| { |
| "sample_rate": 24000, |
| "frame_samples": 1920, |
| "frame_rate": 12.5, |
| "max_context": 512, |
| "prefill_max": 128, |
| "programs": { |
| "flow_lm": { |
| "forward": { |
| "inputs": [ |
| { |
| "shape": [ |
| 1, |
| 128, |
| 512 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1 |
| ], |
| "dtype": "long" |
| }, |
| { |
| "shape": [ |
| 1, |
| 32 |
| ], |
| "dtype": "float" |
| } |
| ], |
| "outputs": [ |
| { |
| "shape": [ |
| 1, |
| 32 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 1 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 1, |
| 512 |
| ], |
| "dtype": "float" |
| } |
| ] |
| } |
| }, |
| "mimi": { |
| "forward": { |
| "inputs": [ |
| { |
| "shape": [ |
| 1, |
| 1, |
| 32 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 512, |
| 16 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 1, |
| 1, |
| 250 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 512, |
| 6 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 256, |
| 6 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 256, |
| 2 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 128, |
| 5 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 128, |
| 2 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 64, |
| 4 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 64, |
| 2 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 64, |
| 2 |
| ], |
| "dtype": "float" |
| } |
| ], |
| "outputs": [ |
| { |
| "shape": [ |
| 1, |
| 1, |
| 1920 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 512, |
| 16 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 1, |
| 1, |
| 250 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 512, |
| 6 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 256, |
| 6 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 256, |
| 2 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 128, |
| 5 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 128, |
| 2 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 64, |
| 4 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 64, |
| 2 |
| ], |
| "dtype": "float" |
| }, |
| { |
| "shape": [ |
| 1, |
| 64, |
| 2 |
| ], |
| "dtype": "float" |
| } |
| ] |
| } |
| } |
| }, |
| "flow_lm": "forward(embeds, input_pos, noise): writes KV positions [input_pos, input_pos + S); outputs are for the last position. Calls: voice prompt (S = 125) at 0, text prompt (S = #tokens) at 125, then one call per frame with S = 1: the first takes phonon_bos_embed, each later one the previous call's next_embeds", |
| "mimi": "forward(latent, *state) -> (pcm, *state): pass the previous call's states back; zeros to start", |
| "mimi_state": [ |
| [ |
| "upsample_partial", |
| [ |
| 1, |
| 512, |
| 16 |
| ] |
| ], |
| [ |
| "k0", |
| [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ] |
| ], |
| [ |
| "v0", |
| [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ] |
| ], |
| [ |
| "k1", |
| [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ] |
| ], |
| [ |
| "v1", |
| [ |
| 1, |
| 8, |
| 250, |
| 64 |
| ] |
| ], |
| [ |
| "kv_seen", |
| [ |
| 1, |
| 1, |
| 1, |
| 250 |
| ] |
| ], |
| [ |
| "rope_offset", |
| [ |
| 1 |
| ] |
| ], |
| [ |
| "init_conv_prev", |
| [ |
| 1, |
| 512, |
| 6 |
| ] |
| ], |
| [ |
| "up1_partial", |
| [ |
| 1, |
| 256, |
| 6 |
| ] |
| ], |
| [ |
| "res1_prev", |
| [ |
| 1, |
| 256, |
| 2 |
| ] |
| ], |
| [ |
| "up2_partial", |
| [ |
| 1, |
| 128, |
| 5 |
| ] |
| ], |
| [ |
| "res2_prev", |
| [ |
| 1, |
| 128, |
| 2 |
| ] |
| ], |
| [ |
| "up3_partial", |
| [ |
| 1, |
| 64, |
| 4 |
| ] |
| ], |
| [ |
| "res3_prev", |
| [ |
| 1, |
| 64, |
| 2 |
| ] |
| ], |
| [ |
| "final_conv_prev", |
| [ |
| 1, |
| 64, |
| 2 |
| ] |
| ] |
| ], |
| "text": "normalize (the official runtime's English rules; identity for plain sentences), then trim, collapse whitespace, capitalize the first letter, append '.' if the text ends with a letter or digit; encode with tokenizer.json without special tokens; embeds[i] = bf16 table row tokens[i] (as f32)", |
| "noise": "32 values per frame, N(0, temperature); 0.3 is the official runtime's default, 0 makes it deterministic", |
| "eos": "stop after frames_after_eos more frames once eos_logit > 4.0 (frames_after_eos = 3 for texts of <= 4 words, else 1); the EOS frame and its tail are decoded", |
| "frame_budget": "ceil((#tokens / 3 + 2) * 12.5) frames at most", |
| "voices": { |
| "default": { |
| "file": "voice_default.f32.bin", |
| "shape": [ |
| 125, |
| 512 |
| ] |
| }, |
| "en_us_easton": { |
| "file": "voice_en_us_easton.f32.bin", |
| "shape": [ |
| 125, |
| 512 |
| ] |
| } |
| }, |
| "conditions": { |
| "num_speakers": "1", |
| "duration_delta": 0.0, |
| "note": "the runtime's defaults, folded into the programs" |
| } |
| } |
|
|
