fix: VRAM eviction cascades through all tiers for large LLM loads

The original eviction logic blocked ASR eviction even when an LLM
genuinely needed all 16GB VRAM (e.g., gpt-oss-20b at 13GB). Now uses
two-pass eviction: first evicts lower/same priority, then cascades to
higher priority as last resort. Added tests for ASR-survives and
full-cascade scenarios.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

kischdle/llmux/llmux/vram_manager.py

@@ -70,20 +70,39 @@ class VRAMManager:
         requesting_priority = _PRIORITY[requesting_type]
 
         # Evict in priority order: lowest first (LLM=0, TTS=1, ASR=2).
-        # Rule: never evict highest-priority tier (ASR) for a lower-priority
-        # request.  Same-priority replacement is always allowed (e.g., old LLM
-        # evicted for new LLM).  Lower-priority models are fair game for any
-        # requester — cascade through them until enough VRAM is freed.
+        #
+        # Rule: never evict a higher-priority model to make room for a
+        # lower-priority one.  E.g., a TTS request must not evict ASR —
+        # it should evict the LLM instead.  But an LLM request CAN cascade
+        # through TTS and ASR as a last resort, because there is nothing
+        # lower to evict.  Same-priority replacement is always allowed.
+        #
+        # Pass 1: evict models with priority <= requesting priority
+        #         (lower or same tier).
+        # Pass 2: if still not enough, evict higher-priority models
+        #         in ascending order (only when the requester has no
+        #         lower-priority alternatives left).
         candidates = sorted(self._loaded.values(), key=lambda s: s.priority)
-        for slot in candidates:
+
+        # Pass 1: evict lower and same priority
+        for slot in list(candidates):
             if self.available_vram_gb >= needed_gb:
                 break
-            # Skip if this slot is the highest-priority tier and the requester
-            # is lower priority.  (Protects ASR from eviction by TTS/LLM.)
-            if slot.priority > requesting_priority and slot.model_type == "asr":
-                continue
-            logger.info(
-                f"Evicting {slot.model_id} ({slot.model_type}, {slot.vram_gb}GB)"
-            )
-            await slot.backend.unload(slot.model_id)
-            del self._loaded[slot.model_id]
+            if slot.priority <= requesting_priority:
+                logger.info(
+                    f"Evicting {slot.model_id} ({slot.model_type}, {slot.vram_gb}GB)"
+                )
+                await slot.backend.unload(slot.model_id)
+                del self._loaded[slot.model_id]
+
+        # Pass 2: evict higher priority as last resort
+        if self.available_vram_gb < needed_gb:
+            candidates = sorted(self._loaded.values(), key=lambda s: s.priority)
+            for slot in list(candidates):
+                if self.available_vram_gb >= needed_gb:
+                    break
+                logger.info(
+                    f"Evicting {slot.model_id} ({slot.model_type}, {slot.vram_gb}GB) [last resort]"
+                )
+                await slot.backend.unload(slot.model_id)
+                del self._loaded[slot.model_id]