docs: add blog post 'Make DONUT go brrr'🚀

content/post/make_donut_go_brr.md

@@ -0,0 +1,78 @@
+---
+title: Make DONUT go brrr🚀
+date: '2025-01-01'
+Description: In this post we explore how convert DONUT to onnx end-to-end.
+Tags: [DONUT, transformers, onnx, inference-optimization]
+Categories: [machine-learning]
+DisableComments: false
+---
+
+I had several weeks off from work and wanted to do a macro cycle to ramp up my onnx skills[^1]. And *boy* did I go on a journey! This blogpost documents the entire process.
+
+## DONUT architecture
+
+TODO:
+
+## converting image processor
+
+TODO:
+
+## converting the tokenizer
+
+TODO:
+
+## converting the encoder
+
+TODO:
+
+## converting the decoder with kv cache
+
+TODO:
+
+## merging the image processor, encoder, decoder without past
+
+TODO:
+
+## implementing beam search
+
+TODO:
+
+## putting it all together
+
+TODO:
+
+## final session
+
+```python
+so = ort.SessionOptions()
+so.register_custom_ops_library(get_library_path())
+so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_EXTENDED
+
+max_length = 4096
+min_length = 1
+num_beams = 1
+num_return_sequences = 1
+repetition_penalty = 0.0
+
+sess_donut = ort.InferenceSession(
+    output_dir / "donut_e2e.onnx",
+    sess_options=so,
+    providers=["CPUExecutionProvider"],
+)
+sess = {
+    "pixel_values": np.array(sample_img),
+    "max_length": np.array([max_length], dtype=np.int32),
+    "min_length": np.array([min_length], dtype=np.int32),
+    "num_beams": np.array([num_beams], dtype=np.int32),
+    "num_return_sequences": np.array([num_return_sequences], dtype=np.int32),
+    "length_penalty": np.array([0.01], dtype=np.float32),
+    "fairseq": [True],
+}
+sess.run(None, ort_inputs)[0]
+```
+
+## go really brr...
+
+TODO: add more optimization tricks
+
+[^1]: Inspired by this [video](https://www.youtube.com/watch?v=SgaN-4po_cA).