add new albert and t5 models

Author: RbRe145

graph_net/test/nlp_model_getter.py

@@ -154,3 +154,93 @@ def get_xlnet_model_and_inputs(model_name, text, dtype):
         enc["attention_mask"] = (input_ids != pad_id).astype("int64")
 
     return model, enc
+
+
+def get_t5_model_and_inputs(model_name, text, dtype):
+    import paddle
+    from paddlenlp.transformers import T5ForConditionalGeneration, T5Tokenizer
+
+    # 1) 分词器（先建 tokenizer 方便取 pad/eos id）
+    tokenizer = T5Tokenizer.from_pretrained(model_name)
+
+    # 2) 编码输入（支持单条或批量 text）
+    enc = tokenizer(
+        text,
+        return_tensors="pd",
+        padding=True,
+        truncation=True,
+        max_length=512,
+    )
+
+    # 补 attention_mask（pad 处为 0，其他为 1）
+    if "attention_mask" not in enc:
+        input_ids = enc["input_ids"]
+        attn_mask = (input_ids != tokenizer.pad_token_id).astype("int64")
+        enc["attention_mask"] = attn_mask
+
+    # 构造 decoder_input_ids：
+    # T5 以 pad_token_id 作为 decoder_start_token_id
+    batch_size = enc["input_ids"].shape[0]
+    decoder_input_ids = paddle.full(
+        shape=[batch_size, 1],
+        fill_value=tokenizer.pad_token_id,
+        dtype="int64",
+    )
+
+    # 3) 加载模型
+    model = T5ForConditionalGeneration.from_pretrained(model_name)
+    if dtype == "float16":
+        model = model.astype(paddle.float16)
+    model.eval()
+
+    # 4) 组装喂给模型的输入
+    inputs = {
+        "input_ids": enc["input_ids"],
+        "attention_mask": enc["attention_mask"],
+        "decoder_input_ids": decoder_input_ids,
+    }
+    return model, inputs
+
+
+def get_albert_model_and_inputs(model_name, text, dtype):
+    """
+    加载 ALBERT backbone（AlbertModel）并构造输入。
+    - model_name 例如: "albert-base-v2", "albert-xxlarge-v1"（PaddleNLP 内置名称）
+    - dtype: "float32" 或 "float16"
+    返回: (model, inputs_dict)
+    """
+    import paddle
+    from paddlenlp.transformers import AlbertConfig, AlbertModel, AlbertTokenizer
+
+    # 1) 读取配置（不触发权重下载）
+    config = AlbertConfig.from_pretrained(model_name)
+
+    # 2) 模型
+    #    若你只需要网络结构，可改成: model = AlbertModel(config)
+    model = AlbertModel(config)
+    if dtype == "float16":
+        model = model.astype(paddle.float16)
+    model.eval()
+
+    # 3) 分词器
+    tokenizer = AlbertTokenizer.from_pretrained(model_name)
+
+    # 若无 pad_token，则回退到 unk_token（ALBERT 没有 eos_token，别设 pad=eos）
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.unk_token
+
+    # 4) 构造输入（支持 str 或 List[str]）
+    enc = tokenizer(
+        text,
+        return_tensors="pd",
+        padding=True,
+        truncation=True,
+        max_length=512,
+    )
+
+    # 显式补 attention_mask（pad 处为 0）
+    if "attention_mask" not in enc:
+        input_ids = enc["input_ids"]
+        enc["attention_mask"] = (input_ids != tokenizer.pad_token_id).astype("int64")
+
+    return model, enc

paddle_samples/PaddleNLP/albert-base-v1/graph_net.json

@@ -0,0 +1,6 @@
+{
+    "framework": "paddle",
+    "model_name": "albert-base-v1",
+    "num_devices_required": 1,
+    "num_nodes_required": 1
+}

paddle_samples/PaddleNLP/albert-base-v1/input_meta.py

@@ -0,0 +1,41 @@
+class Program_weight_tensor_data_0:
+    name = "data_0"
+    shape = [1, 21]
+    dtype = "int64"
+    data = [
+        2,
+        10975,
+        15,
+        51,
+        204,
+        25,
+        1909,
+        9,
+        31,
+        589,
+        2477,
+        88,
+        370,
+        816,
+        2761,
+        17,
+        66,
+        2607,
+        18,
+        9,
+        3,
+    ]
+
+
+class Program_weight_tensor_data_1:
+    name = "data_1"
+    shape = [1, 21]
+    dtype = "int64"
+    data = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
+
+
+class Program_weight_tensor_data_2:
+    name = "data_2"
+    shape = [1, 21]
+    dtype = "int64"
+    data = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]