Upload folder using huggingface_hub

README.md

@@ -0,0 +1,6 @@
+
+This is just the encoder stuffs from "google/t5-v1_1-xl"
+It takes 11GB down to 4GB.
+The script to do the extraction is included here as
+[transform.py](transform.py)
+

config.json

@@ -0,0 +1,31 @@
+{
+  "architectures": [
+    "T5EncoderModel"
+  ],
+  "classifier_dropout": 0.0,
+  "d_ff": 5120,
+  "d_kv": 64,
+  "d_model": 2048,
+  "decoder_start_token_id": 0,
+  "dense_act_fn": "gelu_new",
+  "dropout_rate": 0.1,
+  "eos_token_id": 1,
+  "feed_forward_proj": "gated-gelu",
+  "initializer_factor": 1.0,
+  "is_encoder_decoder": true,
+  "is_gated_act": true,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "t5",
+  "num_decoder_layers": 24,
+  "num_heads": 32,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_max_distance": 128,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.52.4",
+  "use_cache": true,
+  "vocab_size": 32128
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:277bc2373d204e436e9ffaefafa3696b47809a7ae79b25ee284f7eb10fff6e56
+size 4894136632

transform.py

@@ -0,0 +1,21 @@
+#!/bin/env python
+
+# This script extracts the "encoder-only" part from the full t5-xl model
+
+from transformers import T5ForConditionalGeneration, T5EncoderModel
+
+src_model_name = "google/t5-v1_1-xl"
+dst_dir = "./t5-v1_1-xl-encoder-only"
+
+full_model = T5ForConditionalGeneration.from_pretrained(src_model_name)
+# Initialize empty encoder-only model (inherits config, so tokenizer stays compatible)
+encoder_model = T5EncoderModel(full_model.config)
+
+# Get the full state dict, then ditch the parts we dont need
+state_dict = full_model.state_dict()
+encoder_state_dict = {k: v for k, v in state_dict.items() if not k.startswith("decoder.") and not k.startswith("lm_head.")}
+
+encoder_model.load_state_dict(encoder_state_dict)
+
+encoder_model.save_pretrained(dst_dir)
+print(f"Encoder-only model saved to {dst_dir}")