Skip to content

UPGRADE.md

Latest commit

 

History

History
78 lines (59 loc) · 3.8 KB

UPGRADE.md

File metadata and controls

78 lines (59 loc) · 3.8 KB

Upgrading This 48 GB Setup

This backup targets the 48 GB Qwen3.6 MLX 8-bit workflow.

To upgrade without losing the working baseline:

  1. Edit /Users/oiqbal/AndroidStudioProjects/llm-opencode/config/profile-48gb.env.
  2. Mirror OpenCode-facing changes in /Users/oiqbal/AndroidStudioProjects/llm-opencode/config/opencode.json.
  3. Run /Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/validate-profile-sync.sh.
  4. Run /Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/install-opencode-config.sh on the target 48 GB Mac.
  5. Run /Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/ensure-lmstudio-models.sh.
  6. Run /Users/oiqbal/AndroidStudioProjects/llm-opencode/scripts/smoke-test.sh.

Use SMOKE_QWEN_GENERATION=1 only when you deliberately want the slower direct Qwen generation check. The normal smoke follows the production routing design: fast model for no-tool generation, Qwen loaded/configured for coding routes.

Keep these as the 48 GB constraints unless intentionally changing the profile:

  • OpenCode default model id: lmstudio/local-fast
  • OpenCode Qwen coding model id: lmstudio/local-coder
  • LM Studio fast identifier: local-fast
  • LM Studio coding identifier: local-coder
  • Fast model source: mlx-community/Llama-3.2-3B-Instruct-4bit
  • Model source: froggeric/qwen3.6-27b-mlx-8bit
  • Model path: froggeric/Qwen3.6-27B-MLX-8bit
  • Qwen context/output: 32768 / 4096
  • Fast context/output: 32768 / 1024
  • Parallel: 1
  • LM Studio bind: 127.0.0.1:1234

If Qwen 32K context is unstable, reload the same local-coder alias at 24576, then 16384. Keep local-fast loaded separately so OpenCode does not fall back to manual model swapping.

Do not replace Qwen with Gemma or another model in this profile without a separate experiment and smoke test. Extra models do not add MCP capability by themselves; they only change how well a model uses the already configured tools.

When changing agents or commands, keep these loop-prevention rules unless there is a deliberate new design:

  • default_agent stays fast.
  • build.steps stays low, currently 18.
  • plan.steps stays low, currently 10.
  • build denies local_code_index_*, local_dev_tools_*, context7_*, and gh_grep_*.
  • /research is the MCP-backed path.
  • /explain remains fast/no-tool and conservative for project-specific questions.
  • Qwen instructions keep <|think_off|> at the top.

Do not copy model weights into this repo. Store identifiers, source paths, runtime URLs, checksums, and scripts so LM Studio or Lucebox model directories remain the places that own model downloads and updates.

Ubuntu RTX 3090 Profile

For the separate Ubuntu i9 13th gen, 96 GB RAM, RTX 3090 setup, do not overwrite the 48 GB Mac defaults. Use the opt-in profile and config instead:

OPENCODE_BACKUP_PROFILE="$PWD/config/profile-96gb-ubuntu-nvidia.env" \
OPENCODE_BACKUP_CONFIG="$PWD/config/opencode-96gb-ubuntu-nvidia.json" \
scripts/validate-profile-sync.sh

Follow PLAN-96GB-UBUNTU-NVIDIA.md on the Linux target. That profile keeps the same agents, commands, MCPs, RAG, and LSP shape, but changes the primary chat model to lucebox/luce-dflash through the Lucebox DFlash autowake proxy on 127.0.0.1:18080.

Current Ubuntu runtime contract:

  • Lucebox backend: 127.0.0.1:18081, started on demand by the proxy.
  • Target model: Qwen3.6-27B-Q4_K_M.gguf.
  • Draft model: dflash-draft-3.6-q4_k_m.gguf.
  • Context/output: 49152 / 2048.
  • Prefix cache: disabled with --prefix-cache-slots 0.
  • LM Studio: embedding-only in normal mode, rollback chat only with LMSTUDIO_LOAD_CHAT_ROLLBACK=1.

Run Ubuntu validation with the profile explicitly selected:

OPENCODE_BACKUP_PROFILE="$PWD/config/profile-96gb-ubuntu-nvidia.env" \
OPENCODE_BACKUP_CONFIG="$PWD/config/opencode-96gb-ubuntu-nvidia.json" \
scripts/smoke-test.sh