enne2/qwentts.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

logs

Browse Source
This commit is contained in:
Pascal
2026-07-04 22:47:51 +02:00
parent 4057d0331f
commit f84fc05292
12 changed files with 638 additions and 536 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/clone-CPU-BF16.log
+46 -36
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp a62fde6 (2026-05-30)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CPU (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-BF16.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 3322.0 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 334.1 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-BF16.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 48.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=off clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=off clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 3 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922110 0.609857 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067471 0.119239 0.240741 0.344303
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 48.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 41 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.112500 13.760122 -0.329032 2.993861
[Debug] enc-transformer-out: [430, 512] first4: 0.526661 3.208130 -0.164395 0.605249
@@ -63,8 +65,8 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Prompt] Built: 38 ids, N_text=30, N_instruct=0, T_ctx=225, hidden=2048, lang=english (id=2050), speaker=none (id=-1) ref_spk_emb=yes icl=yes
[Debug] prompt-ids: [38] first4: 151644.000000 77091.000000 198.000000 80.000000
[Debug] talker-input-embed: [225, 2048] first4: 0.022162 -0.009036 0.008171 -0.020198
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005331 0.008751 -0.004225 0.001141
[Debug] tts-pad-embed: [2048] first4: -0.005331 0.008751 -0.004225 0.001141
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005340 0.008744 -0.004225 0.001144
[Debug] tts-pad-embed: [2048] first4: -0.005340 0.008744 -0.004225 0.001144
[Debug] spk-emb: [2048] first4: 0.098336 -0.023751 -0.011551 0.006713
[Debug] ref-codes: [16, 215] first4: 1221.000000 2042.000000 509.000000 1522.000000
[Debug] talker-hidden-prefill-l0: [225, 2048] first4: -0.461332 -0.069998 0.016245 -0.039201
@@ -73,7 +75,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -0.770039 0.162546 0.799939 1.111742
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 7.054070 -34.409252 -7.069129 28.624571
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.240086 -1.222942 -0.223565 0.888025
[Debug] talker-logits-prefill: [3072] first4: 1.339085 -7.643277 -7.016098 -10.536522
[Debug] talker-logits-prefill: [3072] first4: 1.339535 -7.642556 -6.988994 -10.529062
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
@@ -91,8 +93,8 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=13 c=122 u=-1.0000000000 subseq=14
[Sample-CP] g=14 c=365 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: 0.073561 -0.034233 -0.008814 0.042004
[Debug] talker-hidden-step1: [2048] first4: 0.568384 2.580451 -0.014018 0.222672
[Debug] next-emb-step0: [2048] first4: 0.073552 -0.034240 -0.008814 0.042007
[Debug] talker-hidden-step1: [2048] first4: 0.576087 2.576824 -0.018687 0.235103
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
@@ -119,7 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001455 -0.002483 -0.002276 -0.001114
[Debug] output-audio: [122880] first4: 0.000529 0.000353 -0.000865 -0.000453
[Perf] PromptBuild 2.9 ms
[Perf] Prefill 389.4 ms (T_ctx prefill)
[Perf] TTFA 417.5 ms (first frame codes)
[Perf] TalkerDecode 2918.4 ms (64 frames, 45.60 ms/frame)
[Perf] CodePredictor 1492.6 ms (23.32 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 974.0 ms
[Perf] Total 5778.7 ms (64 frames, 68.92 ms/frame AR, audio 5.12 s, RTF 1.129)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -140,18 +150,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[GGML] Audio: 122880 samples 24000 Hz 5.12s -> cpp/clone/clone-cpp.wav
[Cossim] PromptIDs exact: 100.00% (38 values)
[Cossim] RefCodes exact: 86.77% (3440 values)
[Cossim] Embed cos: 0.978030 max: 1.8825e-01 mean: 6.5135e-03
[Cossim] TrailingText cos: 1.000000 max: 1.0431e-06 mean: 9.3737e-09
[Cossim] TTSPadEmbed cos: 1.000000 max: 1.0431e-06 mean: 9.3737e-09
[Cossim] L0 cos: 0.989590 max: 7.7387e-01 mean: 8.9714e-03
[Cossim] L7 cos: 0.999993 max: 1.2846e+01 mean: 2.3285e-02
[Cossim] L14 cos: 0.999976 max: 1.2850e+01 mean: 5.2420e-02
[Cossim] L21 cos: 0.999695 max: 1.7467e+01 mean: 2.0203e-01
[Cossim] L27 cos: 0.998834 max: 4.6440e+01 mean: 4.6481e-01
[Cossim] Final cos: 0.993922 max: 1.9975e+01 mean: 1.5114e-01
[Cossim] Logits cos: 0.999807 max: 5.5696e-01 mean: 7.8726e-02
[Cossim] NextEmbStep0 cos: 0.967766 max: 6.3431e-02 mean: 1.2632e-02
[Cossim] TalkerHiddenStep1 cos: 0.998209 max: 8.0196e-01 mean: 1.3068e-01
[Cossim] Embed cos: 0.978030 max: 1.8825e-01 mean: 6.5145e-03
[Cossim] TrailingText cos: 1.000000 max: 3.5048e-05 mean: 2.4958e-06
[Cossim] TTSPadEmbed cos: 1.000000 max: 3.5048e-05 mean: 2.4958e-06
[Cossim] L0 cos: 0.989591 max: 7.7401e-01 mean: 8.9719e-03
[Cossim] L7 cos: 0.999993 max: 1.2846e+01 mean: 2.3289e-02
[Cossim] L14 cos: 0.999976 max: 1.2850e+01 mean: 5.2409e-02
[Cossim] L21 cos: 0.999696 max: 1.7400e+01 mean: 2.0189e-01
[Cossim] L27 cos: 0.998836 max: 4.6276e+01 mean: 4.6497e-01
[Cossim] Final cos: 0.993939 max: 1.9954e+01 mean: 1.5107e-01
[Cossim] Logits cos: 0.999804 max: 5.9451e-01 mean: 7.9720e-02
[Cossim] NextEmbStep0 cos: 0.967765 max: 6.3429e-02 mean: 1.2632e-02
[Cossim] TalkerHiddenStep1 cos: 0.998217 max: 8.0423e-01 mean: 1.3036e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -169,6 +179,6 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkMFA cos: 0.999979 max: 6.8943e-02 mean: 1.0444e-03
[Cossim] SpkASP cos: 0.999999 max: 2.2756e-03 mean: 1.8859e-04
[Cossim] SpeakerEmb cos: 1.000000 max: 6.4747e-04 mean: 1.3471e-04
[Cossim] CodesFull exact: 4.07% (1008 values)
[Cossim] Audio cos: 0.006752
[Cossim] WAV stft_cos: 0.182248 samples: 120960
[Cossim] CodesFull exact: 3.47% (1008 values)
[Cossim] Audio cos: -0.007662
[Cossim] WAV stft_cos: 0.207437 samples: 120960
tests/clone-CPU-F32.log
+33 -23
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp a62fde6 (2026-05-30)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CPU (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-F32.gguf: 480 tensors, data at offset 5954432
[WeightCtx] Loaded 316 tensors, 6643.5 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 668.1 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 45.8 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-F32.gguf: 398 tensors, data at offset 29664
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 96.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=off clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=off clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 45.8 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 5 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922734 0.614916 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067723 0.119136 0.240880 0.344137
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 96.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 35 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.120292 13.742489 -0.348104 3.009387
[Debug] enc-transformer-out: [430, 512] first4: 0.526076 3.125370 -0.189241 0.630764
@@ -73,7 +75,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -0.771753 0.171250 0.791419 1.111589
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 7.133089 -34.474190 -7.108908 28.674923
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.242892 -1.225840 -0.224931 0.890014
[Debug] talker-logits-prefill: [3072] first4: 1.330120 -7.768508 -6.935927 -10.338602
[Debug] talker-logits-prefill: [3072] first4: 1.330122 -7.768507 -6.935926 -10.338600
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
@@ -92,7 +94,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=365 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: 0.073561 -0.034233 -0.008814 0.042004
[Debug] talker-hidden-step1: [2048] first4: 0.584085 2.683545 -0.065920 0.135427
[Debug] talker-hidden-step1: [2048] first4: 0.584086 2.683544 -0.065920 0.135428
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
@@ -119,7 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001458 -0.002491 -0.002285 -0.001123
[Debug] output-audio: [122880] first4: 0.000450 0.000327 -0.000890 -0.000492
[Perf] PromptBuild 5.0 ms
[Perf] Prefill 695.5 ms (T_ctx prefill)
[Perf] TTFA 766.0 ms (first frame codes)
[Perf] TalkerDecode 5506.9 ms (64 frames, 86.04 ms/frame)
[Perf] CodePredictor 3926.2 ms (61.35 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 1029.6 ms
[Perf] Total 11164.8 ms (64 frames, 147.39 ms/frame AR, audio 5.12 s, RTF 2.181)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -141,15 +151,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] PromptIDs exact: 100.00% (38 values)
[Cossim] RefCodes exact: 96.72% (3440 values)
[Cossim] Embed cos: 0.994609 max: 1.5292e-01 mean: 1.6562e-03
[Cossim] TrailingText cos: 1.000000 max: 1.0431e-06 mean: 9.3737e-09
[Cossim] TTSPadEmbed cos: 1.000000 max: 1.0431e-06 mean: 9.3737e-09
[Cossim] TrailingText cos: 1.000000 max: 2.9802e-08 mean: 5.6171e-10
[Cossim] TTSPadEmbed cos: 1.000000 max: 2.9802e-08 mean: 5.6171e-10
[Cossim] L0 cos: 0.997321 max: 7.2147e-01 mean: 2.5377e-03
[Cossim] L7 cos: 0.999998 max: 1.2205e+00 mean: 7.8975e-03
[Cossim] L14 cos: 0.999994 max: 2.2472e+00 mean: 2.0513e-02
[Cossim] L21 cos: 0.999920 max: 1.3671e+01 mean: 8.9237e-02
[Cossim] L27 cos: 0.999724 max: 2.5151e+01 mean: 2.0175e-01
[Cossim] Final cos: 0.998553 max: 1.1352e+01 mean: 6.5864e-02
[Cossim] Logits cos: 0.999966 max: 2.0577e-01 mean: 3.0348e-02
[Cossim] Logits cos: 0.999966 max: 2.0576e-01 mean: 3.0347e-02
[Cossim] NextEmbStep0 cos: 0.967766 max: 6.3431e-02 mean: 1.2632e-02
[Cossim] TalkerHiddenStep1 cos: 0.999113 max: 4.3387e-01 mean: 9.2054e-02
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
@@ -170,5 +180,5 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkASP cos: 1.000000 max: 8.1062e-06 mean: 2.0734e-07
[Cossim] SpeakerEmb cos: 1.000000 max: 1.3351e-05 mean: 2.5133e-07
[Cossim] CodesFull exact: 5.26% (1008 values)
[Cossim] Audio cos: 0.028471
[Cossim] WAV stft_cos: 0.257819 samples: 120960
[Cossim] Audio cos: 0.025914
[Cossim] WAV stft_cos: 0.245369 samples: 120960
tests/clone-CPU-Q4_K_M.log
+67 -57
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp a62fde6 (2026-05-30)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CPU (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-Q4_K_M.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 1012.7 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 109.5 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-Q4_K_M.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 13.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=off clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=off clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 3 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922732 0.615171 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067754 0.119138 0.240900 0.344101
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 13.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 7 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.120292 13.742489 -0.348104 3.009387
[Debug] enc-transformer-out: [430, 512] first4: 0.867046 3.016766 0.059672 0.395116
@@ -63,8 +65,8 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Prompt] Built: 38 ids, N_text=30, N_instruct=0, T_ctx=225, hidden=2048, lang=english (id=2050), speaker=none (id=-1) ref_spk_emb=yes icl=yes
[Debug] prompt-ids: [38] first4: 151644.000000 77091.000000 198.000000 80.000000
[Debug] talker-input-embed: [225, 2048] first4: 0.021947 -0.009189 0.008398 -0.020109
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005388 0.008927 -0.004415 0.000679
[Debug] tts-pad-embed: [2048] first4: -0.005388 0.008927 -0.004415 0.000679
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005240 0.008931 -0.004543 0.000682
[Debug] tts-pad-embed: [2048] first4: -0.005240 0.008931 -0.004543 0.000682
[Debug] spk-emb: [2048] first4: 0.098343 -0.023770 -0.011812 0.006754
[Debug] ref-codes: [16, 215] first4: 1221.000000 2042.000000 696.000000 1522.000000
[Debug] talker-hidden-prefill-l0: [225, 2048] first4: -0.439808 -0.066686 0.001274 0.017172
@@ -73,42 +75,42 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -8.708588 1.644478 1.087476 -0.519538
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 4.580647 -33.465454 -6.676281 23.526917
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.155397 -1.185542 -0.210456 0.727513
[Debug] talker-logits-prefill: [3072] first4: -0.145868 -4.606725 -7.498469 -10.572260
[Debug] talker-logits-prefill: [3072] first4: -0.397744 -4.641821 -7.453447 -10.679588
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
[Sample-CP] g=2 c=417 u=-1.0000000000 subseq=3
[Sample-CP] g=3 c=442 u=-1.0000000000 subseq=4
[Sample-CP] g=4 c=1412 u=-1.0000000000 subseq=5
[Sample-CP] g=4 c=985 u=-1.0000000000 subseq=5
[Sample-CP] g=5 c=1155 u=-1.0000000000 subseq=6
[Sample-CP] g=6 c=815 u=-1.0000000000 subseq=7
[Sample-CP] g=7 c=1304 u=-1.0000000000 subseq=8
[Sample-CP] g=8 c=153 u=-1.0000000000 subseq=9
[Sample-CP] g=9 c=1776 u=-1.0000000000 subseq=10
[Sample-CP] g=10 c=105 u=-1.0000000000 subseq=11
[Sample-CP] g=11 c=587 u=-1.0000000000 subseq=12
[Sample-CP] g=12 c=872 u=-1.0000000000 subseq=13
[Sample-CP] g=13 c=1733 u=-1.0000000000 subseq=14
[Sample-CP] g=14 c=214 u=-1.0000000000 subseq=15
[Sample-CP] g=6 c=470 u=-1.0000000000 subseq=7
[Sample-CP] g=7 c=1311 u=-1.0000000000 subseq=8
[Sample-CP] g=8 c=1551 u=-1.0000000000 subseq=9
[Sample-CP] g=9 c=1952 u=-1.0000000000 subseq=10
[Sample-CP] g=10 c=1492 u=-1.0000000000 subseq=11
[Sample-CP] g=11 c=272 u=-1.0000000000 subseq=12
[Sample-CP] g=12 c=124 u=-1.0000000000 subseq=13
[Sample-CP] g=13 c=408 u=-1.0000000000 subseq=14
[Sample-CP] g=14 c=549 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 417.000000
[Debug] next-emb-step0: [2048] first4: 0.048200 -0.017350 -0.022171 0.064698
[Debug] talker-hidden-step1: [2048] first4: 0.675463 1.937895 0.812008 0.921976
[Debug] next-emb-step0: [2048] first4: 0.035528 0.008631 -0.031989 0.034729
[Debug] talker-hidden-step1: [2048] first4: 0.529886 2.483560 0.540937 0.745365
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
[Sample-CP] g=2 c=324 u=-1.0000000000 subseq=19
[Sample-CP] g=3 c=1210 u=-1.0000000000 subseq=20
[Sample-CP] g=4 c=985 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=1640 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=470 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=1058 u=-1.0000000000 subseq=24
[Sample-CP] g=8 c=103 u=-1.0000000000 subseq=25
[Sample-CP] g=9 c=414 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=248 u=-1.0000000000 subseq=27
[Sample-CP] g=3 c=405 u=-1.0000000000 subseq=20
[Sample-CP] g=4 c=390 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=447 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=1051 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=564 u=-1.0000000000 subseq=24
[Sample-CP] g=8 c=1004 u=-1.0000000000 subseq=25
[Sample-CP] g=9 c=1702 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=437 u=-1.0000000000 subseq=27
[Sample-CP] g=11 c=729 u=-1.0000000000 subseq=28
[Sample-CP] g=12 c=1529 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=1845 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=1435 u=-1.0000000000 subseq=31
[Sample-CP] g=12 c=142 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=328 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=1341 u=-1.0000000000 subseq=31
[Pipeline] Generated 8 frames
[Pipeline] Generated 16 frames
[Pipeline] Generated 24 frames
@@ -119,7 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 417.000000
[Debug] output-audio: [122880] first4: -0.001564 -0.002691 -0.002643 -0.001516
[Debug] output-audio: [122880] first4: -0.000160 -0.000284 -0.001051 -0.000894
[Perf] PromptBuild 2.8 ms
[Perf] Prefill 417.7 ms (T_ctx prefill)
[Perf] TTFA 428.0 ms (first frame codes)
[Perf] TalkerDecode 973.2 ms (64 frames, 15.21 ms/frame)
[Perf] CodePredictor 407.1 ms (6.36 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 991.3 ms
[Perf] Total 2793.5 ms (64 frames, 21.57 ms/frame AR, audio 5.12 s, RTF 0.546)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -140,18 +150,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[GGML] Audio: 122880 samples 24000 Hz 5.12s -> cpp/clone/clone-cpp.wav
[Cossim] PromptIDs exact: 100.00% (38 values)
[Cossim] RefCodes exact: 53.08% (3440 values)
[Cossim] Embed cos: 0.922425 max: 2.2247e-01 mean: 2.1114e-02
[Cossim] TrailingText cos: 0.999969 max: 1.4858e-03 mean: 2.2947e-04
[Cossim] TTSPadEmbed cos: 0.999969 max: 1.4858e-03 mean: 2.2947e-04
[Cossim] L0 cos: 0.963438 max: 9.8654e-01 mean: 2.7587e-02
[Cossim] L7 cos: 0.997432 max: 3.4743e+01 mean: 1.6242e-01
[Cossim] L14 cos: 0.997360 max: 3.4833e+01 mean: 2.4566e-01
[Cossim] L21 cos: 0.995879 max: 3.7708e+01 mean: 7.7289e-01
[Cossim] L27 cos: 0.941278 max: 2.2592e+03 mean: 2.0156e+00
[Cossim] Final cos: 0.953850 max: 6.1624e+01 mean: 5.5407e-01
[Cossim] Logits cos: 0.990367 max: 3.6698e+00 mean: 5.6017e-01
[Cossim] NextEmbStep0 cos: 0.828616 max: 1.3286e-01 mean: 2.9929e-02
[Cossim] TalkerHiddenStep1 cos: 0.960554 max: 3.6704e+00 mean: 6.1223e-01
[Cossim] Embed cos: 0.922423 max: 2.2237e-01 mean: 2.1115e-02
[Cossim] TrailingText cos: 0.999962 max: 1.8994e-03 mean: 2.5071e-04
[Cossim] TTSPadEmbed cos: 0.999962 max: 1.8994e-03 mean: 2.5071e-04
[Cossim] L0 cos: 0.963438 max: 1.1006e+00 mean: 2.7581e-02
[Cossim] L7 cos: 0.997432 max: 3.4743e+01 mean: 1.6241e-01
[Cossim] L14 cos: 0.997360 max: 3.4833e+01 mean: 2.4561e-01
[Cossim] L21 cos: 0.995871 max: 3.7418e+01 mean: 7.7468e-01
[Cossim] L27 cos: 0.944652 max: 2.1706e+03 mean: 2.0094e+00
[Cossim] Final cos: 0.954129 max: 6.0720e+01 mean: 5.5397e-01
[Cossim] Logits cos: 0.989332 max: 3.5730e+00 mean: 5.7743e-01
[Cossim] NextEmbStep0 cos: 0.835192 max: 1.4636e-01 mean: 2.9644e-02
[Cossim] TalkerHiddenStep1 cos: 0.955398 max: 4.9009e+00 mean: 6.4500e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -169,6 +179,6 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkMFA cos: 1.000000 max: 9.8418e-03 mean: 1.3212e-04
[Cossim] SpkASP cos: 1.000000 max: 2.6596e-04 mean: 2.3921e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 2.3174e-04 mean: 1.6094e-05
[Cossim] CodesFull exact: 1.09% (1008 values)
[Cossim] Audio cos: -0.003839
[Cossim] WAV stft_cos: 0.179597 samples: 120960
[Cossim] CodesFull exact: 1.29% (1008 values)
[Cossim] Audio cos: -0.019840
[Cossim] WAV stft_cos: 0.178108 samples: 120960
tests/clone-CPU-Q8_0.log
+66 -56
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp a62fde6 (2026-05-30)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CPU (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-Q8_0.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 1765.0 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 177.5 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-Q8_0.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 25.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=off clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=off clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 3 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922732 0.615171 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067754 0.119138 0.240900 0.344101
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 25.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 8 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.120292 13.742489 -0.348104 3.009387
[Debug] enc-transformer-out: [430, 512] first4: 0.716872 2.959070 -0.221450 0.852651
@@ -63,8 +65,8 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Prompt] Built: 38 ids, N_text=30, N_instruct=0, T_ctx=225, hidden=2048, lang=english (id=2050), speaker=none (id=-1) ref_spk_emb=yes icl=yes
[Debug] prompt-ids: [38] first4: 151644.000000 77091.000000 198.000000 80.000000
[Debug] talker-input-embed: [225, 2048] first4: 0.021991 -0.008996 0.008355 -0.020301
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005270 0.008716 -0.004355 0.001036
[Debug] tts-pad-embed: [2048] first4: -0.005270 0.008716 -0.004355 0.001036
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005262 0.008715 -0.004324 0.001049
[Debug] tts-pad-embed: [2048] first4: -0.005262 0.008715 -0.004324 0.001049
[Debug] spk-emb: [2048] first4: 0.098343 -0.023770 -0.011812 0.006754
[Debug] ref-codes: [16, 215] first4: 1221.000000 2042.000000 509.000000 1522.000000
[Debug] talker-hidden-prefill-l0: [225, 2048] first4: -0.467312 -0.069465 0.014126 -0.037120
@@ -73,42 +75,42 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -1.446977 0.302457 0.675166 0.977161
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 6.696979 -34.327644 -7.302014 28.627213
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.227865 -1.219680 -0.230862 0.887843
[Debug] talker-logits-prefill: [3072] first4: 1.353084 -7.812044 -6.845252 -10.473629
[Debug] talker-logits-prefill: [3072] first4: 1.309391 -7.886597 -6.855894 -10.555799
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
[Sample-CP] g=2 c=400 u=-1.0000000000 subseq=3
[Sample-CP] g=3 c=550 u=-1.0000000000 subseq=4
[Sample-CP] g=4 c=1506 u=-1.0000000000 subseq=5
[Sample-CP] g=5 c=1452 u=-1.0000000000 subseq=6
[Sample-CP] g=6 c=470 u=-1.0000000000 subseq=7
[Sample-CP] g=3 c=240 u=-1.0000000000 subseq=4
[Sample-CP] g=4 c=2010 u=-1.0000000000 subseq=5
[Sample-CP] g=5 c=196 u=-1.0000000000 subseq=6
[Sample-CP] g=6 c=580 u=-1.0000000000 subseq=7
[Sample-CP] g=7 c=1713 u=-1.0000000000 subseq=8
[Sample-CP] g=8 c=1221 u=-1.0000000000 subseq=9
[Sample-CP] g=9 c=18 u=-1.0000000000 subseq=10
[Sample-CP] g=10 c=1143 u=-1.0000000000 subseq=11
[Sample-CP] g=11 c=605 u=-1.0000000000 subseq=12
[Sample-CP] g=12 c=40 u=-1.0000000000 subseq=13
[Sample-CP] g=13 c=1730 u=-1.0000000000 subseq=14
[Sample-CP] g=14 c=552 u=-1.0000000000 subseq=15
[Sample-CP] g=9 c=1796 u=-1.0000000000 subseq=10
[Sample-CP] g=10 c=1228 u=-1.0000000000 subseq=11
[Sample-CP] g=11 c=2015 u=-1.0000000000 subseq=12
[Sample-CP] g=12 c=143 u=-1.0000000000 subseq=13
[Sample-CP] g=13 c=1062 u=-1.0000000000 subseq=14
[Sample-CP] g=14 c=1094 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: 0.037031 -0.042681 0.009044 0.073258
[Debug] talker-hidden-step1: [2048] first4: 0.714530 2.568675 0.176894 0.269541
[Debug] next-emb-step0: [2048] first4: 0.010612 -0.008921 0.074940 0.036301
[Debug] talker-hidden-step1: [2048] first4: 0.118400 2.726040 -0.843629 0.103909
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
[Sample-CP] g=2 c=324 u=-1.0000000000 subseq=19
[Sample-CP] g=3 c=1501 u=-1.0000000000 subseq=20
[Sample-CP] g=4 c=1748 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=20 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=1544 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=1058 u=-1.0000000000 subseq=24
[Sample-CP] g=3 c=350 u=-1.0000000000 subseq=20
[Sample-CP] g=4 c=985 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=853 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=329 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=564 u=-1.0000000000 subseq=24
[Sample-CP] g=8 c=692 u=-1.0000000000 subseq=25
[Sample-CP] g=9 c=223 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=1603 u=-1.0000000000 subseq=27
[Sample-CP] g=9 c=1702 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=437 u=-1.0000000000 subseq=27
[Sample-CP] g=11 c=729 u=-1.0000000000 subseq=28
[Sample-CP] g=12 c=781 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=812 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=803 u=-1.0000000000 subseq=31
[Sample-CP] g=12 c=344 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=85 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=1668 u=-1.0000000000 subseq=31
[Pipeline] Generated 8 frames
[Pipeline] Generated 16 frames
[Pipeline] Generated 24 frames
@@ -119,7 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001637 -0.002874 -0.002587 -0.001074
[Debug] output-audio: [122880] first4: 0.000527 0.000324 -0.000983 -0.000676
[Perf] PromptBuild 2.8 ms
[Perf] Prefill 619.0 ms (T_ctx prefill)
[Perf] TTFA 634.3 ms (first frame codes)
[Perf] TalkerDecode 1668.8 ms (64 frames, 26.07 ms/frame)
[Perf] CodePredictor 651.2 ms (10.17 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 987.3 ms
[Perf] Total 3930.5 ms (64 frames, 36.25 ms/frame AR, audio 5.12 s, RTF 0.768)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -140,18 +150,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[GGML] Audio: 122880 samples 24000 Hz 5.12s -> cpp/clone/clone-cpp.wav
[Cossim] PromptIDs exact: 100.00% (38 values)
[Cossim] RefCodes exact: 90.90% (3440 values)
[Cossim] Embed cos: 0.984762 max: 1.6364e-01 mean: 5.1405e-03
[Cossim] TrailingText cos: 0.999999 max: 2.1233e-04 mean: 4.3863e-05
[Cossim] TTSPadEmbed cos: 0.999999 max: 2.1233e-04 mean: 4.3863e-05
[Cossim] L0 cos: 0.992178 max: 7.7259e-01 mean: 7.3131e-03
[Cossim] L7 cos: 0.999982 max: 5.4102e+00 mean: 2.6651e-02
[Cossim] L14 cos: 0.999969 max: 5.4209e+00 mean: 5.2390e-02
[Cossim] L21 cos: 0.999704 max: 1.7729e+01 mean: 2.0345e-01
[Cossim] L27 cos: 0.998595 max: 1.1723e+02 mean: 4.9341e-01
[Cossim] Final cos: 0.994075 max: 1.6792e+01 mean: 1.5252e-01
[Cossim] Logits cos: 0.999832 max: 7.0396e-01 mean: 6.9169e-02
[Cossim] NextEmbStep0 cos: 0.883720 max: 1.3778e-01 mean: 2.4555e-02
[Cossim] TalkerHiddenStep1 cos: 0.995519 max: 9.0677e-01 mean: 2.0596e-01
[Cossim] Embed cos: 0.984762 max: 1.6364e-01 mean: 5.1406e-03
[Cossim] TrailingText cos: 0.999999 max: 1.9968e-04 mean: 4.5182e-05
[Cossim] TTSPadEmbed cos: 0.999999 max: 1.9968e-04 mean: 4.5182e-05
[Cossim] L0 cos: 0.992177 max: 7.7295e-01 mean: 7.3151e-03
[Cossim] L7 cos: 0.999982 max: 5.4102e+00 mean: 2.6624e-02
[Cossim] L14 cos: 0.999969 max: 5.4209e+00 mean: 5.2229e-02
[Cossim] L21 cos: 0.999712 max: 1.7887e+01 mean: 1.9992e-01
[Cossim] L27 cos: 0.998655 max: 1.1809e+02 mean: 4.8277e-01
[Cossim] Final cos: 0.994339 max: 1.6804e+01 mean: 1.4894e-01
[Cossim] Logits cos: 0.999855 max: 6.7472e-01 mean: 6.7599e-02
[Cossim] NextEmbStep0 cos: 0.829035 max: 1.5489e-01 mean: 2.9164e-02
[Cossim] TalkerHiddenStep1 cos: 0.993414 max: 1.8271e+00 mean: 2.4874e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -169,6 +179,6 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkMFA cos: 1.000000 max: 9.8418e-03 mean: 1.3212e-04
[Cossim] SpkASP cos: 1.000000 max: 2.6596e-04 mean: 2.3921e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 2.3174e-04 mean: 1.6094e-05
[Cossim] CodesFull exact: 1.59% (1008 values)
[Cossim] Audio cos: 0.006366
[Cossim] WAV stft_cos: 0.197239 samples: 120960
[Cossim] CodesFull exact: 2.68% (1008 values)
[Cossim] Audio cos: -0.002547
[Cossim] WAV stft_cos: 0.174666 samples: 120960
tests/clone-CUDA0-BF16.log
+37 -34
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CUDA0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-BF16.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 3322.0 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 334.1 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-BF16.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 48.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 22 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.919922 0.609741 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067580 0.118872 0.240696 0.344030
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 48.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 91 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.111631 13.770183 -0.302467 2.984188
[Debug] enc-transformer-out: [430, 512] first4: 0.531195 3.238480 -0.146394 0.612285
@@ -92,8 +94,9 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=1094 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: 0.010331 -0.009322 0.074922 0.036956
[Debug] talker-hidden-step1: [2048] first4: 0.016130 2.930190 -0.767163 0.054849
[Debug] talker-hidden-step1: [2048] first4: 0.017435 2.931466 -0.766813 0.054457
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
[Sample-CP] g=2 c=324 u=-1.0000000000 subseq=19
@@ -119,15 +122,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001310 -0.002270 -0.002123 -0.000994
[Debug] output-audio: [122880] first4: 0.000465 0.000259 -0.001065 -0.000736
[Perf] PromptBuild 1.2 ms
[Perf] Prefill 13.9 ms (T_ctx prefill)
[Perf] TTFA 25.1 ms (first frame codes)
[Perf] TalkerDecode 253.0 ms (64 frames, 3.95 ms/frame)
[Perf] CodePredictor 462.4 ms (7.23 ms/frame)
[Perf] HostCompose 1.3 ms (c0 sample + next emb)
[Perf] CodecDecode 10.1 ms
[Perf] Total 742.8 ms (64 frames, 11.20 ms/frame AR, audio 5.12 s, RTF 0.145)
[Perf] Prefill 12.8 ms (T_ctx prefill)
[Perf] TTFA 22.0 ms (first frame codes)
[Perf] TalkerDecode 203.8 ms (64 frames, 3.18 ms/frame)
[Perf] CodePredictor 331.0 ms (5.17 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 14.7 ms
[Perf] Total 564.5 ms (64 frames, 8.36 ms/frame AR, audio 5.12 s, RTF 0.110)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -152,14 +155,14 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] TrailingText cos: 1.000000 max: 3.5107e-05 mean: 2.5004e-06
[Cossim] TTSPadEmbed cos: 1.000000 max: 3.5107e-05 mean: 2.5004e-06
[Cossim] L0 cos: 0.984329 max: 8.0274e-01 mean: 1.3008e-02
[Cossim] L7 cos: 0.999981 max: 4.7523e+01 mean: 3.1978e-02
[Cossim] L14 cos: 0.999956 max: 4.7531e+01 mean: 7.0164e-02
[Cossim] L21 cos: 0.999567 max: 4.7566e+01 mean: 2.6764e-01
[Cossim] L27 cos: 0.998445 max: 4.7993e+01 mean: 6.0528e-01
[Cossim] Final cos: 0.991977 max: 1.5199e+01 mean: 1.9679e-01
[Cossim] Logits cos: 0.999456 max: 1.0356e+00 mean: 1.3008e-01
[Cossim] L7 cos: 0.999981 max: 4.7525e+01 mean: 3.1978e-02
[Cossim] L14 cos: 0.999956 max: 4.7533e+01 mean: 7.0164e-02
[Cossim] L21 cos: 0.999567 max: 4.7568e+01 mean: 2.6764e-01
[Cossim] L27 cos: 0.998445 max: 4.7995e+01 mean: 6.0528e-01
[Cossim] Final cos: 0.991977 max: 1.5200e+01 mean: 1.9679e-01
[Cossim] Logits cos: 0.999456 max: 1.0356e+00 mean: 1.3007e-01
[Cossim] NextEmbStep0 cos: 0.829094 max: 1.5468e-01 mean: 2.9154e-02
[Cossim] TalkerHiddenStep1 cos: 0.992003 max: 2.2537e+00 mean: 2.7399e-01
[Cossim] TalkerHiddenStep1 cos: 0.992002 max: 2.2514e+00 mean: 2.7397e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -177,6 +180,6 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkMFA cos: 0.999973 max: 7.5789e-02 mean: 1.1867e-03
[Cossim] SpkASP cos: 0.999999 max: 3.9836e-03 mean: 2.4063e-04
[Cossim] SpeakerEmb cos: 1.000000 max: 1.7091e-03 mean: 2.1396e-04
[Cossim] CodesFull exact: 2.38% (1008 values)
[Cossim] Audio cos: 0.008950
[Cossim] WAV stft_cos: 0.294906 samples: 120960
[Cossim] CodesFull exact: 5.85% (1008 values)
[Cossim] Audio cos: 0.030355
[Cossim] WAV stft_cos: 0.297198 samples: 120960
tests/clone-CUDA0-F32.log
+35 -32
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CUDA0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-F32.gguf: 480 tensors, data at offset 5954432
[WeightCtx] Loaded 316 tensors, 6643.5 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 668.1 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 45.8 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-F32.gguf: 398 tensors, data at offset 29664
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 96.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 45.8 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 4 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922734 0.614915 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067723 0.119136 0.240880 0.344137
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 96.1 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 49 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.134336 13.731120 -0.351051 3.013485
[Debug] enc-transformer-out: [430, 512] first4: 0.525800 3.116711 -0.187038 0.631811
@@ -92,8 +94,9 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=541 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: 0.003672 0.000908 0.039506 0.094594
[Debug] talker-hidden-step1: [2048] first4: 0.754330 2.408696 -0.056724 0.159879
[Debug] talker-hidden-step1: [2048] first4: 0.754625 2.407892 -0.056593 0.158147
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
[Sample-CP] g=2 c=324 u=-1.0000000000 subseq=19
@@ -119,15 +122,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001265 -0.002192 -0.002047 -0.000956
[Perf] PromptBuild 1.2 ms
[Perf] Prefill 28.1 ms (T_ctx prefill)
[Perf] TTFA 48.6 ms (first frame codes)
[Perf] TalkerDecode 369.5 ms (64 frames, 5.77 ms/frame)
[Perf] CodePredictor 559.6 ms (8.74 ms/frame)
[Perf] HostCompose 1.3 ms (c0 sample + next emb)
[Perf] CodecDecode 10.0 ms
[Perf] Total 979.4 ms (64 frames, 14.54 ms/frame AR, audio 5.12 s, RTF 0.191)
[Debug] output-audio: [122880] first4: 0.000739 0.000593 -0.000747 -0.000435
[Perf] PromptBuild 1.4 ms
[Perf] Prefill 27.2 ms (T_ctx prefill)
[Perf] TTFA 37.0 ms (first frame codes)
[Perf] TalkerDecode 311.1 ms (64 frames, 4.86 ms/frame)
[Perf] CodePredictor 419.2 ms (6.55 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 14.6 ms
[Perf] Total 774.6 ms (64 frames, 11.41 ms/frame AR, audio 5.12 s, RTF 0.151)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -154,12 +157,12 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] L0 cos: 0.986912 max: 6.6445e-01 mean: 1.1103e-02
[Cossim] L7 cos: 0.999992 max: 1.2570e+00 mean: 2.7510e-02
[Cossim] L14 cos: 0.999973 max: 2.1782e+00 mean: 5.9937e-02
[Cossim] L21 cos: 0.999719 max: 8.3604e+00 mean: 2.2268e-01
[Cossim] L21 cos: 0.999719 max: 8.3606e+00 mean: 2.2268e-01
[Cossim] L27 cos: 0.998971 max: 2.3989e+01 mean: 5.1174e-01
[Cossim] Final cos: 0.994704 max: 7.3139e+00 mean: 1.6649e-01
[Cossim] Logits cos: 0.999743 max: 5.7111e-01 mean: 8.8930e-02
[Cossim] Final cos: 0.994704 max: 7.3140e+00 mean: 1.6649e-01
[Cossim] Logits cos: 0.999743 max: 5.7110e-01 mean: 8.8929e-02
[Cossim] NextEmbStep0 cos: 0.829216 max: 1.2523e-01 mean: 2.9129e-02
[Cossim] TalkerHiddenStep1 cos: 0.993859 max: 1.4381e+00 mean: 2.4190e-01
[Cossim] TalkerHiddenStep1 cos: 0.993854 max: 1.4329e+00 mean: 2.4202e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -170,7 +173,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SeanetStage1 cos: 0.999865 max: 7.6868e-01 mean: 4.9571e-03
[Cossim] SeanetStage3 cos: 0.999995 max: 1.6565e+00 mean: 3.9086e-02
[Cossim] SeanetOut cos: 0.999963 max: 4.1938e-01 mean: 1.7123e-02
[Cossim] EncTransformer cos: 0.999975 max: 3.4209e-01 mean: 7.3238e-03
[Cossim] EncTransformer cos: 0.999975 max: 3.4209e-01 mean: 7.3239e-03
[Cossim] CodecPreFSQ cos: 0.999972 max: 4.4734e-01 mean: 1.8065e-02
[Cossim] SpkFrontend cos: 1.000000 max: 2.1511e-04 mean: 1.4495e-06
[Cossim] SpkBlock3 cos: 1.000000 max: 3.2926e-04 mean: 4.3974e-06
@@ -178,5 +181,5 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkASP cos: 1.000000 max: 7.9870e-06 mean: 2.2269e-07
[Cossim] SpeakerEmb cos: 1.000000 max: 1.4305e-05 mean: 2.5463e-07
[Cossim] CodesFull exact: 3.27% (1008 values)
[Cossim] Audio cos: 0.053931
[Cossim] WAV stft_cos: 0.375521 samples: 120960
[Cossim] Audio cos: 0.083313
[Cossim] WAV stft_cos: 0.370796 samples: 120960
tests/clone-CUDA0-Q4_K_M.log
+82 -37
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CUDA0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-Q4_K_M.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 1012.7 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 109.5 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-Q4_K_M.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 13.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 27 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922734 0.614915 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067723 0.119136 0.240880 0.344137
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 13.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 44 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.134336 13.731120 -0.351051 3.013485
[Debug] enc-transformer-out: [430, 512] first4: 1.140990 3.713432 0.131578 0.616251
@@ -92,8 +94,9 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=735 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1042.000000 390.000000 881.000000
[Debug] next-emb-step0: [2048] first4: -0.033968 -0.020822 -0.037756 -0.013151
[Debug] talker-hidden-step1: [2048] first4: 1.290885 1.802708 -1.078711 1.471289
[Debug] talker-hidden-step1: [2048] first4: 1.225613 1.784957 -1.039321 1.449351
[Sample] step=1 c0=99 u=-1.0000000000 subseq=16
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Sample-CP] g=0 c=43 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=1859 u=-1.0000000000 subseq=18
[Sample-CP] g=2 c=22 u=-1.0000000000 subseq=19
@@ -101,33 +104,75 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=4 c=915 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=2019 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=329 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=996 u=-1.0000000000 subseq=24
[Sample-CP] g=8 c=1673 u=-1.0000000000 subseq=25
[Sample-CP] g=9 c=1652 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=1615 u=-1.0000000000 subseq=27
[Sample-CP] g=11 c=914 u=-1.0000000000 subseq=28
[Sample-CP] g=12 c=344 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=85 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=1435 u=-1.0000000000 subseq=31
[Sample-CP] g=7 c=241 u=-1.0000000000 subseq=24
[Sample-CP] g=8 c=1004 u=-1.0000000000 subseq=25
[Sample-CP] g=9 c=826 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=1603 u=-1.0000000000 subseq=27
[Sample-CP] g=11 c=1160 u=-1.0000000000 subseq=28
[Sample-CP] g=12 c=1759 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=591 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=901 u=-1.0000000000 subseq=31
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Pipeline] Generated 8 frames
[Pipeline] Generated 16 frames
[Pipeline] Generated 24 frames
[Pipeline] Generated 32 frames
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup reset
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Pipeline] Generated 40 frames
[Pipeline] Generated 48 frames
[Pipeline] Generated 56 frames
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1042.000000 390.000000 881.000000
[Debug] output-audio: [122880] first4: -0.001003 -0.001738 -0.001707 -0.000971
[Perf] PromptBuild 56.7 ms
[Perf] Prefill 15.2 ms (T_ctx prefill)
[Perf] TTFA 81.0 ms (first frame codes)
[Perf] TalkerDecode 199.4 ms (64 frames, 3.12 ms/frame)
[Perf] CodePredictor 450.5 ms (7.04 ms/frame)
[Perf] HostCompose 2.6 ms (c0 sample + next emb)
[Perf] CodecDecode 11.0 ms
[Perf] Total 736.4 ms (64 frames, 10.20 ms/frame AR, audio 5.12 s, RTF 0.144)
[Debug] output-audio: [122880] first4: 0.000087 0.000212 -0.000383 -0.000346
[Perf] PromptBuild 55.2 ms
[Perf] Prefill 14.2 ms (T_ctx prefill)
[Perf] TTFA 77.1 ms (first frame codes)
[Perf] TalkerDecode 231.8 ms (64 frames, 3.62 ms/frame)
[Perf] CodePredictor 358.2 ms (5.60 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 15.8 ms
[Perf] Total 676.2 ms (64 frames, 9.22 ms/frame AR, audio 5.12 s, RTF 0.132)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -159,7 +204,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] Final cos: 0.952287 max: 2.7175e+01 mean: 5.6964e-01
[Cossim] Logits cos: 0.991304 max: 3.3342e+00 mean: 4.9692e-01
[Cossim] NextEmbStep0 cos: 0.709511 max: 1.6319e-01 mean: 3.8936e-02
[Cossim] TalkerHiddenStep1 cos: 0.836398 max: 1.1069e+01 mean: 1.2096e+00
[Cossim] TalkerHiddenStep1 cos: 0.836093 max: 1.1068e+01 mean: 1.2121e+00
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -177,6 +222,6 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkMFA cos: 1.000000 max: 3.8069e-04 mean: 1.3207e-06
[Cossim] SpkASP cos: 1.000000 max: 7.8678e-06 mean: 2.2438e-07
[Cossim] SpeakerEmb cos: 1.000000 max: 1.4305e-05 mean: 2.5522e-07
[Cossim] CodesFull exact: 0.60% (1008 values)
[Cossim] Audio cos: 0.001172
[Cossim] WAV stft_cos: 0.129338 samples: 120960
[Cossim] CodesFull exact: 0.30% (1008 values)
[Cossim] Audio cos: -0.010168
[Cossim] WAV stft_cos: 0.170792 samples: 120960
tests/clone-CUDA0-Q8_0.log
+42 -39
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: CUDA0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-Q8_0.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 1765.0 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 177.5 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-Q8_0.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,23 +22,18 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 25.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 3 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
@@ -50,6 +43,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922734 0.614915 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067723 0.119136 0.240880 0.344137
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 25.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Codec encoder lazy loaded in 8 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.134336 13.731120 -0.351051 3.013485
[Debug] enc-transformer-out: [430, 512] first4: 0.628775 2.909739 -0.209228 0.726215
@@ -92,23 +94,24 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=1681 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: -0.034936 -0.024575 0.029866 0.045185
[Debug] talker-hidden-step1: [2048] first4: 0.923327 2.394357 0.442331 -0.226438
[Debug] talker-hidden-step1: [2048] first4: 0.931840 2.357023 0.427709 -0.264416
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
[Sample-CP] g=2 c=324 u=-1.0000000000 subseq=19
[Sample-CP] g=3 c=350 u=-1.0000000000 subseq=20
[Sample-CP] g=4 c=985 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=853 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=329 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=564 u=-1.0000000000 subseq=24
[Sample-CP] g=3 c=1501 u=-1.0000000000 subseq=20
[Sample-CP] g=4 c=407 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=1452 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=1544 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=1058 u=-1.0000000000 subseq=24
[Sample-CP] g=8 c=692 u=-1.0000000000 subseq=25
[Sample-CP] g=9 c=1702 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=437 u=-1.0000000000 subseq=27
[Sample-CP] g=10 c=362 u=-1.0000000000 subseq=27
[Sample-CP] g=11 c=729 u=-1.0000000000 subseq=28
[Sample-CP] g=12 c=344 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=85 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=1668 u=-1.0000000000 subseq=31
[Sample-CP] g=12 c=781 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=82 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=1435 u=-1.0000000000 subseq=31
[Pipeline] Generated 8 frames
[Pipeline] Generated 16 frames
[Pipeline] Generated 24 frames
@@ -119,15 +122,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001528 -0.002728 -0.002590 -0.001367
[Debug] output-audio: [122880] first4: 0.000596 0.000259 -0.001224 -0.000817
[Perf] PromptBuild 1.4 ms
[Perf] Prefill 15.4 ms (T_ctx prefill)
[Perf] TTFA 25.7 ms (first frame codes)
[Perf] TalkerDecode 226.3 ms (64 frames, 3.54 ms/frame)
[Perf] CodePredictor 449.9 ms (7.03 ms/frame)
[Perf] HostCompose 1.2 ms (c0 sample + next emb)
[Perf] CodecDecode 10.4 ms
[Perf] Total 705.6 ms (64 frames, 10.59 ms/frame AR, audio 5.12 s, RTF 0.138)
[Perf] Prefill 14.6 ms (T_ctx prefill)
[Perf] TTFA 22.3 ms (first frame codes)
[Perf] TalkerDecode 166.6 ms (64 frames, 2.60 ms/frame)
[Perf] CodePredictor 277.8 ms (4.34 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 15.0 ms
[Perf] Total 476.5 ms (64 frames, 6.95 ms/frame AR, audio 5.12 s, RTF 0.093)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -152,14 +155,14 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] TrailingText cos: 0.999999 max: 1.9968e-04 mean: 4.5182e-05
[Cossim] TTSPadEmbed cos: 0.999999 max: 1.9968e-04 mean: 4.5182e-05
[Cossim] L0 cos: 0.984964 max: 7.6776e-01 mean: 1.2815e-02
[Cossim] L7 cos: 0.999978 max: 3.6948e+00 mean: 3.8281e-02
[Cossim] L14 cos: 0.999955 max: 3.7148e+00 mean: 7.4563e-02
[Cossim] L7 cos: 0.999978 max: 3.6929e+00 mean: 3.8281e-02
[Cossim] L14 cos: 0.999955 max: 3.7129e+00 mean: 7.4563e-02
[Cossim] L21 cos: 0.999568 max: 1.7895e+01 mean: 2.7422e-01
[Cossim] L27 cos: 0.998158 max: 9.9299e+01 mean: 6.4856e-01
[Cossim] Final cos: 0.991511 max: 1.5600e+01 mean: 2.0381e-01
[Cossim] Logits cos: 0.998602 max: 1.4375e+00 mean: 2.2099e-01
[Cossim] NextEmbStep0 cos: 0.839291 max: 1.9986e-01 mean: 2.8325e-02
[Cossim] TalkerHiddenStep1 cos: 0.988487 max: 2.5975e+00 mean: 3.2991e-01
[Cossim] TalkerHiddenStep1 cos: 0.988427 max: 2.6641e+00 mean: 3.3037e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -177,6 +180,6 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SpkMFA cos: 1.000000 max: 3.8069e-04 mean: 1.3207e-06
[Cossim] SpkASP cos: 1.000000 max: 7.8678e-06 mean: 2.2438e-07
[Cossim] SpeakerEmb cos: 1.000000 max: 1.4305e-05 mean: 2.5522e-07
[Cossim] CodesFull exact: 2.78% (1008 values)
[Cossim] Audio cos: 0.034690
[Cossim] WAV stft_cos: 0.243983 samples: 120960
[Cossim] CodesFull exact: 2.28% (1008 values)
[Cossim] Audio cos: -0.001204
[Cossim] WAV stft_cos: 0.184588 samples: 120960
tests/clone-Vulkan0-BF16.log
+54 -52
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: Vulkan0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-BF16.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 3322.0 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 334.1 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-BF16.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,6 +22,27 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 4 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.099754 0.858150 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.962617 0.099754 1.239126 0.246363
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.923592 0.609606 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067514 0.118921 0.240827 0.344086
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 48.1 MB into backend
@@ -32,24 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.099754 0.858150 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.962623 0.099754 1.239126 0.246363
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.923592 0.609606 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067512 0.118912 0.240832 0.344075
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[Pipeline] Codec encoder lazy loaded in 44 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.110593 13.765666 -0.317055 2.981106
[Debug] enc-transformer-out: [430, 512] first4: 0.536260 3.195749 -0.156796 0.610111
@@ -65,7 +67,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-input-embed: [225, 2048] first4: 0.022157 -0.009049 0.008161 -0.020214
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005331 0.008751 -0.004225 0.001141
[Debug] tts-pad-embed: [2048] first4: -0.005331 0.008751 -0.004225 0.001141
[Debug] spk-emb: [2048] first4: 0.098348 -0.023729 -0.011668 0.006747
[Debug] spk-emb: [2048] first4: 0.098352 -0.023723 -0.011675 0.006733
[Debug] ref-codes: [16, 215] first4: 1221.000000 2042.000000 509.000000 1522.000000
[Debug] talker-hidden-prefill-l0: [225, 2048] first4: -0.461455 -0.069813 0.016136 -0.039145
[Debug] talker-hidden-prefill-l7: [225, 2048] first4: -1.522913 -0.014861 0.347484 1.775680
@@ -73,7 +75,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -0.769277 0.163230 0.799681 1.111220
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 7.051195 -34.399265 -7.070650 28.631254
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.239991 -1.222603 -0.223616 0.888243
[Debug] talker-logits-prefill: [3072] first4: 0.907549 -7.873562 -6.795165 -10.447464
[Debug] talker-logits-prefill: [3072] first4: 0.882721 -7.881639 -6.796915 -10.414125
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
@@ -92,7 +94,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=348 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: -0.008599 -0.015697 0.093591 0.086640
[Debug] talker-hidden-step1: [2048] first4: 0.463387 2.304747 -0.152458 0.223634
[Debug] talker-hidden-step1: [2048] first4: 0.466706 2.304709 -0.154099 0.228768
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
@@ -119,15 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001651 -0.002839 -0.002671 -0.001491
[Perf] PromptBuild 2.0 ms
[Perf] Prefill 32.0 ms (T_ctx prefill)
[Perf] TTFA 42.0 ms (first frame codes)
[Perf] TalkerDecode 230.8 ms (64 frames, 3.61 ms/frame)
[Perf] CodePredictor 425.1 ms (6.64 ms/frame)
[Perf] HostCompose 1.3 ms (c0 sample + next emb)
[Perf] CodecDecode 21.7 ms
[Perf] Total 713.9 ms (64 frames, 10.27 ms/frame AR, audio 5.12 s, RTF 0.139)
[Debug] output-audio: [122880] first4: 0.000515 0.000357 -0.000915 -0.000575
[Perf] PromptBuild 1.7 ms
[Perf] Prefill 31.0 ms (T_ctx prefill)
[Perf] TTFA 623.5 ms (first frame codes)
[Perf] TalkerDecode 247.6 ms (64 frames, 3.87 ms/frame)
[Perf] CodePredictor 986.5 ms (15.41 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 30.9 ms
[Perf] Total 1299.0 ms (64 frames, 19.29 ms/frame AR, audio 5.12 s, RTF 0.254)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -149,17 +151,17 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] PromptIDs exact: 100.00% (38 values)
[Cossim] RefCodes exact: 76.83% (3440 values)
[Cossim] Embed cos: 0.962709 max: 1.9437e-01 mean: 1.0689e-02
[Cossim] TrailingText cos: 1.000000 max: 5.9605e-08 mean: 5.6559e-10
[Cossim] TTSPadEmbed cos: 1.000000 max: 5.9605e-08 mean: 5.6559e-10
[Cossim] L0 cos: 0.983221 max: 7.7033e-01 mean: 1.3992e-02
[Cossim] L7 cos: 0.999989 max: 1.2872e+01 mean: 3.3232e-02
[Cossim] L14 cos: 0.999964 max: 1.2875e+01 mean: 7.1831e-02
[Cossim] L21 cos: 0.999532 max: 1.6953e+01 mean: 2.8047e-01
[Cossim] L27 cos: 0.998216 max: 4.0677e+01 mean: 6.4098e-01
[Cossim] Final cos: 0.990697 max: 1.5845e+01 mean: 2.0881e-01
[Cossim] Logits cos: 0.999304 max: 1.0959e+00 mean: 1.3300e-01
[Cossim] TrailingText cos: 1.000000 max: 2.9802e-08 mean: 5.4933e-10
[Cossim] TTSPadEmbed cos: 1.000000 max: 2.9802e-08 mean: 5.4933e-10
[Cossim] L0 cos: 0.983221 max: 7.7030e-01 mean: 1.3992e-02
[Cossim] L7 cos: 0.999989 max: 1.2885e+01 mean: 3.3235e-02
[Cossim] L14 cos: 0.999964 max: 1.2887e+01 mean: 7.1843e-02
[Cossim] L21 cos: 0.999531 max: 1.6955e+01 mean: 2.8087e-01
[Cossim] L27 cos: 0.998209 max: 4.0700e+01 mean: 6.4207e-01
[Cossim] Final cos: 0.990658 max: 1.5838e+01 mean: 2.0917e-01
[Cossim] Logits cos: 0.999301 max: 1.1222e+00 mean: 1.3438e-01
[Cossim] NextEmbStep0 cos: 0.818255 max: 1.4225e-01 mean: 3.0455e-02
[Cossim] TalkerHiddenStep1 cos: 0.993909 max: 1.7548e+00 mean: 2.3409e-01
[Cossim] TalkerHiddenStep1 cos: 0.993880 max: 1.7297e+00 mean: 2.3463e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -173,10 +175,10 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] EncTransformer cos: 0.999963 max: 8.5211e-01 mean: 1.0424e-02
[Cossim] CodecPreFSQ cos: 0.999962 max: 4.6668e-01 mean: 2.2088e-02
[Cossim] SpkFrontend cos: 0.999997 max: 2.5180e-02 mean: 9.4975e-04
[Cossim] SpkBlock3 cos: 0.999991 max: 1.0191e-01 mean: 4.1523e-03
[Cossim] SpkMFA cos: 0.999979 max: 7.0797e-02 mean: 1.0370e-03
[Cossim] SpkASP cos: 1.000000 max: 2.3381e-03 mean: 1.5046e-04
[Cossim] SpeakerEmb cos: 1.000000 max: 6.5649e-04 mean: 1.2850e-04
[Cossim] SpkBlock3 cos: 0.999991 max: 1.0191e-01 mean: 4.1519e-03
[Cossim] SpkMFA cos: 0.999979 max: 7.0797e-02 mean: 1.0369e-03
[Cossim] SpkASP cos: 1.000000 max: 2.3533e-03 mean: 1.4947e-04
[Cossim] SpeakerEmb cos: 1.000000 max: 6.7258e-04 mean: 1.2809e-04
[Cossim] CodesFull exact: 2.28% (1008 values)
[Cossim] Audio cos: 0.002591
[Cossim] WAV stft_cos: 0.228070 samples: 120960
[Cossim] Audio cos: 0.003079
[Cossim] WAV stft_cos: 0.208123 samples: 120960
tests/clone-Vulkan0-F32.log
+69 -67
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: Vulkan0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-F32.gguf: 480 tensors, data at offset 5954432
[WeightCtx] Loaded 316 tensors, 6643.5 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 668.1 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 45.8 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-F32.gguf: 398 tensors, data at offset 29664
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,6 +22,27 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 45.8 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 6 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.098113 0.854719 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.956474 0.098113 1.233186 0.246316
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922765 0.615104 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067725 0.119143 0.240895 0.344143
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 96.1 MB into backend
@@ -32,24 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.098113 0.854719 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.956472 0.098113 1.233186 0.246316
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922765 0.615104 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067725 0.119141 0.240898 0.344141
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[Pipeline] Codec encoder lazy loaded in 34 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.138303 13.699992 -0.351357 3.025051
[Debug] enc-transformer-out: [430, 512] first4: 0.526569 3.116264 -0.195330 0.625080
@@ -65,50 +67,50 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-input-embed: [225, 2048] first4: 0.022157 -0.009049 0.008161 -0.020214
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005331 0.008751 -0.004225 0.001141
[Debug] tts-pad-embed: [2048] first4: -0.005331 0.008751 -0.004225 0.001141
[Debug] spk-emb: [2048] first4: 0.098351 -0.023771 -0.011823 0.006760
[Debug] spk-emb: [2048] first4: 0.098352 -0.023771 -0.011820 0.006759
[Debug] ref-codes: [16, 215] first4: 1221.000000 2042.000000 509.000000 1522.000000
[Debug] talker-hidden-prefill-l0: [225, 2048] first4: -0.458786 -0.069777 0.016191 -0.038746
[Debug] talker-hidden-prefill-l7: [225, 2048] first4: -1.521495 -0.008156 0.335267 1.773009
[Debug] talker-hidden-prefill-l14: [225, 2048] first4: -1.292948 0.188377 0.436406 1.739375
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -0.768596 0.168659 0.786884 1.106115
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 7.115746 -34.442368 -7.106302 28.621552
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.242316 -1.224780 -0.224862 0.888410
[Debug] talker-logits-prefill: [3072] first4: 1.007812 -7.740234 -6.858398 -10.242188
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 7.115197 -34.442551 -7.104288 28.617508
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.242297 -1.224786 -0.224798 0.888284
[Debug] talker-logits-prefill: [3072] first4: 1.013672 -7.734375 -6.857910 -10.250000
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
[Sample-CP] g=2 c=400 u=-1.0000000000 subseq=3
[Sample-CP] g=3 c=550 u=-1.0000000000 subseq=4
[Sample-CP] g=4 c=1506 u=-1.0000000000 subseq=5
[Sample-CP] g=5 c=1452 u=-1.0000000000 subseq=6
[Sample-CP] g=5 c=915 u=-1.0000000000 subseq=6
[Sample-CP] g=6 c=470 u=-1.0000000000 subseq=7
[Sample-CP] g=7 c=1713 u=-1.0000000000 subseq=8
[Sample-CP] g=8 c=1221 u=-1.0000000000 subseq=9
[Sample-CP] g=9 c=18 u=-1.0000000000 subseq=10
[Sample-CP] g=10 c=1143 u=-1.0000000000 subseq=11
[Sample-CP] g=11 c=605 u=-1.0000000000 subseq=12
[Sample-CP] g=12 c=40 u=-1.0000000000 subseq=13
[Sample-CP] g=13 c=1730 u=-1.0000000000 subseq=14
[Sample-CP] g=14 c=552 u=-1.0000000000 subseq=15
[Sample-CP] g=9 c=223 u=-1.0000000000 subseq=10
[Sample-CP] g=10 c=1502 u=-1.0000000000 subseq=11
[Sample-CP] g=11 c=1702 u=-1.0000000000 subseq=12
[Sample-CP] g=12 c=1860 u=-1.0000000000 subseq=13
[Sample-CP] g=13 c=1008 u=-1.0000000000 subseq=14
[Sample-CP] g=14 c=127 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: 0.037001 -0.042584 0.009297 0.073590
[Debug] talker-hidden-step1: [2048] first4: 0.657201 2.473228 0.382572 0.284397
[Debug] next-emb-step0: [2048] first4: 0.034006 -0.013787 0.039109 0.091488
[Debug] talker-hidden-step1: [2048] first4: 0.742126 2.656314 -0.143770 -0.182433
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
[Sample-CP] g=2 c=324 u=-1.0000000000 subseq=19
[Sample-CP] g=3 c=1210 u=-1.0000000000 subseq=20
[Sample-CP] g=3 c=350 u=-1.0000000000 subseq=20
[Sample-CP] g=4 c=985 u=-1.0000000000 subseq=21
[Sample-CP] g=5 c=1640 u=-1.0000000000 subseq=22
[Sample-CP] g=5 c=853 u=-1.0000000000 subseq=22
[Sample-CP] g=6 c=329 u=-1.0000000000 subseq=23
[Sample-CP] g=7 c=564 u=-1.0000000000 subseq=24
[Sample-CP] g=8 c=1004 u=-1.0000000000 subseq=25
[Sample-CP] g=8 c=692 u=-1.0000000000 subseq=25
[Sample-CP] g=9 c=1702 u=-1.0000000000 subseq=26
[Sample-CP] g=10 c=437 u=-1.0000000000 subseq=27
[Sample-CP] g=11 c=729 u=-1.0000000000 subseq=28
[Sample-CP] g=12 c=344 u=-1.0000000000 subseq=29
[Sample-CP] g=13 c=85 u=-1.0000000000 subseq=30
[Sample-CP] g=14 c=1435 u=-1.0000000000 subseq=31
[Sample-CP] g=14 c=1668 u=-1.0000000000 subseq=31
[Pipeline] Generated 8 frames
[Pipeline] Generated 16 frames
[Pipeline] Generated 24 frames
@@ -119,15 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001606 -0.002810 -0.002527 -0.001049
[Perf] PromptBuild 2.3 ms
[Perf] Prefill 19.8 ms (T_ctx prefill)
[Perf] TTFA 31.4 ms (first frame codes)
[Perf] TalkerDecode 349.5 ms (64 frames, 5.46 ms/frame)
[Perf] CodePredictor 517.3 ms (8.08 ms/frame)
[Perf] HostCompose 1.3 ms (c0 sample + next emb)
[Perf] CodecDecode 12.8 ms
[Perf] Total 904.1 ms (64 frames, 13.56 ms/frame AR, audio 5.12 s, RTF 0.177)
[Debug] output-audio: [122880] first4: 0.000673 0.000478 -0.000828 -0.000482
[Perf] PromptBuild 21.5 ms
[Perf] Prefill 1361.8 ms (T_ctx prefill)
[Perf] TTFA 1740.1 ms (first frame codes)
[Perf] TalkerDecode 518.2 ms (64 frames, 8.10 ms/frame)
[Perf] CodePredictor 842.7 ms (13.17 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 1373.9 ms
[Perf] Total 4119.4 ms (64 frames, 21.27 ms/frame AR, audio 5.12 s, RTF 0.805)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -149,17 +151,17 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] PromptIDs exact: 100.00% (38 values)
[Cossim] RefCodes exact: 82.41% (3440 values)
[Cossim] Embed cos: 0.972129 max: 1.9482e-01 mean: 8.1111e-03
[Cossim] TrailingText cos: 1.000000 max: 5.9605e-08 mean: 5.6559e-10
[Cossim] TTSPadEmbed cos: 1.000000 max: 5.9605e-08 mean: 5.6559e-10
[Cossim] L0 cos: 0.987066 max: 6.8550e-01 mean: 1.0819e-02
[Cossim] L7 cos: 0.999990 max: 1.4625e+01 mean: 2.7204e-02
[Cossim] L14 cos: 0.999971 max: 1.4619e+01 mean: 5.9834e-02
[Cossim] L21 cos: 0.999707 max: 1.4627e+01 mean: 2.2446e-01
[Cossim] L27 cos: 0.998906 max: 2.5748e+01 mean: 5.2196e-01
[Cossim] Final cos: 0.994367 max: 7.7750e+00 mean: 1.6925e-01
[Cossim] Logits cos: 0.999549 max: 8.4361e-01 mean: 1.2133e-01
[Cossim] NextEmbStep0 cos: 0.883765 max: 1.3749e-01 mean: 2.4548e-02
[Cossim] TalkerHiddenStep1 cos: 0.994868 max: 1.0153e+00 mean: 2.2038e-01
[Cossim] TrailingText cos: 1.000000 max: 2.9802e-08 mean: 5.4933e-10
[Cossim] TTSPadEmbed cos: 1.000000 max: 2.9802e-08 mean: 5.4933e-10
[Cossim] L0 cos: 0.987066 max: 6.8534e-01 mean: 1.0819e-02
[Cossim] L7 cos: 0.999990 max: 1.4636e+01 mean: 2.7206e-02
[Cossim] L14 cos: 0.999971 max: 1.4633e+01 mean: 5.9840e-02
[Cossim] L21 cos: 0.999707 max: 1.4641e+01 mean: 2.2449e-01
[Cossim] L27 cos: 0.998906 max: 2.5746e+01 mean: 5.2221e-01
[Cossim] Final cos: 0.994366 max: 7.7972e+00 mean: 1.6928e-01
[Cossim] Logits cos: 0.999568 max: 8.3676e-01 mean: 1.1911e-01
[Cossim] NextEmbStep0 cos: 0.822712 max: 2.7881e-01 mean: 3.0097e-02
[Cossim] TalkerHiddenStep1 cos: 0.990742 max: 1.7658e+00 mean: 2.9620e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -173,10 +175,10 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] EncTransformer cos: 0.999974 max: 3.3284e-01 mean: 7.6420e-03
[Cossim] CodecPreFSQ cos: 0.999972 max: 4.5137e-01 mean: 1.8013e-02
[Cossim] SpkFrontend cos: 1.000000 max: 2.6600e-03 mean: 1.1964e-04
[Cossim] SpkBlock3 cos: 1.000000 max: 1.2137e-02 mean: 5.2077e-04
[Cossim] SpkBlock3 cos: 1.000000 max: 1.2138e-02 mean: 5.2076e-04
[Cossim] SpkMFA cos: 1.000000 max: 8.9911e-03 mean: 1.3077e-04
[Cossim] SpkASP cos: 1.000000 max: 2.2933e-04 mean: 1.7414e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 7.9088e-05 mean: 1.3439e-05
[Cossim] CodesFull exact: 3.67% (1008 values)
[Cossim] Audio cos: 0.012961
[Cossim] WAV stft_cos: 0.169275 samples: 120960
[Cossim] SpkASP cos: 1.000000 max: 2.2851e-04 mean: 1.7482e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 7.8417e-05 mean: 1.3521e-05
[Cossim] CodesFull exact: 3.17% (1008 values)
[Cossim] Audio cos: 0.023761
[Cossim] WAV stft_cos: 0.225096 samples: 120960
tests/clone-Vulkan0-Q4_K_M.log
+54 -52
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: Vulkan0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-Q4_K_M.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 1012.7 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 109.5 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-Q4_K_M.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,6 +22,27 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 4 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.098113 0.854719 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.956470 0.098113 1.233186 0.246316
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922765 0.615104 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067722 0.119141 0.240900 0.344142
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 13.6 MB into backend
@@ -32,24 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.098113 0.854719 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.956472 0.098113 1.233186 0.246316
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922765 0.615104 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067725 0.119141 0.240898 0.344141
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[Pipeline] Codec encoder lazy loaded in 9 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.138303 13.699992 -0.351357 3.025051
[Debug] enc-transformer-out: [430, 512] first4: 0.887039 3.463154 0.047173 0.245495
@@ -65,7 +67,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-input-embed: [225, 2048] first4: 0.022060 -0.009214 0.008459 -0.019803
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005416 0.008943 -0.004432 0.000657
[Debug] tts-pad-embed: [2048] first4: -0.005416 0.008943 -0.004432 0.000657
[Debug] spk-emb: [2048] first4: 0.098351 -0.023771 -0.011823 0.006760
[Debug] spk-emb: [2048] first4: 0.098350 -0.023773 -0.011820 0.006758
[Debug] ref-codes: [16, 215] first4: 1221.000000 2042.000000 696.000000 1522.000000
[Debug] talker-hidden-prefill-l0: [225, 2048] first4: -0.437574 -0.067419 0.007751 0.014628
[Debug] talker-hidden-prefill-l7: [225, 2048] first4: -8.835860 0.790539 0.495832 0.094764
@@ -73,7 +75,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -8.652716 1.697395 1.098899 -0.791332
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 4.921398 -31.754539 -6.813190 21.707518
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.166135 -1.119393 -0.213715 0.667947
[Debug] talker-logits-prefill: [3072] first4: -0.132812 -5.521973 -7.044922 -10.550781
[Debug] talker-logits-prefill: [3072] first4: -0.151123 -5.516602 -7.045898 -10.537109
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
@@ -92,7 +94,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=1047 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 417.000000
[Debug] next-emb-step0: [2048] first4: 0.005714 0.034105 -0.000204 0.073642
[Debug] talker-hidden-step1: [2048] first4: 0.586295 2.651410 0.840972 0.821965
[Debug] talker-hidden-step1: [2048] first4: 0.590231 2.653425 0.827661 0.831199
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
@@ -119,15 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 417.000000
[Debug] output-audio: [122880] first4: -0.001181 -0.002022 -0.001853 -0.000890
[Perf] PromptBuild 2.7 ms
[Perf] Prefill 17.1 ms (T_ctx prefill)
[Perf] TTFA 26.5 ms (first frame codes)
[Perf] TalkerDecode 159.3 ms (64 frames, 2.49 ms/frame)
[Perf] CodePredictor 343.7 ms (5.37 ms/frame)
[Perf] HostCompose 2.6 ms (c0 sample + next emb)
[Perf] CodecDecode 14.3 ms
[Perf] Total 540.5 ms (64 frames, 7.90 ms/frame AR, audio 5.12 s, RTF 0.106)
[Debug] output-audio: [122880] first4: -0.000101 0.000003 -0.000779 -0.000457
[Perf] PromptBuild 1050.5 ms
[Perf] Prefill 1572.6 ms (T_ctx prefill)
[Perf] TTFA 2788.5 ms (first frame codes)
[Perf] TalkerDecode 269.8 ms (64 frames, 4.22 ms/frame)
[Perf] CodePredictor 478.1 ms (7.47 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 29.4 ms
[Perf] Total 3410.1 ms (64 frames, 11.69 ms/frame AR, audio 5.12 s, RTF 0.666)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -151,15 +153,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] Embed cos: 0.917128 max: 2.2245e-01 mean: 2.1853e-02
[Cossim] TrailingText cos: 0.999967 max: 1.5749e-03 mean: 2.3645e-04
[Cossim] TTSPadEmbed cos: 0.999967 max: 1.5749e-03 mean: 2.3645e-04
[Cossim] L0 cos: 0.961488 max: 8.1836e-01 mean: 2.8364e-02
[Cossim] L7 cos: 0.997433 max: 3.9399e+01 mean: 1.6337e-01
[Cossim] L14 cos: 0.997359 max: 3.8767e+01 mean: 2.4736e-01
[Cossim] L21 cos: 0.995877 max: 3.9008e+01 mean: 7.7251e-01
[Cossim] L27 cos: 0.956635 max: 1.7506e+03 mean: 1.9749e+00
[Cossim] Final cos: 0.955792 max: 3.3331e+01 mean: 5.4933e-01
[Cossim] Logits cos: 0.992697 max: 2.9549e+00 mean: 4.4853e-01
[Cossim] L0 cos: 0.961486 max: 8.1847e-01 mean: 2.8364e-02
[Cossim] L7 cos: 0.997433 max: 3.9308e+01 mean: 1.6337e-01
[Cossim] L14 cos: 0.997359 max: 3.8675e+01 mean: 2.4737e-01
[Cossim] L21 cos: 0.995877 max: 3.8922e+01 mean: 7.7256e-01
[Cossim] L27 cos: 0.956688 max: 1.7483e+03 mean: 1.9751e+00
[Cossim] Final cos: 0.955767 max: 3.3655e+01 mean: 5.4944e-01
[Cossim] Logits cos: 0.992621 max: 2.9744e+00 mean: 4.5055e-01
[Cossim] NextEmbStep0 cos: 0.819095 max: 1.4988e-01 mean: 3.0364e-02
[Cossim] TalkerHiddenStep1 cos: 0.957774 max: 3.9701e+00 mean: 6.2679e-01
[Cossim] TalkerHiddenStep1 cos: 0.957568 max: 3.9459e+00 mean: 6.2838e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -171,12 +173,12 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] SeanetStage3 cos: 0.999995 max: 1.4657e+00 mean: 3.8722e-02
[Cossim] SeanetOut cos: 0.999963 max: 4.0791e-01 mean: 1.6984e-02
[Cossim] EncTransformer cos: 0.992979 max: 3.5659e+00 mean: 1.6375e-01
[Cossim] CodecPreFSQ cos: 0.996263 max: 2.2478e+00 mean: 2.4022e-01
[Cossim] CodecPreFSQ cos: 0.996263 max: 2.2477e+00 mean: 2.4022e-01
[Cossim] SpkFrontend cos: 1.000000 max: 2.6600e-03 mean: 1.1964e-04
[Cossim] SpkBlock3 cos: 1.000000 max: 1.2137e-02 mean: 5.2077e-04
[Cossim] SpkMFA cos: 1.000000 max: 8.9911e-03 mean: 1.3077e-04
[Cossim] SpkASP cos: 1.000000 max: 2.2933e-04 mean: 1.7414e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 7.9088e-05 mean: 1.3439e-05
[Cossim] CodesFull exact: 1.88% (1008 values)
[Cossim] Audio cos: -0.006096
[Cossim] WAV stft_cos: 0.202987 samples: 120960
[Cossim] SpkBlock3 cos: 1.000000 max: 1.2137e-02 mean: 5.2085e-04
[Cossim] SpkMFA cos: 1.000000 max: 8.9911e-03 mean: 1.3079e-04
[Cossim] SpkASP cos: 1.000000 max: 2.2919e-04 mean: 1.7381e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 8.0518e-05 mean: 1.3621e-05
[Cossim] CodesFull exact: 2.68% (1008 values)
[Cossim] Audio cos: -0.013397
[Cossim] WAV stft_cos: 0.197850 samples: 120960
tests/clone-Vulkan0-Q8_0.log
+53 -51
View File
@@ -1,19 +1,17 @@
[Qwen] qwentts.cpp 199a658 (2026-05-31)
[Qwen] qwentts.cpp 3e3b6c1 (2026-07-04)
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/qwentts.cpp/build/libggml-cuda.so
load_backend: loaded CUDA backend from /mnt/workspace/git/qwentts.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/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-cpu-zen4.so
load_backend: loaded Vulkan backend from /mnt/workspace/git/qwentts.cpp/build/libggml-vulkan.so
load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libggml-cpu-zen4.so
[Load] Talker backend: Vulkan0 (CPU threads: 16)
[GGUF] ../models/qwen-talker-1.7b-base-Q8_0.gguf: 480 tensors, data at offset 5954496
[WeightCtx] Loaded 316 tensors, 1765.0 MB into backend
[Talker] Loaded: 28 layers, hidden 2048, heads 16/8, head_dim 128, FFN 6144, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 88 tensors, 177.5 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj linear
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] ../models/qwen-tokenizer-12hz-Q8_0.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
@@ -24,6 +22,27 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[WeightCtx] Loaded 76 tensors, 34.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=2048 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[Pipeline] Speaker encoder lazy loaded in 4 ms
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.098113 0.854719 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.956470 0.098113 1.233186 0.246316
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922765 0.615104 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067722 0.119141 0.240900 0.344142
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 25.6 MB into backend
@@ -32,24 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[WAV] Read buffer: 380672 samples, 22050 Hz, 1 ch, 16 bit
[Audio-Resample] 22050 Hz -> 24000 Hz, 380672 samples...
[Audio-Resample] Done: 380672 -> 414337 samples
[Debug] mel-spk: [1618, 128] first4: -6.330339 -6.170074 -6.011195 -6.243161
[Debug] mel-hann: [1024] first4: 0.000000 0.000009 0.000038 0.000085
[Debug] mel-basis: [128, 513] first4: 0.000000 0.033550 0.008569 0.000000
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.098113 0.854719 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.956472 0.098113 1.233186 0.246316
[Debug] spk-mfa: [1618, 1536] first4: 0.000000 0.922765 0.615104 0.000000
[Debug] spk-asp: [1, 3072] first4: 0.067725 0.119141 0.240898 0.344141
[SpkExtract] Extracted 2048-dim embedding (414337 samples, padded 415105)
[Pipeline] Codec encoder lazy loaded in 10 ms
[Debug] audio-input: [412800] first4: -0.000546 -0.001162 -0.000871 0.000340
[Debug] seanet-out: [430, 512] first4: -0.138303 13.699992 -0.351357 3.025051
[Debug] enc-transformer-out: [430, 512] first4: 0.633121 3.145168 -0.143299 0.564666
@@ -65,7 +67,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-input-embed: [225, 2048] first4: 0.021991 -0.008996 0.008355 -0.020301
[Debug] trailing-text-hidden: [1, 2048] first4: -0.005262 0.008715 -0.004324 0.001049
[Debug] tts-pad-embed: [2048] first4: -0.005262 0.008715 -0.004324 0.001049
[Debug] spk-emb: [2048] first4: 0.098351 -0.023771 -0.011823 0.006760
[Debug] spk-emb: [2048] first4: 0.098350 -0.023773 -0.011820 0.006758
[Debug] ref-codes: [16, 215] first4: 1221.000000 2042.000000 509.000000 1522.000000
[Debug] talker-hidden-prefill-l0: [225, 2048] first4: -0.463469 -0.072789 0.015072 -0.037665
[Debug] talker-hidden-prefill-l7: [225, 2048] first4: -2.203363 0.166743 0.191879 1.621024
@@ -73,7 +75,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Debug] talker-hidden-prefill-l21: [225, 2048] first4: -1.465379 0.288384 0.656437 1.015133
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 6.677526 -34.405407 -7.383842 28.511698
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.227110 -1.221941 -0.233353 0.883898
[Debug] talker-logits-prefill: [3072] first4: 1.240234 -7.818359 -7.036621 -10.630859
[Debug] talker-logits-prefill: [3072] first4: 1.264160 -7.826172 -7.041016 -10.638672
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Sample-CP] g=0 c=1408 u=-1.0000000000 subseq=1
[Sample-CP] g=1 c=1824 u=-1.0000000000 subseq=2
@@ -92,7 +94,7 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Sample-CP] g=14 c=516 u=-1.0000000000 subseq=15
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] next-emb-step0: [2048] first4: 0.021546 -0.036059 0.003111 0.073181
[Debug] talker-hidden-step1: [2048] first4: 0.368557 2.554765 -0.075612 0.176847
[Debug] talker-hidden-step1: [2048] first4: 0.372459 2.555521 -0.076819 0.171399
[Sample] step=1 c0=1902 u=-1.0000000000 subseq=16
[Sample-CP] g=0 c=1924 u=-1.0000000000 subseq=17
[Sample-CP] g=1 c=2025 u=-1.0000000000 subseq=18
@@ -119,15 +121,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Pipeline] Generated 64 frames
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: -0.001815 -0.003137 -0.002853 -0.001444
[Perf] PromptBuild 3.1 ms
[Perf] Prefill 17.1 ms (T_ctx prefill)
[Perf] TTFA 27.0 ms (first frame codes)
[Perf] TalkerDecode 181.1 ms (64 frames, 2.83 ms/frame)
[Perf] CodePredictor 347.4 ms (5.43 ms/frame)
[Perf] HostCompose 1.4 ms (c0 sample + next emb)
[Perf] CodecDecode 13.7 ms
[Perf] Total 564.9 ms (64 frames, 8.28 ms/frame AR, audio 5.12 s, RTF 0.110)
[Debug] output-audio: [122880] first4: 0.000658 0.000394 -0.000943 -0.000566
[Perf] PromptBuild 1303.3 ms
[Perf] Prefill 1974.9 ms (T_ctx prefill)
[Perf] TTFA 3439.7 ms (first frame codes)
[Perf] TalkerDecode 196.2 ms (64 frames, 3.07 ms/frame)
[Perf] CodePredictor 488.2 ms (7.63 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 30.3 ms
[Perf] Total 3994.3 ms (64 frames, 10.70 ms/frame AR, audio 5.12 s, RTF 0.780)
[WAV] Wrote cpp/clone/clone-cpp.wav: 122880 samples, 24000 Hz, mono S16
[Pipeline] Wrote 122880 samples (5.12 s) -> cpp/clone/clone-cpp.wav
[Input] Prompt: 100 chars: qwentts.cpp is a minimal C++17 and GGML port of Qwen3-TTS fo...
@@ -151,15 +153,15 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] Embed cos: 0.970174 max: 1.9001e-01 mean: 8.9738e-03
[Cossim] TrailingText cos: 0.999999 max: 1.9968e-04 mean: 4.5182e-05
[Cossim] TTSPadEmbed cos: 0.999999 max: 1.9968e-04 mean: 4.5182e-05
[Cossim] L0 cos: 0.985827 max: 7.7032e-01 mean: 1.1813e-02
[Cossim] L7 cos: 0.999979 max: 5.4026e+00 mean: 3.6053e-02
[Cossim] L14 cos: 0.999957 max: 5.4250e+00 mean: 7.0620e-02
[Cossim] L21 cos: 0.999606 max: 1.8725e+01 mean: 2.5042e-01
[Cossim] L27 cos: 0.998546 max: 5.4983e+01 mean: 5.6850e-01
[Cossim] Final cos: 0.992857 max: 1.9049e+01 mean: 1.7996e-01
[Cossim] Logits cos: 0.999582 max: 7.9829e-01 mean: 1.0569e-01
[Cossim] L0 cos: 0.985826 max: 7.7038e-01 mean: 1.1813e-02
[Cossim] L7 cos: 0.999979 max: 5.4011e+00 mean: 3.6051e-02
[Cossim] L14 cos: 0.999957 max: 5.4243e+00 mean: 7.0612e-02
[Cossim] L21 cos: 0.999605 max: 1.8733e+01 mean: 2.5057e-01
[Cossim] L27 cos: 0.998543 max: 5.3200e+01 mean: 5.6948e-01
[Cossim] Final cos: 0.992844 max: 1.9032e+01 mean: 1.8018e-01
[Cossim] Logits cos: 0.999587 max: 7.9291e-01 mean: 1.0517e-01
[Cossim] NextEmbStep0 cos: 0.848337 max: 2.7267e-01 mean: 2.8037e-02
[Cossim] TalkerHiddenStep1 cos: 0.993807 max: 1.5080e+00 mean: 2.3942e-01
[Cossim] TalkerHiddenStep1 cos: 0.993831 max: 1.5021e+00 mean: 2.3902e-01
[Cossim] MelHann cos: 1.000000 max: 1.7881e-07 mean: 3.2043e-08
[Cossim] MelBasis cos: 1.000000 max: 1.7881e-07 mean: 3.9379e-10
[Cossim] MelMag cos: 1.000000 max: 3.1307e-05 mean: 5.6396e-07
@@ -173,10 +175,10 @@ load_backend: loaded CPU backend from /mnt/workspace/qwentts.cpp/build/libggml-c
[Cossim] EncTransformer cos: 0.999932 max: 3.3955e-01 mean: 1.5945e-02
[Cossim] CodecPreFSQ cos: 0.999951 max: 4.5186e-01 mean: 2.6992e-02
[Cossim] SpkFrontend cos: 1.000000 max: 2.6600e-03 mean: 1.1964e-04
[Cossim] SpkBlock3 cos: 1.000000 max: 1.2137e-02 mean: 5.2077e-04
[Cossim] SpkMFA cos: 1.000000 max: 8.9911e-03 mean: 1.3077e-04
[Cossim] SpkASP cos: 1.000000 max: 2.2933e-04 mean: 1.7414e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 7.9088e-05 mean: 1.3439e-05
[Cossim] CodesFull exact: 3.87% (1008 values)
[Cossim] Audio cos: -0.021260
[Cossim] WAV stft_cos: 0.319623 samples: 120960
[Cossim] SpkBlock3 cos: 1.000000 max: 1.2137e-02 mean: 5.2085e-04
[Cossim] SpkMFA cos: 1.000000 max: 8.9911e-03 mean: 1.3079e-04
[Cossim] SpkASP cos: 1.000000 max: 2.2919e-04 mean: 1.7381e-05
[Cossim] SpeakerEmb cos: 1.000000 max: 8.0518e-05 mean: 1.3621e-05
[Cossim] CodesFull exact: 2.38% (1008 values)
[Cossim] Audio cos: 0.013854
[Cossim] WAV stft_cos: 0.183103 samples: 120960