pin transformers to PR #45547 merge commit (auto-detect disable_mmap on hf-mount)

app.py

@@ -1,8 +1,13 @@
 import os
+import sys
 import time
 import traceback
 
+sys.stdout.reconfigure(line_buffering=True)
+sys.stderr.reconfigure(line_buffering=True)
+
 T_START = time.time()
+print(f"[app] __main__ start t={T_START:.2f}", flush=True)
 
 LOAD_STRATEGY = os.environ.get("LOAD_STRATEGY", "normal")
 MODEL_ID = "google/gemma-4-E2B-it"
@@ -19,9 +24,12 @@ STATS = {
 
 
 def _profile():
+    print("[app] profile: importing torch...", flush=True)
     import torch
+    print("[app] profile: importing transformers...", flush=True)
     import transformers
     from transformers import AutoModelForImageTextToText, AutoProcessor
+    print(f"[app] profile: torch={torch.__version__}, transformers={transformers.__version__}", flush=True)
 
     STATS["transformers_version"] = transformers.__version__
     STATS["torch_version"] = torch.__version__
@@ -33,11 +41,14 @@ def _profile():
     t_imports_done = time.time()
     STATS["imports_seconds"] = t_imports_done - T_START
 
+    print(f"[app] loading processor from {MODEL_SOURCE!r}", flush=True)
     t0 = time.time()
     processor = AutoProcessor.from_pretrained(MODEL_SOURCE)
     t1 = time.time()
+    print(f"[app] processor loaded in {t1-t0:.2f}s", flush=True)
     STATS["processor_load_seconds"] = t1 - t0
 
+    print(f"[app] loading model from {MODEL_SOURCE!r}", flush=True)
     t2 = time.time()
     model = AutoModelForImageTextToText.from_pretrained(
         MODEL_SOURCE,
@@ -45,6 +56,7 @@ def _profile():
         device_map="auto",
     )
     t3 = time.time()
+    print(f"[app] model loaded in {t3-t2:.2f}s", flush=True)
     STATS["model_load_seconds"] = t3 - t2
     STATS["total_load_seconds"] = t3 - t0
 
@@ -57,10 +69,12 @@ def _profile():
         return_tensors="pt",
     ).to(model.device)
 
+    print("[app] generating...", flush=True)
     t4 = time.time()
     with torch.inference_mode():
         out = model.generate(**inputs, max_new_tokens=64, do_sample=False)
     t5 = time.time()
+    print(f"[app] generate done in {t5-t4:.2f}s", flush=True)
     STATS["predict_seconds"] = t5 - t4
 
     new_tokens = out[0][inputs["input_ids"].shape[1]:]