evaluation
mudler
committed
53e7f5a
docs: long-form VAD validation for Ultra and Redux
mudler
committed
daae921
docs: correct Ultra and Redux claims, sharpen the VAD reading
mudler
committed
6a7ae55
loader: validate packed ternary tensors before reading them
mudler
committed
d038c7e
vad: harden the config, the segmenter and the frame offsets
mudler
committed
82788d2
docs: scalar-kernel fallback, benchmark provenance, batch and two-speaker checks
mudler
committed
a8682ce
harden: scan all tensors for packed ternary, guard VAD options, check VAD frame at load
mudler
committed
f545dc4