feat(audio8_tts): Falcon-H1 0.1B (Mamba2 + attention) support and codec/AR performance - #444

Open

gqf2008 wants to merge 17 commits into

0xShug0:mainfrom

gqf2008:feat/audio8-tts-falcon-h1-01b

Commits on Sep 4, 2026

feat(audio8_tts): Falcon-H1 0.1B (Mamba2 + attention) stateful forward port
gqf2008
committed
993a9a2
chore(ggml-metal): renormalize ggml-metal.metal CRLF -> LF
gqf2008
committed
0f8b614
fix: Falcon-H1 state corruption and numerical stability
gqf2008
committed
720c3b5
docs: record Falcon-H1 0.1B port status and open issues
gqf2008
committed
332a777
docs(audio8_tts): mark Falcon-H1 open issues resolved with root causes
gqf2008
committed
01e3e0d
perf(audio8_tts): run fast AR graph on a dedicated CPU backend
gqf2008
committed
e44fc03
feat(audio8_tts): time codec decode graph build vs compute
gqf2008
committed
e81db68
perf(conv): Metal fast path for stride-1 conv1d on time-fast layouts
gqf2008
committed
24ec127
perf(audio8_tts): run Falcon-H1 per-token step graph on the CPU backend
gqf2008
committed
7757f0e
perf(audio8_tts): keep Falcon KV cache in padded host buffers, write slots in-graph
gqf2008
committed
f281804
perf(audio8_tts): reuse the Falcon step graph per KV capacity bucket
gqf2008
committed
c5b3186
perf(audio8_tts): accumulate conv1d per-tap GEMMs in-place via ggml_mul_mat_acc
gqf2008
committed
24a245b
perf: fused snake1d activation op (GGML_OP_SNAKE_1D) for audio8_tts codec
gqf2008
committed
c08b443
perf(audio8_tts): zero-copy Falcon step state and constants on CPU backend
gqf2008
committed
855b7ef
perf(audio8_tts): run decoder blocks in channel-fast layout on Metal
gqf2008
committed
b2b9708
fix(ggml-metal): ssm_scan reduction reads garbage when sgptg < NW
gqf2008
committed
71325ea
fix(audio8_tts): Falcon-H1 0.1B logits parity and long-run stability
gqf2008
committed
fcba3a4

Navigation Menu

feat(audio8_tts): Falcon-H1 0.1B (Mamba2 + attention) support and codec/AR performance - #444

feat(audio8_tts): Falcon-H1 0.1B (Mamba2 + attention) support and codec/AR performance#444
gqf2008 wants to merge 17 commits into
0xShug0:mainfrom
gqf2008:feat/audio8-tts-falcon-h1-01b

Commits on Sep 4, 2026

feat(audio8_tts): Falcon-H1 0.1B (Mamba2 + attention) stateful forward port

chore(ggml-metal): renormalize ggml-metal.metal CRLF -> LF

fix: Falcon-H1 state corruption and numerical stability

docs: record Falcon-H1 0.1B port status and open issues

docs(audio8_tts): mark Falcon-H1 open issues resolved with root causes

perf(audio8_tts): run fast AR graph on a dedicated CPU backend

feat(audio8_tts): time codec decode graph build vs compute

perf(conv): Metal fast path for stride-1 conv1d on time-fast layouts

perf(audio8_tts): run Falcon-H1 per-token step graph on the CPU backend

perf(audio8_tts): keep Falcon KV cache in padded host buffers, write slots in-graph

perf(audio8_tts): reuse the Falcon step graph per KV capacity bucket

perf(audio8_tts): accumulate conv1d per-tap GEMMs in-place via ggml_mul_mat_acc

perf: fused snake1d activation op (GGML_OP_SNAKE_1D) for audio8_tts codec

perf(audio8_tts): zero-copy Falcon step state and constants on CPU backend

perf(audio8_tts): run decoder blocks in channel-fast layout on Metal

fix(ggml-metal): ssm_scan reduction reads garbage when sgptg < NW

fix(audio8_tts): Falcon-H1 0.1B logits parity and long-run stability