FreeVideo 0.4.30

THIRD_PARTY_NOTICES.md

@@ -20,6 +20,9 @@ FreeVideo code uses [Apache-2.0](LICENSE). Its dependencies and models retain th
 | [ONNX Runtime](https://github.com/microsoft/onnxruntime) 1.28.2 (CPU), © Microsoft Corporation; its third-party notices are shipped inside the sherpa-onnx wheel | MIT |
 | [Eigen](https://gitlab.com/libeigen/eigen) 5.0.1, compiled into sherpa-onnx and ONNX Runtime; source: https://gitlab.com/libeigen/eigen/-/tree/5.0.1 | MPL-2.0 |
 | [Silero VAD](https://github.com/snakers4/silero-vad) v5.0 voice-activity model, © 2020-present Silero Team (optional download with subtitles, license file downloaded with the model; the onnx-community export is downloaded with the music model) | [MIT](freevideo_engine/licenses/Silero-VAD-MIT.txt) |
+| [ACE-Step 1.5](https://github.com/ace-step/ACE-Step-1.5) XL turbo music model, © 2026 ACEStep (optional download for music; inference subset adapted in `freevideo_engine/music/acestep_model.py`) | [MIT](freevideo_engine/licenses/ACE-Step-MIT.txt) |
+| [vector-quantize-pytorch](https://github.com/lucidrains/vector-quantize-pytorch), © 2020 Phil Wang (residual FSQ decoding adapted in `freevideo_engine/music/acestep_model.py`) | [MIT](freevideo_engine/licenses/Vector-Quantize-PyTorch-MIT.txt) |
+| [Qwen3-Embedding-0.6B](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B), © Alibaba Cloud (text encoder of the music model, downloaded with it) | [Apache-2.0](freevideo_engine/licenses/Qwen3-Embedding-Apache-2.0.txt) |
 | [NVIDIA NeMo stt_en_conformer_ctc_small](https://huggingface.co/nvidia/stt_en_conformer_ctc_small), © NVIDIA, exported to ONNX and quantized to int8 by the sherpa-onnx project, changed from the original; not endorsed by NVIDIA (optional download for English subtitles; [notice](freevideo_engine/licenses/NeMo-stt_en_conformer_ctc_small-NOTICE.txt)) | [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/) |
 | [OpenAI Whisper tiny](https://github.com/openai/whisper), © 2022 OpenAI, exported to ONNX and quantized to int8 by the sherpa-onnx project (optional download for spoken-language detection) | [MIT](freevideo_engine/licenses/Whisper-MIT.txt) |
 | [SenseVoiceSmall](https://huggingface.co/FunAudioLLM/SenseVoiceSmall) by FunASR / FunAudioLLM (Alibaba Group), converted to ONNX and quantized to int8 by the sherpa-onnx project, used unchanged under its original name (optional download for multilingual subtitles; [notice](freevideo_engine/licenses/SenseVoice-NOTICE.txt)) | [FunASR Model Open Source License Agreement v1.1](freevideo_engine/licenses/FunASR-MODEL_LICENSE-v1.1.txt) |

freevideo.ps1

@@ -137,10 +137,14 @@ try {
                   if ($pythonExit -eq 0) { $installed = $true; break }
                   if (-not (Test-FreeVideoSlowDownloadFailure $pythonOutput)) { $allNetworkFailures = $false }
                   Write-Host "Python install source $attempt returned exit code $pythonExit. Details retained in $bootstrapLog"
+                  $pythonReason = Get-FreeVideoFailureReason $pythonOutput
+                  if ($pythonReason) { Write-Host "uv reported: $pythonReason" }
                 }
               }
             }
-            if (-not $installed) { throw "Python installation failed (uv exit $pythonExit). Check your network or connection mode in Downloads. See $bootstrapLog and retry." }
+            $pythonDetail = ''
+            if ($pythonReason) { $pythonDetail = " uv reported: $pythonReason." }
+            if (-not $installed) { throw "Python installation failed (uv exit $pythonExit).$pythonDetail Check your network or connection mode in Downloads. See $bootstrapLog and retry." }
             $freevideoPython = (& $uv python find --managed-python $versions.python | Select-Object -Last 1).Trim()
             if ($LASTEXITCODE -ne 0 -or -not (Test-Path -LiteralPath $freevideoPython)) { throw 'Managed Python could not be located.' }
         } finally {

freevideo_engine/activation_staging.py

@@ -1,6 +1,75 @@
 """Contiguous D2H head groups with double-buffered CPU projection tiles."""
+import collections
 import torch
 
+# Every refused locked-page copy of this process, residual stream and head
+# readouts alike, so a report shows whether the fallback fired, which copy was
+# refused and what CUDA said. The newest entries are kept; the count is exact.
+_REFUSED = collections.deque(maxlen=32)
+_refused_total = 0
+
+
+def record_refusal(source, copy, error):
+    global _refused_total
+    lines = str(error).strip().splitlines()
+    _REFUSED.append(dict(source=source, copy=copy, error=(lines[0] if lines else '')[:200]))
+    _refused_total += 1
+
+
+def refusals_mark():
+    return _refused_total
+
+
+def refusals_since(mark):
+    """Refusals recorded after `mark`, as plain dicts for the sampling metrics."""
+    count = min(_refused_total - mark, len(_REFUSED))
+    return [dict(row) for row in list(_REFUSED)[len(_REFUSED) - count:]] if count > 0 else []
+
+
+class LockedCopies:
+    """Asynchronous copies through page-locked buffers that continue in pageable memory.
+
+    Windows can refuse a queued copy with "invalid argument" when the system
+    takes locked pages back. The queued work is finished first, so a sticky
+    error is raised as itself rather than hidden by the retry. The copy then
+    goes through a pageable temporary, which holds the same bytes, and the
+    remaining copies of these buffers skip the locked path. A failed device
+    allocation is not a refusal: it is raised for the caller to recover from,
+    and the buffers stay locked.
+    """
+
+    def __init__(self):
+        self.locked = True
+        self.fallbacks = 0
+
+    def _leave(self, device, copy, error):
+        record_refusal('readout', copy, error)
+        if device.type == 'cuda':
+            torch.cuda.synchronize(device)
+        self.locked = False
+        self.fallbacks += 1
+
+    def to_host(self, destination, source):
+        if self.locked:
+            try:
+                destination.copy_(source, non_blocking=True)
+                return
+            except torch.cuda.OutOfMemoryError:
+                raise
+            except RuntimeError as error:
+                self._leave(source.device, 'to_host', error)
+        destination.copy_(source.to('cpu'))
+
+    def to_device(self, source, device):
+        if self.locked:
+            try:
+                return source.to(device, non_blocking=True)
+            except torch.cuda.OutOfMemoryError:
+                raise
+            except RuntimeError as error:
+                self._leave(torch.device(device), 'to_device', error)
+        return source.clone().to(device)
+
 
 class HeadReadouts:
     def __init__(self, rows, head_widths, dtype, tile_rows):
@@ -11,14 +80,15 @@ def __init__(self, rows, head_widths, dtype, tile_rows):
         self.groups = [torch.empty((self.capacity, width), dtype=dtype, pin_memory=True) for width in head_widths]
         self.tiles = [torch.empty((tile_rows, sum(head_widths)), dtype=dtype, pin_memory=True) for _ in range(2)]
         self.events = [None, None]
+        self.copies = LockedCopies()
 
     def supports(self, rows, head_widths, dtype, tile_rows):
         return rows <= self.capacity and tuple(head_widths) == self.widths and dtype == self.dtype and tile_rows == self.tile_rows
 
     def store(self, group, values):
         if values.shape[1] != self.widths[group] or not values.is_contiguous():
             raise ValueError('Head readouts must be contiguous [rows, head channels]')
-        self.groups[group][:len(values)].copy_(values, non_blocking=True)
+        self.copies.to_host(self.groups[group][:len(values)], values)
 
     def project(self, module, output, rows):
         # The caller synchronizes once after all D2H groups have been written.
@@ -33,7 +103,7 @@ def project(self, module, output, rows):
             for group, width in zip(self.groups, self.widths):
                 tile[:, offset:offset + width].copy_(group[start:start + count])
                 offset += width
-            device_values = tile.to(output.device, non_blocking=True)
+            device_values = self.copies.to_device(tile, output.device)
             done = torch.cuda.Event()
             done.record(stream)
             self.events[slot] = done

freevideo_engine/adaptive.py

@@ -363,14 +363,18 @@ def nonlocal_exhausted(failure):
     Page-locked host memory counts against that budget. A 20 s request failed
     with non-local usage at 13.14 of 13.24 GiB, 10.68 GiB of VRAM free and no
     caching-allocator OOM; the failed allocation itself is not in the usage.
+    A 1920x1088 request failed with usage at 13.9 of 15.2 GiB: the allocation
+    that did not fit was the 2 GiB residual buffer, so the margin is the pin
+    reserve rather than one GiB.
     """
     gpu = failure.get('gpu') if isinstance(failure.get('gpu'), dict) else {}
     values = [gpu.get(key) for key in ('nonlocal_budget_bytes', 'nonlocal_usage_bytes',
                                        'local_budget_bytes', 'local_usage_bytes')]
     if not all(type(value) is int and value >= 0 for value in values) or not values[0]:
         return False
     budget, usage, local_budget, local_usage = values
-    return (usage >= budget - 2**30 and local_usage <= local_budget - 2 * 2**30
+    from .windows_gpu_memory import NONLOCAL_PIN_RESERVE
+    return (usage >= budget - NONLOCAL_PIN_RESERVE and local_usage <= local_budget - 2 * 2**30
             and not gpu.get('allocator_ooms'))
 
 

freevideo_engine/agent/comfy.py

@@ -354,7 +354,8 @@ def image_start(self, action, value):
             code = getattr(error, 'code', str(error))
             code = code if code in IMAGE_ERRORS else 'invalid_request'
             raise JobError(code, IMAGE_ERRORS[code], retryable=code in IMAGE_WAIT_ERRORS,
-                           kind='image', report_id=getattr(error, 'report_id', None)) from None
+                           kind='image', report_id=getattr(error, 'report_id', None),
+                           other_usage=getattr(error, 'status', {}).get('other_usage')) from None
 
     def image_status(self, key):
         return self.image.status(key)

freevideo_engine/agent/errors.py

@@ -12,6 +12,7 @@
 REQUEST_ID = ContextVar('agent_request_id', default=None)
 # The work session a token request belongs to (sessions.py); jobs and productions record it.
 SESSION = ContextVar('agent_session', default=None)
+_UNREAD_USAGE = object()
 # The operation digest a confirmation approves: method, route, path values and body without ref.
 CONFIRM = ContextVar('agent_confirm', default=None)
 # code: HTTP status, retryable, public message, concrete recovery action
@@ -128,6 +129,8 @@
 
 # Static bilingual image reasons; backend exception strings never enter the API.
 IMAGE_REASONS = {
+    'small_memory': ('This Mac has less than 16 GB of memory and cannot generate pictures.',
+                     '这台 Mac 的内存不足 16 GB，无法生成图片。'),
     'no_memory': ('Not enough GPU memory, even at the smallest size. Close other apps using the GPU, then try again.',
                   '显存不够，最小尺寸也没能生成。关掉其他占用显卡的程序后再试。'),
     'busy_video': ('A video is generating. Try again when it finishes.', '视频正在生成，完成后再试。'),
@@ -164,6 +167,7 @@ def image_failure(code, report_id=None):
 # Recovery is a function of the committed task and phase, not just an error code.
 IMAGE_ERRORS = {
     **LORA_MESSAGE_EN,
+    'small_memory': IMAGE_REASONS['small_memory'][0],
     'invalid_request': 'The picture request contains invalid fields.',
     'invalid_prompt': 'The picture prompt must contain text and be at most 2000 characters.',
     'invalid_dimensions': 'Picture width and height must be integers from 256 to 2048.',
@@ -413,7 +417,26 @@ def failure_reason_en(error):
 
 
 def reasons(error):
-    return dict(error_zh=failure_reason(error), error_en=failure_reason_en(error))
+    from ..other_usage import fact, join
+    result = dict(error_zh=failure_reason(error), error_en=failure_reason_en(error))
+    for key, zh in (('error_zh', True), ('error_en', False)):
+        suffix = fact(error.get('other_usage'), zh=zh)
+        if suffix and suffix not in result[key]:
+            result[key] = join(result[key], suffix, zh=zh)
+    return result
+
+
+def capture_other_usage(error):
+    from ..other_usage import at_failure, fact, join, MEMORY_ERRORS
+    if error and error.get('code') in MEMORY_ERRORS and 'other_usage' not in error:
+        error = dict(error, other_usage=at_failure(error['code']))
+    if error and error.get('other_usage') and error.get('reason'):
+        error = dict(error, reason=dict(error['reason']))
+        for lang in ('en', 'zh'):
+            suffix = fact(error['other_usage'], zh=lang == 'zh')
+            if suffix and suffix not in error['reason'].get(lang, ''):
+                error['reason'][lang] = join(error['reason'].get(lang, ''), suffix, zh=lang == 'zh')
+    return error
 
 
 def with_reason(error):
@@ -422,7 +445,7 @@ def with_reason(error):
 
 def build(code, *, details=None, request=None, validation=False, message=None, next_action=None,
           action_required=True, retry_after_seconds=None, owner=None, kind=None, stage='request', resumable=True,
-          report_id=None, memory=None):
+          report_id=None, memory=None, other_usage=_UNREAD_USAGE):
     code = code if code in ERRORS else 'internal_error'
     _, retryable, default_message, _ = ERRORS[code]
     action = recovery(code, kind=kind, stage=stage, resumable=resumable)
@@ -438,6 +461,9 @@ def build(code, *, details=None, request=None, validation=False, message=None, n
         result.update(image_failure(code, report_id))
         if code == 'no_memory' and memory == 'host':
             result['memory'] = 'host'
+    if other_usage is not _UNREAD_USAGE:
+        result['other_usage'] = other_usage
+    result = capture_other_usage(result)
     if action_required:
         result['next'] = next_action if validation and next_action else action
     if code == 'busy':

freevideo_engine/agent/jobs.py

@@ -88,7 +88,8 @@ def view(self):
         return errors.build(self.code, details=self.details.get('details'),
                             owner=self.details.get('owner'),
                             retry_after_seconds=self.details.get('retry_after_seconds'),
-                            kind=self.details.get('kind'), report_id=self.details.get('report_id'))
+                            kind=self.details.get('kind'), report_id=self.details.get('report_id'),
+                            **({'other_usage': self.details['other_usage']} if 'other_usage' in self.details else {}))
 
 
 class ManagedByProduction(JobError):
@@ -681,16 +682,19 @@ def touch(self, row, **changes):
 
     def finish(self, row, status, *, result=None, error=None, save=True):
         if error is not None:
+            error = errors.capture_other_usage(error)
             code = error.get('code', 'internal_error')
             report_id = error.get('report_id')
             if code == 'invalid_json' and row['kind'] in ('vlm', 'enhance'):
                 code = 'vlm_invalid_json'
-            reason_zh, reason_en = errors.failure_reason(dict(error, code=code)), errors.failure_reason_en(dict(error, code=code))
+            failure_reasons = errors.reasons(dict(error, code=code))
+            reason_zh, reason_en = failure_reasons['error_zh'], failure_reasons['error_en']
             if code != error.get('code') or not error.get('request_id') or row['kind'] == 'image':
                 error = errors.build(code, details=error.get('details'),
                     request=row.get('request_id'), action_required=code != 'cancelled',
                     kind=row['kind'], stage='terminal', resumable=not row['spec'].get('refine_from'),
-                    report_id=error.get('report_id'), memory=error.get('memory'))
+                    report_id=error.get('report_id'), memory=error.get('memory'),
+                    other_usage=error.get('other_usage'))
                 if row['kind'] == 'image':
                     reason_zh, reason_en = error['error_zh'], error['error_en']
         if error is not None:
@@ -1235,7 +1239,8 @@ def view(self, row, *, detail=True):
             if row['kind'] == 'image':
                 result['error'] = errors.build(result['error']['code'], kind='image', stage='terminal',
                     details=result['error'].get('details'), request=row.get('request_id'),
-                    action_required=result['error']['code'] != 'cancelled', report_id=row['error'].get('report_id'))
+                    action_required=result['error']['code'] != 'cancelled', report_id=row['error'].get('report_id'),
+                    other_usage=row['error'].get('other_usage'))
             else:
                 result['error'] = dict(result['error'], **errors.reasons(row['error']))
                 report_id = row['error'].get('report_id')
@@ -2098,7 +2103,8 @@ def poll_image(self, row):
             code = failure.get('code') if isinstance(failure, dict) else failure
             code = code or 'generation_failed'
             if not self.wait_image(row, code):
-                self.finish(row, 'failed', error=dict(code=code, report_id=state.get('report_id'), memory=state.get('memory')))
+                self.finish(row, 'failed', error=dict(code=code, report_id=state.get('report_id'), memory=state.get('memory'),
+                                                     other_usage=state.get('other_usage')))
 
     # Local VLM -----------------------------------------------------------------------------------------
     def start_vlm(self, row):

freevideo_engine/agent/schema.py

@@ -1538,6 +1538,9 @@ def openapi():
                            'description': 'Image diagnostic report ID, or null when no report exists.'},
                 request_id=string(), docs=string(), retry_after_seconds=number(0), card={'type': 'object'}, job=JOB,
                 project={'type': 'object'}, current_revision=number(1, integer=True),
+                other_usage=nullable(obj(dict(vram_gib=number(0), ram_gib=number(0),
+                    kinds={'type': 'array', 'items': enum(('gpu', 'ram'))}, signature=string()),
+                    ('vram_gib', 'ram_gib', 'kinds', 'signature'))),
                 error_zh=string(), error_en=string()),
                 ('code', 'message', 'details', 'retryable', 'request_id', 'docs'))}))