Modal: switch L4->T4 (L4 pool constrained/Pending); T4+enforce-eager fits 4B

modal_serve_qwen3b.py

@@ -23,7 +23,9 @@ import modal
 MODEL_NAME = "nvidia/Nemotron-Mini-4B-Instruct"
 SERVED_NAME = "nvidia/nemotron-mini-4b-instruct"  # must match app.py ADVISORY_MODEL
 
-GPU_CONFIG = "L4"          # 24GB — ample for a 4B model (~8.4GB FP16)
+GPU_CONFIG = "T4"          # 16GB — fits the 4B (~8.4GB FP16); biggest/most-available pool.
+#                            (L4 was stuck "Pending" — constrained. T4 + --enforce-eager
+#                             below is the reliable combo for a 4B on Turing.)
 VLLM_PORT = 8000
 MINUTES = 60