initial release

This commit is contained in:
civ
2026-07-05 18:11:23 +07:00
commit 8fb29dac70
2105 changed files with 499091 additions and 0 deletions
+143
View File
@@ -0,0 +1,143 @@
[Quant] Q4_K_M -> ../models/omnivoice-base-Q4_K_M.gguf + ../models/omnivoice-tokenizer-Q4_K_M.gguf
[Input] Prompt: 102 chars: omnivoice.cpp is a minimal C++17 and GGML port of OmniVoice ...
[Input] Instruct: male, young adult, moderate pitch
[Input] Language: French
[Input] Seed: 42
Loading weights: 0%| | 0/313 [00:00<?, ?it/s]
Loading weights: 100%|██████████| 313/313 [00:00<00:00, 27715.28it/s]
Loading weights: 0%| | 0/527 [00:00<?, ?it/s]
Loading weights: 100%|██████████| 527/527 [00:00<00:00, 31183.32it/s]
[CLI] Seed: 42
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 97247 MiB):
Device 0: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, compute capability 12.0, VMM: yes, VRAM: 97247 MiB
load_backend: loaded CUDA backend from /mnt/workspace/omnivoice.cpp/build/libggml-cuda.so
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = NVIDIA RTX PRO 6000 Blackwell Workstation Edition (NVIDIA) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
load_backend: loaded Vulkan backend from /mnt/workspace/omnivoice.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/omnivoice.cpp/build/libggml-cpu-zen4.so
[Load] LM backend: CPU (CPU threads: 16)
[GGUF] ../models/omnivoice-base-Q4_K_M.gguf: 312 tensors, data at offset 5339136
[LM-Load] Loaded: 28L H=1024 FFN=3072 Nh=16 Nkv=8 D=128 theta=1000000 eps=1e-06 | audio: K=8 V=1025 mask_id=1024
[Qwen3] Attn: Q+K fused, V separate
[Qwen3] MLP: gate+up fused
[WeightCtx] Loaded 312 tensors, 383.5 MB into backend
[Load] Flash attention disabled
[BPE] Loaded from GGUF: 151676 vocab, 151387 merges, eos_id=151643
[BPE] Registered 7 OmniVoice special tokens (total specials=8)
[GGUF] ../models/omnivoice-tokenizer-Q4_K_M.gguf: 486 tensors, data at offset 43680
[WeightCtx] Loaded 52 tensors, 11.1 MB into backend
[DAC] Loaded: 5 blocks, upsample 960x (8*5*4*2*3), 24 kHz mono out, weights 38.7 MB
[DAC-Enc] Loaded: 5 blocks, downsample 960x (8*5*4*2*3), 24 kHz mono in, weights 98.0 MB
[Sem-Enc] Loaded: 2 blocks, hidden=768, no temporal downsample, weights 28.1 MB
[HuBERT-Feat] Loaded: 7 conv layers, cumul stride 320, weights 8.0 MB
[HuBERT-Proj] Loaded: LN(512) + Linear(512 -> 768), weights 0.2 MB
[HuBERT-EncInit] Loaded: pos_conv k=128 g=16 + LN(768), weights 9.0 MB
[HuBERT-Stack] Loaded: 12 layers, hidden=768 heads=12 ffn=3072, weights 47.8 MB
[PipelineCodec] Loaded codec: sr=24000 hop=960 backend=CPU
[TTS-Long] Single-shot path: T=163 frames (6.52s), threshold=750 frames
[Prompt] Built: B'=2 K=8 S=205 N1=12 N2=30 Sref=0 Stgt=163 c_len=205 u_len=163 denoise=0
[Debug] prompt-cond-ids: [205] first4: 151670.000000 1626.000000 151671.000000 151672.000000
[Debug] prompt-uncond-ids: [205] first4: 1024.000000 1024.000000 1024.000000 1024.000000
[TTS] Prompt: B'=2 K=8 S=205 c_len=205 u_len=163
[MaskGIT] Start: T=163 K=8 S=205 num_step=32 guidance=2.00 t_shift=0.100 layer_pen=5.00 cls_t=0.00 pos_t=5.00
[Debug] lm-hidden-step0-cond-embed: [205, 1024] first4: -0.032717 -0.045180 -0.006232 0.006232
[Debug] lm-hidden-step0-cond-l0: [205, 1024] first4: 0.154686 0.026378 -0.185436 -3.165199
[Debug] lm-hidden-step0-cond-l1: [205, 1024] first4: 0.437802 0.166829 -0.134803 -4.232050
[Debug] lm-hidden-step0-cond-l2: [205, 1024] first4: 0.362836 0.387874 -0.106459 -4.173686
[Debug] lm-hidden-step0-cond-l3: [205, 1024] first4: 0.306157 0.463569 0.102114 -3.870833
[Debug] lm-hidden-step0-cond-l4: [205, 1024] first4: 0.042881 0.650357 0.153386 -3.869005
[Debug] lm-hidden-step0-cond-l5: [205, 1024] first4: 0.206240 0.633691 0.103334 -3.664940
[Debug] lm-hidden-step0-cond-l6: [205, 1024] first4: 0.485093 0.551843 -0.019555 -3.264849
[Debug] lm-hidden-step0-cond-l13: [205, 1024] first4: 1.280035 1.342078 0.369379 -4.710588
[Debug] lm-hidden-step0-cond-l14: [205, 1024] first4: 1.235396 1.730470 0.270711 -4.348973
[Debug] lm-hidden-step0-cond-l15: [205, 1024] first4: 0.762168 2.424932 0.479876 -4.697135
[Debug] lm-hidden-step0-cond-l16: [205, 1024] first4: 0.691627 3.194687 0.777326 -3.734025
[Debug] lm-hidden-step0-cond-l17: [205, 1024] first4: -0.326634 4.616097 -0.344347 -2.091247
[Debug] lm-hidden-step0-cond-l18: [205, 1024] first4: 0.122559 5.398581 -1.973707 -0.736050
[Debug] lm-hidden-step0-cond-l19: [205, 1024] first4: 0.397533 4.904632 0.790809 -1.144656
[Debug] lm-hidden-step0-cond-l20: [205, 1024] first4: 0.298477 11.882438 -3.561796 -0.085610
[Debug] lm-hidden-step0-cond-l1-norm1: [205, 1024] first4: 0.020594 0.005074 0.019820 0.166612
[Debug] lm-hidden-step0-cond-l1-attn: [205, 1024] first4: 0.219669 -0.057559 0.416608 -0.603003
[Debug] lm-hidden-step0-cond-l1-norm2: [205, 1024] first4: 0.187479 -0.021981 -0.150537 -4.956777
[Debug] lm-hidden-step0-cond-l1-mlp: [205, 1024] first4: 0.063447 0.198009 -0.365975 -0.463848
[Debug] lm-hidden-step0-cond: [205, 1024] first4: 0.277472 5.364466 35.501648 0.088516
[Debug] lm-hidden-step0-uncond-embed: [205, 1024] first4: 0.004852 0.004656 0.011258 0.007240
[Debug] lm-hidden-step0-uncond-l0: [205, 1024] first4: -0.085264 0.050480 0.237071 -2.684460
[Debug] lm-hidden-step0-uncond-l1: [205, 1024] first4: -0.332236 0.444955 0.257447 -3.125254
[Debug] lm-hidden-step0-uncond-l2: [205, 1024] first4: -0.468589 0.796536 0.267579 -3.503743
[Debug] lm-hidden-step0-uncond-l3: [205, 1024] first4: -0.148804 0.885914 0.134291 -3.467972
[Debug] lm-hidden-step0-uncond-l4: [205, 1024] first4: -0.283279 0.627035 -0.036198 -3.553025
[Debug] lm-hidden-step0-uncond-l5: [205, 1024] first4: -0.225921 1.179644 0.038044 -3.990106
[Debug] lm-hidden-step0-uncond-l6: [205, 1024] first4: -0.314412 0.870721 0.027771 -4.692402
[Debug] lm-hidden-step0-uncond-l13: [205, 1024] first4: 2.514897 1.465800 0.011024 -12.771043
[Debug] lm-hidden-step0-uncond-l14: [205, 1024] first4: 2.601196 2.329514 -0.056488 -14.936769
[Debug] lm-hidden-step0-uncond-l15: [205, 1024] first4: 3.449160 3.286325 0.180722 -19.210449
[Debug] lm-hidden-step0-uncond-l16: [205, 1024] first4: 4.349464 1.904123 -1.049485 -15.553217
[Debug] lm-hidden-step0-uncond-l17: [205, 1024] first4: 4.022482 3.321661 -2.405239 -12.912841
[Debug] lm-hidden-step0-uncond-l18: [205, 1024] first4: 2.874889 3.653088 2.802409 -11.111227
[Debug] lm-hidden-step0-uncond-l19: [205, 1024] first4: 3.302193 7.646958 8.948181 -6.603518
[Debug] lm-hidden-step0-uncond-l20: [205, 1024] first4: 5.849554 16.194990 0.903898 -4.651946
[Debug] lm-hidden-step0-uncond-l1-norm1: [205, 1024] first4: -0.011855 0.010139 -0.026462 0.147568
[Debug] lm-hidden-step0-uncond-l1-attn: [205, 1024] first4: -0.248962 0.399358 0.359951 -0.046009
[Debug] lm-hidden-step0-uncond-l1-norm2: [205, 1024] first4: -0.180156 0.341313 -0.418447 -3.865834
[Debug] lm-hidden-step0-uncond-l1-mlp: [205, 1024] first4: 0.001989 -0.004882 -0.339575 -0.394785
[Debug] lm-hidden-step0-uncond: [205, 1024] first4: 0.806666 46.954834 128.707962 0.587398
[Debug] lm-logits-step0-cond: [8, 163, 1025] first4: 17.425339 11.735054 16.922569 16.372063
[Debug] lm-logits-step0-uncond: [8, 163, 1025] first4: 17.462112 11.404427 15.982658 15.724873
[Debug] mg-pred-tokens-step0: [8, 163] first4: 61.000000 489.000000 599.000000 885.000000
[Debug] mg-scores-step0: [8, 163] first4: -2.178911 -1.816939 -2.602834 -1.702895
[Debug] mg-log-probs-step0: [8, 163, 1025] first4: -10.229448 -15.184931 -8.778849 -9.914797
[MaskGIT-Step] 1/32 demask=5 remaining=1299
[MaskGIT-Step] 2/32 demask=5 remaining=1294
[MaskGIT-Step] 3/32 demask=5 remaining=1289
[MaskGIT-Step] 4/32 demask=6 remaining=1283
[MaskGIT-Step] 5/32 demask=6 remaining=1277
[MaskGIT-Step] 6/32 demask=6 remaining=1271
[MaskGIT-Step] 7/32 demask=7 remaining=1264
[MaskGIT-Step] 8/32 demask=7 remaining=1257
[MaskGIT-Step] 9/32 demask=8 remaining=1249
[MaskGIT-Step] 10/32 demask=8 remaining=1241
[MaskGIT-Step] 11/32 demask=9 remaining=1232
[MaskGIT-Step] 12/32 demask=9 remaining=1223
[MaskGIT-Step] 13/32 demask=10 remaining=1213
[MaskGIT-Step] 14/32 demask=11 remaining=1202
[MaskGIT-Step] 15/32 demask=12 remaining=1190
[MaskGIT-Step] 16/32 demask=13 remaining=1177
[MaskGIT-Step] 17/32 demask=15 remaining=1162
[MaskGIT-Step] 18/32 demask=16 remaining=1146
[MaskGIT-Step] 19/32 demask=18 remaining=1128
[MaskGIT-Step] 20/32 demask=21 remaining=1107
[MaskGIT-Step] 21/32 demask=23 remaining=1084
[MaskGIT-Step] 22/32 demask=27 remaining=1057
[MaskGIT-Step] 23/32 demask=31 remaining=1026
[MaskGIT-Step] 24/32 demask=36 remaining=990
[MaskGIT-Step] 25/32 demask=43 remaining=947
[MaskGIT-Step] 26/32 demask=52 remaining=895
[MaskGIT-Step] 27/32 demask=64 remaining=831
[MaskGIT-Step] 28/32 demask=80 remaining=751
[MaskGIT-Step] 29/32 demask=105 remaining=646
[MaskGIT-Step] 30/32 demask=142 remaining=504
[MaskGIT-Step] 31/32 demask=204 remaining=300
[MaskGIT-Step] 32/32 demask=300 remaining=0
[MaskGIT] Total LM forward: 10652.87 ms across 32 steps (avg 332.90 ms/step)
[Debug] mg-tokens: [8, 163] first4: 511.000000 439.000000 668.000000 291.000000
[TTS] Decode: K=8 T=163 expected_samples=156480
[Debug] output-audio: [156480] first4: -0.033257 -0.033647 -0.031930 -0.030415
[TTS-Long] Post-proc: 156480 -> 161280 samples (6.72s at 24000 Hz, ref_rms=-1.0000)
[WAV] Wrote cpp/tts-cpp.wav: 161280 samples, 24000 Hz, mono F32
[OmniVoice-TTS] TTS: wrote cpp/tts-cpp.wav (161280 samples @ 24000 Hz, 6.72 s)
[Python] Audio: 161280 samples 6.72s -> python/tts-python.wav
[GGML] Cmd: ../build/omnivoice-tts --model ../models/omnivoice-base-Q4_K_M.gguf --codec ../models/omnivoice-tokenizer-Q4_K_M.gguf --seed 42 --instruct male, young adult, moderate pitch --lang French --format wav32 --dump cpp --no-fa -o cpp/tts-cpp.wav
[GGML] Audio: 161280 samples 24000 Hz 6.72s -> cpp/tts-cpp.wav
[Cossim] PromptIDs cond exact: 100.00% uncond exact: 100.00%
[Cossim] Embed cond cos: 0.999831 max: 1.7486e-02 mean: 1.3334e-03 uncond cos: 0.999835 max: 1.2504e-02 mean: 1.2963e-03
[Cossim] L0 cond cos: 0.999842 max: 3.4227e-01 mean: 1.1238e-02 uncond cos: 0.999707 max: 4.2769e-01 mean: 1.4919e-02
[Cossim] L6 cond cos: 0.999434 max: 1.7328e+00 mean: 3.3462e-02 uncond cos: 0.998964 max: 1.0225e+00 mean: 5.0820e-02
[Cossim] L13 cond cos: 0.999322 max: 6.9382e+00 mean: 8.4511e-02 uncond cos: 0.999234 max: 3.2781e+00 mean: 8.6687e-02
[Cossim] L20 cond cos: 0.999123 max: 6.0707e+01 mean: 6.2127e-01 uncond cos: 0.999165 max: 8.5036e+00 mean: 4.0812e-01
[Cossim] Final cond cos: 0.999904 max: 2.6531e+01 mean: 3.4373e-02 uncond cos: 0.999972 max: 2.4236e+00 mean: 3.0973e-02
[Cossim] Logits cond cos: 0.999936 max: 7.2435e+00 mean: 7.7442e-01 uncond cos: 0.999935 max: 6.2314e+00 mean: 8.1805e-01
[Cossim] Step0Tokens exact: 16.18% diffs: 1093
[Cossim] Step0Scores cos: 0.963855 max_abs_diff: 4.685781 mean_abs_diff: 1.445358