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

abi, pipeline, cli: conform qwentts on omnivoice convention

Browse Source
This commit is contained in:
Pascal
2026-05-14 21:42:52 +02:00
parent 259e7059f6
commit 38bf6d762a
11 changed files with 602 additions and 278 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/abi-c.c
+49 -19
View File
@@ -1,4 +1,4 @@
/* tests/abi-c.c : link-only ABI smoke test for qwen.h.
/* tests/abi-c.c: link-only ABI smoke test for qwen.h.
*
* Compiled in pure C99 with -Wall -Werror -pedantic. The purpose of this
* test is NOT to run a full synthesis (no GGUF loaded, no model required);
@@ -22,11 +22,23 @@
#include <stdlib.h>
#include <string.h>
static bool stub_cancel(void * ud) {
(void) ud;
return false;
}
static bool stub_on_chunk(const float * samples, int n_samples, void * ud) {
(void) samples;
(void) n_samples;
(void) ud;
return true;
}
/* Counter incremented by the stub log callback. The probe checks that at
* least one log line was routed through the callback by triggering a
* qt_init failure (which emits a [qwen] ERROR line via qt_log). */
static int g_log_lines = 0;
static enum qt_log_level g_last_log_level = QT_LOG_DEBUG;
* qt_init failure (which emits a [Qwen] ERROR line via qt_log). */
static int g_log_lines = 0;
static enum qt_log_level g_last_log_level = QT_LOG_DEBUG;
static char g_last_log_msg[512] = { 0 };
static void stub_log(enum qt_log_level level, const char * msg, void * user_data) {
@@ -55,29 +67,44 @@ int main(void) {
struct qt_tts_params params;
qt_tts_default_params(&params);
/* Sanity-check a few default values, including the abi_version. */
if (params.max_new_tokens != 2048 || params.temperature != 0.9f) {
/* Sanity-check a few default values, including the abi_version and
* the new use_fa / clamp_fp16 / on_chunk / chunk_duration_sec slots. */
if (params.max_new_tokens != 2048 || params.chunk_duration_sec <= 0.0f) {
fprintf(stderr, "[Probe] default values do not match\n");
return 1;
}
if (iparams.abi_version != QT_ABI_VERSION || params.abi_version != QT_ABI_VERSION) {
fprintf(stderr, "[Probe] abi_version not set by qwen_*_default_params\n");
fprintf(stderr, "[Probe] abi_version not set by qt_*_default_params\n");
return 1;
}
if (!iparams.use_fa || iparams.clamp_fp16) {
fprintf(stderr, "[Probe] init_params defaults do not match (use_fa=true, clamp_fp16=false)\n");
return 1;
}
if (QT_CODEC_SAMPLE_RATE != 24000) {
fprintf(stderr, "[Probe] QT_CODEC_SAMPLE_RATE is not 24000\n");
return 1;
}
/* Touch every output struct field so the compiler validates the
* layout end-to-end without ever needing a model. */
/* Touch every reference-pointer field, every callback typedef and
* every output struct field so the compiler validates the layout
* end-to-end without ever needing a model. */
params.cancel = stub_cancel;
params.cancel_user_data = NULL;
params.on_chunk = stub_on_chunk;
params.on_chunk_user_data = NULL;
struct qt_audio audio = { 0 };
qt_audio_free(&audio);
/* Install the log callback before the failing init so the [qwen]
/* Install the log callback before the failing init so the [Qwen]
* ERROR line lands on stub_log instead of stderr. */
qt_log_set(stub_log, NULL);
/* Call every entry through its early-return path. qt_init returns
* NULL on missing talker_path / codec_path, qt_synthesize fails on
* NULL handle, qt_free is safe on NULL. None of these load a model,
* but the linker must resolve every name to satisfy the call. */
* NULL on missing paths, qt_synthesize / qt_duration_sec_to_tokens
* fail on NULL handle, qt_free is safe on NULL. None of these load a
* model, but the linker must resolve every name to satisfy the call. */
struct qt_context * dummy = qt_init(NULL);
if (dummy != NULL) {
fprintf(stderr, "[Probe] qt_init(NULL) was supposed to return NULL\n");
@@ -102,17 +129,14 @@ int main(void) {
return 6;
}
if (g_last_log_level != QT_LOG_ERROR) {
fprintf(stderr, "[Probe] last log level was %d, expected %d\n", (int) g_last_log_level,
(int) QT_LOG_ERROR);
fprintf(stderr, "[Probe] last log level was %d, expected %d\n", (int) g_last_log_level, (int) QT_LOG_ERROR);
return 7;
}
printf("[Probe] qt_log_set routed %d line(s), last: '%s'\n", g_log_lines, g_last_log_msg);
printf("[Probe] qt_last_error reads '%s'\n", err);
/* abi_version validation : a struct claiming a future ABI must be
* rejected up front, before any allocation. Both paths are filled
* with placeholders so the NULL guard does not short-circuit the
* abi_version branch. */
* rejected up front, before any allocation. */
struct qt_init_params future_iparams;
qt_init_default_params(&future_iparams);
future_iparams.talker_path = "irrelevant.gguf";
@@ -132,8 +156,14 @@ int main(void) {
return 3;
}
int frames = qt_duration_sec_to_tokens(NULL, 1.0f);
if (frames < 1) {
fprintf(stderr, "[Probe] qt_duration_sec_to_tokens returned %d, expected >= 1\n", frames);
return 4;
}
/* Restore the default stderr fallback before exit so the trailing
* [qwen] log lines from the cleanup paths land where the user
* [Qwen] log lines from the cleanup paths land where the user
* expects them. */
qt_log_set(NULL, NULL);
tests/debug-clone-cossim.py
+7 -10
View File
@@ -13,7 +13,7 @@ length, the shorter one padded with tts_pad / truncated as needed.
Cote Python the speaker embedding is captured directly via
model.extract_speaker_embedding, and the reference codec frames via
model.speech_tokenizer.encode. Both intermediates land as speaker-emb.bin
model.speech_tokenizer.encode. Both intermediates land as spk-emb.bin
and ref-codes.bin and are compared against the C++ side dumps emitted
by pipeline-tts.cpp when --ref-wav and --ref-text are set.
@@ -38,9 +38,6 @@ CKPT = "../checkpoints/Qwen3-TTS-12Hz-1.7B-Base"
DUMP_CPP = "cpp/clone"
DUMP_PT = "python/clone"
DEFAULT_REF_AUDIO = "../examples/freeman.wav"
DEFAULT_REF_TEXT = "../examples/freeman.txt"
# Mode B adds two pre-talker stages to the standard list : the speaker
# embedding extracted from the reference audio (ECAPA forward, projected to
# talker hidden), and the reference codec frames at 12.5 Hz. Plus three
@@ -65,7 +62,7 @@ STAGES_CLONE = cc.STAGES_STANDARD + [
("SpkBlock3", "spk-block3.bin"),
("SpkMFA", "spk-mfa.bin"),
("SpkASP", "spk-asp.bin"),
("SpeakerEmb", "speaker-emb.bin"),
("SpeakerEmb", "spk-emb.bin"),
]
def install_clone_hooks(model, dump_dir):
@@ -271,9 +268,9 @@ def dump_mel_mag_python(ref_wav, dump_dir):
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--prompt", default="../examples/prompt.txt")
ap.add_argument("--ref-wav", default=DEFAULT_REF_AUDIO,
ap.add_argument("--ref-wav", default="../examples/freeman.wav",
help="reference WAV path for voice cloning")
ap.add_argument("--ref-text-file", default=DEFAULT_REF_TEXT,
ap.add_argument("--ref-text", default="../examples/freeman.txt",
help="path to a UTF-8 file with the transcript of ref-wav")
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--lang", default="english")
@@ -297,7 +294,7 @@ def main():
with open(args.prompt, "r", encoding="utf-8") as f:
text = f.read().strip()
with open(args.ref_text_file, "r", encoding="utf-8") as f:
with open(args.ref_text, "r", encoding="utf-8") as f:
ref_text = f.read().strip()
print(f"[Input] Prompt: {len(text)} chars: {text[:60]}{'...' if len(text) > 60 else ''}")
print(f"[Input] RefAudio: {args.ref_wav}")
@@ -356,7 +353,7 @@ def main():
# Extract speaker embedding via ECAPA forward, projected to talker hidden.
spk_emb = model.extract_speaker_embedding(audio=ref_wav, sr=ref_sr)
print(f"[Python] SpeakerEmb shape: {tuple(spk_emb.shape)} dtype: {spk_emb.dtype}")
cc.save_dump(os.path.join(DUMP_PT, "speaker-emb.bin"), spk_emb)
cc.save_dump(os.path.join(DUMP_PT, "spk-emb.bin"), spk_emb)
# Encode the reference audio to 16 codebook codes at 12.5 Hz. The encode
# call returns shape [T_codec, K=16] after the internal transpose, while
@@ -460,7 +457,7 @@ def main():
"--seed", str(args.seed),
"--text", text,
"--ref-wav", args.ref_wav,
"--ref-text", ref_text,
"--ref-text", args.ref_text,
"--lang", args.lang,
"--max-new", str(args.max_new_tokens),
"--dump", DUMP_CPP,
tests/debug-customvoice-cossim.py
+1 -3
View File
@@ -35,12 +35,10 @@ CKPT = "../checkpoints/Qwen3-TTS-12Hz-1.7B-CustomVoice"
DUMP_CPP = "cpp/customvoice"
DUMP_PT = "python/customvoice"
DEFAULT_SPEAKER = "vivian"
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--prompt", default="../examples/prompt.txt")
ap.add_argument("--speaker", default=DEFAULT_SPEAKER,
ap.add_argument("--speaker", default="vivian",
help="speaker preset key (lowercase), validated by the model")
ap.add_argument("--instruct", default="",
help="optional style instruction, empty disables the instruct prefix")
tests/debug-tts-cossim.py
+1 -3
View File
@@ -32,12 +32,10 @@ CKPT = "../checkpoints/Qwen3-TTS-12Hz-1.7B-VoiceDesign"
DUMP_CPP = "cpp/tts"
DUMP_PT = "python/tts"
DEFAULT_INSTRUCT = "male, young adult, moderate pitch"
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--prompt", default="../examples/prompt.txt")
ap.add_argument("--instruct", default=DEFAULT_INSTRUCT,
ap.add_argument("--instruct", default="male, young adult, moderate pitch",
help="natural language style instruction")
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--lang", default="english")