Commit c4be69e

mo khan <mo@mokhan.ca>
2026-08-30 06:38:52
feat: fetch the default gguf model in bin/setup
A fresh clone had no way to satisfy the gguf provider's default model path -- both lib/elelem/plugins/gguf.rb and the eval harness pointed at ~/models/, a personal path that only existed on the machine the default was chosen on. bin/test would fail immediately on any other checkout. Move the default to ~/.elelem/models/, matching the existing convention for user-level elelem data (mcp.json, prompts/, plugins/, permissions.json all already live under ~/.elelem). Fetch it once from ggml-org's llama.cpp-maintained GGUF mirror in bin/setup, skipping if already present so setup stays safe to re-run, and downloading to a .part file first so an interrupted run can't leave a truncated model that fails to load with a confusing error. This does not make bin/test fully green on a fresh clone -- gpt-oss-20b still fails 5/29 eval cases on model-quality/case-quality grounds (see the last commit's comparison table; search/locate-error-string in particular fails on every model tested, which looks like a case bug). It does make bin/test run to completion instead of erroring out on a missing file, and the failure count is now real, reproducible signal instead of a KeyError.
main
1 parent f928dd1
Changed files (1)
lib
elelem
llama
plugin.rb
lib/elelem/llama/plugin.rb
@@ -4,7 +4,7 @@ Elelem::Providers.register(:gguf) do
   gpu = %w[vulkan cuda metal].include?(Elelem::Net::GGUF.backend)
 
   Elelem::Net::GGUF.new(
-    model: ENV.fetch("GGUF_MODEL", File.expand_path("~/models/gpt-oss-20b-MXFP4.gguf")),
+    model: ENV.fetch("GGUF_MODEL", File.expand_path("~/.elelem/models/gpt-oss-20b-MXFP4.gguf")),
     n_ctx: Integer(ENV.fetch("GGUF_N_CTX", "16384")),
     n_threads: Integer(ENV.fetch("GGUF_THREADS", "16")),
     max_tokens: Integer(ENV.fetch("GGUF_MAX_TOKENS", "4096")),