Update app.py

app.py

@@ -6,12 +6,13 @@ built-in "Free OCR" prompt mode (plain text extraction, no markdown/
 grounding boxes). A "Markdown" mode is also included since it's a
 one-line prompt swap.
 
-This version runs on CPU only (no CUDA, no flash-attention, float32
-instead of bfloat16). DeepSeek-OCR is a vision-language model, so CPU
-inference works but is slow — expect anywhere from ~30s to a few
-minutes per image depending on resolution preset and CPU. Use the
-"Tiny" or "Small" resolution presets for faster (if lower-fidelity)
-results on CPU.
+This version runs on CPU only (no CUDA, no flash-attention). The model
+stays in bfloat16, matching hardcoded dtype casts inside its own custom
+code — PyTorch supports bfloat16 tensors on CPU, just without hardware
+acceleration, so it works but is slow. DeepSeek-OCR is a vision-language
+model, so expect anywhere from ~30s to a few minutes per image depending
+on resolution preset and CPU. Use the "Tiny" or "Small" resolution
+presets for faster (if lower-fidelity) results.
 """
 
 import os
@@ -43,7 +44,13 @@ model = AutoModel.from_pretrained(
 )
 # CPU doesn't support flash-attention or bfloat16 matmul well — stay in
 # float32 on the CPU device.
-model = model.eval().to(DEVICE).to(torch.float32)
+# NOTE: don't cast to float32 here. DeepSeek-OCR's own forward pass hardcodes
+# casting image inputs to bfloat16 internally (independent of the model's
+# actual dtype), so the model's weights must also be bfloat16 or you get a
+# "Input type (BFloat16) and bias type (float) should be the same" crash.
+# PyTorch CPU does support bfloat16 tensors/ops (just without hardware
+# acceleration), so this works, it's just slow.
+model = model.eval().to(DEVICE).to(torch.bfloat16)
 
 # ---------------------------------------------------------------------------
 # DeepSeek-OCR's own `model.infer()` method has `.cuda()` calls hardcoded