Add Llama Flax Implementation (#24587)

* Copies `modeling_flax_gpt_neo.py` to start

* MLP Block. WIP Attention and Block

* Adds Flax implementation of `LlamaMLP`
Validated with in-file test.
Some slight numeric differences, but assuming it isn't an issue

* Adds `FlaxLlamaRMSNorm` layer
`flax.linen` includes `RMSNorm` layer but not necessarily in all
versions. Hence, we add in-file.

* Adds FlaxLlamaAttention
Copied from GPT-J as it has efficient caching implementation as well as
rotary embeddings.
Notice numerically different, but not by a huge amount. Needs
investigating

* Adds `FlaxLlamaDecoderLayer`
numerically inaccurate, debugging..

* debugging rotary mismatch
gptj uses interleaved whilst llama uses contiguous
i think they match now but still final result is wrong.
maybe drop back to just debugging attention layer?

* fixes bug with decoder layer
still somewhat numerically inaccurate, but close enough for now

* adds markers for what to implement next
the structure here diverges a lot from the PT version.
not a big fan of it, but just get something working for now

* implements `FlaxLlamaBlockCollection`]
tolerance must be higher than expected, kinda disconcerting

* Adds `FlaxLlamaModule`
equivalent PyTorch model is `LlamaModel`
yay! a language model🤗

* adds `FlaxLlamaForCausalLMModule`
equivalent to `LlamaForCausalLM`
still missing returning dict or tuple, will add later

* start porting pretrained wrappers
realised it probably needs return dict as a prereq

* cleanup, quality, style

* readds `return_dict` and model output named tuples

* (tentatively) pretrained wrappers work 🔥

* fixes numerical mismatch in `FlaxLlamaRMSNorm`
seems `jax.lax.rsqrt` does not match `torch.sqrt`.
manually computing `1 / jax.numpy.sqrt` results in matching values.

* [WIP] debugging numerics

* numerical match
I think issue was accidental change of backend. forcing CPU fixes test.
We expect some mismatch on GPU.

* adds in model and integration tests for Flax Llama
summary of failing:
- mul invalid combination of dimensions
- one numerical mismatch
- bf16 conversion (maybe my local backend issue)
- params are not FrozenDict

* adds missing TYPE_CHECKING import and `make fixup`

* adds back missing docstrings
needs review on quality of docstrings, not sure what is required.
Furthermore, need to check if `CHECKPOINT_FOR_DOC` is valid. See TODO

* commenting out equivalence test as can just use common

* debugging

* Fixes bug where mask and pos_ids were swapped in pretrained models
This results in all tests passing now 🔥



* cleanup of modeling file

* cleanup of test file

* Resolving simpler review comments

* addresses more minor review comments

* fixing introduced pytest errors from review

* wip additional slow tests

* wip tests
need to grab a GPU machine to get real logits for comparison
otherwise, slow tests should be okay

* `make quality`, `make style`

* adds slow integration tests
- checking logits
- checking hidden states
- checking generation outputs

* `make fix-copies`

* fix mangled function following `make fix-copies`

* adds missing type checking imports

* fixes missing parameter checkpoint warning

* more finegrained 'Copied from' tags
avoids issue of overwriting `LLAMA_INPUTS_DOCSTRING`

* swaps import guards
??? how did these get swapped initially?

* removing `inv_freq` again as pytorch version has now removed

* attempting to get CI to pass

* adds doc entries for llama flax models

* fixes typo in __init__.py imports

* adds back special equivalence tests
these come from the gpt neo flax tests. there is special behaviour for these models that needs to override the common version

* overrides tests with dummy to see if CI passes
need to fill in these tests later

* adds my contribution to docs

* `make style; make quality`

* replaces random masking with fixed to work with flax version

* `make quality; make style`

* Update src/transformers/models/llama/modeling_flax_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update src/transformers/models/llama/modeling_flax_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update src/transformers/models/llama/modeling_flax_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update src/transformers/models/llama/modeling_flax_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update src/transformers/models/llama/modeling_flax_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update src/transformers/models/llama/modeling_flax_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* updates `x`->`tensor` in `rotate_half`

* addresses smaller review comments

* Update docs/source/en/model_doc/llama.md
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* adds integration test class

* adds `dtype` to rotary embedding to cast outputs

* adds type to flax llama rotary layer

* `make style`

* `make fix-copies`

* Apply suggestions from code review
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* applies suggestions from review

* Update modeling_flax_llama.py

* `make fix-copies`

* Update tests/models/llama/test_modeling_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Update src/transformers/models/llama/modeling_flax_llama.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* fixes shape mismatch in FlaxLlamaMLP

* applies some suggestions from reviews

* casts attn output logits to f32 regardless of dtype

* adds attn bias using `LlamaConfig.attention_bias`

* adds Copied From comments to Flax Llama test

* mistral and persimmon test change -copy from llama

* updates docs index

* removes Copied from in tests

it was preventing `make fix-copies` from succeeding

* quality and style

* ignores FlaxLlama input docstring

* adds revision to `_CHECKPOINT_FOR_DOC`

* repo consistency and quality

* removes unused import

* removes copied from from Phi test

now diverges from llama tests following FlaxLlama changes

* adds `_REAL_CHECKPOINT_FOR_DOC`

* removes refs from pr tests

* reformat to make ruff happy

---------
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

docs/source/en/index.md

-<!--Copyright 2020 The HuggingFace Team. All rights reserved.
+        <!--Copyright 2020 The HuggingFace Team. All rights reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
 the License. You may obtain a copy of the License at
@@ -94,7 +94,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                       [CLIPSeg](model_doc/clipseg)                       |       ✅        |         ❌         |      ❌      |
 |                          [CLVP](model_doc/clvp)                          |       ✅        |         ❌         |      ❌      |
 |                       [CodeGen](model_doc/codegen)                       |       ✅        |         ❌         |      ❌      |
-|                    [CodeLlama](model_doc/code_llama)                     |       ✅        |         ❌         |      ❌      |
+|                    [CodeLlama](model_doc/code_llama)                     |       ✅        |         ❌         |      ✅      |
 |              [Conditional DETR](model_doc/conditional_detr)              |       ✅        |         ❌         |      ❌      |
 |                      [ConvBERT](model_doc/convbert)                      |       ✅        |         ✅         |      ❌      |
 |                      [ConvNeXT](model_doc/convnext)                      |       ✅        |         ✅         |      ❌      |
@@ -167,8 +167,8 @@ Flax), PyTorch, and/or TensorFlow.
 |                           [LED](model_doc/led)                           |       ✅        |         ✅         |      ❌      |
 |                         [LeViT](model_doc/levit)                         |       ✅        |         ❌         |      ❌      |
 |                          [LiLT](model_doc/lilt)                          |       ✅        |         ❌         |      ❌      |
-|                         [LLaMA](model_doc/llama)                         |       ✅        |         ❌         |      ❌      |
-|                        [Llama2](model_doc/llama2)                        |       ✅        |         ❌         |      ❌      |
+|                         [LLaMA](model_doc/llama)                         |       ✅        |         ❌         |      ✅      |
+|                        [Llama2](model_doc/llama2)                        |       ✅        |         ❌         |      ✅      |
 |                    [Longformer](model_doc/longformer)                    |       ✅        |         ✅         |      ❌      |
 |                        [LongT5](model_doc/longt5)                        |       ✅        |         ❌         |      ✅      |
 |                          [LUKE](model_doc/luke)                          |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/llama.md

@@ -50,6 +50,9 @@ come in several checkpoints they each contain a part of each weight of the model
 
 - The LLaMA tokenizer is a BPE model based on [sentencepiece](https://github.com/google/sentencepiece). One quirk of sentencepiece is that when decoding a sequence, if the first token is the start of the word (e.g. "Banana"), the tokenizer does not prepend the prefix space to the string.
 
+This model was contributed by [zphang](https://huggingface.co/zphang) with contributions from [BlackSamorez](https://huggingface.co/BlackSamorez). The code of the implementation in Hugging Face is based on GPT-NeoX [here](https://github.com/EleutherAI/gpt-neox). The original code of the authors can be found [here](https://github.com/facebookresearch/llama). The Flax version of the implementation was contributed by [afmck](https://huggingface.co/afmck) with the code in the implementation based on Hugging Face's Flax GPT-Neo.
+
+
 Based on the original LLaMA model, Meta AI has released some follow-up works:
 
 - **Llama2**: Llama2 is an improved version of Llama with some architectural tweaks (Grouped Query Attention), and is pre-trained on 2Trillion tokens. Refer to the documentation of Llama2 which can be found [here](llama2).
@@ -112,3 +115,13 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
 
 [[autodoc]] LlamaForSequenceClassification
     - forward
+
+## FlaxLlamaModel
+
+[[autodoc]] FlaxLlamaModel
+    - __call__
+
+## FlaxLlamaForCausalLM
+
+[[autodoc]] FlaxLlamaForCausalLM
+    - __call__

src/transformers/__init__.py

@@ -4554,6 +4554,7 @@ else:
         ["FlaxGPTNeoForCausalLM", "FlaxGPTNeoModel", "FlaxGPTNeoPreTrainedModel"]
     )
     _import_structure["models.gptj"].extend(["FlaxGPTJForCausalLM", "FlaxGPTJModel", "FlaxGPTJPreTrainedModel"])
+    _import_structure["models.llama"].extend(["FlaxLlamaForCausalLM", "FlaxLlamaModel", "FlaxLlamaPreTrainedModel"])
     _import_structure["models.longt5"].extend(
         [
             "FlaxLongT5ForConditionalGeneration",
@@ -8631,6 +8632,11 @@ if TYPE_CHECKING:
             FlaxGPTJModel,
             FlaxGPTJPreTrainedModel,
         )
+        from .models.llama import (
+            FlaxLlamaForCausalLM,
+            FlaxLlamaModel,
+            FlaxLlamaPreTrainedModel,
+        )
         from .models.longt5 import (
             FlaxLongT5ForConditionalGeneration,
             FlaxLongT5Model,

src/transformers/modeling_flax_utils.py

@@ -1267,7 +1267,9 @@ def overwrite_call_docstring(model_class, docstring):
     model_class.__call__ = add_start_docstrings_to_model_forward(docstring)(model_class.__call__)
 
 
-def append_call_sample_docstring(model_class, checkpoint, output_type, config_class, mask=None, revision=None):
+def append_call_sample_docstring(
+    model_class, checkpoint, output_type, config_class, mask=None, revision=None, real_checkpoint=None
+):
     model_class.__call__ = copy_func(model_class.__call__)
     model_class.__call__ = add_code_sample_docstrings(
         checkpoint=checkpoint,
@@ -1275,6 +1277,7 @@ def append_call_sample_docstring(model_class, checkpoint, output_type, config_cl
         config_class=config_class,
         model_cls=model_class.__name__,
         revision=revision,
+        real_checkpoint=real_checkpoint,
     )(model_class.__call__)
 
 

src/transformers/models/auto/modeling_flax_auto.py

@@ -43,6 +43,7 @@ FLAX_MODEL_MAPPING_NAMES = OrderedDict(
         ("gpt2", "FlaxGPT2Model"),
         ("gpt_neo", "FlaxGPTNeoModel"),
         ("gptj", "FlaxGPTJModel"),
+        ("llama", "FlaxLlamaModel"),
         ("longt5", "FlaxLongT5Model"),
         ("marian", "FlaxMarianModel"),
         ("mbart", "FlaxMBartModel"),
@@ -146,6 +147,7 @@ FLAX_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("gpt2", "FlaxGPT2LMHeadModel"),
         ("gpt_neo", "FlaxGPTNeoForCausalLM"),
         ("gptj", "FlaxGPTJForCausalLM"),
+        ("llama", "FlaxLlamaForCausalLM"),
         ("opt", "FlaxOPTForCausalLM"),
         ("roberta", "FlaxRobertaForCausalLM"),
         ("roberta-prelayernorm", "FlaxRobertaPreLayerNormForCausalLM"),

src/transformers/models/bloom/modeling_bloom.py

@@ -489,7 +489,7 @@ class BloomPreTrainedModel(PreTrainedModel):
 
     @staticmethod
     def _convert_to_bloom_cache(
-        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]]
+        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]],
     ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
         """
         Converts the cache to the format expected by Bloom, i.e. to tuple(tuple([batch_size * num_heads, ...]))

src/transformers/models/fuyu/image_processing_fuyu.py

@@ -54,7 +54,7 @@ logger = logging.get_logger(__name__)
 
 
 def make_list_of_list_of_images(
-    images: Union[List[List[ImageInput]], List[ImageInput], ImageInput]
+    images: Union[List[List[ImageInput]], List[ImageInput], ImageInput],
 ) -> List[List[ImageInput]]:
     if is_valid_image(images):
         return [[images]]

src/transformers/models/llama/__init__.py

@@ -16,6 +16,7 @@ from typing import TYPE_CHECKING
 from ...utils import (
     OptionalDependencyNotAvailable,
     _LazyModule,
+    is_flax_available,
     is_sentencepiece_available,
     is_tokenizers_available,
     is_torch_available,
@@ -55,6 +56,14 @@ else:
         "LlamaForSequenceClassification",
     ]
 
+try:
+    if not is_flax_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_flax_llama"] = ["FlaxLlamaForCausalLM", "FlaxLlamaModel", "FlaxLlamaPreTrainedModel"]
+
 
 if TYPE_CHECKING:
     from .configuration_llama import LLAMA_PRETRAINED_CONFIG_ARCHIVE_MAP, LlamaConfig
@@ -83,6 +92,14 @@ if TYPE_CHECKING:
     else:
         from .modeling_llama import LlamaForCausalLM, LlamaForSequenceClassification, LlamaModel, LlamaPreTrainedModel
 
+    try:
+        if not is_flax_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_flax_llama import FlaxLlamaForCausalLM, FlaxLlamaModel, FlaxLlamaPreTrainedModel
+
 
 else:
     import sys

src/transformers/models/llama/modeling_flax_llama.py

+# coding=utf-8
+# Copyright 2023 Meta AI, EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Flax LLaMA model."""
+from functools import partial
+from typing import Optional, Tuple
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
+from flax.linen.attention import dot_product_attention_weights
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax import lax
+
+from ...modeling_flax_outputs import FlaxBaseModelOutput, FlaxCausalLMOutput
+from ...modeling_flax_utils import ACT2FN, FlaxPreTrainedModel, append_call_sample_docstring
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_llama import LlamaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+_CHECKPOINT_FOR_DOC = "afmck/testing-llama-tiny"
+_REAL_CHECKPOINT_FOR_DOC = "openlm-research/open_llama_3b_v2"
+
+LLAMA_START_DOCSTRING = r"""
+
+    This model inherits from [`FlaxPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a Flax Linen
+    [flax.nn.Module](https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html) subclass. Use it as a
+    regular Flax Module and refer to the Flax documentation for all matter related to general usage and behavior.
+
+    Finally, this model supports inherent JAX features such as:
+
+    - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+    - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+    - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+    - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+    Parameters:
+        config ([`LlamaConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~FlaxPreTrainedModel.from_pretrained`] method to load the model weights.
+        dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
+            The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16`, or
+            `jax.numpy.bfloat16`.
+
+            This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
+            specified all the computation will be performed with the given `dtype`.
+
+            **Note that this only specifies the dtype of the computation and does not influence the dtype of model
+            parameters.**
+
+            If you wish to change the dtype of the model parameters, see [`~FlaxPreTrainedModel.to_fp16`] and
+            [`~FlaxPreTrainedModel.to_bf16`].
+"""
+
+LLAMA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`numpy.ndarray` of shape `(batch_size, input_ids_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`numpy.ndarray` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Dict[str, np.ndarray]`, *optional*, returned by `init_cache` or when passing previous `past_key_values`):
+            Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
+            auto-regressive decoding. Pre-computed key and value hidden-states are of shape *[batch_size, max_length]*.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+def create_sinusoidal_positions(num_pos, dim):
+    inv_freq = 1.0 / (10000 ** (np.arange(0, dim, 2) / dim))
+    freqs = np.einsum("i , j -> i j", np.arange(num_pos), inv_freq).astype("float32")
+
+    emb = np.concatenate((freqs, freqs), axis=-1)
+    out = np.concatenate((np.sin(emb)[:, None, :], np.cos(emb)[:, None, :]), axis=-1)
+    return jnp.array(out[:, :, :num_pos])
+
+
+def rotate_half(tensor):
+    """Rotates half the hidden dims of the input."""
+    rotate_half_tensor = jnp.concatenate(
+        (-tensor[..., tensor.shape[-1] // 2 :], tensor[..., : tensor.shape[-1] // 2]), axis=-1
+    )
+    return rotate_half_tensor
+
+
+def apply_rotary_pos_emb(tensor, sin_pos, cos_pos):
+    return (tensor * cos_pos) + (rotate_half(tensor) * sin_pos)
+
+
+class FlaxLlamaRMSNorm(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.epsilon = self.config.rms_norm_eps
+        self.weight = self.param("weight", lambda _, shape: jnp.ones(shape), self.config.hidden_size)
+
+    def __call__(self, hidden_states):
+        variance = jnp.asarray(hidden_states, dtype=jnp.float32)
+        variance = jnp.power(variance, 2)
+        variance = variance.mean(-1, keepdims=True)
+        # use `jax.numpy.sqrt` as `jax.lax.rsqrt` does not match `torch.rsqrt`
+        hidden_states = hidden_states / jnp.sqrt(variance + self.epsilon)
+
+        return self.weight * jnp.asarray(hidden_states, dtype=self.dtype)
+
+
+class FlaxLlamaRotaryEmbedding(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        head_dim = self.config.hidden_size // self.config.num_attention_heads
+        self.sincos = create_sinusoidal_positions(self.config.max_position_embeddings, head_dim)
+
+    def __call__(self, key, query, position_ids):
+        sincos = self.sincos[position_ids]
+        sin_pos, cos_pos = jnp.split(sincos, 2, axis=-1)
+
+        key = apply_rotary_pos_emb(key, sin_pos, cos_pos)
+        query = apply_rotary_pos_emb(query, sin_pos, cos_pos)
+
+        key = jnp.asarray(key, dtype=self.dtype)
+        query = jnp.asarray(query, dtype=self.dtype)
+
+        return key, query
+
+
+class FlaxLlamaAttention(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+    causal: bool = True
+    is_cross_attention: bool = False
+
+    def setup(self):
+        config = self.config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        self.attention_softmax_in_fp32 = self.dtype is not jnp.float32
+
+        dense = partial(
+            nn.Dense,
+            self.embed_dim,
+            use_bias=config.attention_bias,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
+        )
+
+        self.q_proj, self.k_proj, self.v_proj = dense(), dense(), dense()
+        self.o_proj = dense()
+
+        self.causal_mask = make_causal_mask(jnp.ones((1, config.max_position_embeddings), dtype="bool"), dtype="bool")
+        self.rotary_emb = FlaxLlamaRotaryEmbedding(config, dtype=self.dtype)
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.num_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.embed_dim,))
+
+    @nn.compact
+    # Copied from transformers.models.gpt_neo.modeling_flax_gpt_neo.FlaxGPTNeoSelfAttention._concatenate_to_cache
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slighly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask,
+        position_ids,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        query = self.q_proj(hidden_states)
+        key = self.k_proj(hidden_states)
+        value = self.v_proj(hidden_states)
+
+        query = self._split_heads(query)
+        key = self._split_heads(key)
+        value = self._split_heads(value)
+
+        key, query = self.rotary_emb(key, query, position_ids)
+
+        query_length, key_length = query.shape[1], key.shape[1]
+
+        if self.has_variable("cache", "cached_key"):
+            mask_shift = self.variables["cache"]["cache_index"]
+            max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+            causal_mask = lax.dynamic_slice(
+                self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
+            )
+        else:
+            causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+
+        batch_size = hidden_states.shape[0]
+        causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+        attention_mask = combine_masks(attention_mask, causal_mask)
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.has_variable("cache", "cached_key") or init_cache:
+            key, value, attention_mask = self._concatenate_to_cache(key, value, query, attention_mask)
+
+        # transform boolean mask into float mask
+        attention_bias = lax.select(
+            attention_mask > 0,
+            jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+            jnp.full(attention_mask.shape, jnp.finfo(self.dtype).min).astype(self.dtype),
+        )
+
+        # usual dot product attention
+        attention_dtype = jnp.float32 if self.attention_softmax_in_fp32 else self.dtype
+        attn_weights = dot_product_attention_weights(
+            query,
+            key,
+            bias=attention_bias,
+            deterministic=deterministic,
+            dtype=attention_dtype,
+        )
+
+        if self.attention_softmax_in_fp32:
+            attn_weights = attn_weights.astype(self.dtype)
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value)
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.o_proj(attn_output)
+
+        outputs = (attn_output, attn_weights) if output_attentions else (attn_output,)
+        return outputs
+
+
+class FlaxLlamaMLP(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        embed_dim = self.config.hidden_size
+        inner_dim = self.config.intermediate_size if self.config.intermediate_size is not None else 4 * embed_dim
+
+        kernel_init = jax.nn.initializers.normal(self.config.initializer_range)
+        self.act = ACT2FN[self.config.hidden_act]
+
+        self.gate_proj = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, kernel_init=kernel_init)
+        self.down_proj = nn.Dense(embed_dim, use_bias=False, dtype=self.dtype, kernel_init=kernel_init)
+        self.up_proj = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, kernel_init=kernel_init)
+
+    def __call__(self, hidden_states):
+        up_proj_states = self.up_proj(hidden_states)
+        gate_states = self.act(self.gate_proj(hidden_states))
+
+        hidden_states = self.down_proj(up_proj_states * gate_states)
+        return hidden_states
+
+
+class FlaxLlamaDecoderLayer(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.input_layernorm = FlaxLlamaRMSNorm(self.config, dtype=self.dtype)
+        self.self_attn = FlaxLlamaAttention(self.config, dtype=self.dtype)
+        self.post_attention_layernorm = FlaxLlamaRMSNorm(self.config, dtype=self.dtype)
+        self.mlp = FlaxLlamaMLP(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        position_ids=None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        outputs = self.self_attn(
+            hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+        )
+        # residual connection
+        attn_output = outputs[0]
+        hidden_states = residual + attn_output
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        # residual connection
+        hidden_states = residual + hidden_states
+
+        return (hidden_states,) + outputs[1:]
+
+
+# Copied from transformers.models.gpt_neo.modeling_flax_gpt_neo.FlaxGPTNeoPreTrainedModel with GPTNeo->Llama, GPT_NEO->LLAMA, transformer->model
+class FlaxLlamaPreTrainedModel(FlaxPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = LlamaConfig
+    base_model_prefix = "model"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+        input_shape: Tuple = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        random_params = self.module.init(rngs, input_ids, attention_mask, position_ids, return_dict=False)["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unfreeze(random_params))
+            params = flatten_dict(unfreeze(params))
+            for missing_key in self._missing_keys:
+                params[missing_key] = random_params[missing_key]
+            self._missing_keys = set()
+            return freeze(unflatten_dict(params))
+        else:
+            return random_params
+
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length))
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return unfreeze(init_variables["cache"])
+
+    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        params: dict = None,
+        past_key_values: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        batch_size, sequence_length = input_ids.shape
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `position_ids` when passing `past_key_values`.")
+
+            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be changed by FlaxLlamaAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        outputs = self.module.apply(
+            inputs,
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            False,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+            mutable=mutable,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past_key_values = outputs
+            outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past_key_values = outputs
+            outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        return outputs
+
+
+class FlaxLlamaLayerCollection(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.blocks = [
+            FlaxLlamaDecoderLayer(self.config, dtype=self.dtype, name=str(i))
+            for i in range(self.config.num_hidden_layers)
+        ]
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        position_ids=None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = False,
+    ):
+        all_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        for block in self.blocks:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            layer_outputs = block(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                deterministic=deterministic,
+                init_cache=init_cache,
+                output_attentions=output_attentions,
+            )
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions += (layer_outputs[1],)
+
+        # this contains possible `None` values - `FlaxLlamaModule` will filter them out
+        outputs = (hidden_states, all_hidden_states, all_attentions)
+
+        return outputs
+
+
+class FlaxLlamaModule(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.hidden_size = self.config.hidden_size
+        embedding_init = jax.nn.initializers.normal(stddev=self.config.initializer_range)
+        self.embed_tokens = nn.Embed(
+            self.config.vocab_size,
+            self.hidden_size,
+            embedding_init=embedding_init,
+            dtype=self.dtype,
+        )
+        self.layers = FlaxLlamaLayerCollection(self.config, dtype=self.dtype)
+        self.norm = FlaxLlamaRMSNorm(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        deterministic=True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        input_embeds = self.embed_tokens(input_ids.astype("i4"))
+
+        outputs = self.layers(
+            input_embeds,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        hidden_states = self.norm(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = outputs[1] + (hidden_states,)
+            outputs = (hidden_states, all_hidden_states) + outputs[2:]
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=outputs[1],
+            attentions=outputs[-1],
+        )
+
+
+@add_start_docstrings(
+    "The bare Llama Model transformer outputting raw hidden-states without any specific head on top.",
+    LLAMA_START_DOCSTRING,
+)
+class FlaxLlamaModel(FlaxLlamaPreTrainedModel):
+    module_class = FlaxLlamaModule
+
+
+append_call_sample_docstring(
+    FlaxLlamaModel,
+    _CHECKPOINT_FOR_DOC,
+    FlaxBaseModelOutput,
+    _CONFIG_FOR_DOC,
+    real_checkpoint=_REAL_CHECKPOINT_FOR_DOC,
+)
+
+
+class FlaxLlamaForCausalLMModule(nn.Module):
+    config: LlamaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.model = FlaxLlamaModule(self.config, dtype=self.dtype)
+        self.lm_head = nn.Dense(
+            self.config.vocab_size,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+        )
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        outputs = self.model(
+            input_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        lm_logits = self.lm_head(hidden_states)
+
+        if not return_dict:
+            return (lm_logits,) + outputs[1:]
+
+        return FlaxCausalLMOutput(logits=lm_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions)
+
+
+@add_start_docstrings(
+    """
+    The Llama Model transformer with a language modeling head (linear layer) on top.
+    """,
+    LLAMA_START_DOCSTRING,
+)
+# Copied from transformers.models.gptj.modeling_flax_gptj.FlaxGPTJForCausalLM with GPTJ->Llama
+class FlaxLlamaForCausalLM(FlaxLlamaPreTrainedModel):
+    module_class = FlaxLlamaForCausalLMModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jax.Array] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since Llama uses a causal mask, those positions are masked anyways.
+        # Thus we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    FlaxLlamaForCausalLM,
+    _CHECKPOINT_FOR_DOC,
+    FlaxCausalLMOutput,
+    _CONFIG_FOR_DOC,
+    real_checkpoint=_REAL_CHECKPOINT_FOR_DOC,
+)

src/transformers/models/mpt/modeling_mpt.py

@@ -265,7 +265,7 @@ class MptPreTrainedModel(PreTrainedModel):
 
     @staticmethod
     def _convert_to_mpt_cache(
-        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]]
+        past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]],
     ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
         """
         Converts the cache to the format expected by Mpt, i.e. to tuple(tuple([batch_size * num_heads, ...]))

src/transformers/utils/dummy_flax_objects.py

@@ -800,6 +800,27 @@ class FlaxGPTJPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxLlamaForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxLlamaModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxLlamaPreTrainedModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxLongT5ForConditionalGeneration(metaclass=DummyObject):
     _backends = ["flax"]
 

tests/models/llama/test_modeling_flax_llama.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+import numpy as np
+
+from transformers import LlamaConfig, is_flax_available, is_tokenizers_available
+from transformers.testing_utils import require_flax, slow
+
+from ...generation.test_flax_utils import FlaxGenerationTesterMixin
+from ...test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
+
+
+if is_flax_available():
+    import jax.numpy as jnp
+
+    from transformers.models.llama.modeling_flax_llama import FlaxLlamaForCausalLM, FlaxLlamaModel
+
+
+if is_tokenizers_available():
+    from transformers import LlamaTokenizerFast
+
+
+class FlaxLlamaModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=2,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=False,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=16,
+        num_hidden_layers=2,
+        num_attention_heads=2,
+        intermediate_size=64,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        window_size=7,
+        initializer_range=0.02,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.window_size = window_size
+        self.initializer_range = initializer_range
+        self.scope = None
+        self.bos_token_id = vocab_size - 1
+        self.eos_token_id = vocab_size - 1
+        self.pad_token_id = vocab_size - 1
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = np.tril(np.ones((self.batch_size, self.seq_length)))
+
+        config = LlamaConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            use_cache=True,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+        return (config, input_ids, input_mask)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": attention_mask}
+        return config, inputs_dict
+
+    def check_use_cache_forward(self, model_class_name, config, input_ids, attention_mask):
+        max_decoder_length = 20
+        model = model_class_name(config)
+
+        past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length)
+        attention_mask = jnp.ones((input_ids.shape[0], max_decoder_length), dtype="i4")
+
+        position_ids = jnp.broadcast_to(
+            jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1)
+        )
+        outputs_cache = model(
+            input_ids[:, :-1],
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4")
+        outputs_cache_next = model(
+            input_ids[:, -1:],
+            attention_mask=attention_mask,
+            past_key_values=outputs_cache.past_key_values,
+            position_ids=position_ids,
+        )
+
+        outputs = model(input_ids)
+
+        diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
+        self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
+
+    def check_use_cache_forward_with_attn_mask(self, model_class_name, config, input_ids, attention_mask):
+        max_decoder_length = 20
+        model = model_class_name(config)
+
+        attention_mask_cache = jnp.concatenate(
+            [attention_mask, jnp.zeros((attention_mask.shape[0], max_decoder_length - attention_mask.shape[1]))],
+            axis=-1,
+        )
+
+        past_key_values = model.init_cache(input_ids.shape[0], max_decoder_length)
+        position_ids = jnp.broadcast_to(
+            jnp.arange(input_ids.shape[-1] - 1)[None, :], (input_ids.shape[0], input_ids.shape[-1] - 1)
+        )
+
+        outputs_cache = model(
+            input_ids[:, :-1],
+            attention_mask=attention_mask_cache,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+        position_ids = jnp.array(input_ids.shape[0] * [[input_ids.shape[-1] - 1]], dtype="i4")
+        outputs_cache_next = model(
+            input_ids[:, -1:],
+            past_key_values=outputs_cache.past_key_values,
+            attention_mask=attention_mask_cache,
+            position_ids=position_ids,
+        )
+
+        outputs = model(input_ids, attention_mask=attention_mask)
+
+        diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
+        self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
+
+
+@require_flax
+class FlaxLlamaModelTest(FlaxModelTesterMixin, FlaxGenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (FlaxLlamaModel, FlaxLlamaForCausalLM) if is_flax_available() else ()
+    all_generative_model_classes = (FlaxLlamaForCausalLM,) if is_flax_available() else ()
+
+    def setUp(self):
+        self.model_tester = FlaxLlamaModelTester(self)
+
+    def test_use_cache_forward(self):
+        for model_class_name in self.all_model_classes:
+            config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs()
+            self.model_tester.check_use_cache_forward(model_class_name, config, input_ids, attention_mask)
+
+    def test_use_cache_forward_with_attn_mask(self):
+        for model_class_name in self.all_model_classes:
+            config, input_ids, attention_mask = self.model_tester.prepare_config_and_inputs()
+            self.model_tester.check_use_cache_forward_with_attn_mask(
+                model_class_name, config, input_ids, attention_mask
+            )
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_class_name in self.all_model_classes:
+            model = model_class_name.from_pretrained("openlm-research/open_llama_3b_v2", from_pt=True)
+            outputs = model(np.ones((1, 1)))
+            self.assertIsNotNone(outputs)
+
+
+@slow
+@require_flax
+class FlaxLlamaIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.model_id = "openlm-research/open_llama_3b_v2"
+        self.model = FlaxLlamaForCausalLM.from_pretrained(self.model_id, from_pt=True)
+        self.test_batch = jnp.arange(32).reshape(4, 8) + 1911
+
+    def test_model_logits(self):
+        flax_logits = self.model(self.test_batch).logits
+
+        # fmt: off
+        EXPECTED_LOGITS = [-74.4243, -74.0680, -65.2507, -79.1658, -77.7460, -69.2379, -86.4588, -84.8933, -77.8456]
+        EXPECTED_MIN, EXPECTED_MAX, EXPECTED_MEAN = -96.9952
+        EXPECTED_MAX = -18.4571
+        EXPECTED_MEAN = -65.0608
+        # fmt: on
+
+        self.assertTrue(np.allclose(flax_logits[0, :3, :3].flatten(), EXPECTED_LOGITS, atol=1e-4))
+        self.assertAlmostEqual(flax_logits.min(), EXPECTED_MIN, places=3)
+        self.assertAlmostEqual(flax_logits.max(), EXPECTED_MAX, places=3)
+        self.assertAlmostEqual(flax_logits.mean(), EXPECTED_MEAN, places=3)
+
+    def test_model_hidden_states(self):
+        flax_hidden_states = self.model(self.test_batch, output_hidden_states=True).hidden_states
+        flax_hidden_means = [h.mean() for h in flax_hidden_states]
+
+        # fmt: off
+        EXPECTED_HIDDEN_MEANS = [
+            -0.00007,-0.00049,-0.00169,-0.00253,-0.00271,
+            -0.00290,-0.00252,0.00230,0.00230,0.00198,
+            0.00196,0.00174,0.00246,0.00205,0.00242,
+            0.00171,0.00092,0.00054,0.00102,0.00024,
+            0.00029,0.00037,-0.00101,-0.00062,-0.00341,-0.00636,-0.00357
+        ]
+        # fmt: on
+
+        self.assertTrue(np.allclose(flax_hidden_means, EXPECTED_HIDDEN_MEANS, atol=1e-4))
+
+    def test_generated_text(self):
+        tokenizer = LlamaTokenizerFast.from_pretrained(self.model_id)
+        tokenizer.pad_token_id = 2
+        test_batch = ["Aloha, World! ", "2 + 2 = ", "Paris is the capital of ", "我很高興認識"]
+
+        inputs = tokenizer(test_batch, return_tensors="np", truncation=True, padding=True)
+        generated_ids = self.model.generate(**inputs, max_length=15).sequences
+        generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+
+        # fmt: off
+        EXPECTED_GENERATION = [
+            "Aloha, World! 201",
+            "2 + 2 = 4\n2",
+            "Paris is the capital of Île-",
+            "我很高興認識你，我"
+        ]
+        # fmt: on
+
+        self.assertListEqual(generated_text, EXPECTED_GENERATION)

tests/models/llama/test_modeling_llama.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 """ Testing suite for the PyTorch LLaMA model. """
 
-
 import unittest
 
 import pytest
@@ -33,7 +32,7 @@ from transformers.testing_utils import (
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
-from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
 
 
@@ -105,7 +104,7 @@ class LlamaModelTester:
 
         input_mask = None
         if self.use_input_mask:
-            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+            input_mask = torch.tril(torch.ones(self.batch_size, self.seq_length))
 
         token_type_ids = None
         if self.use_token_type_ids:

tests/models/mistral/test_modeling_mistral.py

@@ -34,7 +34,7 @@ from transformers.testing_utils import (
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
-from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
 
 
@@ -107,7 +107,7 @@ class MistralModelTester:
 
         input_mask = None
         if self.use_input_mask:
-            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+            input_mask = torch.tril(torch.ones(self.batch_size, self.seq_length))
 
         token_type_ids = None
         if self.use_token_type_ids:

tests/models/persimmon/test_modeling_persimmon.py

@@ -32,7 +32,7 @@ from transformers.testing_utils import (
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
-from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
 
 
@@ -104,7 +104,7 @@ class PersimmonModelTester:
 
         input_mask = None
         if self.use_input_mask:
-            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+            input_mask = torch.tril(torch.ones(self.batch_size, self.seq_length))
 
         token_type_ids = None
         if self.use_token_type_ids:

tests/models/phi/test_modeling_phi.py

@@ -38,7 +38,6 @@ if is_torch_available():
     )
 
 
-# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester with Llama->Phi
 class PhiModelTester:
     def __init__(
         self,

utils/check_docstrings.py

@@ -233,6 +233,8 @@ OBJECTS_TO_IGNORE = [
     "FlaxGPTJModel",
     "FlaxGPTNeoForCausalLM",
     "FlaxGPTNeoModel",
+    "FlaxLlamaForCausalLM",
+    "FlaxLlamaModel",
     "FlaxMBartForConditionalGeneration",
     "FlaxMBartForQuestionAnswering",
     "FlaxMBartForSequenceClassification",