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

tests: refresh the clone grid logs

Browse Source
This commit is contained in:
Pascal
2026-07-30 23:05:37 +02:00
parent 9a9a425260
commit abab6b3bf3
12 changed files with 236 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/clone-CPU-BF16.log
+20 -10
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -51,7 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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
[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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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=382.1 ms build=2.1 ms (stall for 0 active slots)
[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
@@ -93,14 +93,14 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 2.1 ms
[Perf] Prefill 382.1 ms (T_ctx prefill)
[Perf] TTFA 1482.8 ms (first frame codes)
[Perf] TalkerDecode 2862.8 ms (64 frames, 44.73 ms/frame)
[Perf] CodePredictor 1428.0 ms (22.31 ms/frame)
[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 2259.3 ms
[Perf] Total 8010.2 ms (64 frames, 67.05 ms/frame AR, audio 5.12 s, RTF 1.564)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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)
tests/clone-CPU-F32.log
+20 -10
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -33,7 +33,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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
[Pipeline] Speaker encoder lazy loaded in 7 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
@@ -51,7 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 31 ms
[Pipeline] Codec encoder lazy loaded in 37 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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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: 1.330122 -7.768507 -6.935926 -10.338600
[Batch] Admit slot=0 T_ctx=225 prefill=633.9 ms build=3.7 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=630.2 ms build=3.7 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: 0.073561 -0.034233 -0.008814 0.042004
@@ -94,13 +94,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000352 0.000178 -0.001044 -0.000571
[Perf] PromptBuild 3.7 ms
[Perf] Prefill 633.9 ms (T_ctx prefill)
[Perf] TTFA 1757.1 ms (first frame codes)
[Perf] TalkerDecode 5437.1 ms (64 frames, 84.95 ms/frame)
[Perf] CodePredictor 3884.8 ms (60.70 ms/frame)
[Perf] Prefill 630.2 ms (T_ctx prefill)
[Perf] TTFA 1882.9 ms (first frame codes)
[Perf] TalkerDecode 5451.6 ms (64 frames, 85.18 ms/frame)
[Perf] CodePredictor 3921.6 ms (61.27 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 2192.1 ms
[Perf] Total 13219.1 ms (64 frames, 145.66 ms/frame AR, audio 5.12 s, RTF 2.582)
[Perf] CodecDecode 3868.2 ms
[Perf] Total 15071.1 ms (64 frames, 146.46 ms/frame AR, audio 5.12 s, RTF 2.944)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Cossim] CodesFull exact: 5.26% (1008 values)
[Cossim] Audio cos: 0.031659
[Cossim] WAV stft_cos: 0.252827 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-F32.gguf --codec ../models/qwen-tokenizer-12hz-F32.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 3.7 ms
[Perf Fused] Prefill 638.6 ms (T_ctx prefill)
[Perf Fused] TTFA 5259.6 ms (first frame codes)
[Perf Fused] TalkerDecode 5480.6 ms (64 frames, 85.63 ms/frame)
[Perf Fused] CodePredictor 5685.1 ms (88.83 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 3345.6 ms
[Perf Fused] Total 16337.0 ms (64 frames, 174.46 ms/frame AR, audio 5.12 s, RTF 3.191)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-CPU-Q4_K_M.log
+19 -9
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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: -0.397744 -4.641821 -7.453447 -10.679588
[Batch] Admit slot=0 T_ctx=225 prefill=413.0 ms build=2.9 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=412.8 ms build=3.3 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 417.000000
[Debug] next-emb-step0: [2048] first4: 0.035528 0.008631 -0.031989 0.034729
@@ -93,14 +93,14 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 417.000000
[Debug] output-audio: [122880] first4: -0.000340 -0.000466 -0.001126 -0.000988
[Perf] PromptBuild 2.9 ms
[Perf] Prefill 413.0 ms (T_ctx prefill)
[Perf] TTFA 1499.0 ms (first frame codes)
[Perf] TalkerDecode 958.0 ms (64 frames, 14.97 ms/frame)
[Perf] CodePredictor 368.4 ms (5.76 ms/frame)
[Perf] PromptBuild 3.3 ms
[Perf] Prefill 412.8 ms (T_ctx prefill)
[Perf] TTFA 1490.2 ms (first frame codes)
[Perf] TalkerDecode 938.1 ms (64 frames, 14.66 ms/frame)
[Perf] CodePredictor 441.2 ms (6.89 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 2213.5 ms
[Perf] Total 5033.0 ms (64 frames, 20.73 ms/frame AR, audio 5.12 s, RTF 0.983)
[Perf] CodecDecode 2230.9 ms
[Perf] Total 5102.6 ms (64 frames, 21.55 ms/frame AR, audio 5.12 s, RTF 0.997)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Cossim] CodesFull exact: 1.29% (1008 values)
[Cossim] Audio cos: -0.017629
[Cossim] WAV stft_cos: 0.176947 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-Q4_K_M.gguf --codec ../models/qwen-tokenizer-12hz-Q4_K_M.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 2.8 ms
[Perf Fused] Prefill 435.3 ms (T_ctx prefill)
[Perf Fused] TTFA 4974.7 ms (first frame codes)
[Perf Fused] TalkerDecode 999.3 ms (64 frames, 15.61 ms/frame)
[Perf Fused] CodePredictor 1991.8 ms (31.12 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 3364.7 ms
[Perf Fused] Total 7934.7 ms (64 frames, 46.74 ms/frame AR, audio 5.12 s, RTF 1.550)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-CPU-Q8_0.log
+19 -9
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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: 1.309391 -7.886597 -6.855894 -10.555799
[Batch] Admit slot=0 T_ctx=225 prefill=543.0 ms build=2.8 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=540.7 ms build=2.7 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: 0.010612 -0.008921 0.074940 0.036301
@@ -93,14 +93,14 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000360 0.000157 -0.001131 -0.000804
[Perf] PromptBuild 2.8 ms
[Perf] Prefill 543.0 ms (T_ctx prefill)
[Perf] TTFA 1633.2 ms (first frame codes)
[Perf] TalkerDecode 1643.9 ms (64 frames, 25.69 ms/frame)
[Perf] CodePredictor 625.3 ms (9.77 ms/frame)
[Perf] PromptBuild 2.7 ms
[Perf] Prefill 540.7 ms (T_ctx prefill)
[Perf] TTFA 1601.1 ms (first frame codes)
[Perf] TalkerDecode 1627.0 ms (64 frames, 25.42 ms/frame)
[Perf] CodePredictor 680.9 ms (10.64 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 2213.9 ms
[Perf] Total 6105.7 ms (64 frames, 35.46 ms/frame AR, audio 5.12 s, RTF 1.193)
[Perf] CodecDecode 2225.9 ms
[Perf] Total 6123.9 ms (64 frames, 36.06 ms/frame AR, audio 5.12 s, RTF 1.196)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Cossim] CodesFull exact: 2.68% (1008 values)
[Cossim] Audio cos: -0.001410
[Cossim] WAV stft_cos: 0.175402 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-Q8_0.gguf --codec ../models/qwen-tokenizer-12hz-Q8_0.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 2.9 ms
[Perf Fused] Prefill 553.9 ms (T_ctx prefill)
[Perf Fused] TTFA 5066.1 ms (first frame codes)
[Perf Fused] TalkerDecode 1652.7 ms (64 frames, 25.82 ms/frame)
[Perf Fused] CodePredictor 2223.8 ms (34.75 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 3329.9 ms
[Perf Fused] Total 8905.9 ms (64 frames, 60.57 ms/frame AR, audio 5.12 s, RTF 1.739)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-CUDA0-BF16.log
+20 -10
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -33,7 +33,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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
[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
@@ -51,7 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 90 ms
[Pipeline] Codec encoder lazy loaded in 65 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: 0.536580 3.214924 -0.152684 0.613806
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 7.045298 -34.398609 -7.069448 28.624956
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.239789 -1.222574 -0.223577 0.888043
[Debug] talker-logits-prefill: [3072] first4: 0.944606 -7.870981 -6.684478 -10.343767
[Batch] Admit slot=0 T_ctx=225 prefill=22.6 ms build=1.3 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=12.3 ms build=1.3 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: 0.013383 -0.012534 0.088655 0.033408
@@ -95,13 +95,13 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000387 0.000201 -0.001097 -0.000789
[Perf] PromptBuild 1.3 ms
[Perf] Prefill 22.6 ms (T_ctx prefill)
[Perf] TTFA 852.9 ms (first frame codes)
[Perf] TalkerDecode 166.7 ms (64 frames, 2.61 ms/frame)
[Perf] CodePredictor 257.6 ms (4.02 ms/frame)
[Perf] Prefill 12.3 ms (T_ctx prefill)
[Perf] TTFA 812.1 ms (first frame codes)
[Perf] TalkerDecode 167.0 ms (64 frames, 2.61 ms/frame)
[Perf] CodePredictor 271.1 ms (4.24 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 42.9 ms
[Perf] Total 1313.4 ms (64 frames, 6.63 ms/frame AR, audio 5.12 s, RTF 0.257)
[Perf] CodecDecode 32.8 ms
[Perf] Total 1273.8 ms (64 frames, 6.85 ms/frame AR, audio 5.12 s, RTF 0.249)
[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...
@@ -154,3 +154,13 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Cossim] CodesFull exact: 2.08% (1008 values)
[Cossim] Audio cos: 0.019831
[Cossim] WAV stft_cos: 0.189216 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 1.3 ms
[Perf Fused] Prefill 13.7 ms (T_ctx prefill)
[Perf Fused] TTFA 844.4 ms (first frame codes)
[Perf Fused] TalkerDecode 167.8 ms (64 frames, 2.62 ms/frame)
[Perf Fused] CodePredictor 335.9 ms (5.25 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 70.3 ms
[Perf Fused] Total 1338.5 ms (64 frames, 7.87 ms/frame AR, audio 5.12 s, RTF 0.261)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-CUDA0-F32.log
+19 -9
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -51,7 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 31 ms
[Pipeline] Codec encoder lazy loaded in 30 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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 7.131224 -34.471714 -7.108000 28.674042
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.242833 -1.225771 -0.224906 0.890001
[Debug] talker-logits-prefill: [3072] first4: 1.171040 -7.796190 -6.822366 -10.286754
[Batch] Admit slot=0 T_ctx=225 prefill=27.2 ms build=1.3 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=27.5 ms build=1.3 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: 0.003672 0.000908 0.039506 0.094594
@@ -95,13 +95,13 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000518 0.000403 -0.000870 -0.000545
[Perf] PromptBuild 1.3 ms
[Perf] Prefill 27.2 ms (T_ctx prefill)
[Perf] TTFA 767.1 ms (first frame codes)
[Perf] TalkerDecode 275.4 ms (64 frames, 4.30 ms/frame)
[Perf] CodePredictor 348.0 ms (5.44 ms/frame)
[Perf] Prefill 27.5 ms (T_ctx prefill)
[Perf] TTFA 777.0 ms (first frame codes)
[Perf] TalkerDecode 276.4 ms (64 frames, 4.32 ms/frame)
[Perf] CodePredictor 359.2 ms (5.61 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 33.1 ms
[Perf] Total 1417.0 ms (64 frames, 9.74 ms/frame AR, audio 5.12 s, RTF 0.277)
[Perf] CodecDecode 33.2 ms
[Perf] Total 1437.4 ms (64 frames, 9.93 ms/frame AR, audio 5.12 s, RTF 0.281)
[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...
@@ -154,3 +154,13 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Cossim] CodesFull exact: 3.27% (1008 values)
[Cossim] Audio cos: 0.077050
[Cossim] WAV stft_cos: 0.373618 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-F32.gguf --codec ../models/qwen-tokenizer-12hz-F32.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 1.3 ms
[Perf Fused] Prefill 27.0 ms (T_ctx prefill)
[Perf Fused] TTFA 848.7 ms (first frame codes)
[Perf Fused] TalkerDecode 276.8 ms (64 frames, 4.32 ms/frame)
[Perf Fused] CodePredictor 428.6 ms (6.70 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 72.5 ms
[Perf Fused] Total 1543.1 ms (64 frames, 11.02 ms/frame AR, audio 5.12 s, RTF 0.301)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-CUDA0-Q4_K_M.log
+21 -11
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -33,7 +33,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 21 ms
[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
@@ -51,7 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 47 ms
[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: -0.134336 13.731120 -0.351051 3.013485
[Debug] enc-transformer-out: [430, 512] first4: 1.140990 3.713432 0.131578 0.616251
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 4.146309 -31.912390 -6.918371 22.418171
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.140643 -1.130366 -0.218058 0.693131
[Debug] talker-logits-prefill: [3072] first4: -0.461190 -5.760653 -7.184410 -10.423358
[Batch] Admit slot=0 T_ctx=225 prefill=15.1 ms build=1.5 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=14.6 ms build=1.4 ms (stall for 0 active slots)
[Sample] step=0 c0=27 u=-1.0000000000 subseq=0
[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
@@ -94,14 +94,14 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1042.000000 390.000000 881.000000
[Debug] output-audio: [122880] first4: -0.000098 0.000072 -0.000411 -0.000387
[Perf] PromptBuild 1.5 ms
[Perf] Prefill 15.1 ms (T_ctx prefill)
[Perf] TTFA 778.4 ms (first frame codes)
[Perf] TalkerDecode 100.8 ms (64 frames, 1.58 ms/frame)
[Perf] CodePredictor 191.8 ms (3.00 ms/frame)
[Perf] PromptBuild 1.4 ms
[Perf] Prefill 14.6 ms (T_ctx prefill)
[Perf] TTFA 739.1 ms (first frame codes)
[Perf] TalkerDecode 101.2 ms (64 frames, 1.58 ms/frame)
[Perf] CodePredictor 205.7 ms (3.21 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 33.0 ms
[Perf] Total 1099.0 ms (64 frames, 4.57 ms/frame AR, audio 5.12 s, RTF 0.215)
[Perf] CodecDecode 33.1 ms
[Perf] Total 1071.7 ms (64 frames, 4.80 ms/frame AR, audio 5.12 s, RTF 0.209)
[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...
@@ -154,3 +154,13 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Cossim] CodesFull exact: 0.30% (1008 values)
[Cossim] Audio cos: -0.004365
[Cossim] WAV stft_cos: 0.172113 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-Q4_K_M.gguf --codec ../models/qwen-tokenizer-12hz-Q4_K_M.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 1.5 ms
[Perf Fused] Prefill 15.4 ms (T_ctx prefill)
[Perf Fused] TTFA 800.0 ms (first frame codes)
[Perf Fused] TalkerDecode 102.3 ms (64 frames, 1.60 ms/frame)
[Perf Fused] CodePredictor 271.1 ms (4.24 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 70.0 ms
[Perf Fused] Total 1164.6 ms (64 frames, 5.84 ms/frame AR, audio 5.12 s, RTF 0.227)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-CUDA0-Q8_0.log
+20 -10
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -33,7 +33,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 17 ms
[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
@@ -51,7 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 46 ms
[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: -0.134336 13.731120 -0.351051 3.013485
[Debug] enc-transformer-out: [430, 512] first4: 0.628775 2.909739 -0.209228 0.726215
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Debug] talker-hidden-prefill-l27: [225, 2048] first4: 6.677233 -34.349281 -7.298450 28.605591
[Debug] talker-hidden-prefill-final: [225, 2048] first4: 0.227255 -1.220783 -0.230812 0.887415
[Debug] talker-logits-prefill: [3072] first4: 1.706040 -8.254342 -7.318320 -10.816376
[Batch] Admit slot=0 T_ctx=225 prefill=14.7 ms build=1.4 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=15.2 ms build=1.4 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: -0.034936 -0.024575 0.029866 0.045185
@@ -95,13 +95,13 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000485 0.000126 -0.001332 -0.000870
[Perf] PromptBuild 1.4 ms
[Perf] Prefill 14.7 ms (T_ctx prefill)
[Perf] TTFA 777.8 ms (first frame codes)
[Perf] TalkerDecode 131.1 ms (64 frames, 2.05 ms/frame)
[Perf] CodePredictor 193.6 ms (3.03 ms/frame)
[Perf] Prefill 15.2 ms (T_ctx prefill)
[Perf] TTFA 747.6 ms (first frame codes)
[Perf] TalkerDecode 129.8 ms (64 frames, 2.03 ms/frame)
[Perf] CodePredictor 227.5 ms (3.55 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 33.1 ms
[Perf] Total 1130.6 ms (64 frames, 5.08 ms/frame AR, audio 5.12 s, RTF 0.221)
[Perf] CodecDecode 33.2 ms
[Perf] Total 1130.3 ms (64 frames, 5.58 ms/frame AR, audio 5.12 s, RTF 0.221)
[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...
@@ -154,3 +154,13 @@ ggml_backend_cuda_graph_compute: CUDA graph warmup complete
[Cossim] CodesFull exact: 2.28% (1008 values)
[Cossim] Audio cos: -0.001822
[Cossim] WAV stft_cos: 0.187675 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-Q8_0.gguf --codec ../models/qwen-tokenizer-12hz-Q8_0.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 1.4 ms
[Perf Fused] Prefill 15.5 ms (T_ctx prefill)
[Perf Fused] TTFA 813.7 ms (first frame codes)
[Perf Fused] TalkerDecode 127.5 ms (64 frames, 1.99 ms/frame)
[Perf Fused] CodePredictor 273.4 ms (4.27 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 69.7 ms
[Perf Fused] Total 1206.1 ms (64 frames, 6.27 ms/frame AR, audio 5.12 s, RTF 0.236)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-Vulkan0-BF16.log
+19 -9
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -51,7 +51,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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
[Pipeline] Codec encoder lazy loaded in 45 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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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: 0.882721 -7.881639 -6.796915 -10.414125
[Batch] Admit slot=0 T_ctx=225 prefill=31.5 ms build=1.9 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=31.5 ms build=2.0 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: -0.008599 -0.015697 0.093591 0.086640
@@ -93,14 +93,14 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000396 0.000229 -0.000999 -0.000637
[Perf] PromptBuild 1.9 ms
[Perf] PromptBuild 2.0 ms
[Perf] Prefill 31.5 ms (T_ctx prefill)
[Perf] TTFA 1379.9 ms (first frame codes)
[Perf] TalkerDecode 211.8 ms (64 frames, 3.31 ms/frame)
[Perf] CodePredictor 955.2 ms (14.93 ms/frame)
[Perf] TTFA 778.4 ms (first frame codes)
[Perf] TalkerDecode 192.6 ms (64 frames, 3.01 ms/frame)
[Perf] CodePredictor 330.0 ms (5.16 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 65.3 ms
[Perf] Total 2025.2 ms (64 frames, 18.24 ms/frame AR, audio 5.12 s, RTF 0.396)
[Perf] CodecDecode 63.1 ms
[Perf] Total 1358.0 ms (64 frames, 8.17 ms/frame AR, audio 5.12 s, RTF 0.265)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Cossim] CodesFull exact: 2.28% (1008 values)
[Cossim] Audio cos: 0.001872
[Cossim] WAV stft_cos: 0.214056 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 1.9 ms
[Perf Fused] Prefill 40.9 ms (T_ctx prefill)
[Perf Fused] TTFA 8280.4 ms (first frame codes)
[Perf Fused] TalkerDecode 190.4 ms (64 frames, 2.97 ms/frame)
[Perf Fused] CodePredictor 482.8 ms (7.54 ms/frame)
[Perf Fused] HostCompose 0.2 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 7488.6 ms
[Perf Fused] Total 8945.6 ms (64 frames, 10.52 ms/frame AR, audio 5.12 s, RTF 1.747)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-Vulkan0-F32.log
+20 -10
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -33,7 +33,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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 9 ms
[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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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
[Batch] Admit slot=0 T_ctx=225 prefill=1396.2 ms build=21.3 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=19.9 ms build=2.3 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: 0.034006 -0.013787 0.039109 0.091488
@@ -93,14 +93,14 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000497 0.000311 -0.000955 -0.000572
[Perf] PromptBuild 21.3 ms
[Perf] Prefill 1396.2 ms (T_ctx prefill)
[Perf] TTFA 4781.5 ms (first frame codes)
[Perf] TalkerDecode 475.1 ms (64 frames, 7.42 ms/frame)
[Perf] CodePredictor 801.3 ms (12.52 ms/frame)
[Perf] PromptBuild 2.3 ms
[Perf] Prefill 19.9 ms (T_ctx prefill)
[Perf] TTFA 784.0 ms (first frame codes)
[Perf] TalkerDecode 310.7 ms (64 frames, 4.85 ms/frame)
[Perf] CodePredictor 423.3 ms (6.61 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 1394.4 ms
[Perf] Total 7095.9 ms (64 frames, 19.95 ms/frame AR, audio 5.12 s, RTF 1.386)
[Perf] CodecDecode 61.3 ms
[Perf] Total 1572.4 ms (64 frames, 11.47 ms/frame AR, audio 5.12 s, RTF 0.307)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Cossim] CodesFull exact: 3.17% (1008 values)
[Cossim] Audio cos: 0.024399
[Cossim] WAV stft_cos: 0.226009 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-F32.gguf --codec ../models/qwen-tokenizer-12hz-F32.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 26.2 ms (T_ctx prefill)
[Perf Fused] TTFA 2364.0 ms (first frame codes)
[Perf Fused] TalkerDecode 313.8 ms (64 frames, 4.90 ms/frame)
[Perf Fused] CodePredictor 580.1 ms (9.06 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 1592.4 ms
[Perf Fused] Total 3248.9 ms (64 frames, 13.97 ms/frame AR, audio 5.12 s, RTF 0.635)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-Vulkan0-Q4_K_M.log
+19 -9
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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: -0.151123 -5.516602 -7.045898 -10.537109
[Batch] Admit slot=0 T_ctx=225 prefill=1555.1 ms build=1041.5 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=16.9 ms build=2.8 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 417.000000
[Debug] next-emb-step0: [2048] first4: 0.005714 0.034105 -0.000204 0.073642
@@ -93,14 +93,14 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 417.000000
[Debug] output-audio: [122880] first4: -0.000352 -0.000173 -0.000826 -0.000555
[Perf] PromptBuild 1041.5 ms
[Perf] Prefill 1555.1 ms (T_ctx prefill)
[Perf] TTFA 4465.0 ms (first frame codes)
[Perf] TalkerDecode 229.4 ms (64 frames, 3.58 ms/frame)
[Perf] CodePredictor 431.4 ms (6.74 ms/frame)
[Perf] PromptBuild 2.8 ms
[Perf] Prefill 16.9 ms (T_ctx prefill)
[Perf] TTFA 759.2 ms (first frame codes)
[Perf] TalkerDecode 116.2 ms (64 frames, 1.82 ms/frame)
[Perf] CodePredictor 253.3 ms (3.96 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 63.1 ms
[Perf] Total 5038.9 ms (64 frames, 10.33 ms/frame AR, audio 5.12 s, RTF 0.984)
[Perf] CodecDecode 62.8 ms
[Perf] Total 1185.9 ms (64 frames, 5.78 ms/frame AR, audio 5.12 s, RTF 0.232)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Cossim] CodesFull exact: 2.38% (1008 values)
[Cossim] Audio cos: -0.015329
[Cossim] WAV stft_cos: 0.211505 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-Q4_K_M.gguf --codec ../models/qwen-tokenizer-12hz-Q4_K_M.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 2.7 ms
[Perf Fused] Prefill 24.1 ms (T_ctx prefill)
[Perf Fused] TTFA 1973.6 ms (first frame codes)
[Perf Fused] TalkerDecode 116.6 ms (64 frames, 1.82 ms/frame)
[Perf Fused] CodePredictor 400.2 ms (6.25 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 1229.7 ms
[Perf Fused] Total 2483.6 ms (64 frames, 8.08 ms/frame AR, audio 5.12 s, RTF 0.485)
[Cossim] CodesFullFused exact: 100.00% (1024 values)
tests/clone-Vulkan0-Q8_0.log
+20 -10
View File
@@ -1,4 +1,4 @@
[Qwen] qwentts.cpp 710a52a (2026-07-25)
[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
@@ -76,7 +76,7 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[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: 1.264160 -7.826172 -7.041016 -10.638672
[Batch] Admit slot=0 T_ctx=225 prefill=1960.4 ms build=1281.9 ms (stall for 0 active slots)
[Batch] Admit slot=0 T_ctx=225 prefill=17.0 ms build=2.8 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: 0.021546 -0.036059 0.003111 0.073181
@@ -93,14 +93,14 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Pipeline] Generation done : 64 frames
[Debug] codes-full: [64, 16] first4: 27.000000 1408.000000 1824.000000 400.000000
[Debug] output-audio: [122880] first4: 0.000434 0.000233 -0.001032 -0.000635
[Perf] PromptBuild 1281.9 ms
[Perf] Prefill 1960.4 ms (T_ctx prefill)
[Perf] TTFA 5250.2 ms (first frame codes)
[Perf] TalkerDecode 158.5 ms (64 frames, 2.48 ms/frame)
[Perf] CodePredictor 439.0 ms (6.86 ms/frame)
[Perf] HostCompose 0.1 ms (c0 sample + next emb)
[Perf] CodecDecode 64.0 ms
[Perf] Total 5757.1 ms (64 frames, 9.34 ms/frame AR, audio 5.12 s, RTF 1.124)
[Perf] PromptBuild 2.8 ms
[Perf] Prefill 17.0 ms (T_ctx prefill)
[Perf] TTFA 741.4 ms (first frame codes)
[Perf] TalkerDecode 137.6 ms (64 frames, 2.15 ms/frame)
[Perf] CodePredictor 271.2 ms (4.24 ms/frame)
[Perf] HostCompose 0.2 ms (c0 sample + next emb)
[Perf] CodecDecode 62.6 ms
[Perf] Total 1208.2 ms (64 frames, 6.39 ms/frame AR, audio 5.12 s, RTF 0.236)
[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...
@@ -153,3 +153,13 @@ load_backend: loaded CPU backend from /mnt/workspace/git/qwentts.cpp/build/libgg
[Cossim] CodesFull exact: 2.38% (1008 values)
[Cossim] Audio cos: 0.012436
[Cossim] WAV stft_cos: 0.190492 samples: 120960
[GGML] Cmd: ../build/qwen-tts --model ../models/qwen-talker-1.7b-base-Q8_0.gguf --codec ../models/qwen-tokenizer-12hz-Q8_0.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 2.7 ms
[Perf Fused] Prefill 15.8 ms (T_ctx prefill)
[Perf Fused] TTFA 1607.6 ms (first frame codes)
[Perf Fused] TalkerDecode 136.0 ms (64 frames, 2.13 ms/frame)
[Perf Fused] CodePredictor 419.3 ms (6.55 ms/frame)
[Perf Fused] HostCompose 0.1 ms (c0 sample + next emb)
[Perf Fused] CodecDecode 841.1 ms
[Perf Fused] Total 2137.6 ms (64 frames, 8.68 ms/frame AR, audio 5.12 s, RTF 0.417)
[Cossim] CodesFullFused exact: 100.00% (1024 values)