FireRedTTS3 Web (ONNX, WebGPU)

λΈŒλΌμš°μ €μ—μ„œ λ°”λ‘œ μ‹€ν–‰λ˜λŠ” FireRedTTS3 μŒμ„± 볡제 ONNX λͺ¨λΈμž…λ‹ˆλ‹€.

이 λͺ¨λΈλ“€μ€ onnxruntime-web의 WebGPU μ‹€ν–‰μš©μž…λ‹ˆλ‹€. κ°€μ€‘μΉ˜λŠ” μž‘κ²Œ μ €μž₯ν•˜κ³  연산은 fp32둜 ν•©λ‹ˆλ‹€.

파일 λ‚΄μš© μ €μž₯ 방식
backbone_q8.onnx (+.data) Qwen3 1.7B λ°±λ³Έ, KV μΊμ‹œ ν•œ μŠ€ν… MatMulNBits 8λΉ„νŠΈ
flow_step_n10_s8.onnx DiT 흐름 ν—€λ“œ 10μŠ€ν…(CFG 포함)κ³Ό PatchEncoder int8 μ €μž₯ β†’ λ‘œλ“œν•  λ•Œ fp32둜
decoder_s8.onnx RedAE 디코더 (latent β†’ μŒμ„± ν”„λ ˆμž„) int8 μ €μž₯
embed_q8.bin, embed_scale.f32 ν…μŠ€νŠΈ μž„λ² λ”©, ν–‰λ§ˆλ‹€ int8κ³Ό 배율 int8
encoder_s16.onnx RedAE 인코더 (λͺ©μ†Œλ¦¬ λ“±λ‘μš©) fp16 μ €μž₯
prompt_b8.onnx ν”„λ‘¬ν”„νŠΈ PatchEncoder와 ν™”μž 투영 (λ“±λ‘μš©) 32개 블둝 λ‹¨μœ„ int8
campp.onnx CAM++ ν™”μž μž„λ² λ”© (λ“±λ‘μš©) fp32

int8둜 μ €μž₯ν•œ κ·Έλž˜ν”„λŠ” μ„Έμ…˜ μ˜΅μ…˜ session.disable_quant_qdq = "1"을 μΌœμ•Ό λ‘œλ“œν•  λ•Œ fp32둜 νŽΌμ³μ§‘λ‹ˆλ‹€.

원본 λͺ¨λΈκ³Ό λΌμ΄μ„ μŠ€: FireRedTTS3, Apache-2.0.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for kairess/fireredtts3-web

Quantized
(3)
this model