[Qwen] qwentts.cpp 9a9a425 (2026-07-30)
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/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 | fp4: 0 | warp size: 32 | shared memory: 49152 | int dot: 1 | matrix cores: NV_coopmat2
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
[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
[WeightCtx] Loaded 93 tensors, 58.1 MB into backend
[Transformer] Loaded: 8 layers, hidden 512, heads 16/16, head_dim 64, FFN 1024, RoPE theta 10000, sliding window 72
[WeightCtx] Loaded 22 tensors, 48.1 MB into backend
[Upsample] Loaded: 2 blocks (2x ratio per block), channels 1024, dwconv kernel 7
[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, 1 sets -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16, 1 sets -> 0 MB
[Pipeline] Loaded: arch=1b7 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=deferred speakers=0 fa=off clamp_fp16=off max_batch=1
[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
[Debug] mel-mag: [1618, 513] first4: 0.047604 0.042322 0.042187 0.066630
[Debug] spk-frontend: [1618, 512] first4: 0.000000 0.099754 0.858149 0.000000
[Debug] spk-block3: [1618, 512] first4: 1.958561 0.099754 1.238315 0.246145
[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 42 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
[Debug] codec-pre-fsq: [215, 512] first4: -3.826889 -6.990853 -0.834646 2.924343
[Debug] seanet-init: [412800, 64] first4: -0.015825 -0.054180 -0.117431 0.038554
[Debug] seanet-resnet0: [412800, 64] first4: 0.057527 -0.220565 -0.137270 -0.250936
[Debug] seanet-stage0: [103200, 128] first4: 0.233115 -0.323956 0.247371 0.205229
[Debug] seanet-stage1: [20640, 256] first4: -0.363234 0.317898 -0.574110 0.579073
[Debug] seanet-stage3: [430, 1024] first4: -2.301637 -4.978660 -6.998715 -9.010848
[Pipeline] ICL ref_codes: 215 frames at 12.5 Hz (412800 audio samples)
[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: -0.005340 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
[Debug] talker-hidden-prefill-l7: [225, 2048] first4: -1.523166 -0.015666 0.347224 1.775774
[Debug] talker-hidden-prefill-l14: [225, 2048] first4: -1.294858 0.181118 0.448874 1.742527
[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: 1.339535 -7.642556 -6.988994 -10.529062
[Batch] Admit slot=0 T_ctx=225 prefill=384.4 ms build=2.2 ms (stall for 0 active slots)
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[Debug] codes-step0: [16] first4: 27.000000 1408.000000 1824.000000 400.000000
[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
[Pipeline] Generated 8 frames (slot 0)
[Pipeline] Generated 16 frames (slot 0)
[Pipeline] Generated 24 frames (slot 0)
[Pipeline] Generated 32 frames (slot 0)
[Pipeline] Generated 40 frames (slot 0)
[Pipeline] Generated 48 frames (slot 0)
[Pipeline] Generated 56 frames (slot 0)
[Pipeline] Generated 64 frames (slot 0)
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000365 0.000165 -0.001043 -0.000566
[Perf] PromptBuild 2.2 ms
[Perf] Prefill 384.4 ms (T_ctx prefill)
[Perf] TTFA 1557.6 ms (first frame codes)
[Perf] TalkerDecode 2869.0 ms (64 frames, 44.83 ms/frame)
[Perf] CodePredictor 1616.7 ms (25.26 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 2212.0 ms
[Perf] Total 8230.2 ms (64 frames, 70.09 ms/frame AR, audio 5.12 s, RTF 1.607)
[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...
[Input] RefAudio: ../examples/freeman.wav
[Input] RefText: 213 chars: If you go into different cultures, they have different conce...
[Input] Lang: english Seed: 42 MaxNewTokens: 64
[Input] Mode: greedy ICL
[Python] Device: cuda
[Python] RefWav: 414337 samples 24000 Hz 17.26s
[Python] SpeakerEmb shape: (2048,) dtype: torch.float32
[Python] RefCodes shape: (16, 215) (K, T_codec)
[Python] InputIds shape: (1, 38)
[Python] RefIds shape: (1, 53)
[Python] Codes shape: (63, 16) (T_frames, num_code_groups)
[Python] Audio: 120960 samples 24000 Hz 5.04s -> python/clone/clone-python.wav
[Quant] BF16 -> ../models/qwen-talker-1.7b-base-BF16.gguf + ../models/qwen-tokenizer-12hz-BF16.gguf
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-BF16.gguf --codec ../models/qwen-tokenizer-12hz-BF16.gguf --seed 42 --ref-wav ../examples/freeman.wav --ref-text ../examples/freeman.txt --lang english --max-new 64 --dump cpp/clone -o cpp/clone/clone-cpp.wav --greedy
[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: 6.5145e-03
[Cossim] TrailingText cos: 1.000000 max: 3.5048e-05 mean: 2.4957e-06
[Cossim] TTSPadEmbed cos: 1.000000 max: 3.5048e-05 mean: 2.4957e-06
[Cossim] L0 cos: 0.989591 max: 7.7401e-01 mean: 8.9719e-03
[Cossim] L7 cos: 0.999993 max: 1.2845e+01 mean: 2.3289e-02
[Cossim] L14 cos: 0.999976 max: 1.2849e+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: 5.9450e-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
[Cossim] MelSpk cos: 1.000000 max: 2.5492e-03 mean: 9.3754e-06
[Cossim] SeanetInit cos: 1.000000 max: 9.1240e-04 mean: 1.6087e-05
[Cossim] SeanetResnet0 cos: 0.999973 max: 1.9622e-03 mean: 1.4500e-04
[Cossim] SeanetStage0 cos: 0.999996 max: 3.9780e-02 mean: 3.7406e-04
[Cossim] SeanetStage1 cos: 0.999972 max: 7.4337e-01 mean: 2.1076e-03
[Cossim] SeanetStage3 cos: 0.999999 max: 9.8260e-01 mean: 2.5173e-02
[Cossim] SeanetOut cos: 0.999991 max: 8.8695e-02 mean: 9.9127e-03
[Cossim] EncTransformer cos: 0.999988 max: 9.6033e-01 mean: 6.5646e-03
[Cossim] CodecPreFSQ cos: 0.999991 max: 1.2993e-01 mean: 1.1939e-02
[Cossim] SpkFrontend cos: 0.999997 max: 2.5180e-02 mean: 9.4975e-04
[Cossim] SpkBlock3 cos: 0.999991 max: 1.0583e-01 mean: 4.2477e-03
[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: 3.47% (1008 values)
[Cossim] Audio cos: -0.006883
[Cossim] WAV stft_cos: 0.209811 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-BF16.gguf --codec ../models/qwen-tokenizer-12hz-BF16.gguf --seed 42 --ref-wav ../examples/freeman.wav --ref-text ../examples/freeman.txt --lang english --max-new 64 --dump cpp/clone-fused -o - --greedy --codec-fused
[Perf Fused] PromptBuild 2.4 ms
[Perf Fused] Prefill 396.5 ms (T_ctx prefill)
[Perf Fused] TTFA 4942.8 ms (first frame codes)
[Perf Fused] TalkerDecode 2892.3 ms (64 frames, 45.19 ms/frame)
[Perf Fused] CodePredictor 3165.5 ms (49.46 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 3339.6 ms
[Perf Fused] Total 10948.3 ms (64 frames, 94.66 ms/frame AR, audio 5.12 s, RTF 2.138)
[Cossim] CodesFullFused exact: 100.00% (1024 values)