[Wave 3] Medium — TR mirror: docs/guides/troubleshooting.md #627

Wave 3 · 7 finding(s) in 7 fix package(s). English/Turkish mirror drift for this document pair; fix the side the finding names.

Findings

• DOC-TR06-024 (TR mirror, Medium) — TR mirror drift in docs/guides/troubleshooting.md: Effective batch size is preserved ... optimization trajectory therefore differs slightly
  severity Medium
  Where. EN docs/guides/troubleshooting.md:64-66 · TR docs/guides/troubleshooting-tr.md:65-68
  English text. Effective batch size is preserved ... optimization trajectory therefore differs slightly
  Turkish text. Etkili batch size denemeler arasında korunur ... yörünge ... biraz farklıdır
  Which is correct. Neither fully; batch size is preserved, but restart is not 'slight'.
  Proposed correction (Turkish text for the -tr page). Etkili batch size korunur (batch yarıya iner, gradient accumulation iki katına çıkar). Ancak yeniden denemede optimizer ve LR scheduler adım 0'dan başlar, ağırlıklar ise OOM anındaki halini korur; yörünge kesintisiz bir koşudan belirgin biçimde farklı olabilir. Tekrar üretilebilirlik için batch size'ı elle sabitleyin.

• DOC-TR06-025 (TR mirror, Medium) — TR mirror drift in docs/guides/troubleshooting.md: disable mixed precision by ensuring bf16: false and fp16: false (these are ForgeLM defaults)
  severity Medium · already tracked: KNOWN-OPEN C31-DOC-SCHEMA-PROSE docs/roadmap/phase-16-trust-surface-hardening.md:785-786
  Where. EN docs/guides/troubleshooting.md:109 · TR docs/guides/troubleshooting-tr.md:111-112
  English text. disable mixed precision by ensuring bf16: false and fp16: false (these are ForgeLM defaults)
  Turkish text. bf16: false ve fp16: false olduğunu doğrulayın (bunlar ForgeLM varsayılanlarıdır)
  Which is correct. Neither; no such config fields, precision is auto-selected.
  Proposed correction (Turkish text for the -tr page). bf16/fp16 YAML anahtarı değildir; ForgeLM karma hassasiyeti donanıma göre otomatik seçer (destekleniyorsa bf16, değilse CUDA'da fp16, CPU'da kapalı). Sorun sürerse öğrenme oranını düşürün ve model.bnb_4bit_compute_dtype: "float32" (VRAM maliyetiyle) deneyin.

• DOC-TR06-026 (TR mirror, Medium) — TR mirror drift in docs/guides/troubleshooting.md: training.gradient_checkpointing: true
  severity Medium · already tracked: KNOWN-OPEN C31-DOC-SCHEMA-PROSE docs/roadmap/phase-16-trust-surface-hardening.md:785-786; docs/roadmap/risks-and-decisions.md:184
  Where. EN docs/guides/troubleshooting.md:159-164 · TR docs/guides/troubleshooting-tr.md:166-171
  English text. training.gradient_checkpointing: true
  Turkish text. training: gradient_checkpointing: true
  Which is correct. Neither; field is rejected (exit 1).
  Proposed correction (Turkish text for the -tr page). Gradient checkpointing CUDA'da ForgeLM tarafından otomatik açılır; training.gradient_checkpointing bir YAML alanı değildir ve yazılırsa config doğrulaması hata verir. Bu maddeyi kaldırın veya 'zaten varsayılan olarak açık' diye değiştirin.

• DOC-TR06-027 (TR mirror, Medium) — TR mirror drift in docs/guides/troubleshooting.md: retries are handled at the HTTP-client layer
  severity Medium
  Where. EN docs/guides/troubleshooting.md:192-195 · TR docs/guides/troubleshooting-tr.md:200-203
  English text. retries are handled at the HTTP-client layer
  Turkish text. retry'lar HTTP-client katmanında ele alınır
  Which is correct. Neither; no retries exist.
  Proposed correction (Turkish text for the -tr page). SyntheticConfig ayrı retry/batch ayarı sunmaz ve başarısız bir API çağrısı otomatik yeniden denenmez; zaman aşımı/oran sınırı için api_timeout ve api_delay kullanın, komutu gerekirse yeniden çalıştırın.

• DOC-TR06-028 (TR mirror, Medium) — TR mirror drift in docs/guides/troubleshooting.md: use_dora: true and use_rslora: true together (still works, auto-normalizes)
  severity Medium
  Where. EN docs/guides/troubleshooting.md:266-270 · TR docs/guides/troubleshooting-tr.md:284-287
  English text. use_dora: true and use_rslora: true together (still works, auto-normalizes)
  Turkish text. aynı blok
  Which is correct. Neither; combination is rejected.
  Proposed correction (Turkish text for the -tr page). # Deprecated (hâlâ çalışır, uyarı atar): tek bayrak\nlora:\n use_dora: true # otomatik method: "dora"\n# use_dora ve use_rslora birlikte verilirse config doğrulaması hata verir (karşılıklı dışlayıcı).

• DOC-TR06-029 (TR mirror, Medium) — TR mirror drift in docs/guides/troubleshooting.md: FileNotFoundError: DeepSpeed preset 'zero2' not found
  severity Medium
  Where. EN docs/guides/troubleshooting.md:343-345 · TR docs/guides/troubleshooting-tr.md:361-373
  English text. FileNotFoundError: DeepSpeed preset 'zero2' not found
  Turkish text. aynı
  Which is correct. Neither; ConfigError, exit 1; presets are not shipped in the wheel.
  Proposed correction (Turkish text for the -tr page). text\nConfigError: DeepSpeed preset 'zero2' not found at <yol>. ...\n\nPreset'ler yalnızca kaynak depodaki configs/deepspeed/ altındadır (pip paketine dahil değildir); depo kökünden çalıştırın veya distributed.deepspeed_config için kendi JSON dosyanızın mutlak yolunu verin. Çıkış kodu 1'dir.

• DOC-TR06-030 (TR mirror, Medium) — TR mirror drift in docs/guides/troubleshooting.md: JSON diagnostics: ... --output-format json 2>error.log
  severity Medium
  Where. EN docs/guides/troubleshooting.md:404 · TR docs/guides/troubleshooting-tr.md:426
  English text. JSON diagnostics: ... --output-format json 2>error.log
  Turkish text. JSON tanılama: ... --output-format json 2>error.log
  Which is correct. Neither; JSON goes to stdout, 2> captures the log stream.
  Proposed correction (Turkish text for the -tr page). - JSON tanılama: forgelm --config job.yaml --output-format json > result.json 2> run.log (JSON zarfı stdout'a, log satırları stderr'e yazılır)

Done when

• Every finding above is fixed or closed with a reason (tick it here).
• Code fixes carry a test; doc fixes update the Turkish mirror in the same PR.

_{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.}