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/tests/test_vram_manager.py

@@ -65,19 +65,53 @@ async def test_evict_llm_first(manager):
 
 
 @pytest.mark.asyncio
-async def test_evict_cascade_for_large_llm(manager):
+async def test_evict_cascade_asr_survives(manager):
+    """When LLM fits alongside ASR after evicting LLM+TTS, ASR survives."""
     backend = FakeBackend()
     await manager.load_model("cohere-transcribe", model_type="asr", vram_gb=4.0, backend=backend)
     await manager.load_model("chatterbox-multilingual", model_type="tts", vram_gb=2.0, backend=backend)
     await manager.load_model("qwen3.5-4b", model_type="llm", vram_gb=4.0, backend=backend)
-    # 10 GB used. gpt-oss-20b needs 12GB. Evict LLM(4)->free=10. Evict TTS(2)->free=12. Load.
-    await manager.load_model("gpt-oss-20b", model_type="llm", vram_gb=12.0, backend=backend)
+    # 10 GB used. Need 12GB. Evict LLM(4)->free=10. Evict TTS(2)->free=12. ASR+12=16, fits.
+    await manager.load_model("large-llm", model_type="llm", vram_gb=12.0, backend=backend)
     assert not manager.is_loaded("qwen3.5-4b")
     assert not manager.is_loaded("chatterbox-multilingual")
-    assert manager.is_loaded("cohere-transcribe")  # ASR survives if possible
+    assert manager.is_loaded("cohere-transcribe")  # ASR survives
+    assert manager.is_loaded("large-llm")
+
+
+@pytest.mark.asyncio
+async def test_evict_cascade_full_for_huge_llm(manager):
+    """When LLM is too large to fit alongside ASR, everything gets evicted."""
+    backend = FakeBackend()
+    await manager.load_model("cohere-transcribe", model_type="asr", vram_gb=4.0, backend=backend)
+    await manager.load_model("chatterbox-multilingual", model_type="tts", vram_gb=2.0, backend=backend)
+    await manager.load_model("qwen3.5-4b", model_type="llm", vram_gb=4.0, backend=backend)
+    # 10 GB used. gpt-oss-20b needs 13GB. Evict LLM(4)->free=10. TTS(2)->free=12. ASR(4)->free=16. Load alone.
+    await manager.load_model("gpt-oss-20b", model_type="llm", vram_gb=13.0, backend=backend)
+    assert not manager.is_loaded("qwen3.5-4b")
+    assert not manager.is_loaded("chatterbox-multilingual")
+    assert not manager.is_loaded("cohere-transcribe")  # ASR evicted as last resort
     assert manager.is_loaded("gpt-oss-20b")
 
 
+@pytest.mark.asyncio
+async def test_tts_cannot_evict_asr(manager):
+    """TTS request must not evict ASR — it evicts LLM instead."""
+    backend = FakeBackend()
+    await manager.load_model("cohere-transcribe", model_type="asr", vram_gb=4.0, backend=backend)
+    await manager.load_model("qwen3.5-9b-fp8", model_type="llm", vram_gb=9.0, backend=backend)
+    # 13GB used, 3GB free. TTS needs 2GB — fits! Load alongside.
+    await manager.load_model("chatterbox", model_type="tts", vram_gb=2.0, backend=backend)
+    assert manager.is_loaded("cohere-transcribe")
+    assert manager.is_loaded("qwen3.5-9b-fp8")
+    assert manager.is_loaded("chatterbox")
+    # Now replace TTS with a bigger one that needs eviction
+    # 15GB used, 1GB free. New TTS needs 2GB. Evict old TTS(2)->free=3. Load.
+    await manager.load_model("chatterbox-ml", model_type="tts", vram_gb=2.0, backend=backend)
+    assert manager.is_loaded("cohere-transcribe")  # ASR must survive
+    assert manager.is_loaded("chatterbox-ml")
+
+
 @pytest.mark.asyncio
 async def test_asr_evicts_llm_not_reversed(manager):
     """When ASR request arrives and LLM is loaded, evict LLM (lower priority)."""