Upgrade qwentts.cpp to September 28 revision and ABI v5

.github/workflows/metal-wheel.yml

@@ -7,7 +7,7 @@ name: Apple Silicon Metal wheel
       qwentts_ref:
         description: qwentts.cpp revision (must match the binding's verified ABI)
         type: string
-        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
+        default: 6fae92914045cd83364d2845ceaa0f7969727319
   workflow_call:
     inputs:
       pypi:
@@ -16,7 +16,7 @@ name: Apple Silicon Metal wheel
         default: false
       qwentts_ref:
         type: string
-        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
+        default: 6fae92914045cd83364d2845ceaa0f7969727319
 
 permissions:
   contents: read
@@ -26,7 +26,7 @@ jobs:
     runs-on: macos-14
     env:
       MACOSX_DEPLOYMENT_TARGET: "14.0"
-      QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3
         with:

.github/workflows/publish-hf-wheels.yml

@@ -6,7 +6,7 @@ name: Publish Hugging Face Wheels
       qwentts_ref:
         description: qwentts.cpp commit, tag, or branch to build
         required: true
-        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
+        default: 6fae92914045cd83364d2845ceaa0f7969727319
       hf_repo_id:
         description: Hugging Face dataset repo for wheel hosting
         required: true
@@ -17,13 +17,13 @@ permissions:
 
 env:
   HF_WHEEL_REPO_ID: ${{ inputs.hf_repo_id || vars.HF_WHEEL_REPO_ID || 'andito/qwentts-cpp-python-wheels' }}
-  QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+  QWENTTS_REF: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
 
 jobs:
   build-metal-wheel:
     uses: ./.github/workflows/metal-wheel.yml
     with:
-      qwentts_ref: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+      qwentts_ref: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
 
   build-wheels:
     name: Linux ${{ matrix.arch }} CUDA ${{ matrix.cuda_version }} HF wheel

.github/workflows/publish.yml

@@ -9,7 +9,7 @@ name: Publish
       qwentts_ref:
         description: qwentts.cpp commit, tag, or branch to build
         required: true
-        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
+        default: 6fae92914045cd83364d2845ceaa0f7969727319
 
 permissions:
   contents: read
@@ -19,7 +19,7 @@ jobs:
     uses: ./.github/workflows/metal-wheel.yml
     with:
       pypi: true
-      qwentts_ref: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+      qwentts_ref: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
 
   build-wheels:
     name: Linux ${{ matrix.arch }} CUDA 12.8 wheel
@@ -46,7 +46,7 @@ jobs:
       QWENTTS_CPP_BACKEND: cuda
       QWENTTS_CPP_BUILD_JOBS: ${{ matrix.build_jobs }}
       QWENTTS_CPP_WHEEL_BUILD_TAG: 1cu128
-      QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
       CUDA_ARCHITECTURES: ${{ matrix.cuda_architectures }}
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3

.github/workflows/wheels.yml

@@ -6,13 +6,13 @@ name: qwentts-cpp-python wheels
       qwentts_ref:
         description: qwentts.cpp commit, tag, or branch to build
         required: true
-        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
+        default: 6fae92914045cd83364d2845ceaa0f7969727319
 
 jobs:
   macos-metal:
     uses: ./.github/workflows/metal-wheel.yml
     with:
-      qwentts_ref: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+      qwentts_ref: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
 
   linux-cuda:
     name: Linux ${{ matrix.arch }} CUDA ${{ matrix.cuda_version }} wheel
@@ -91,7 +91,7 @@ jobs:
       QWENTTS_CPP_BACKEND: cuda
       QWENTTS_CPP_BUILD_JOBS: ${{ matrix.build_jobs }}
       QWENTTS_CPP_WHEEL_BUILD_TAG: ${{ matrix.wheel_build_tag }}
-      QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
       CUDA_ARCHITECTURES: ${{ matrix.cuda_architectures }}
     steps:
       - uses: actions/checkout@v4
@@ -175,7 +175,7 @@ jobs:
       QWENTTS_CPP_BACKEND: cpu
       QWENTTS_CPP_BUILD_JOBS: "2"
       QWENTTS_CPP_WHEEL_BUILD_TAG: 1cpu
-      QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || '6fae92914045cd83364d2845ceaa0f7969727319' }}
     steps:
       - uses: actions/checkout@v4
 

README.md

@@ -143,8 +143,23 @@ python -m twine check --strict wheelhouse/*.whl
 ### Native ABI compatibility
 
 The CI wheel build defaults to qwentts.cpp
-`7df559a8ca25f66fee02970514ebe5f01dee9055`, which retains ABI v2 and includes
-the latest static-graph, streaming-decode, and widened voice-route changes.
+`6fae92914045cd83364d2845ceaa0f7969727319` (September 28, 2026), which requires
+ABI v5 and includes the upstream batched compute worker, codec memory
+controls, language/model queries, and latest ggml fork update.
+
+ABI v5 changes the initialization and synthesis struct layouts; libraries
+built from the previous ABI v2 pin must be rebuilt. Python's `do_sample=False`
+and `subtalker_do_sample=False` now select greedy decoding by setting the
+corresponding native temperature to zero. Buffered codec chunk size is set
+once with `QwenTTS(..., codec_chunk_sec=24.0)` or
+`QwenTTS.from_pretrained(..., codec_chunk_sec=24.0)`, rather than on
+`synthesize()`. Zero selects the native default. The codec derives its own
+left context; `codec_left_context_sec` is no longer accepted. Streaming's
+`codec_chunk_sec` still controls Python output packet size independently.
+The wrapper continues to serialize requests on each context (native
+`max_batch=1`). `tts.language_names()` lists the model's languages (synthesis
+also accepts `lang="auto"`), and `tts.model_type()` reports `base`,
+`custom_voice`, or `voice_design`.
 
 The loader verifies this native revision before calling functions that write
 ctypes parameter buffers. Upstream does not expose an ABI-version or struct-size
@@ -229,15 +244,15 @@ limit flushes any remaining audio as a short packet, including utterances
 shorter than the requested first packet. Cancellation or errors discard the
 unfinished packet; closing the iterator requests native cancellation.
 
-Packet assembly happens in Python, using the existing verified ABI v2 library.
+Packet assembly happens in Python, using the verified ABI v5 library.
 The native decoder still emits its fixed 1→2→4→8-frame ramp, then 8-frame
 chunks. Consequently, first packets of 1, 2, 4, and 8 frames become available
 after native output has reached 1, 3, 7, and 15 frames respectively (or earlier
 at end-of-speech). Packet boundaries preserve every PCM sample but do not
 change native decode scheduling. Later packets may become available together
 when a native callback spans several packet boundaries; this is not a timed
-playback scheduler. `codec_left_context_sec` is ignored by the stateful native
-stream. Buffered `synthesize()` retains its native codec chunking behavior.
+playback scheduler. Buffered `synthesize()` uses the native codec chunk size
+configured when the context is created; the native codec derives its left context.
 
 `last_stream_profile` keeps `first_callback_*` and `callback_count` for raw
 native callbacks. `first_packet_ready_ms`, `first_packet_audio_s`, and
@@ -281,7 +296,7 @@ the larger first packet is assembled by the binding.
 
 ## Cached voice references
 
-qwentts.cpp ABI v2 can skip reference WAV encoding for Base voice cloning by
+qwentts.cpp can skip reference WAV encoding for Base voice cloning by
 passing precomputed latents:
 
 - `.spk`: raw float32 speaker embedding from `qwen-codec --talker`

scripts/build_native.py

@@ -4,7 +4,6 @@
 import argparse
 from contextlib import contextmanager
 import os
-import re
 import shutil
 import shlex
 import subprocess
@@ -32,7 +31,7 @@ def metal_shader_compatibility(source: Path, enabled: bool):
     if not enabled:
         yield
         return
-    shader = source / "ggml/src/ggml-metal/ggml-metal.metal"
+    shader = source / "ggml/src/ggml-metal/kernels/unary.metal"
     original = shader.read_text()
     old = "dst_ptr[i0] = (T) args.val;"
     new = "dst_ptr[i0] = (T) ((TC) args.val);"
@@ -83,59 +82,22 @@ def native_logging_compatibility(source: Path):
 
 @contextmanager
 def native_diagnostic_compatibility(source: Path):
-    """Route the pinned source's remaining direct diagnostics through qt_log.
+    """Keep a missing tokenizer visible at error log level.
 
-    These headers predate qwentts.cpp's callback API. Keep severity when
-    converting them, and restore the checkout once the wheel is built.
+    Upstream routes diagnostics through qt_log now, but reports this load
+    failure at INFO. Restore the checkout after building.
     """
-    expected = {
-        "audio-io.h": 10, "audio-resample.h": 2, "bpe.h": 9,
-        "code-predictor-forward.h": 5, "code-predictor-weights.h": 4,
-        "convnext-block.h": 3, "dac-decoder-v2.h": 2,
-        "encoder-downsample.h": 2, "encoder-transformer.h": 3,
-        "gguf-weights.h": 9, "graph-arena.h": 1, "kv-cache.h": 3,
-        "prompt-builder.h": 13, "quantizer-decode.h": 4,
-        "quantizer-encode.h": 2, "rvq-file.h": 7, "seanet-encoder.h": 3,
-        "speaker-encoder-extract.h": 6, "speaker-encoder-weights.h": 3,
-        "talker-forward.h": 7, "talker-weights.h": 3,
-        "tokenizer-transformer.h": 3, "wav.h": 7, "weight-ctx.h": 2,
-    }
-    pattern = re.compile(r"fprintf\(stderr,\s*(.*?)\);", re.DOTALL)
-    originals = {}
+    path = source / "src/bpe.h"
+    original = path.read_text()
+    old = 'qt_log(QT_LOG_INFO, "[BPE] Tokenizer not found in %s", gguf_path);'
+    new = 'qt_log(QT_LOG_ERROR, "[BPE] Tokenizer not found in %s", gguf_path);'
+    if original.count(old) != 1:
+        raise SystemExit("Native tokenizer diagnostic changed; review the logging compatibility fix")
     try:
-        for name, count in expected.items():
-            path = source / "src" / name
-            original = path.read_text()
-            if len(pattern.findall(original)) != count or not original.startswith("#pragma once\n"):
-                raise SystemExit(f"Native diagnostics changed in {name}; review the logging compatibility fix")
-
-            def replace(match):
-                args = match.group(1)
-                format_match = re.search(r'"((?:[^"\\]|\\.)*)"', args)
-                if format_match is None:
-                    raise SystemExit(f"Native diagnostic format changed in {name}")
-                message = format_match.group(1).lower()
-                if "warning" in message or "no spk_enc." in message:
-                    level = "QT_LOG_WARN"
-                elif any(word in message for word in (
-                    "fatal", "failed", "cannot", "oom", "unsupported",
-                    "not a valid", "not found", "no audio data", "unknown format",
-                )):
-                    level = "QT_LOG_ERROR"
-                else:
-                    level = "QT_LOG_INFO"
-                # qt_log and the Python trampoline each add the line ending.
-                args = args.replace(r'\n"', '"')
-                return f"qt_log({level}, {args});"
-
-            transformed = pattern.sub(replace, original)
-            transformed = transformed.replace("#pragma once\n", '#pragma once\n#include "qt-error.h"\n', 1)
-            originals[path] = original
-            path.write_text(transformed)
+        path.write_text(original.replace(old, new))
         yield
     finally:
-        for path, original in originals.items():
-            path.write_text(original)
+        path.write_text(original)
 
 
 def find_first(root: Path, patterns: list[str]) -> Path | None: