docs(tutorials): fix 2 cross-tutorial inconsistencies surfaced by a holistic audit

[wanshuiyin]

Holistic collection audit → 2 cross-tutorial fixes

A "look back" audit of the whole 28-tutorial collection — structural (inventory, counts, links, EN/ZH parity, gate, audit-chain) + a cross-tutorial consistency pass over shared concepts (FlashAttention, attention O(L²), KV cache/MLA, RoPE, Pre/Post-LN, μP, AdamW, MoE routing, GQA/MQA, quantization).

Result: the corpus is healthy and highly consistent — structurally all green (28× five-piece complete, 57 OK gate, EN/ZH parity 28/28), and 8/10 shared concepts fully aligned with 5/5 cross-references valid. Two reconcilable inconsistencies were found and are fixed here (both ZH + EN editions):

1. Warmup takeaway (normalization §5.2) — Medium

normalization_init_tutorial flatly stated "Pre-LN → warmup-free" (and "GPT/LLaMA are uniformly Pre-LN"), which reads as the opposite of optimizer_lr_schedule_tutorial §7.1 ("all schedules should be preceded by warmup", whose reason 1 — Adam early-step v̂ variance — is architecture-independent and applies to Pre-LN too). Real Pre-LN LLMs (GPT-3/LLaMA) do warm up.
Fix: a clarifying Note in §5.2 — Pre-LN removes only the gradient-imbalance reason for warmup; modern Pre-LN LLMs still warm up for Adam-variance / large-batch reasons (cross-refs optimizer §7.1). Both tutorials were individually correct; now they read consistently.

2. DeepSeek-V2 MLA compression ratio — Low

The same model appeared as three numbers across two tutorials:

• kv_cache §4: 64× (counts only the bare latent d_c=512)
• long_context §9.6 + table: 57× (the correct full cache d_c+d_r=576) ✓
• long_context §9 intro: a loose ≈50×

Fix: aligned all to the full-cache ~57× (32768/576 ≈ 57), with the bare-latent 64× kept as an explicit caveat. The 57× already in long_context §9.6/§12.3 is the technically-correct figure.

Mechanics

Orchestrator-verified (arithmetic + warmup reasoning re-checked — this was a factual/clarity correction, not a new claim). The original gating reviews remain valid; the 6 HTMLs were re-rendered and the 6 sidecars' source_sha256 refreshed (+ a post_ship_review note on each documenting the fix).

Gate

python3 tools/verify_reviews.py --mode strict --reproduce → 57 OK · 0 FAIL · 3 EXEMPT · PASS.

🤖 Generated with Claude Code