[Sync] llm stuff

2026-09-17 20:58:21 +02:00
parent ed7e34552d
commit d8f6b30e2c
52 changed files with 2233 additions and 237 deletions
@@ -0,0 +1,64 @@
+# WM HOST OVERLAY — full-file copy of common/.config/aichat/config.yaml with the default model
+# flipped to the duskadiy client: wm is not on the home LAN, so the `local` endpoint
+# (192.168.0.204) is unreachable here (kept below for VPN use). Stowed with --override so it
+# shadows the common file on wm only. When the common config changes, mirror the change here.
+# see https://github.com/sigoden/aichat/blob/main/config.example.yaml
+keybindings: vi
+editor: nvim
+model: duskadiy:Qwen3-Coder-30B-Instruct-UD-Q3_K_XL
+
+# Sessions: persist REPL sessions and keep more history before summarizing.
+# The default compress_threshold (4000) summarizes far too early for 24k+ windows.
+save_session: true
+compress_threshold: 16000
+
+# REPL prompts show live context usage (needs max_input_tokens, set per model below)
+left_prompt:  '{color.green}{?session {session}{?role /}}{role}{color.cyan}{?rag @{rag}}{color.reset}> '
+right_prompt: '{color.purple}{?session {consume_tokens}/{max_input_tokens} }{color.reset}'
+
+# NOTE: temperature/top_p are intentionally unset — the LAN server applies tuned
+# per-model sampling via --jinja (e.g. GLM 0.6/0.95); a global value would clobber it.
+
+# max_input_tokens = real ctx-size (from the router) minus output headroom.
+clients:
+# LAN access (fast; only reachable on the home network)
+- type: openai-compatible
+  name: local
+  api_base: http://192.168.0.204:11343/v1
+  models: &lan_models
+  # Speeds are benched decode t/s (clean-night sweeps) — see fl/.config/llamacpp/README.md roster.
+  - name: Qwen3-Coder-30B-Instruct-UD-Q3_K_XL
+    max_input_tokens: 30000            # ctx 32768 · ~30 t/s — main agent coder
+  - name: Qwen3-Coder-Next-UD-IQ3_XXS
+    max_input_tokens: 128000           # ctx 131072 · ~16 t/s — long sessions (128k)
+  - name: Qwen3.6-35B-A3B-MTP-UD-IQ3_XXS
+    max_input_tokens: 22000            # ctx 24576 · ~36 t/s — daily driver
+    supports_vision: true
+  - name: Qwen3.6-35B-A3B-Thinking
+    max_input_tokens: 22000            # ctx 24576 · ~39 t/s — hard problems
+    supports_vision: true
+  - name: Qwen3.5-9B-UD-Q6_K_XL
+    max_input_tokens: 30000            # ctx 32768 · ~32 t/s — quick tasks
+    supports_vision: true
+  - name: Qwen3.8-27B-UD-IQ3_XXS
+    max_input_tokens: 22000            # ctx 24576 · ~15 t/s — hybrid reasoner, quiet desktop only
+    supports_vision: true
+  - name: gemma-4-26B-A4B-it-UD-IQ4_XS
+    max_input_tokens: 22000            # ctx 24576 · ~34 t/s — quality generalist, best vision
+    supports_vision: true
+  - name: gemma-4-E4B-it-UD-Q8_K_XL
+    max_input_tokens: 62000            # ctx 65536 · ~57 t/s — fast generalist, long docs
+    supports_vision: true
+  - name: GLM-4.7-Flash-UD-Q4_K_XL
+    max_input_tokens: 22000            # ctx 24576 · ~21 t/s — quality coder
+  - name: gpt-oss-20b
+    max_input_tokens: 62000            # ctx 65536 · ~38 t/s — fast reasoning + tools
+  - name: gpt-oss-20b-low
+    max_input_tokens: 62000            # ctx 65536 · ~37 t/s — snappy answers
+# Remote access via duskadiy.com (reachable from anywhere; requires an API key —
+# $DUSKADIY_API_KEY, from the gitignored conf.d/secrets.fish). The LAN endpoint above is keyless.
+# Same server, same model ids — reuse the list above via a YAML anchor.
+- type: openai-compatible
+  name: duskadiy
+  api_base: https://llm.duskadiy.com/api/v1
+  models: *lan_models