scott 51188ca973 [dev-fp16] Only convert T3 to fp16, leave s3gen/ve in fp32 ...

s3gen.speaker_encoder (CAMPPlus xvector) hardcodes float32 inputs in
its inference() method, causing dtype mismatch when weights are fp16.
T3 (the autoregressive GPT-2-medium LLM) has no such constraint and
is the token-generation bottleneck that benefits most from fp16.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-04-05 20:41:24 -04:00

.gitea/workflows

Build dev-fp16 branch to :dev-fp16 tag (not :latest)

2026-04-05 20:34:54 -04:00

.gitignore

Initial implementation: Chatterbox TTS with ROCm and Wyoming

2026-04-05 09:51:09 -04:00

config.py

Multi-pass warmup and smaller chunk_size to fix HA timeout

2026-04-05 15:04:46 -04:00

config.yaml

Multi-pass warmup and smaller chunk_size to fix HA timeout

2026-04-05 15:04:46 -04:00

docker-compose.yml

Switch to ROCm 6.1 + torch 2.5.1 to fix MIOpen workspace=0 slowness

2026-04-05 17:27:21 -04:00

Dockerfile.rocm

Switch to ROCm 6.1 + torch 2.5.1 to fix MIOpen workspace=0 slowness

2026-04-05 17:27:21 -04:00

engine.py

[dev-fp16] Only convert T3 to fp16, leave s3gen/ve in fp32

2026-04-05 20:41:24 -04:00

main.py

Cache voice conditionals and add FP16 autocast

2026-04-05 20:22:13 -04:00

perth_stub.py

Replace resemble-perth with a no-op stub

2026-04-05 11:16:08 -04:00

requirements-rocm-init.txt

Switch to ROCm 6.1 + torch 2.5.1 to fix MIOpen workspace=0 slowness

2026-04-05 17:27:21 -04:00

requirements-rocm.txt

Fix safetensors version conflict with transformers 4.46.3

2026-04-05 17:34:23 -04:00

wyoming_handler.py

Add per-event logging and top-level exception catching

2026-04-05 17:10:32 -04:00

wyoming_voices.py

Add supports_synthesize_streaming=True to TtsProgram

2026-04-05 16:52:02 -04:00

Description

No description provided

213 KiB

Languages

Python 100%