Skip to content

Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.

Carry the host-memory shortage from the image worker to the page

Bowen Xue548a933384b8
diff --git a/freevideo_engine/diagnostic_resources.py b/freevideo_engine/diagnostic_resources.py
index c53a765..6444a85 100644
--- a/freevideo_engine/diagnostic_resources.py
+++ b/freevideo_engine/diagnostic_resources.py
@@ -52,13 +52,17 @@ def commit_exhausted(error):
Rust-backed file readers raise a plain OSError carrying only the localized
system message plus "(os error 1455)", so the numeric attributes are unset.
+ A different numeric code wins over the text, which may be a file name.
"""
seen = set()
while isinstance(error, BaseException) and id(error) not in seen:
seen.add(id(error))
- if getattr(error, 'winerror', None) == 1455 or re.search(r'(?:os error|winerror)\s*1455\b', str(error), re.I):
+ code = getattr(error, 'winerror', None)
+ if code == 1455:
return True
- error = error.__cause__ or error.__context__
+ if code is None and re.search(r'(?:os error|winerror)\s*1455\b', str(error), re.I):
+ return True
+ error = error.__cause__ or (None if error.__suppress_context__ else error.__context__)
return False
diff --git a/freevideo_engine/image_gen/failure.py b/freevideo_engine/image_gen/failure.py
index 4920049..9101364 100644
--- a/freevideo_engine/image_gen/failure.py
+++ b/freevideo_engine/image_gen/failure.py
@@ -56,7 +56,12 @@ _MPS_MEMORY = re.compile(r'MPS backend out of memory', re.I)
_LIMIT = 16 * 1024
# On a Mac the GPU's memory is the machine's: the same failure, in its words.
UNIFIED_MESSAGE_ZH = {'no_memory': '内存不够,最小尺寸也没能生成。关掉其他程序后再试。'}
-COMMIT_MESSAGE_ZH = '系统的虚拟内存不够,图像生成没能载入模型。关掉其他程序,或在系统设置里调大虚拟内存后再试。'
+HOST_MESSAGE_ZH = '电脑内存或虚拟内存不够,模型没能载入。关掉其他程序,或调大系统的虚拟内存后再试。'
+
+
+def host_memory_shortage(error):
+ """The machine's own memory ran out, not the card's."""
+ return commit_exhausted(error)
class FailureError(ValueError):
@@ -77,7 +82,7 @@ def classify_exception(error, *, stage='start', cancelled=False):
seen.add(id(error))
chain.append(error)
error = error.__cause__ or (None if error.__suppress_context__ else error.__context__)
- if commit_exhausted(chain[0]) if chain else False:
+ if chain and commit_exhausted(chain[0]):
return 'no_memory'
for item in chain:
names = {cls.__name__ for cls in type(item).__mro__}
diff --git a/freevideo_engine/image_gen/image_runner.py b/freevideo_engine/image_gen/image_runner.py
index 70862b3..e4e11ca 100644
--- a/freevideo_engine/image_gen/image_runner.py
+++ b/freevideo_engine/image_gen/image_runner.py
@@ -283,15 +283,7 @@ class ImageRunner:
from ..system import system_memory
memory = system_memory()
free = memory.get('physical_available_bytes') or memory['available_bytes']
- held = self.dit_stack.host_bytes + self.te_stack.host_bytes
- host_available = free + held
- # Windows charges every committed page to the paging file, so free RAM alone
- # can promise more than the commit limit leaves. A system-managed paging file
- # also grows on demand, so the working margin is not held against the headroom:
- # only a request whose unpinned weights cannot fit in it is refused early.
- commit = memory.get('commit_available_bytes')
- if commit is not None:
- host_available = min(host_available, commit + held + image_policy.MIN_HOST_WORKING)
+ host_available = free + self.dit_stack.host_bytes + self.te_stack.host_bytes
if self.full_card or self.mapped:
host_available = None
except Exception:
diff --git a/freevideo_engine/image_gen/service.py b/freevideo_engine/image_gen/service.py
index 9595ca5..3255b9f 100644
--- a/freevideo_engine/image_gen/service.py
+++ b/freevideo_engine/image_gen/service.py
@@ -18,8 +18,7 @@ import uuid
from . import image_pack
from ..download_settings import observed_speed
-from ..diagnostic_resources import commit_exhausted
-from .failure import (COMMIT_MESSAGE_ZH, Context, ERRORS, FailureError, MESSAGE_ZH, NO_REPORT, STAGES, UNIFIED_MESSAGE_ZH,
+from .failure import (Context, HOST_MESSAGE_ZH, host_memory_shortage, ERRORS, FailureError, MESSAGE_ZH, NO_REPORT, STAGES, UNIFIED_MESSAGE_ZH,
classify_exception)
BODY_LIMIT = 64 * 1024
@@ -749,8 +748,9 @@ class Service:
result['message_zh'] = MESSAGE_ZH[code]
if context.machine.get('device_backend') == 'mps' and code in UNIFIED_MESSAGE_ZH:
result['message_zh'] = UNIFIED_MESSAGE_ZH[code]
- if code == 'no_memory' and error is not None and commit_exhausted(error):
- result['message_zh'] = COMMIT_MESSAGE_ZH
+ if code == 'no_memory' and (isinstance(row, dict) and row.get('memory') == 'host'
+ or error is not None and host_memory_shortage(error)):
+ result.update(memory='host', message_zh=HOST_MESSAGE_ZH)
if code not in NO_REPORT:
if video_busy is None:
try:
diff --git a/freevideo_engine/image_gen/worker.py b/freevideo_engine/image_gen/worker.py
index 469acf6..f9c565a 100644
--- a/freevideo_engine/image_gen/worker.py
+++ b/freevideo_engine/image_gen/worker.py
@@ -9,7 +9,7 @@ import sys
import threading
import time
-from .failure import Context, classify_exception, exception_details, safe_plan
+from .failure import Context, classify_exception, exception_details, host_memory_shortage, safe_plan
class Inbox:
@@ -170,6 +170,8 @@ def serve(pack_dir, *, device='cuda', idle_seconds=45., stream=None, output=None
result = dict(phase='error', code=code, stage=context.timing.stage, reloaded=reload,
seconds=context.timing.seconds(), stages=context.timing.snapshot(), plan=plan,
error=exception_details(error, context.redactor, code))
+ if code == 'no_memory' and host_memory_shortage(error):
+ result['memory'] = 'host'
try:
result.update(attention_report(runner, context.redactor))
except Exception:
diff --git a/web/image_generation.js b/web/image_generation.js
index f98d334..8bba343 100644
--- a/web/image_generation.js
+++ b/web/image_generation.js
@@ -176,6 +176,9 @@ export function imageGenError(code, info, t) {
};
const say = (en, zh) => t(...memoryWords([en, zh], info?.unified));
if (code === 'pack_missing') return imagePackUnavailableMessage(info, t);
+ if (code === 'no_memory' && info?.memory === 'host')
+ return t('Not enough system memory or virtual memory to load the model. Close other programs, or enlarge the system virtual memory, then try again.',
+ '电脑内存或虚拟内存不够,模型没能载入。关掉其他程序,或调大系统的虚拟内存后再试。');
if (code === 'no_memory') {
// Where it ran out decides what helps: the reference pictures are read at a fixed size,
// so a smaller canvas does not change that stage.
@@ -211,7 +214,7 @@ export const REPORT_LABELS = Object.freeze({
// Shared by the editor and the other image entry points. Only the opaque report
// ID goes in the request; raw service messages are never shown in the page.
-export function renderImageFailure(host, {error, report_id, stage, fitted, info} = {}, t,
+export function renderImageFailure(host, {error, report_id, stage, fitted, info, memory} = {}, t,
{productVersion = () => null, fetchApi = globalThis.fetch.bind(globalThis)} = {}) {
host.replaceChildren();
if (!error || error === 'cancelled') return null;
@@ -222,7 +225,7 @@ export function renderImageFailure(host, {error, report_id, stage, fitted, info}
}
const card = doc.createElement('div'); card.className = 'fv-result-links';
const reason = doc.createElement('div'); reason.className = 'fv-image-failure-reason';
- reason.setAttribute('role', 'status'); reason.textContent = imageGenError(error, {stage, fitted, unified: info?.unified_memory}, t);
+ reason.setAttribute('role', 'status'); reason.textContent = imageGenError(error, {stage, fitted, memory, unified: info?.unified_memory}, t);
card.append(reason); host.append(card);
if (!report_id) return card;
const fetchReport = async () => {