Make model tiny (172M -> 4.5M params)

config.json

@@ -1,11 +1,13 @@
 {
-  "_name_or_path": "google/mt5-small",
+  "_name_or_path": "hf-internal-testing/tiny-random-mt5",
   "architectures": [
     "MT5Model"
   ],
   "d_ff": 1024,
-  "d_kv": 64,
-  "d_model": 512,
+  "d_kv": 4,
+  "d_model": 16,
+  "decoder_attention_heads": 4,
+  "decoder_ffn_dim": 4,
   "decoder_start_token_id": 0,
   "dense_act_fn": "gelu_new",
   "dropout_rate": 0.1,
@@ -17,15 +19,15 @@
   "layer_norm_epsilon": 1e-06,
   "model_type": "mt5",
   "num_decoder_layers": 8,
-  "num_heads": 6,
-  "num_layers": 8,
+  "num_heads": 4,
+  "num_layers": 2,
   "pad_token_id": 0,
   "relative_attention_max_distance": 128,
   "relative_attention_num_buckets": 32,
   "tie_word_embeddings": false,
   "tokenizer_class": "T5Tokenizer",
   "torch_dtype": "float32",
-  "transformers_version": "4.21.1",
+  "transformers_version": "4.22.0.dev0",
   "use_cache": true,
   "vocab_size": 250112
 }

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3c5b7434d0b5e2ea1766572e8c70f235239199c0581cc5339b0a9238e80cb97f
-size 688538311
+oid sha256:728f60112bd2377b5ff8a823d8f8b1d0b18283ca8aae6b171590e81c1707b213
+size 18092403