fix(worker+api): batch title vfb alignment, race-free config, auto-queue (3 bugs) #2097

Merged

xiaoxia merged 4 commits from fix/gpu-direct-title-vfb-alignment into develop 2026-09-29 19:51:40 +08:00

Conversation 2 Commits 4 Files Changed 13

+467 -541

4 Commits

Author	SHA1	Message	Date
CI Bot	efb7fa5729	style: auto-format with black + isort + ruff + prettier [skip ci-format-check]	2026-09-29 11:17:23 +00:00
saas-backend-agent	5e1520230f	test(api+worker): align unit tests with #2098 soft user limit + override refactor ... - Bug B: check_queue_limits no longer raises UserPendingLimitExceeded (user limit is soft); only global GLOBAL_PENDING_LIMIT hard-rejects with 503. - safe_enqueue_generation_task user over-limit is warning-only (pre/post enqueue); GlobalQueueFull still hard-rejects/rolls-back. - Preview route (/generation/preview) user-pending pre-check removed to match batch/generation/retry routes; only global 503 pre-check kept. - Route enqueue-loop: UserPendingLimitExceeded except branch removed; all-failed rate-limit response only triggers on GlobalQueueFull (503). - Tests rewritten to reflect new semantics (197 queue/preview/batch/pipeline tests pass). - test_worker_generate_video_task_binding: replaced legacy _sync_task_config_to_plan helper test with _build_task_config_override + _download_voice_for_task (Bug A refactor). - USER_PENDING_LIMIT constant verified at 20 (soft cap).	2026-09-29 19:06:12 +08:00
saas-backend-agent	96bcec5fdd	fix(worker+api): resolve batch title race + auto-queue pending tasks (#2098) ... Two batch-rendering bugs found from staging logs (6-video batch 17:30-17:43): Bug A (P1) — title_config race when same plan renders concurrently: _sync_task_config_to_plan wrote each task's title_config/bgm/resolution to the shared plan.config row, then RenderAdapter.render_plan read back from plan.config during rendering. If two tasks sharing the same plan (e.g. batch retry, preview+render, retry+new) ran concurrently, Task B's write could overwrite Task A's title before Task A's ffmpeg read it, producing videos with the wrong title text/style. Fix: stop mutating plan.config in the worker render path. Introduce task_config_override threading through render_plan / _do_render / UnifiedRenderService; URS reads config via _effective_config() which does a deep-copy plan.config merged with per-task override (title/bgm/ export). _build_task_config_override builds the override dict (with field-name normalization: font_size→size, font_color→color) from task_info, and _download_voice_for_task handles voiceover download independently. plan.config is left as-authored by the API writeback; concurrent tasks no longer race on it. Files: - apps/worker/video_processing/unified_render_service.py: add override_config param + _cfg_section/_effective_config helpers; replace all render-time self.plan.config reads (title/subtitle/bgm/ export/TTS) with _effective_config(). - apps/worker/video_processing/render_adapter.py: render_plan/_do_render/ _prepare_bgm accept task_config_override, merge into bgm/export reads, forward to URS. - apps/worker/worker_app/tasks/generation.py: replace _sync_task_config_to_plan (which wrote plan.config) with _build_task_config_override + _download_voice_for_task; pass task_config_override to render_plan. Bug B (P0) — USER_PENDING_LIMIT=3 returns HTTP 429, blocks batch submit: Pre-check in create_generation_task rejected the whole batch with 429 USER_QUEUE_FULL once user had ≥3 pending tasks; UI showed 'wait ~4 min' and prevented any more submissions. Users expect to submit a batch of 6 and have them queue naturally (worker concurrency=2). Fix: - USER_PENDING_LIMIT 3→20 (soft cap for abuse protection, supports typical batch sizes of 6-10 with headroom). - Remove user-level 429 rejection from pre-check, retry endpoint, single-task confirm endpoint, and safe_enqueue (user path now logs a warning and continues to enqueue). Global GLOBAL_PENDING_LIMIT=20 is retained as a hard 503 system-busy guard. - safe_enqueue post-enqueue check: user-over only logs, does not fail the task or raise. - Batch loop UserPendingLimitExceeded except branch now treats it as a successful enqueue (should not trigger in practice). Files: - apps/api/app/core/task_enqueue.py: limit 3→20; user checks log-only. - apps/api/app/api/routes/generation_tasks.py: remove user pre-check 429; soften retry/single/batch except branches. Stacked on PR #2095/#2097 which already fixed title baseline/width-scaling and default-constant alignment with CPU video_filter_builder. Tests: - tests/unit/test_gpu_direct_pipeline.py: 30 passed - tests/unit/test_render_adapter_pure.py: 18 passed - ruff check/format clean; py_compile clean	2026-09-29 18:31:05 +08:00
saas-backend-agent	dfc5e5a5b6	fix(worker): align gpu-direct title defaults (size/margin/bold) with CPU vfb path ... PR #2095 fixed title baseline positioning and width-scaling but didn't fully align default constants with the CPU video_filter_builder path, causing GPU rendered titles to appear slightly smaller / higher / bolder-differently than the CPU/ASS preview the template was authored against. This shows up in both single and batch GPU renders since every variant shares the same gpu_direct pipeline. Root causes in PR #2095 defaults: - Default title size 36@720p vs config_schemas DEFAULT 48 / vfb default 48 - Top/bottom margin 40@720p (PAD16+margin24) vs vfb _scale_title_len(50) - Subtitle bottom margin 60@720p vs vfb 50 - Faux bold used same-color 1px border (white-on-white invisible for default white text; red-on-red for colored) vs vfb black 2px (intentional choice per #2001 to avoid double-print/halo artifact) - margin_top was treated as absolute y offset, now correctly added on top of base margin (matches vfb additive semantics) - subtitle stroke/bold block referenced uninitialized s_borderw (ruff F821) — added proper init + stroke parsing consistent with title block Changes: - TITLE_DEFAULT_MARGIN_TOP/BOTTOM = 50 (was 24+PAD=40) - SUBTITLE_DEFAULT_MARGIN_BOTTOM = 50 (was 60) - TITLE_FAUX_BOLD_WIDTH = 2, border color = #000000 (was 1, same as text) - default title size 48@720p (was 36) - margin_top from cfg added on top of base 50 (additive, same as vfb) - subtitle: init s_borderw/s_border_color, parse stroke dict before bold check - bottom position: y = h - th - margin_bottom (correct baseline; previously used margin_top variable which was misleading but numerically equivalent before margin split; now uses explicit margin_bottom) Tests updated to new expected scaled values (1280x720 scale=1.778: fontsize 85, y=89, bold borderw=4 black; 1080x1920 vertical: fontsize 72, y=75; margin_top=100 user offset => y=267). Added test_title_bold_false_disables_faux_bold. Batch investigation notes (separate findings, not bugs in this PR): - per-variant plan config is correctly deep-copied via clone_plan_for_variant (config=dict(source.config or {})); title text per-variant via titles[] override; voice per-variant independent download with #1749 strict guards; bgm merged via merge_bgm_config — batch passthrough chain is correct. - Worker generation concurrency = 2 (compose.yml default); USER_PENDING_LIMIT=3, 6 tasks will enqueue in two waves (first 3 → then next 3 as workers free up); GLOBAL_PENDING_LIMIT allows it. This is expected behavior, not a bug. - Legacy key mismatch: subtitle_render_engine.py reads plan_config['subtitle_config'] (snake_case) but normalize_plan_config writes 'subtitle' short key. This file is not imported by unified_render_service (which uses render_subtitles.py with explicit subtitle_config= parameter), so no runtime impact; left for separate cleanup. - Frontend usePlanConfigLoader.ts reads config.title_config snake_case while backend writes 'title' short key — frontend-only issue for subsequent edit sessions, out of backend scope.	2026-09-29 18:08:18 +08:00