Switch larger-family control to root-loaded Qwen3 8B

app.py

@@ -37,9 +37,13 @@ model = None
 # larger-family path loaded this model inside the decorated call and exhausted
 # that reservation before producing an artifact.
 GTE_MODEL_PATH = Path("/models/gte-qwen2-1.5b")
-gte_tokenizer = AutoTokenizer.from_pretrained(GTE_MODEL_PATH, padding_side="left")
-gte_model = AutoModel.from_pretrained(
-    GTE_MODEL_PATH,
+gte_tokenizer = None
+gte_model = None
+
+QWEN8_MODEL_PATH = Path("/models/qwen3-embedding-8b")
+qwen8_tokenizer = AutoTokenizer.from_pretrained(QWEN8_MODEL_PATH, padding_side="left")
+qwen8_model = AutoModel.from_pretrained(
+    QWEN8_MODEL_PATH,
     dtype=torch.bfloat16,
     trust_remote_code=True,
 ).to("cuda").eval()
@@ -287,24 +291,14 @@ def _run_family_bf16_control(family: str) -> dict:
     torch.cuda.reset_peak_memory_stats()
     torch.cuda.synchronize()
     allocation_before = int(torch.cuda.memory_allocated())
-    if family == "gte-qwen2-1.5b":
-        family_tokenizer = gte_tokenizer
-        family_model = gte_model
+    if family == "qwen3-embedding-8b":
+        family_tokenizer = qwen8_tokenizer
+        family_model = qwen8_model
         load_seconds = 0.0
         loading_strategy = "root-module-cuda-emulation"
         owns_model = False
     else:
-        load_started = time.perf_counter()
-        family_tokenizer = AutoTokenizer.from_pretrained(
-            spec["path"], padding_side="left"
-        )
-        family_model = AutoModel.from_pretrained(
-            spec["path"], dtype=torch.bfloat16, trust_remote_code=True
-        ).to("cuda").eval()
-        torch.cuda.synchronize()
-        load_seconds = time.perf_counter() - load_started
-        loading_strategy = "in-call-cuda-placement"
-        owns_model = True
+        raise RuntimeError("GTE 1.5B BF16 control is offline during the 8B phase")
     allocation_after_load = int(torch.cuda.memory_allocated())
 
     encode_started = time.perf_counter()