feat(router): Phase 3 - Dynamic GPU Weighting via Health Scoring

router/router.py

@@ -211,6 +211,23 @@ def is_gpu_busy(model):
     max_c = GPU_MAX_CONCURRENT.get(model, 1)
     return active >= max_c
 
+
+
+# Phase 3: Dynamic GPU Weighting (Health Score)
+def gpu_health_score(model):
+    """Score a GPU based on VRAM, temperature, and load. Lower = better."""
+    h = check_gpu_health(model, sidecar_timeout=1.5, gpu_timeout=1)
+    if h.get("status") == "down":
+        return 999  # never pick down GPUs
+    vram_pct = h.get("vram_pct", 50)
+    temp_c = h.get("temp_c", 50)
+    active = gpu_active_count(model)
+    max_c = GPU_MAX_CONCURRENT.get(model, 1)
+    load_pct = (active / max_c) * 100
+    # Score: lower = better (more headroom, cooler, less loaded)
+    score = vram_pct * 0.4 + max(temp_c - 30, 0) * 0.3 + load_pct * 0.3
+    return score
+
 def select_best_gpu(candidates, reason, agent=""):
     """Pick best GPU, spreading agents across GPUs to prevent hotspots."""
     # Count how many distinct agents are on each GPU
@@ -222,17 +239,19 @@ def select_best_gpu(candidates, reason, agent=""):
                 if r.get("agent_gpu:" + ak["agent"] + ":" + m):
                     count += 1
             gpu_agent_counts[m] = count
+    # Phase 3: Sort candidates by health score before selection
+    sorted_candidates = sorted(candidates, key=gpu_health_score)
     # First pass: prefer GPUs with 0 other agents (fresh GPU for this agent)
-    for m in candidates:
+    for m in sorted_candidates:
         if not is_gpu_busy(m) and gpu_agent_counts.get(m, 0) == 0:
             return {"model": m, "reason": reason}
     # Second pass: prefer GPU this agent is NOT already on (skip own GPU)
     if agent:
-        for m in candidates:
+        for m in sorted_candidates:
             if not is_gpu_busy(m) and not r.get("agent_gpu:" + agent + ":" + m):
                 return {"model": m, "reason": reason}
     # Third pass: any non-busy GPU
-    for m in candidates:
+    for m in sorted_candidates:
         if not is_gpu_busy(m):
             return {"model": m, "reason": reason}
     # All busy — pick least loaded