Drop hub kernels (no speedup); 90s GPU chunks

app.py

@@ -158,15 +158,8 @@ END_TOKENS = ("<|im_end|>", "<|endoftext|>")
 processor = AutoProcessor.from_pretrained(MODEL_ID)
 tokenizer = processor.tokenizer
 model = AutoModelForImageTextToText.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to("cuda").eval()
-try:  # Hub kernels (causal-conv1d, fused norms/activations) for faster decoding; the eager model works without them
-    from kernels import Mode
 
-    model.set_use_kernels(True, mode=Mode.INFERENCE)
-    print("hub kernels enabled")
-except Exception as e:
-    print(f"hub kernels unavailable ({type(e).__name__}: {e}); using the eager PyTorch paths")
-
-GPU_CHUNK_SECONDS = 100  # one GPU call generates for at most this long; long turns continue in the next call
+GPU_CHUNK_SECONDS = 90  # one GPU call generates for at most this long; long turns continue in the next call
 
 
 @spaces.GPU(duration=120)

requirements.txt

@@ -1,5 +1,4 @@
-transformers==5.22.2
-kernels>=0.17.2,<0.18
+transformers>=5.17.0
 torchvision
 accelerate
 flash-linear-attention