newgrit1004

newgrit1004 / qwen3-tts-triton

Public

Triton kernel fusion for Qwen3-TTS 1.7B inference acceleration — RMSNorm, SwiGLU, M-RoPE, Norm+Residual

46
1
100% credibility
Found Mar 24, 2026 at 46 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

This project accelerates text-to-speech generation up to 5x faster using clever speed tricks while matching original audio quality, complete with easy setup, sample comparisons, and live testing dashboard.

How It Works

1
🔍 Find faster voices

You hear about a simple way to make AI turn text into super quick, natural-sounding speech.

2
📥 Get the tool

Download it easily with one quick step on your computer.

3
🪟 Open the dashboard

Launch a friendly window showing voice examples and tools.

4
🎧 Listen and compare

Play speech samples side-by-side to hear how much smoother and faster they are.

5
✍️ Try your own words

Type what you want to say, pick a voice, and create speech right away.

6
⚡ Feel the speed boost

Watch voices appear in seconds, way quicker than normal, with perfect quality.

🎉 Voices ready anytime

Now you can make amazing, instant speech for stories, videos, or fun whenever you want.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 46 to 46 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is qwen3-tts-triton?

This Python package accelerates Qwen3-TTS 1.7B inference up to 5x using Triton kernels fused for key ops like RMSNorm and SwiGLU, pip-installable with zero extra VRAM. It offers four modes—baseline PyTorch, Triton-only, CUDA Graph faster mode, and hybrid—plus drop-in patching for any Qwen3 setup and a Streamlit dashboard for live comparisons. Run `pip install qwen3-tts-triton` then `TritonRunner().generate(text="hello")` for instant speedup on NVIDIA CUDA GPUs.

Why is it gaining traction?

Unlike generic Triton GitHub tutorials or symbolic kernels, it delivers plug-and-play acceleration for a hot 1.7B TTS model, with hybrid mode hitting 4.7x RTF on RTX 5090 while passing 3-tier verification (kernel parity, model sim, E2E quality). ComfyUI nodes, Windows WSL support, kernel profiling tools, and PyTorch benchmarks make it dead simple to test Triton kernel cache hits versus baselines—no serving infra needed.

Who should use this?

TTS pipeline builders generating batch audio for apps or voice cloning in ComfyUI workflows. NVIDIA GPU users tweaking Qwen3 for low-latency inference, like real-time synthesis in Python scripts or OpenAI-style agents. Devs profiling Triton kernels on CUDA for similar 1.7B models.

Verdict

Solid alpha for TTS acceleration—46 stars and 1.0% credibility reflect early days, but comprehensive tests, benches, and docs make it low-risk to pip kernels into your Triton GitHub repo experiments. Grab it if Qwen3 speed is your bottleneck.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.