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-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