Skip to content

Fix PocketTTS prompt graph overflow - #1

Open
RANGROO wants to merge 36 commits into
abeiro:devfrom
RANGROO:codex/pockettts-prompt-graph-capacity
Open

Fix PocketTTS prompt graph overflow#1
RANGROO wants to merge 36 commits into
abeiro:devfrom
RANGROO:codex/pockettts-prompt-graph-capacity

Conversation

@RANGROO

@RANGROO RANGROO commented Jul 29, 2026

Copy link
Copy Markdown

Problem

PocketTTS aborts the entire audio.cpp server when long or token-dense prompts overflow the fixed 32,768-node FlowLM prompt graph. The crash occurs while adding per-step key/value cache transfer nodes.

Change

  • increase only the PocketTTS FlowLM prompt graph capacity to 262,144 nodes
  • keep the normal generation-step graph unchanged
  • document why prompt processing needs the additional headroom

No API, configuration, model, or migration changes are included.

Validation

  • built audiocpp_server with CUDA 12.8 (scripts/build_linux.sh --backend cuda --native-cpu OFF --target audiocpp_server -j 8)
  • default alba voice: HTTP 200, valid WAV
  • Herika cloned voice reference: HTTP 200, valid WAV
  • former crash reproducer (220 repeated Japanese characters): HTTP 200, valid WAV; server remained healthy
  • 640-character dialogue chunk: HTTP 200, valid WAV
  • five sequential requests: all HTTP 200 with valid WAVs
  • two concurrent clients: both HTTP 200 with valid WAVs; server remained healthy

Local deployment

The CUDA build was deployed to /home/dwemer/audio.cpp/build/bin/audiocpp_server and verified through the installed port 8086. HerikaServer TTS Studio reports PocketTTS (audio.cpp) Connected.

Limits

  • CUDA was validated on an RTX 4090.
  • CPU-only execution was not rebuilt or runtime-tested.

abeiro and others added 30 commits June 28, 2026 21:20
Change size_t to std::size_t in noise.h, fill Debian compilation issue
- pack QKV and gate/up projections and use fused SwiGLU in the shared Qwen decoder path
- use grouped FlashAttention, F16 bucketed KV caches, direct set-rows updates, and CUDA-friendly RoPE/view/set-rows graphs
- retain and reuse cloned-reference KV prefixes so repeated server requests prefill only their text suffix
- add graph and F16 KV correctness coverage plus fixed-seed request-level CUDA benchmark/parity tooling
- document the optimized Higgs runtime controls and reproducible validation flow
…rmance

Optimize Higgs Audio v3 inference
# Conflicts:
#	README.md
#	src/framework/runtime/registry.cpp
@RANGROO
RANGROO marked this pull request as ready for review July 29, 2026 05:29
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants