Sep 16

1. 21:07

   revert: ubatch volta a 2048 (4096 mediu ~3% PIOR em A/B controlado); mantem guarda batch>=ubatch

   Erinaldordso
   Verified signature

   89b8453
2. 20:56

   fix: BATCH_SIZE 2048->4096 (deve ser >= ubatch; evita clamp silencioso) + guarda de coerencia

   Erinaldordso
   Verified signature

   a3b8ca7
3. 20:42

   perf: ubatch 2048->4096 (prefill/group-gemm MoE) + --numa opt-in + diag NUMA

   Erinaldordso
   Verified signature

   47520c2
4. 20:33

   diag: captura uso real de AMX (load_tensors) e system_info do llama.cpp

   Erinaldordso
   Verified signature

   0a72b37
5. 20:28

   perf: diagnostico AMX/VNNI/NUMA + backend de CPU + cache-ram 32GB (ctx 128k)

   Erinaldordso
   Verified signature

   5de0eb7
6. 18:33

   feat: troca para Qwen3-Coder-30B-A3B (MoE, mesma arquitetura) + ctx 131072 + sampling do autor do modelo

   Erinaldordso
   Verified signature

   2a0c9d8
7. 17:20

   feat: CTX_SIZE 32768 -> 40960 (max nativo do modelo, sem YaRN) + YaRN opt-in para 131072

   Erinaldordso
   Verified signature

   7ab0dce
8. 17:06

   tooling: patch_app.py escreve LF (local == publicado byte a byte)

   Erinaldordso
   Verified signature

   d5ed9ba
9. 09:04

   fix: remove --cache-reuse (suspeito da instabilidade de tool calling em surtos); cache exato do --cache-prompt segue ativo

   Erinaldordso
   Verified signature

   d8402ae
10. 08:44

    fix: reverte --no-warmup (causava ~2min de tool calling quebrado apos restart); mantem cache-ram e gate do autotune

    Erinaldordso
    Verified signature

    46d88b9
11. 08:26

    fix: autotune escolhia config pior (TTFT +21s); remove spec n-gram inerte; mantem cache-ram

    Erinaldordso
    Verified signature

    b4ee86a
12. 05:37

    perf: flash-attn on, prompt cache, thread clamp, code-gen defaults⁠

    Erinaldordso
    3b4fb97