[fsmt test] basic config test with online model + super tiny model (#7860)

* basic config test with online model

* typo

* style

* better test

scripts/fsmt/fsmt-make-tiny-model.py

@@ -1,10 +1,19 @@
 #!/usr/bin/env python
 # coding: utf-8
 
-# this script creates a tiny model that is useful inside tests, when we just want to test that the machinery works,
-# without needing to the check the quality of the outcomes.
-# it will be used then as "stas/tiny-wmt19-en-de"
+# This script creates a super tiny model that is useful inside tests, when we just want to test that
+# the machinery works, without needing to the check the quality of the outcomes.
+#
+# This version creates a tiny model through reduction of a normal pre-trained model, but keeping the
+# full vocab, merges file, and thus also resulting in a larger model due to a large vocab size.
+# This gives ~3MB in total for all files.
+#
+# If you want a 50 times smaller than this see `fsmt-make-super-tiny-model.py`, which is slightly more complicated
+#
+#
+# It will be used then as "stas/tiny-wmt19-en-de"
 
+# Build
 from transformers import FSMTTokenizer, FSMTConfig, FSMTForConditionalGeneration
 mname = "facebook/wmt19-en-de"
 tokenizer = FSMTTokenizer.from_pretrained(mname)
@@ -18,16 +27,20 @@ config.update(dict(
 
 tiny_model = FSMTForConditionalGeneration(config)
 print(f"num of params {tiny_model.num_parameters()}")
-# Test it
+
+# Test
 batch = tokenizer.prepare_seq2seq_batch(["Making tiny model"])
 outputs = tiny_model(**batch, return_dict=True)
 
-print(len(outputs.logits[0]))
+print("test output:", len(outputs.logits[0]))
+
 # Save
 mname_tiny = "tiny-wmt19-en-de"
 tiny_model.half() # makes it smaller
 tiny_model.save_pretrained(mname_tiny)
 tokenizer.save_pretrained(mname_tiny)
 
+print(f"Generated {mname_tiny}")
+
 # Upload
 # transformers-cli upload tiny-wmt19-en-de