docs/usermanuals/en/operations/troubleshooting.md: broken example — Slow training: enable model.use_unsloth: true; check distributed.cpu_offload: false; 'Mixed-precision misconfigured (use bfloat… (DOC-D3B04B-008) #226

Wave 1 · theme D1-10: Documentation-only fixes · highest severity High · 1 finding(s) · effort docs fix

Summary

Users following the advice get a config validation error (exit 1) or look for a precision key that does not exist.

Primary finding

DOC-D3B04B-008 — docs/usermanuals/en/operations/troubleshooting.md: broken example — Slow training: enable model.use_unsloth: true; check distributed.cpu_offload: false; 'Mixed-precision misconfigured (use bfloat…

severity High

Where. docs/usermanuals/en/operations/troubleshooting.md:69-76 (Turkish mirror: docs/usermanuals/tr/operations/troubleshooting.md:72-75) · type: broken example

The document says. Slow training: enable model.use_unsloth: true; check distributed.cpu_offload: false; 'Mixed-precision misconfigured (use bfloat16, not float16)'.

What the code does. model.use_unsloth and distributed.cpu_offload do not exist and fail validation (verified). Unsloth is model.backend: "unsloth". Offload is distributed.fsdp_offload or the DeepSpeed preset. Training mixed precision is not configurable: bf16 is chosen automatically when CUDA supports it, otherwise fp16 (forgelm/trainer.py:516-517).

Code references. forgelm/config.py:250, forgelm/trainer.py:516-517

User impact. Users following the advice get a config validation error (exit 1) or look for a precision key that does not exist.

Proposed correction. 1. Not using Unsloth (model.backend: "unsloth", Linux + CUDA). 3. CPU offload enabled (distributed.fsdp_offload: true or a ZeRO-3-offload DeepSpeed preset). 4. GPU without bf16 support — ForgeLM picks bf16 automatically when available, else fp16.

Turkish mirror. the Turkish page has the same error; fix both in the same PR

Acceptance criteria

• Each statement is re-checked against the code at fix time.
• Documentation listed above is corrected in English and in its Turkish mirror in the same PR.
• Every finding in this issue is resolved or explicitly closed with a reason.

_{Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.}