current org

tasks/FRE-9.yaml

@@ -0,0 +1,46 @@
+---
+date: 2026-03-08
+day_of_week: Sunday
+task_id: FRE-9
+title: Fix TTS Generation Bug in AudiobookPipeline
+status: done
+company_id: FrenoCorp
+objective: Resolve CUDA/meta tensor error in TTS generation stage to enable working pipeline
+context: |
+  - Product: AudiobookPipeline using Qwen3-TTS 1.7B VoiceDesign model
+  - MVP deadline: April 4, 2026 (4 weeks from today)
+  - Pipeline works through segmentation but fails at generation with "Tensor.item() cannot be called on meta tensors" error
+  - Intern Pan assigned to this task by CEO
+  - Codebase located at /home/mike/code/AudiobookPipeline/
+  - TTS model wrapper at /home/mike/code/AudiobookPipeline/src/generation/tts_model.py
+  - Batch processor at /home/mike/code/AudiobookPipeline/src/generation/batch_processor.py
+issue_type: bug
+priority: high
+assignee: intern
+parent_task: null
+goal_id: MVP_Pipeline_Working
+blocking_tasks:
+  - FRE-10 (MVP Development)
+  - FRE-11 (Testing & QA)
+expected_outcome: |
+  - TTS generation stage completes successfully
+  - Full pipeline processes an epub to MP3 without errors
+  - Audio output meets quality standards (-23 LUFS, proper sample rate)
+  - Mock mode works for testing without GPU
+acceptance_criteria:
+  - Run `make test` passes all tests including generation tests
+  - CLI can process sample.epub and produce output.mp3
+  - No CUDA/meta tensor errors in logs
+  - Generation time under 2x baseline (with mock) or reasonable with real model
+
+notes:
+  - Root cause: device_map="auto" resulted in meta tensors when GPU unavailable
+  - Fix added GPU detection with CPU fallback in tts_model.py:125-146
+  - Added validation to reject models loaded on meta device
+  - Fixed test infrastructure: PYTHONPATH in Makefile, renamed duplicate test file
+  - All 669 tests now pass
+
+links:
+  strategic_plan: /home/mike/code/FrenoCorp/STRATEGIC_PLAN.md
+  technical_architecture: /home/mike/code/FrenoCorp/technical-architecture.md
+  codebase: /home/mike/code/AudiobookPipeline/