[Feature] Add llama2 native implements (#235)

* add llama2 native implements

* rename configs/eval_llama_7b.py

---------
Co-authored-by: zhoufengzhe <zhoufengzhe@pjlab.org.cn>

configs/eval_llama2_7b.py

+from mmengine.config import read_base
+
+with read_base():
+    from .datasets.collections.base_medium_llama import piqa_datasets, siqa_datasets
+    from .models.llama2_7b import models
+
+
+datasets = [*piqa_datasets, *siqa_datasets]

configs/models/llama2_13b.py

+from opencompass.models import Llama2
+
+# Please follow the instruction in the Meta AI website https://github.com/facebookresearch/llama
+# and download the LLaMA-2 model and tokenizer to the path './models/llama2/llama/'.
+#
+# The LLaMA requirement is also needed to be installed.
+#
+# git clone https://github.com/facebookresearch/llama.git
+# cd llama
+# pip install -e .
+
+models = [
+    dict(
+        abbr="llama-2-13b",
+        type=Llama2,
+        path="./models/llama2/llama/llama-2-13b/",
+        tokenizer_path="./models/llama2/llama/tokenizer.model",
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=16,
+        run_cfg=dict(num_gpus=2, num_procs=2),
+    ),
+]

configs/models/llama2_70b.py

+from opencompass.models import Llama2
+
+# Please follow the instruction in the Meta AI website https://github.com/facebookresearch/llama
+# and download the LLaMA-2 model and tokenizer to the path './models/llama2/llama/'.
+#
+# The LLaMA requirement is also needed to be installed.
+#
+# git clone https://github.com/facebookresearch/llama.git
+# cd llama
+# pip install -e .
+
+models = [
+    dict(
+        abbr="llama-2-70b",
+        type=Llama2,
+        path="./models/llama2/llama/llama-2-70b/",
+        tokenizer_path="./models/llama2/llama/tokenizer.model",
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=16,
+        run_cfg=dict(num_gpus=8, num_procs=8),
+    ),
+]

configs/models/llama2_7b.py

+from opencompass.models import Llama2
+
+# Please follow the instruction in the Meta AI website https://github.com/facebookresearch/llama
+# and download the LLaMA-2 model and tokenizer to the path './models/llama2/llama/'.
+#
+# The LLaMA requirement is also needed to be installed.
+#
+# git clone https://github.com/facebookresearch/llama.git
+# cd llama
+# pip install -e .
+
+models = [
+    dict(
+        abbr="llama-2-7b",
+        type=Llama2,
+        path="./models/llama2/llama/llama-2-7b/",
+        tokenizer_path="./models/llama2/llama/tokenizer.model",
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=16,
+        run_cfg=dict(num_gpus=1, num_procs=1),
+    ),
+]

configs/models/llama_13b.py

+from opencompass.models import Llama2
+
+# Please follow the instruction in the Meta AI website https://github.com/facebookresearch/llama/tree/llama_v1
+# and download the LLaMA model and tokenizer to the path './models/llama/'.
+#
+# The LLaMA requirement is also needed to be installed.
+# *Note* that the LLaMA-2 branch is fully compatible with LLAMA-1, and the LLaMA-2 branch is used here.
+#
+# git clone https://github.com/facebookresearch/llama.git
+# cd llama
+# pip install -e .
+
+models = [
+    dict(
+        abbr="llama-13b",
+        type=Llama2,
+        path="./models/llama/13B/",
+        tokenizer_path="./models/llama/tokenizer.model",
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=16,
+        run_cfg=dict(num_gpus=2, num_procs=2),
+    ),
+]

configs/models/llama_30b.py

+from opencompass.models import Llama2
+
+# Please follow the instruction in the Meta AI website https://github.com/facebookresearch/llama/tree/llama_v1
+# and download the LLaMA model and tokenizer to the path './models/llama/'.
+#
+# The LLaMA requirement is also needed to be installed.
+# *Note* that the LLaMA-2 branch is fully compatible with LLAMA-1, and the LLaMA-2 branch is used here.
+#
+# git clone https://github.com/facebookresearch/llama.git
+# cd llama
+# pip install -e .
+
+models = [
+    dict(
+        abbr="llama-30b",
+        type=Llama2,
+        path="./models/llama/30B/",
+        tokenizer_path="./models/llama/tokenizer.model",
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=16,
+        run_cfg=dict(num_gpus=4, num_procs=4),
+    ),
+]

configs/models/llama_65b.py

+from opencompass.models import Llama2
+
+# Please follow the instruction in the Meta AI website https://github.com/facebookresearch/llama/tree/llama_v1
+# and download the LLaMA model and tokenizer to the path './models/llama/'.
+#
+# The LLaMA requirement is also needed to be installed.
+# *Note* that the LLaMA-2 branch is fully compatible with LLAMA-1, and the LLaMA-2 branch is used here.
+#
+# git clone https://github.com/facebookresearch/llama.git
+# cd llama
+# pip install -e .
+
+models = [
+    dict(
+        abbr="llama-65b",
+        type=Llama2,
+        path="./models/llama/65B/",
+        tokenizer_path="./models/llama/tokenizer.model",
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=16,
+        run_cfg=dict(num_gpus=8, num_procs=8),
+    ),
+]

configs/models/llama_7b.py

+from opencompass.models import Llama2
+
+# Please follow the instruction in the Meta AI website https://github.com/facebookresearch/llama/tree/llama_v1
+# and download the LLaMA model and tokenizer to the path './models/llama/'.
+#
+# The LLaMA requirement is also needed to be installed.
+# *Note* that the LLaMA-2 branch is fully compatible with LLAMA-1, and the LLaMA-2 branch is used here.
+#
+# git clone https://github.com/facebookresearch/llama.git
+# cd llama
+# pip install -e .
+
+models = [
+    dict(
+        abbr="llama-7b",
+        type=Llama2,
+        path="./models/llama/7B/",
+        tokenizer_path="./models/llama/tokenizer.model",
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=16,
+        run_cfg=dict(num_gpus=1, num_procs=1),
+    ),
+]

opencompass/models/__init__.py

@@ -4,5 +4,5 @@ from .glm import GLM130B  # noqa: F401, F403
 from .huggingface import HuggingFace  # noqa: F401, F403
 from .huggingface import HuggingFaceCausalLM  # noqa: F401, F403
 from .intern_model import InternLM  # noqa: F401, F403
-from .llama2 import Llama2Chat  # noqa: F401, F403
+from .llama2 import Llama2, Llama2Chat  # noqa: F401, F403
 from .openai_api import OpenAI  # noqa: F401

opencompass/models/llama2.py

 from typing import Dict, List, Optional, Union
 
+import torch
+
 from opencompass.models.base import BaseModel
 from opencompass.models.base_api import APITemplateParser
 from opencompass.utils.logging import get_logger
@@ -8,6 +10,94 @@ from opencompass.utils.prompt import PromptList
 PromptType = Union[PromptList, str]
 
 
+class Llama2(BaseModel):
+    """LLaMA-2 model wrapper
+    https://github.com/facebookresearch/llama/tree/main.
+
+    Args:
+        path (str): path to the model directory
+        max_seq_len (int): max sequence length
+        max_batch_size (int): max batch size
+        tokenizer_only (bool): whether to load tokenizer only
+        tokenizer_path (str): path to the tokenizer directory
+        meta_template (dict): meta template for the model
+    """
+
+    def __init__(
+        self,
+        path: str,
+        max_seq_len: int = 2048,
+        max_batch_size: int = 16,
+        tokenizer_only: bool = False,
+        tokenizer_path: Optional[str] = None,
+        meta_template: Optional[Dict] = None,
+    ):  # noqa
+        if tokenizer_only:
+            self._load_tokenizer(tokenizer_path=tokenizer_path)
+        else:
+            self._load_model(path=path,
+                             max_seq_len=max_seq_len,
+                             max_batch_size=max_batch_size,
+                             tokenizer_path=tokenizer_path)
+        self.max_seq_len = max_seq_len
+        self.template_parser = APITemplateParser(meta_template)
+        self.logger = get_logger()
+
+    def _load_model(self,
+                    path: str,
+                    max_seq_len: int,
+                    max_batch_size: int,
+                    tokenizer_path: Optional[str] = None):
+        from llama import Llama
+        self.generator = Llama.build(path, tokenizer_path, max_seq_len,
+                                     max_batch_size)
+        self.tokenizer = self.generator.tokenizer
+        self.model = self.generator.model
+
+    def _load_tokenizer(self, tokenizer_path: str):
+        from llama import Tokenizer
+        self.tokenizer = Tokenizer(tokenizer_path)
+
+    def generate(self, inputs: List[str], max_out_len: int) -> List[str]:
+        out = self.generator.text_completion(
+            inputs,
+            temperature=0,
+            max_gen_len=max_out_len,
+        )
+        return [i['generation'] for i in out]
+
+    def get_ppl(self,
+                inputs: List[str],
+                mask_length: Optional[List[int]] = None) -> List[float]:
+        assert mask_length is None, 'mask_length is not supported'
+        bsz = len(inputs)
+        params = self.model.params
+        assert bsz <= params.max_batch_size, (bsz, params.max_batch_size)
+        # tokenize
+        prompt_tokens = [self.tokenizer.encode(x, True, False) for x in inputs]
+        max_prompt_size = max([len(t) for t in prompt_tokens])
+        total_len = min(params.max_seq_len, max_prompt_size)
+        tokens = torch.zeros((bsz, total_len)).cuda().long()
+        for k, t in enumerate(prompt_tokens):
+            num_token = min(total_len, len(t))
+            tokens[k, :num_token] = torch.tensor(t[-num_token:]).long()
+        # forward
+        outputs = self.model.forward(tokens, 0)
+        # compute ppl
+        shift_logits = outputs[..., :-1, :].contiguous()
+        shift_labels = tokens[..., 1:].contiguous()
+        shift_logits = shift_logits.view(-1, shift_logits.size(-1))
+        shift_labels = shift_labels.view(-1)
+        loss_fct = torch.nn.CrossEntropyLoss(reduction='none', ignore_index=0)
+        loss = loss_fct(shift_logits, shift_labels).view(bsz, -1)
+        lens = (tokens != 0).sum(-1).cpu().numpy()
+        ce_loss = loss.sum(-1).cpu().detach().numpy() / lens
+        return ce_loss
+
+    def get_token_len(self, prompt: str) -> int:
+        return len(self.tokenizer.encode(prompt, True, True))
+
+
 class Llama2Chat(BaseModel):
     """LLaMA-2 chat model wrapper
     https://github.com/facebookresearch/llama/tree/main.