refactor: LiteLLM unified routing via JeanOrchestrator
- All models now route through single jean-proxy:3012 → orchestrator:8080 - Removed jean-proxy-moe (no longer needed with auto-switch orchestrator) - Added qwen3.6-moe and qwen3.6-moe-reason model routes - Updated architecture comments
This commit is contained in:
+33
-12
@@ -2,21 +2,16 @@
|
||||
# LiteLLM Proxy — Routing pour Effi
|
||||
# Endpoint OpenAI-compatible sur :4000 (host network)
|
||||
#
|
||||
# Backend unique : Jean (llama.cpp/ROCm sur PC gaming .11:8080)
|
||||
# Backend unique : JeanOrchestrator (port 8080 sur PC gaming .11)
|
||||
# via jean-proxy :3012 + WOL auto si PC eteint.
|
||||
#
|
||||
# CONTRAINTE VRAM : la RX 7800 XT a 16 Go. Le modele Qwen3.6-27B IQ4_XS
|
||||
# occupe ~15.3 Go. Ollama ne peut PAS cohabiter (llama3.1-8b = ~5 Go
|
||||
# supplémentaires -> OOM). On abandonne Ollama et on expose 3 profils
|
||||
# du meme modele Qwen3.6, avec ou sans raisonnement (toggle natif
|
||||
# chat_template_kwargs.enable_thinking):
|
||||
# - qwen3.6-reason : raisonnement ON (tâches complexes, code, analyse)
|
||||
# - qwen3.6-fast : raisonnement OFF (chat rapide, mail, extraction simple)
|
||||
# - qwen3.6-27b : alias (= reason, défaut)
|
||||
# L'orchestrateur gere automatiquement le switch entre :
|
||||
# - Qwen3.6-35B MoE I-Nano (defaut, 11.7 Go, 100k ctx, ~70 tok/s)
|
||||
# - Qwen3.6-27B IQ4_XS (code/analyse, 14.6 Go, 55k ctx, ~60 tok/s)
|
||||
#
|
||||
# Latence observee (campagne de tests 2026-06-27) :
|
||||
# reason : TTFT 11-38s, 60-75 tok/s (le modele reflechit avant de repondre)
|
||||
# fast : TTFT 0.4-1s, 60-75 tok/s (reponse quasi immediate)
|
||||
# Tous les modeles passent par le meme proxy :3012.
|
||||
# L'orchestrateur detecte le modele demande et charge le bon LLM.
|
||||
# Retour automatique au MoE apres 15 min sans 27B.
|
||||
# ═══════════════════════════════════════════════════════════
|
||||
|
||||
model_list:
|
||||
@@ -72,6 +67,32 @@ model_list:
|
||||
chat_template_kwargs:
|
||||
enable_thinking: false
|
||||
|
||||
# ═════════════════════════════════════════════════════════
|
||||
# Profil 3 — Qwen3.6-35B MoE APEX I-Nano (chat Hermes, usage quotidien)
|
||||
# Modele MoE : 35B total / 3B actifs, 11.7 Go, cache KV 5x plus petit
|
||||
# que le 27B -> contexte 100k possible. Cohabite avec ComfyUI.
|
||||
# Route via le meme proxy :3012 que le 27B (orchestrateur fait le switch).
|
||||
# ═════════════════════════════════════════════════════════
|
||||
- model_name: qwen3.6-moe
|
||||
litellm_params:
|
||||
model: openai/qwen3.6-moe
|
||||
api_base: http://localhost:3012/v1
|
||||
api_key: "null"
|
||||
timeout: 600
|
||||
extra_body:
|
||||
chat_template_kwargs:
|
||||
enable_thinking: false
|
||||
|
||||
- model_name: qwen3.6-moe-reason
|
||||
litellm_params:
|
||||
model: openai/qwen3.6-moe-reason
|
||||
api_base: http://localhost:3012/v1
|
||||
api_key: "null"
|
||||
timeout: 600
|
||||
extra_body:
|
||||
chat_template_kwargs:
|
||||
enable_thinking: true
|
||||
|
||||
# ═════════════════════════════════════════════════════════
|
||||
# Backend Ollama — COMMENTE (contention VRAM, voir entete)
|
||||
# Pour reactiver si dispo (autre machine / GPU dedie), decommenter
|
||||
|
||||
Reference in New Issue
Block a user