The original failure error was invisible: the 500-char truncation was consumed entirely by ggml-metal's informational init log (tensor API disabled for pre-M5 and pre-A19 devices — a normal log, not an error), leaving the actual crash reason cut off. A previous fix incorrectly treated this message as a GPU failure and added a CPU-only retry; that assumption was wrong and has been removed.

Changes

server/src/services/acestep.ts

• Error truncation 500 → 2000 chars so real failure reasons are visible
• Removed incorrect Metal tensor API auto-retry (-ngl 0 fallback) — the ggml-metal init message is informational; the GPU works correctly on M1–M4
• runBinary gains an onLine callback — stderr is split into lines and streamed to the caller in real time; full stderr is still accumulated for error reporting
• makeLmProgressHandler — parses ace-qwen3 stderr into job.stage / job.progress (0–50%):
  • [Phase1] step N … tok/s → 0–28%
  • [Phase1] Decode → 30%
  • [Decode] step N … total codes … tok/s → 30–50% (budget from [Phase2] max_tokens)
• makeDitVaeProgressHandler — parses dit-vae stderr into job.stage / job.progress (50–100%):
  • [DiT] step N/M → 50–85%
  • [DiT] Total generation → 85%
  • [VAE] Tiled decode N tiles / Tiled decode done → 85–98%
• Named progress-budget constants (PROGRESS_LM_PHASE1_MAX, PROGRESS_LM_PHASE2_END, PROGRESS_DIT_END, PROGRESS_VAE_END) make the 0–100 allocation explicit

The frontend's existing job-status polling picks up stage and progress automatically — no frontend changes needed.

.env.example

Removed incorrect Apple Silicon Metal retry guidance; kept DIT_VAE_EXTRA_ARGS / ACE_QWEN3_EXTRA_ARGS with a corrected general-purpose description.

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.