Add new CrossEncoder model

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,402 @@
+---
+tags:
+- sentence-transformers
+- cross-encoder
+- reranker
+- generated_from_trainer
+- dataset_size:3190
+- loss:ListNetLoss
+base_model: jinaai/jina-reranker-v2-base-multilingual
+pipeline_tag: text-ranking
+library_name: sentence-transformers
+---
+
+# CrossEncoder based on jinaai/jina-reranker-v2-base-multilingual
+
+This is a [Cross Encoder](https://www.sbert.net/docs/cross_encoder/usage/usage.html) model finetuned from [jinaai/jina-reranker-v2-base-multilingual](https://huggingface.co/jinaai/jina-reranker-v2-base-multilingual) using the [sentence-transformers](https://www.SBERT.net) library. It computes scores for pairs of texts, which can be used for text reranking and semantic search.
+
+## Model Details
+
+### Model Description
+- **Model Type:** Cross Encoder
+- **Base model:** [jinaai/jina-reranker-v2-base-multilingual](https://huggingface.co/jinaai/jina-reranker-v2-base-multilingual) <!-- at revision 2f894e63642a95228da19cdd583cd2309983c867 -->
+- **Maximum Sequence Length:** 1024 tokens
+- **Number of Output Labels:** 1 label
+<!-- - **Training Dataset:** Unknown -->
+<!-- - **Language:** Unknown -->
+<!-- - **License:** Unknown -->
+
+### Model Sources
+
+- **Documentation:** [Sentence Transformers Documentation](https://sbert.net)
+- **Documentation:** [Cross Encoder Documentation](https://www.sbert.net/docs/cross_encoder/usage/usage.html)
+- **Repository:** [Sentence Transformers on GitHub](https://github.com/UKPLab/sentence-transformers)
+- **Hugging Face:** [Cross Encoders on Hugging Face](https://huggingface.co/models?library=sentence-transformers&other=cross-encoder)
+
+## Usage
+
+### Direct Usage (Sentence Transformers)
+
+First install the Sentence Transformers library:
+
+```bash
+pip install -U sentence-transformers
+```
+
+Then you can load this model and run inference.
+```python
+from sentence_transformers import CrossEncoder
+
+# Download from the 🤗 Hub
+model = CrossEncoder("Pranjal2002/jina_finance_v2")
+# Get scores for pairs of texts
+pairs = [
+    ['What consolidation trends among competitors are highlighted in disclosures affecting Regions Financial Corporation’s regional banking operations?', '10-K'],
+    ['What consolidation trends among competitors are highlighted in disclosures affecting Regions Financial Corporation’s regional banking operations?', 'Earnings'],
+    ['What consolidation trends among competitors are highlighted in disclosures affecting Regions Financial Corporation’s regional banking operations?', 'DEF14A'],
+    ['What consolidation trends among competitors are highlighted in disclosures affecting Regions Financial Corporation’s regional banking operations?', '8-K'],
+    ['What consolidation trends among competitors are highlighted in disclosures affecting Regions Financial Corporation’s regional banking operations?', '10-Q'],
+]
+scores = model.predict(pairs)
+print(scores.shape)
+# (5,)
+
+# Or rank different texts based on similarity to a single text
+ranks = model.rank(
+    'What consolidation trends among competitors are highlighted in disclosures affecting Regions Financial Corporation’s regional banking operations?',
+    [
+        '10-K',
+        'Earnings',
+        'DEF14A',
+        '8-K',
+        '10-Q',
+    ]
+)
+# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]
+```
+
+<!--
+### Direct Usage (Transformers)
+
+<details><summary>Click to see the direct usage in Transformers</summary>
+
+</details>
+-->
+
+<!--
+### Downstream Usage (Sentence Transformers)
+
+You can finetune this model on your own dataset.
+
+<details><summary>Click to expand</summary>
+
+</details>
+-->
+
+<!--
+### Out-of-Scope Use
+
+*List how the model may foreseeably be misused and address what users ought not to do with the model.*
+-->
+
+<!--
+## Bias, Risks and Limitations
+
+*What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model.*
+-->
+
+<!--
+### Recommendations
+
+*What are recommendations with respect to the foreseeable issues? For example, filtering explicit content.*
+-->
+
+## Training Details
+
+### Training Dataset
+
+#### Unnamed Dataset
+
+* Size: 3,190 training samples
+* Columns: <code>query</code>, <code>docs</code>, and <code>labels</code>
+* Approximate statistics based on the first 1000 samples:
+  |         | query                                                                                            | docs                               | labels                             |
+  |:--------|:-------------------------------------------------------------------------------------------------|:-----------------------------------|:-----------------------------------|
+  | type    | string                                                                                           | list                               | list                               |
+  | details | <ul><li>min: 55 characters</li><li>mean: 103.12 characters</li><li>max: 180 characters</li></ul> | <ul><li>size: 5 elements</li></ul> | <ul><li>size: 5 elements</li></ul> |
+* Samples:
+  | query                                                                                                                                                                          | docs                                                       | labels                       |
+  |:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:-----------------------------------------------------------|:-----------------------------|
+  | <code>What year over year growth rate was shown for paid memberships in the same table</code>                                                                                  | <code>['10-Q', '10-K', '8-K', 'Earnings', 'DEF14A']</code> | <code>[4, 3, 2, 1, 0]</code> |
+  | <code>How did non‑GAAP EPS growth align with the incentive metrics set for management?</code>                                                                                  | <code>['DEF14A', '8-K', '10-K', '10-Q', 'Earnings']</code> | <code>[2, 1, 0, 0, 0]</code> |
+  | <code>What questions were raised regarding Xcel Energy Inc.’s risk factors and mitigation plans related to the integration of renewable energy sources into their grid?</code> | <code>['10-K', 'Earnings', '8-K', '10-Q', 'DEF14A']</code> | <code>[4, 3, 2, 1, 0]</code> |
+* Loss: [<code>ListNetLoss</code>](https://sbert.net/docs/package_reference/cross_encoder/losses.html#listnetloss) with these parameters:
+  ```json
+  {
+      "activation_fn": "torch.nn.modules.linear.Identity",
+      "mini_batch_size": null
+  }
+  ```
+
+### Evaluation Dataset
+
+#### Unnamed Dataset
+
+* Size: 798 evaluation samples
+* Columns: <code>query</code>, <code>docs</code>, and <code>labels</code>
+* Approximate statistics based on the first 798 samples:
+  |         | query                                                                                            | docs                               | labels                             |
+  |:--------|:-------------------------------------------------------------------------------------------------|:-----------------------------------|:-----------------------------------|
+  | type    | string                                                                                           | list                               | list                               |
+  | details | <ul><li>min: 53 characters</li><li>mean: 102.91 characters</li><li>max: 179 characters</li></ul> | <ul><li>size: 5 elements</li></ul> | <ul><li>size: 5 elements</li></ul> |
+* Samples:
+  | query                                                                                                                                                          | docs                                                       | labels                       |
+  |:---------------------------------------------------------------------------------------------------------------------------------------------------------------|:-----------------------------------------------------------|:-----------------------------|
+  | <code>What consolidation trends among competitors are highlighted in disclosures affecting Regions Financial Corporation’s regional banking operations?</code> | <code>['10-K', 'Earnings', 'DEF14A', '8-K', '10-Q']</code> | <code>[4, 3, 2, 1, 0]</code> |
+  | <code>How does Pentair manage equity award burn rate or share pool availability?</code>                                                                        | <code>['10-K', 'DEF14A', '10-Q', 'Earnings', '8-K']</code> | <code>[4, 3, 2, 1, 0]</code> |
+  | <code>What key takeaways emerged from Valero Energy Corporation’s most recent earnings announcement?</code>                                                    | <code>['10-Q', '10-K', 'Earnings', '8-K', 'DEF14A']</code> | <code>[4, 3, 2, 1, 0]</code> |
+* Loss: [<code>ListNetLoss</code>](https://sbert.net/docs/package_reference/cross_encoder/losses.html#listnetloss) with these parameters:
+  ```json
+  {
+      "activation_fn": "torch.nn.modules.linear.Identity",
+      "mini_batch_size": null
+  }
+  ```
+
+### Training Hyperparameters
+#### Non-Default Hyperparameters
+
+- `eval_strategy`: steps
+- `per_device_train_batch_size`: 4
+- `per_device_eval_batch_size`: 4
+- `gradient_accumulation_steps`: 2
+- `learning_rate`: 2e-05
+- `num_train_epochs`: 5
+- `warmup_steps`: 100
+- `bf16`: True
+- `load_best_model_at_end`: True
+- `optim`: adamw_torch
+
+#### All Hyperparameters
+<details><summary>Click to expand</summary>
+
+- `overwrite_output_dir`: False
+- `do_predict`: False
+- `eval_strategy`: steps
+- `prediction_loss_only`: True
+- `per_device_train_batch_size`: 4
+- `per_device_eval_batch_size`: 4
+- `per_gpu_train_batch_size`: None
+- `per_gpu_eval_batch_size`: None
+- `gradient_accumulation_steps`: 2
+- `eval_accumulation_steps`: None
+- `torch_empty_cache_steps`: None
+- `learning_rate`: 2e-05
+- `weight_decay`: 0.0
+- `adam_beta1`: 0.9
+- `adam_beta2`: 0.999
+- `adam_epsilon`: 1e-08
+- `max_grad_norm`: 1.0
+- `num_train_epochs`: 5
+- `max_steps`: -1
+- `lr_scheduler_type`: linear
+- `lr_scheduler_kwargs`: {}
+- `warmup_ratio`: 0.0
+- `warmup_steps`: 100
+- `log_level`: passive
+- `log_level_replica`: warning
+- `log_on_each_node`: True
+- `logging_nan_inf_filter`: True
+- `save_safetensors`: True
+- `save_on_each_node`: False
+- `save_only_model`: False
+- `restore_callback_states_from_checkpoint`: False
+- `no_cuda`: False
+- `use_cpu`: False
+- `use_mps_device`: False
+- `seed`: 42
+- `data_seed`: None
+- `jit_mode_eval`: False
+- `use_ipex`: False
+- `bf16`: True
+- `fp16`: False
+- `fp16_opt_level`: O1
+- `half_precision_backend`: auto
+- `bf16_full_eval`: False
+- `fp16_full_eval`: False
+- `tf32`: None
+- `local_rank`: 0
+- `ddp_backend`: None
+- `tpu_num_cores`: None
+- `tpu_metrics_debug`: False
+- `debug`: []
+- `dataloader_drop_last`: False
+- `dataloader_num_workers`: 0
+- `dataloader_prefetch_factor`: None
+- `past_index`: -1
+- `disable_tqdm`: False
+- `remove_unused_columns`: True
+- `label_names`: None
+- `load_best_model_at_end`: True
+- `ignore_data_skip`: False
+- `fsdp`: []
+- `fsdp_min_num_params`: 0
+- `fsdp_config`: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
+- `fsdp_transformer_layer_cls_to_wrap`: None
+- `accelerator_config`: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
+- `parallelism_config`: None
+- `deepspeed`: None
+- `label_smoothing_factor`: 0.0
+- `optim`: adamw_torch
+- `optim_args`: None
+- `adafactor`: False
+- `group_by_length`: False
+- `length_column_name`: length
+- `ddp_find_unused_parameters`: None
+- `ddp_bucket_cap_mb`: None
+- `ddp_broadcast_buffers`: False
+- `dataloader_pin_memory`: True
+- `dataloader_persistent_workers`: False
+- `skip_memory_metrics`: True
+- `use_legacy_prediction_loop`: False
+- `push_to_hub`: False
+- `resume_from_checkpoint`: None
+- `hub_model_id`: None
+- `hub_strategy`: every_save
+- `hub_private_repo`: None
+- `hub_always_push`: False
+- `hub_revision`: None
+- `gradient_checkpointing`: False
+- `gradient_checkpointing_kwargs`: None
+- `include_inputs_for_metrics`: False
+- `include_for_metrics`: []
+- `eval_do_concat_batches`: True
+- `fp16_backend`: auto
+- `push_to_hub_model_id`: None
+- `push_to_hub_organization`: None
+- `mp_parameters`: 
+- `auto_find_batch_size`: False
+- `full_determinism`: False
+- `torchdynamo`: None
+- `ray_scope`: last
+- `ddp_timeout`: 1800
+- `torch_compile`: False
+- `torch_compile_backend`: None
+- `torch_compile_mode`: None
+- `include_tokens_per_second`: False
+- `include_num_input_tokens_seen`: False
+- `neftune_noise_alpha`: None
+- `optim_target_modules`: None
+- `batch_eval_metrics`: False
+- `eval_on_start`: False
+- `use_liger_kernel`: False
+- `liger_kernel_config`: None
+- `eval_use_gather_object`: False
+- `average_tokens_across_devices`: False
+- `prompts`: None
+- `batch_sampler`: batch_sampler
+- `multi_dataset_batch_sampler`: proportional
+- `router_mapping`: {}
+- `learning_rate_mapping`: {}
+
+</details>
+
+### Training Logs
+| Epoch    | Step     | Training Loss | Validation Loss |
+|:--------:|:--------:|:-------------:|:---------------:|
+| 0.1253   | 50       | 1.7131        | -               |
+| 0.2506   | 100      | 1.5888        | -               |
+| 0.3759   | 150      | 1.49          | -               |
+| 0.5013   | 200      | 1.4408        | 1.4397          |
+| 0.6266   | 250      | 1.4225        | -               |
+| 0.7519   | 300      | 1.4216        | -               |
+| 0.8772   | 350      | 1.4329        | -               |
+| 1.0025   | 400      | 1.3996        | 1.4083          |
+| 1.1278   | 450      | 1.4126        | -               |
+| 1.2531   | 500      | 1.4002        | -               |
+| 1.3784   | 550      | 1.4098        | -               |
+| 1.5038   | 600      | 1.3692        | 1.4042          |
+| 1.6291   | 650      | 1.3784        | -               |
+| 1.7544   | 700      | 1.4014        | -               |
+| 1.8797   | 750      | 1.3815        | -               |
+| 2.0050   | 800      | 1.3982        | 1.3910          |
+| 2.1303   | 850      | 1.3864        | -               |
+| 2.2556   | 900      | 1.3983        | -               |
+| 2.3810   | 950      | 1.3662        | -               |
+| 2.5063   | 1000     | 1.3747        | 1.3968          |
+| 2.6316   | 1050     | 1.3739        | -               |
+| 2.7569   | 1100     | 1.3687        | -               |
+| 2.8822   | 1150     | 1.3858        | -               |
+| 3.0075   | 1200     | 1.3847        | 1.3897          |
+| 3.1328   | 1250     | 1.3684        | -               |
+| 3.2581   | 1300     | 1.3787        | -               |
+| 3.3835   | 1350     | 1.3612        | -               |
+| 3.5088   | 1400     | 1.3906        | 1.3920          |
+| 3.6341   | 1450     | 1.3838        | -               |
+| 3.7594   | 1500     | 1.3817        | -               |
+| 3.8847   | 1550     | 1.3615        | -               |
+| **4.01** | **1600** | **1.3978**    | **1.3892**      |
+| 4.1353   | 1650     | 1.3793        | -               |
+| 4.2607   | 1700     | 1.3753        | -               |
+| 4.3860   | 1750     | 1.3847        | -               |
+| 4.5113   | 1800     | 1.3857        | 1.3887          |
+| 4.6366   | 1850     | 1.3583        | -               |
+| 4.7619   | 1900     | 1.3644        | -               |
+| 4.8872   | 1950     | 1.3696        | -               |
+
+* The bold row denotes the saved checkpoint.
+
+### Framework Versions
+- Python: 3.12.11
+- Sentence Transformers: 5.1.0
+- Transformers: 4.56.1
+- PyTorch: 2.8.0+cu126
+- Accelerate: 1.10.1
+- Datasets: 4.0.0
+- Tokenizers: 0.22.0
+
+## Citation
+
+### BibTeX
+
+#### Sentence Transformers
+```bibtex
+@inproceedings{reimers-2019-sentence-bert,
+    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
+    author = "Reimers, Nils and Gurevych, Iryna",
+    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
+    month = "11",
+    year = "2019",
+    publisher = "Association for Computational Linguistics",
+    url = "https://arxiv.org/abs/1908.10084",
+}
+```
+
+#### ListNetLoss
+```bibtex
+@inproceedings{cao2007learning,
+    title={Learning to Rank: From Pairwise Approach to Listwise Approach},
+    author={Cao, Zhe and Qin, Tao and Liu, Tie-Yan and Tsai, Ming-Feng and Li, Hang},
+    booktitle={Proceedings of the 24th international conference on Machine learning},
+    pages={129--136},
+    year={2007}
+}
+```
+
+<!--
+## Glossary
+
+*Clearly define terms in order to be accessible across audiences.*
+-->
+
+<!--
+## Model Card Authors
+
+*Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction.*
+-->
+
+<!--
+## Model Card Contact
+
+*Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors.*
+-->

block.py

@@ -0,0 +1,470 @@
+# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/block.py
+# Commit id: abbc1311731867310635f9edc2a9ec18317c8c48
+
+# Copyright (c) 2024, Tri Dao.
+
+from functools import partial
+from typing import Optional
+
+import torch
+import torch.fx
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+
+from .mha import MHA
+from .mlp import Mlp
+
+try:
+    from flash_attn.ops.triton.layer_norm import layer_norm_fn, RMSNorm
+except ImportError:
+    layer_norm_fn, RMSNorm = None, None
+
+
+def stochastic_depth(
+    input: Tensor, p: float, mode: str, training: bool = True
+) -> Tensor:
+    """
+    Implements the Stochastic Depth from `"Deep Networks with Stochastic Depth"
+    <https://arxiv.org/abs/1603.09382>`_ used for randomly dropping residual
+    branches of residual architectures.
+    Args:
+        input (Tensor[N, ...]): The input tensor or arbitrary dimensions with the first one
+                    being its batch i.e. a batch with ``N`` rows.
+        p (float): probability of the input to be zeroed.
+        mode (str): ``"batch"`` or ``"row"``.
+                    ``"batch"`` randomly zeroes the entire input, ``"row"`` zeroes
+                    randomly selected rows from the batch.
+        training: apply stochastic depth if is ``True``. Default: ``True``
+    Returns:
+        Tensor[N, ...]: The randomly zeroed tensor.
+    """
+    if p < 0.0 or p > 1.0:
+        raise ValueError(f"drop probability has to be between 0 and 1, but got {p}")
+    if mode not in ["batch", "row"]:
+        raise ValueError(f"mode has to be either 'batch' or 'row', but got {mode}")
+    if not training or p == 0.0:
+        return input
+
+    survival_rate = 1.0 - p
+    if mode == "row":
+        size = [input.shape[0]] + [1] * (input.ndim - 1)
+    else:
+        size = [1] * input.ndim
+    noise = torch.empty(size, dtype=input.dtype, device=input.device)
+    noise = noise.bernoulli_(survival_rate)
+    if survival_rate > 0.0:
+        noise.div_(survival_rate)
+    return input * noise
+
+
+torch.fx.wrap("stochastic_depth")
+
+
+class StochasticDepth(nn.Module):
+    """
+    See :func:`stochastic_depth`.
+    """
+
+    def __init__(self, p: float, mode: str) -> None:
+        super().__init__()
+        self.p = p
+        self.mode = mode
+
+    def forward(self, input: Tensor) -> Tensor:
+        return stochastic_depth(input, self.p, self.mode, self.training)
+
+    def __repr__(self) -> str:
+        s = f"{self.__class__.__name__}(p={self.p}, mode={self.mode})"
+        return s
+
+        
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim,
+        mixer_cls=None,
+        mlp_cls=None,
+        norm_cls=nn.LayerNorm,
+        dropout_cls=nn.Dropout,
+        prenorm=True,
+        resid_dropout1=0.0,
+        resid_dropout2=0.0,
+        drop_path1=0.0,
+        drop_path2=0.0,
+        fused_dropout_add_ln=False,
+        return_residual=False,
+        residual_in_fp32=False,
+        sequence_parallel=False,
+        mark_shared_params=False,
+    ):
+        """
+        For prenorm=True, this Block has a slightly different structure compared to a regular
+        prenorm Transformer block.
+        The standard block is: LN -> MHA -> Dropout -> Add -> LN -> MLP -> Dropout -> Add.
+        [Ref: https://arxiv.org/abs/2002.04745]
+        Here we have: Dropout -> Add -> LN -> MHA -> Dropout -> Add -> LN -> MLP, returning both
+        the hidden_states (output of the MLP) and the residual.
+        This is for performance reasons, as we can fuse the dropout, add and LayerNorm.
+        The residual needs to be provided (except for the very first block).
+
+        For prenorm=False, this Block has the same structure as a regular postnorm Transformer
+        block: MHA -> Dropout -> Add -> LN -> MLP -> Dropout -> Add -> LN.
+
+        return_residual: whether each of the sub-layers (mixer and mlp) will return the residual.
+        This is for performance reason: for post-norm architecture, returning the input allows us
+        to fuse the backward of nn.Linear with the residual connection.
+        """
+        super().__init__()
+        self.prenorm = prenorm
+        self.fused_dropout_add_ln = fused_dropout_add_ln
+        self.return_residual = return_residual
+        self.residual_in_fp32 = residual_in_fp32
+        if self.residual_in_fp32:
+            assert self.prenorm, "residual_in_fp32 is only compatible with prenorm=True"
+        if mixer_cls is None:
+            mixer_cls = partial(MHA, num_heads=dim // 64)
+        if mlp_cls is None:
+            mlp_cls = partial(Mlp, hidden_features=4 * dim)
+        self.mixer = mixer_cls(dim)
+        self.dropout1 = dropout_cls(resid_dropout1)
+        self.drop_path1 = StochasticDepth(drop_path1, mode="row")
+        self.norm1 = norm_cls(dim)
+        self.mlp = mlp_cls(dim)
+        if not isinstance(self.mlp, nn.Identity):
+            self.dropout2 = dropout_cls(resid_dropout2)
+            self.drop_path2 = StochasticDepth(drop_path2, mode="row")
+            self.norm2 = norm_cls(dim)
+
+        if self.fused_dropout_add_ln:
+            assert layer_norm_fn is not None, "Triton is not installed"
+            assert isinstance(self.norm1, (nn.LayerNorm, RMSNorm)) and isinstance(
+                self.dropout1, nn.Dropout
+            )
+
+        # TD [2023-01-07]: TODO: During training, if sequence_parallel is False and dropout != 0.0,
+        # then the input to each worker in the tensor parallel group will be different.
+        # This would produce wrong outputs? Somehow we'd need to sync the RNG state across workers.
+        # For now this is not an issue because we always use sequence_parallel=True during training
+        # and only use sequence_parallel=False during inference.
+
+        # Mark the norm parameters as "sequence_parallel" so that we run all-reduce on their grads.
+        if sequence_parallel:
+            for p in self.norm1.parameters():
+                p._sequence_parallel = True
+            if hasattr(self, "norm2"):
+                for p in self.norm2.parameters():
+                    p._sequence_parallel = True
+        # Mark the norm parameters as "shared_params" so that we sync their values at init.
+        if mark_shared_params:
+            for p in self.norm1.parameters():
+                p._shared_params = True
+            if hasattr(self, "norm2"):
+                for p in self.norm2.parameters():
+                    p._shared_params = True
+
+    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
+        return self.mixer.allocate_inference_cache(
+            batch_size, max_seqlen, dtype=dtype, **kwargs
+        )
+
+    def forward(
+        self,
+        hidden_states: Tensor,
+        residual: Optional[Tensor] = None,
+        mixer_subset=None,
+        mixer_kwargs=None,
+    ):
+        r"""Pass the input through the encoder layer.
+
+        Args:
+            hidden_states: the sequence to the encoder layer (required).
+            residual: if postnorm, residual=None, If prenorm, hidden_states = Attn/MLP(LN(residual))
+            mixer_subset: for cross-attention only. If not None, will take a subset of x
+                before applying the query projection. Useful for e.g., ViT where we only care
+                about the CLS token in the last layer.
+        """
+        if self.prenorm:
+            if not self.fused_dropout_add_ln:
+                dropped = self.drop_path1(self.dropout1(hidden_states))
+                residual = (dropped + residual) if residual is not None else dropped
+                hidden_states = self.norm1(residual.to(dtype=self.norm1.weight.dtype))
+                if self.residual_in_fp32:
+                    residual = residual.to(torch.float32)
+            else:
+                if self.drop_path1.p == 0 or not self.training:
+                    rowscale1 = None
+                else:
+                    rowscale1 = self.drop_path1(
+                        torch.ones(
+                            hidden_states.shape[:-1],
+                            device=hidden_states.device,
+                            dtype=hidden_states.dtype,
+                        )
+                    )
+                hidden_states, residual = layer_norm_fn(
+                    hidden_states,
+                    self.norm1.weight,
+                    self.norm1.bias,
+                    residual=residual,
+                    eps=self.norm1.eps,
+                    dropout_p=self.dropout1.p if self.training else 0.0,
+                    rowscale=rowscale1,
+                    prenorm=True,
+                    residual_in_fp32=self.residual_in_fp32,
+                    is_rms_norm=isinstance(self.norm1, RMSNorm),
+                )
+            if mixer_kwargs is None:
+                mixer_kwargs = {}
+            if mixer_subset is not None:
+                mixer_kwargs["mixer_subset"] = mixer_subset
+            hidden_states = self.mixer(hidden_states, **mixer_kwargs)
+            if mixer_subset is not None:
+                residual = residual[:, mixer_subset]
+            if not isinstance(self.mlp, nn.Identity):
+                if not self.fused_dropout_add_ln:
+                    dropped = self.drop_path2(self.dropout2(hidden_states))
+                    residual = (dropped + residual) if residual is not None else dropped
+                    hidden_states = self.norm2(
+                        residual.to(dtype=self.norm2.weight.dtype)
+                    )
+                    if self.residual_in_fp32:
+                        residual = residual.to(torch.float32)
+                else:
+                    if self.drop_path2.p == 0 or not self.training:
+                        rowscale2 = None
+                    else:
+                        rowscale2 = self.drop_path2(
+                            torch.ones(
+                                hidden_states.shape[:-1],
+                                device=hidden_states.device,
+                                dtype=hidden_states.dtype,
+                            )
+                        )
+                    hidden_states, residual = layer_norm_fn(
+                        hidden_states,
+                        self.norm2.weight,
+                        self.norm2.bias,
+                        residual=residual,
+                        eps=self.norm2.eps,
+                        dropout_p=self.dropout2.p if self.training else 0.0,
+                        rowscale=rowscale2,
+                        prenorm=True,
+                        residual_in_fp32=self.residual_in_fp32,
+                        is_rms_norm=isinstance(self.norm2, RMSNorm),
+                    )
+                hidden_states = self.mlp(hidden_states)
+            return hidden_states, residual
+        else:
+            assert residual is None
+            mixer_out = self.mixer(
+                hidden_states, **(mixer_kwargs if mixer_kwargs is not None else {})
+            )
+            if self.return_residual:  # mixer out is actually a pair here
+                mixer_out, hidden_states = mixer_out
+            if not self.fused_dropout_add_ln:
+                hidden_states = self.norm1(
+                    (self.drop_path1(self.dropout1(mixer_out)) + hidden_states).to(
+                        dtype=self.norm1.weight.dtype
+                    )
+                )
+            else:
+                if self.drop_path1.p == 0 or not self.training:
+                    rowscale1 = None
+                else:
+                    rowscale1 = self.drop_path1(
+                        torch.ones(
+                            mixer_out.shape[:-1],
+                            device=mixer_out.device,
+                            dtype=mixer_out.dtype,
+                        )
+                    )
+                hidden_states = layer_norm_fn(
+                    mixer_out,
+                    self.norm1.weight,
+                    self.norm1.bias,
+                    residual=hidden_states,
+                    eps=self.norm1.eps,
+                    dropout_p=self.dropout1.p if self.training else 0.0,
+                    rowscale=rowscale1,
+                    prenorm=False,
+                    is_rms_norm=isinstance(self.norm1, RMSNorm),
+                )
+            if not isinstance(self.mlp, nn.Identity):
+                mlp_out = self.mlp(hidden_states)
+                if self.return_residual:  # mlp out is actually a pair here
+                    mlp_out, hidden_states = mlp_out
+                if not self.fused_dropout_add_ln:
+                    hidden_states = self.norm2(
+                        (self.drop_path2(self.dropout2(mlp_out)) + hidden_states).to(
+                            dtype=self.norm2.weight.dtype
+                        )
+                    )
+                else:
+                    if self.drop_path2.p == 0 or not self.training:
+                        rowscale2 = None
+                    else:
+                        rowscale2 = self.drop_path2(
+                            torch.ones(
+                                mlp_out.shape[:-1],
+                                device=mlp_out.device,
+                                dtype=mlp_out.dtype,
+                            )
+                        )
+                    hidden_states = layer_norm_fn(
+                        mlp_out,
+                        self.norm2.weight,
+                        self.norm2.bias,
+                        residual=hidden_states,
+                        eps=self.norm2.eps,
+                        dropout_p=self.dropout2.p if self.training else 0.0,
+                        rowscale=rowscale2,
+                        prenorm=False,
+                        is_rms_norm=isinstance(self.norm2, RMSNorm),
+                    )
+            return hidden_states
+
+
+class ParallelBlock(nn.Module):
+    """The attention (mixer) and MLP blocks are done in parallel, similar to GPT-J, GPT-NeoX,
+    and PaLM.
+    """
+
+    def __init__(
+        self,
+        dim,
+        mixer_cls=None,
+        mlp_cls=None,
+        norm_cls=nn.LayerNorm,
+        dropout_cls=nn.Dropout,
+        resid_dropout1=0.0,
+        resid_dropout2=0.0,
+        tied_norm=False,
+        fused_dropout_add_ln=False,
+        residual_in_fp32=False,
+        sequence_parallel=False,
+        mark_shared_params=False,
+    ):
+        """
+        This Block has a slightly different structure compared to a regular
+        prenorm Transformer block.
+        The standard block is: LN -> MHA / MLP -> Dropout -> Add.
+        [Ref: https://arxiv.org/abs/2002.04745]
+        Here we have: Dropout -> Add -> LN -> MHA / MLP, returning both
+        the hidden_states (output1 of the MHA / MLP) and the residual.
+        This is for performance reasons, as we can fuse the dropout, add and LayerNorm.
+        The residual needs to be provided (except for the very first block).
+        """
+        super().__init__()
+        self.tied_norm = tied_norm
+        self.fused_dropout_add_ln = fused_dropout_add_ln
+        self.residual_in_fp32 = residual_in_fp32
+        if mixer_cls is None:
+            mixer_cls = partial(MHA, num_heads=dim // 64)
+        if mlp_cls is None:
+            mlp_cls = partial(Mlp, hidden_features=4 * dim)
+        self.mixer = mixer_cls(dim)
+        self.dropout1 = dropout_cls(resid_dropout1)
+        self.norm1 = norm_cls(dim)
+        self.mlp = mlp_cls(dim)
+        self.dropout2 = dropout_cls(resid_dropout2)
+        if not self.tied_norm:
+            self.norm2 = norm_cls(dim)
+
+        if self.fused_dropout_add_ln:
+            assert layer_norm_fn is not None, "Triton is not installed"
+            assert isinstance(self.norm1, (nn.LayerNorm, RMSNorm)) and isinstance(
+                self.dropout1, nn.Dropout
+            )
+
+        # TD [2023-01-07]: TODO: During training, if sequence_parallel is False and dropout != 0.0,
+        # then the input to each worker in the tensor parallel group will be different.
+        # This would produce wrong outputs? Somehow we'd need to sync the RNG state across workers.
+        # For now this is not an issue because we always use sequence_parallel=True during training
+        # and only use sequence_parallel=False during inference.
+
+        # Mark the norm parameters as "sequence_parallel" so that we run all-reduce on their grads.
+        if sequence_parallel:
+            for p in self.norm1.parameters():
+                p._sequence_parallel = True
+            if hasattr(self, "norm2"):
+                for p in self.norm2.parameters():
+                    p._sequence_parallel = True
+        # Mark the norm parameters as "shared_params" so that we sync their values at init.
+        if mark_shared_params:
+            for p in self.norm1.parameters():
+                p._shared_params = True
+            if hasattr(self, "norm2"):
+                for p in self.norm2.parameters():
+                    p._shared_params = True
+
+    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
+        return self.mixer.allocate_inference_cache(
+            batch_size, max_seqlen, dtype=dtype, **kwargs
+        )
+
+    def forward(
+        self,
+        hidden_states1: Tensor,
+        hidden_states2: Optional[Tensor] = None,
+        residual: Optional[Tensor] = None,
+        mixer_kwargs=None,
+    ):
+        r"""Pass the input through the encoder layer.
+
+        Args:
+            hidden_states1: the output of the previous attention (mixer) or embedding layer.
+            hidden_states2: the output of the previous MLP layer (if None, will use hidden_states1).
+            residual.
+        """
+        # TODO: Ideally we should only do the allgather / allreduce once for
+        # the Linear to MLP & Attention
+        if not self.fused_dropout_add_ln:
+            dropped1 = self.dropout1(hidden_states1)
+            # For the very 1st block, we only want 1 dropout, not two different dropouts
+            if hidden_states2 is not None:
+                dropped2 = self.dropout2(hidden_states2)
+                residual = (
+                    (residual + dropped1 + dropped2)
+                    if residual is not None
+                    else dropped1 + dropped2
+                )
+            else:
+                residual = (residual + dropped1) if residual is not None else dropped1
+            hidden_states1 = self.norm1(residual.to(dtype=self.norm1.weight.dtype))
+            hidden_states2 = (
+                self.norm2(residual.to(dtype=self.norm2.weight.dtype))
+                if not self.tied_norm
+                else hidden_states1
+            )
+            if self.residual_in_fp32:
+                residual = residual.to(torch.float32)
+        else:
+            weight2, bias2 = (
+                (self.norm2.weight, self.norm2.bias)
+                if not self.tied_norm
+                else (None, None)
+            )
+            hidden_states1, *rest, residual = layer_norm_fn(
+                hidden_states1,
+                self.norm1.weight,
+                self.norm1.bias,
+                residual=residual,
+                x1=hidden_states2,
+                weight1=weight2,
+                bias1=bias2,
+                eps=self.norm1.eps,
+                dropout_p=self.dropout1.p if self.training else 0.0,
+                prenorm=True,
+                residual_in_fp32=self.residual_in_fp32,
+                is_rms_norm=isinstance(self.norm1, RMSNorm),
+            )
+            if self.tied_norm:
+                hidden_states2 = hidden_states1
+            else:
+                (hidden_states2,) = rest
+        if mixer_kwargs is None:
+            mixer_kwargs = {}
+        hidden_states1 = self.mixer(hidden_states1, **mixer_kwargs)
+        hidden_states2 = self.mlp(hidden_states2)
+        return hidden_states1, hidden_states2, residual

config.json

@@ -0,0 +1,51 @@
+{
+  "architectures": [
+    "XLMRobertaForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "auto_map": {
+    "AutoConfig": "configuration_xlm_roberta.XLMRobertaFlashConfig",
+    "AutoModel": "modeling_xlm_roberta.XLMRobertaModel",
+    "AutoModelForSequenceClassification": "modeling_xlm_roberta.XLMRobertaForSequenceClassification"
+  },
+  "bos_token_id": 0,
+  "classifier_dropout": null,
+  "dtype": "bfloat16",
+  "emb_pooler": null,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "id2label": {
+    "0": "LABEL_0"
+  },
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "label2id": {
+    "LABEL_0": 0
+  },
+  "layer_norm_eps": 1e-05,
+  "load_trained_adapters": false,
+  "lora_adaptations": null,
+  "lora_alpha": 1,
+  "lora_dropout_p": 0.0,
+  "lora_main_params_trainable": false,
+  "lora_rank": 4,
+  "matryoshka_dimensions": null,
+  "max_position_embeddings": 1026,
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "output_past": true,
+  "pad_token_id": 1,
+  "position_embedding_type": "absolute",
+  "sentence_transformers": {
+    "activation_fn": "torch.nn.modules.activation.Sigmoid",
+    "version": "5.1.0"
+  },
+  "transformers_version": "4.56.1",
+  "truncate_dim": null,
+  "type_vocab_size": 1,
+  "use_cache": false,
+  "use_flash_attn": true,
+  "vocab_size": 250002
+}

configuration_xlm_roberta.py

@@ -0,0 +1,69 @@
+from transformers import PretrainedConfig
+import torch
+
+class XLMRobertaFlashConfig(PretrainedConfig):
+    def __init__(
+            self,
+            vocab_size=30522,
+            hidden_size=768,
+            num_hidden_layers=12,
+            num_attention_heads=12,
+            intermediate_size=3072,
+            hidden_act="gelu",
+            hidden_dropout_prob=0.1,
+            attention_probs_dropout_prob=0.1,
+            max_position_embeddings=512,
+            type_vocab_size=2,
+            initializer_range=0.02,
+            layer_norm_eps=1e-12,
+            pad_token_id=1,
+            bos_token_id=0,
+            eos_token_id=2,
+            position_embedding_type="absolute",
+            use_cache=True,
+            classifier_dropout=None,
+            lora_adaptations=None,
+            lora_rank=4,
+            lora_dropout_p=0.0,
+            lora_alpha=1,
+            lora_main_params_trainable=False,
+            load_trained_adapters=False,
+            use_flash_attn=True,
+            torch_dtype=None,
+            emb_pooler=None,
+            matryoshka_dimensions=None,
+            truncate_dim=None,
+            **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+
+        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.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout
+        self.load_trained_adapters = load_trained_adapters
+        self.lora_adaptations = lora_adaptations
+        self.lora_rank = lora_rank
+        self.lora_dropout_p = lora_dropout_p
+        self.lora_alpha = lora_alpha
+        self.lora_main_params_trainable = lora_main_params_trainable
+        self.use_flash_attn = use_flash_attn
+        self.emb_pooler = emb_pooler
+        self.matryoshka_dimensions = matryoshka_dimensions
+        self.truncate_dim = truncate_dim
+        if torch_dtype and hasattr(torch, torch_dtype) and type(getattr(torch, torch_dtype)) is torch.dtype:
+            self.torch_dtype = getattr(torch, torch_dtype)
+        else:
+            self.torch_dtype = torch_dtype

embedding.py

@@ -0,0 +1,62 @@
+# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/embedding.py
+# Commit id: f1a73d074002226c42ce65a1df170ecff9f022c0
+
+# Copyright (c) 2022, Tri Dao.
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from torch import Tensor
+
+from transformers.models.xlm_roberta.modeling_xlm_roberta import create_position_ids_from_input_ids
+
+
+class XLMRobertaEmbeddings(nn.Module):
+    def __init__(
+        self,
+        embed_dim,
+        vocab_size,
+        max_position_embeddings,
+        type_vocab_size,
+        padding_idx=None,
+        device=None,
+        dtype=None,
+    ):
+        """
+        If max_position_embeddings <= 0, there's no position embeddings
+        If type_vocab_size <= 0, there's no token type embeddings
+        """
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        self.word_embeddings = nn.Embedding(
+            vocab_size, embed_dim, padding_idx=padding_idx, **factory_kwargs
+        )
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        if self.max_position_embeddings > 0:
+            self.position_embeddings = nn.Embedding(
+                max_position_embeddings, embed_dim, **factory_kwargs
+            )
+        if self.type_vocab_size > 0:
+            self.token_type_embeddings = nn.Embedding(type_vocab_size, embed_dim, **factory_kwargs)
+
+    def forward(self, input_ids, position_ids=None, token_type_ids=None):
+        """
+        input_ids: (batch, seqlen)
+        position_ids: (batch, seqlen)
+        token_type_ids: (batch, seqlen)
+        """
+        batch_size, seqlen = input_ids.shape
+        embeddings = self.word_embeddings(input_ids)
+        if self.max_position_embeddings > 0:
+            if position_ids is None:
+                position_ids = create_position_ids_from_input_ids(input_ids, padding_idx=self.word_embeddings.padding_idx).to(input_ids.device)
+                # position_ids = torch.arange(seqlen, dtype=torch.long, device=input_ids.device)
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings = embeddings + position_embeddings
+        if self.type_vocab_size > 0:
+            if token_type_ids is None:
+                token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
+            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            embeddings = embeddings + token_type_embeddings
+        return embeddings

mha.py

@@ -0,0 +1,662 @@
+# Copyright (c) 2023, Tri Dao.
+# Adapted from https://github.com/Dao-AILab/flash-attention/pull/556
+
+import math
+from functools import partial
+
+import torch
+import torch.nn as nn
+from einops import rearrange, repeat
+
+try:
+    from flash_attn import (
+        flash_attn_kvpacked_func,
+        flash_attn_qkvpacked_func,
+        flash_attn_varlen_kvpacked_func,
+        flash_attn_varlen_qkvpacked_func,
+        flash_attn_with_kvcache,
+    )
+except ImportError:
+    flash_attn_varlen_qkvpacked_func, flash_attn_varlen_kvpacked_func = None, None
+    flash_attn_qkvpacked_func, flash_attn_kvpacked_func = None, None
+    flash_attn_with_kvcache = None
+
+try:
+    from flash_attn.ops.fused_dense import ColumnParallelLinear, FusedDense, RowParallelLinear
+except ImportError:
+    FusedDense, ColumnParallelLinear, RowParallelLinear = None, None, None
+
+
+class FlashSelfAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(
+        self,
+        causal=False,
+        softmax_scale=None,
+        attention_dropout=0.0,
+        window_size=(-1, -1),
+        deterministic=False,
+    ):
+        super().__init__()
+        assert flash_attn_varlen_qkvpacked_func is not None, "FlashAttention is not installed"
+        assert flash_attn_qkvpacked_func is not None, "FlashAttention is not installed"
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+        self.window_size = window_size
+        self.deterministic = deterministic
+
+    def forward(self, qkv, causal=None, cu_seqlens=None, max_seqlen=None):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            qkv: The tensor containing the query, key, and value.
+                If cu_seqlens is None and max_seqlen is None, then qkv has shape (B, S, 3, H, D).
+                If cu_seqlens is not None and max_seqlen is not None, then qkv has shape
+                (total, 3, H, D), where total is the sum of the sequence lengths in the batch.
+            causal: if passed, will override self.causal
+            cu_seqlens: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
+                of the sequences in the batch, used to index into qkv.
+            max_seqlen: int. Maximum sequence length in the batch.
+        Returns:
+        --------
+            out: (total, H, D) if cu_seqlens is not None and max_seqlen is not None,
+                else (B, S, H, D).
+        """
+        assert qkv.dtype in [torch.float16, torch.bfloat16]
+        assert qkv.is_cuda
+        causal = self.causal if causal is None else causal
+        unpadded = cu_seqlens is not None
+
+        if unpadded:
+            assert cu_seqlens.dtype == torch.int32
+            assert max_seqlen is not None
+            assert isinstance(max_seqlen, int)
+            return flash_attn_varlen_qkvpacked_func(
+                qkv,
+                cu_seqlens,
+                max_seqlen,
+                self.drop.p if self.training else 0.0,
+                softmax_scale=self.softmax_scale,
+                causal=causal,
+                alibi_slopes=None,
+                window_size=self.window_size,
+                deterministic=self.deterministic,
+            )
+        else:
+            return flash_attn_qkvpacked_func(
+                qkv,
+                self.drop.p if self.training else 0.0,
+                softmax_scale=self.softmax_scale,
+                causal=causal,
+                alibi_slopes=None,
+                window_size=self.window_size,
+                deterministic=self.deterministic,
+            )
+
+
+class FlashCrossAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(
+        self,
+        causal=False,
+        softmax_scale=None,
+        attention_dropout=0.0,
+        window_size=(-1, -1),
+        deterministic=False,
+    ):
+        super().__init__()
+        assert flash_attn_varlen_kvpacked_func is not None, "FlashAttention is not installed"
+        assert flash_attn_kvpacked_func is not None, "FlashAttention is not installed"
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+        self.window_size = window_size
+        self.deterministic = deterministic
+
+    def forward(
+        self,
+        q,
+        kv,
+        causal=None,
+        cu_seqlens=None,
+        max_seqlen=None,
+        cu_seqlens_k=None,
+        max_seqlen_k=None,
+    ):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q: The tensor containing the query. (B, Sq, H, D)
+            kv: The tensor containing the key and value. (B, Sk, 2, H_k, D)
+            causal: if passed, will override self.causal
+            cu_seqlens: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
+                of the sequences in the batch, used to index into q.
+            max_seqlen: int. Maximum sequence length in the batch of q.
+            cu_seqlens_k: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
+                of the sequences in the batch, used to index into kv.
+            max_seqlen_k: int. Maximum sequence length in the batch of k and v.
+        """
+        assert q.dtype in [torch.float16, torch.bfloat16]
+        assert q.is_cuda and kv.is_cuda
+        causal = self.causal if causal is None else causal
+        unpadded = cu_seqlens is not None
+
+        if unpadded:
+            assert cu_seqlens.dtype == torch.int32
+            assert max_seqlen is not None
+            assert isinstance(max_seqlen, int)
+            assert cu_seqlens_k is not None
+            assert cu_seqlens_k.dtype == torch.int32
+            assert max_seqlen_k is not None
+            assert isinstance(max_seqlen, int)
+            return flash_attn_varlen_kvpacked_func(
+                q,
+                kv,
+                cu_seqlens,
+                cu_seqlens_k,
+                max_seqlen,
+                max_seqlen_k,
+                self.drop.p if self.training else 0.0,
+                softmax_scale=self.softmax_scale,
+                causal=causal,
+                alibi_slopes=None,
+                window_size=self.window_size,
+                deterministic=self.deterministic,
+            )
+        else:
+            batch_size, seqlen_q = q.shape[0], q.shape[1]
+            seqlen_k = kv.shape[1]
+            assert kv.shape[0] == batch_size and kv.shape[4] == q.shape[3]
+            return flash_attn_kvpacked_func(
+                q,
+                kv,
+                self.drop.p if self.training else 0.0,
+                causal=causal,
+                softmax_scale=self.softmax_scale,
+                alibi_slopes=None,
+                window_size=self.window_size,
+                deterministic=self.deterministic,
+            )
+
+
+class SelfAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0):
+        super().__init__()
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    def forward(self, qkv, causal=None, key_padding_mask=None):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            qkv: The tensor containing the query, key, and value. (B, S, 3, H, D)
+            causal: if passed, will override self.causal
+            key_padding_mask: boolean mask to apply to the attention weights. True means to keep,
+                False means to mask out. (B, S)
+        """
+        batch_size, seqlen = qkv.shape[0], qkv.shape[1]
+        causal = self.causal if causal is None else causal
+        q, k, v = qkv.unbind(dim=2)
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
+        if key_padding_mask is not None:
+            padding_mask = torch.full(
+                (batch_size, seqlen), -10000.0, dtype=scores.dtype, device=scores.device
+            )
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
+        if causal:
+            # "triu_tril_cuda_template" not implemented for 'BFloat16'
+            # So we have to construct the mask in float
+            causal_mask = torch.triu(
+                torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1
+            )
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + causal_mask.to(dtype=scores.dtype)
+        attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
+        attention_drop = self.drop(attention)
+        output = torch.einsum("bhts,bshd->bthd", attention_drop, v)
+        return output
+
+
+class CrossAttention(nn.Module):
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0):
+        super().__init__()
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.drop = nn.Dropout(attention_dropout)
+
+    def forward(self, q, kv, causal=None, key_padding_mask=None):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q: The tensor containing the query. (B, Sq, H, D)
+            kv: The tensor containing the key and value. (B, Sk, 2, H_k, D)
+            causal: if passed, will override self.causal
+            key_padding_mask: boolean mask to apply to the attention weights. True means to keep,
+                False means to mask out. (B, Sk)
+        """
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        causal = self.causal if causal is None else causal
+        seqlen_k = kv.shape[1]
+        assert kv.shape[0] == batch_size and kv.shape[4] == q.shape[3]
+        if kv.shape[3] != q.shape[2]:  # MQA/GQA
+            kv = repeat(kv, "... hkv d -> ... (hkv g) d", g=q.shape[2] // kv.shape[3])
+        k, v = kv.unbind(dim=2)
+        softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+        scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
+        if key_padding_mask is not None:
+            padding_mask = torch.full(
+                (batch_size, seqlen_k), -10000.0, dtype=scores.dtype, device=scores.device
+            )
+            padding_mask.masked_fill_(key_padding_mask, 0.0)
+            # TD [2022-09-30]: Adding is faster than masked_fill_ (idk why, just better kernel I guess)
+            scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
+        if causal:
+            # causal mask needs to take into account the difference between seqlen_q and seqlen_k
+            row_idx = rearrange(
+                torch.arange(seqlen_q, device=q.device, dtype=torch.long), "s -> s 1"
+            )
+            col_idx = torch.arange(seqlen_k, device=kv.device, dtype=torch.long)
+            sk = (
+                seqlen_k
+                if key_padding_mask is None
+                else rearrange(key_padding_mask.sum(-1), "b -> b 1 1 1")
+            )
+            causal_mask = col_idx > row_idx + sk - seqlen_q
+            scores = scores.masked_fill(causal_mask, -10000.0)
+        attention = torch.softmax(scores, dim=-1, dtype=v.dtype)
+        attention_drop = self.drop(attention)
+        output = torch.einsum("bhts,bshd->bthd", attention_drop, v)
+        return output
+
+
+class LinearResidual(nn.Linear):
+    """Wrap nn.Linear to return the residual as well. For compatibility with FusedDense."""
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        return super().forward(input), input
+
+
+def _update_kv_cache(kv, inference_params, layer_idx):
+    """kv: (batch_size, seqlen, 2, nheads, head_dim) or (batch_size, 1, 2, nheads, head_dim)"""
+    # Pre-allocate memory for key-values for inference.
+    num_heads, head_dim = kv.shape[-2:]
+    if layer_idx not in inference_params.key_value_memory_dict:
+        kv_cache = torch.empty(
+            inference_params.max_batch_size,
+            inference_params.max_seqlen,
+            2,
+            num_heads,
+            head_dim,
+            dtype=kv.dtype,
+            device=kv.device,
+        )
+        inference_params.key_value_memory_dict[layer_idx] = kv_cache
+    else:
+        kv_cache = inference_params.key_value_memory_dict[layer_idx]
+    # Adjust key and value for inference
+    batch_start = inference_params.batch_size_offset
+    batch_end = batch_start + kv.shape[0]
+    sequence_start = inference_params.seqlen_offset
+    sequence_end = sequence_start + kv.shape[1]
+    assert batch_end <= kv_cache.shape[0]
+    assert sequence_end <= kv_cache.shape[1]
+    assert kv_cache is not None
+    kv_cache[batch_start:batch_end, sequence_start:sequence_end, ...] = kv
+    return kv_cache[batch_start:batch_end, :sequence_end, ...]
+
+
+class MHA(nn.Module):
+    """Multi-head self-attention and cross-attention"""
+
+    def __init__(
+        self,
+        embed_dim,
+        num_heads,
+        num_heads_kv=None,
+        cross_attn=False,
+        qkv_proj_bias=True,
+        out_proj_bias=True,
+        dropout=0.0,
+        softmax_scale=None,
+        causal=False,
+        layer_idx=None,
+        dwconv=False,
+        window_size=(-1, -1),
+        fused_bias_fc=False,
+        use_flash_attn=False,
+        return_residual=False,
+        checkpointing=False,
+        device=None,
+        dtype=None,
+    ) -> None:
+        """
+        num_heads_kv: can be used to toggle MQA / GQA. If None, use num_heads.
+        return_residual: whether to return the input x along with the output. This is for
+            performance reason: for post-norm architecture, returning the input allows us
+            to fuse the backward of nn.Linear with the residual connection.
+        """
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.cross_attn = cross_attn
+        self.causal = causal
+        self.layer_idx = layer_idx
+        self.dwconv = dwconv
+        self.use_flash_attn = use_flash_attn
+        self.return_residual = return_residual
+        self.checkpointing = checkpointing
+
+        if window_size != (-1, -1):
+            assert use_flash_attn, "Local (sliding window) attention code path requires flash_attn"
+
+        self.num_heads = num_heads
+        self.num_heads_kv = num_heads_kv if num_heads_kv is not None else num_heads
+        assert (
+            self.num_heads % self.num_heads_kv == 0
+        ), "num_heads must be divisible by num_heads_kv"
+        assert self.embed_dim % num_heads == 0, "embed_dim must be divisible by num_heads"
+        self.head_dim = self.embed_dim // num_heads
+        qkv_dim = self.head_dim * (self.num_heads + 2 * self.num_heads_kv)
+        kv_dim = 2 * self.head_dim * self.num_heads_kv
+
+        if fused_bias_fc and FusedDense is None:
+            raise ImportError("fused_dense is not installed")
+        linear_cls = nn.Linear if not fused_bias_fc else FusedDense
+        linear_resid_cls = (
+            LinearResidual if not fused_bias_fc else partial(FusedDense, return_residual=True)
+        )
+        wqkv_cls = linear_cls if not self.return_residual else linear_resid_cls
+        inner_attn_cls = (
+            partial(FlashSelfAttention, window_size=window_size)
+            if use_flash_attn
+            else SelfAttention
+        )
+        inner_cross_attn_cls = (
+            partial(FlashCrossAttention, window_size=window_size)
+            if use_flash_attn
+            else CrossAttention
+        )
+        if not self.cross_attn:
+            self.Wqkv = wqkv_cls(embed_dim, qkv_dim, bias=qkv_proj_bias, **factory_kwargs)
+        else:
+            self.Wq = linear_cls(embed_dim, embed_dim, bias=qkv_proj_bias, **factory_kwargs)
+            self.Wkv = wqkv_cls(embed_dim, kv_dim, bias=qkv_proj_bias, **factory_kwargs)
+        if self.dwconv:
+            if self.num_heads_kv == self.num_heads:
+                self.dwconv_qkv = nn.Conv1d(
+                    qkv_dim, qkv_dim, kernel_size=3, padding=2, groups=qkv_dim
+                )
+            else:
+                self.dwconv_q = nn.Conv1d(
+                    embed_dim, embed_dim, kernel_size=3, padding=2, groups=embed_dim
+                )
+                self.dwconv_kv = nn.Conv1d(kv_dim, kv_dim, kernel_size=3, padding=2, groups=kv_dim)
+        self.inner_attn = inner_attn_cls(
+            causal=causal,
+            softmax_scale=softmax_scale,
+            attention_dropout=dropout,
+        )
+        self.inner_cross_attn = inner_cross_attn_cls(
+            causal=causal, softmax_scale=softmax_scale, attention_dropout=dropout
+        )
+        self.out_proj = linear_cls(embed_dim, embed_dim, bias=out_proj_bias, **factory_kwargs)
+
+    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None):
+        dtype = self.out_proj.weight.dtype if dtype is None else dtype
+        device = self.out_proj.weight.device
+        return torch.empty(
+            batch_size,
+            max_seqlen,
+            2,
+            self.num_heads_kv,
+            self.head_dim,
+            dtype=dtype,
+            device=device,
+        )
+
+    def _update_kv_cache(self, kv, inference_params):
+        """kv: (batch_size, seqlen, 2, nheads, head_dim) or (batch_size, 1, 2, nheads, head_dim)"""
+        assert not self.dwconv, "Generation does not support dwconv yet"
+        assert self.layer_idx is not None, "Generation requires layer_idx in the constructor"
+        return _update_kv_cache(kv, inference_params, self.layer_idx)
+
+    def _apply_rotary_update_kvcache_attention(self, q, kv, inference_params):
+        """
+        Fast path that combine 3 steps: apply rotary to Q and K, update kv cache, and apply attention.
+        q: (batch_size, seqlen_q, nheads, head_dim)
+        kv: (batch_size, seqlen_k, 2, nheads_kv, head_dim)
+        """
+        assert inference_params is not None and inference_params.seqlen_offset > 0
+        assert self.use_flash_attn
+        batch = q.shape[0]
+        kv_cache = inference_params.key_value_memory_dict[self.layer_idx][:batch]
+        cache_seqlens = (
+            inference_params.lengths_per_sample[:batch]
+            if inference_params.lengths_per_sample is not None
+            else inference_params.seqlen_offset
+        )
+        context = flash_attn_with_kvcache(
+            q,
+            kv_cache[:, :, 0],
+            kv_cache[:, :, 1],
+            kv[:, :, 0],
+            kv[:, :, 1],
+            cache_seqlens=cache_seqlens,
+            softmax_scale=self.inner_cross_attn.softmax_scale,
+            causal=self.inner_cross_attn.causal,
+            rotary_interleaved=False,
+            alibi_slopes=None,
+        )
+        return context
+
+    def _update_kvcache_attention(self, q, kv, inference_params):
+        """Write kv to inference_params, then do attention"""
+        if (
+            inference_params.seqlen_offset == 0
+            or flash_attn_with_kvcache is None
+            or not self.use_flash_attn
+        ):
+            # TODO: this only uses seqlen_offset and not lengths_per_sample.
+            kv = self._update_kv_cache(kv, inference_params)
+            return self.inner_cross_attn(q, kv)
+        else:
+            batch = q.shape[0]
+            kv_cache = inference_params.key_value_memory_dict[self.layer_idx][:batch]
+            cache_seqlens = (
+                inference_params.lengths_per_sample[:batch]
+                if inference_params.lengths_per_sample is not None
+                else inference_params.seqlen_offset
+            )
+            return flash_attn_with_kvcache(
+                q,
+                kv_cache[:, :, 0],
+                kv_cache[:, :, 1],
+                kv[:, :, 0],
+                kv[:, :, 1],
+                cache_seqlens=cache_seqlens,
+                softmax_scale=self.inner_cross_attn.softmax_scale,
+                causal=self.inner_cross_attn.causal,
+                alibi_slopes=None,
+            )
+
+    def forward(
+        self,
+        x,
+        x_kv=None,
+        key_padding_mask=None,
+        cu_seqlens=None,
+        max_seqlen=None,
+        mixer_subset=None,
+        inference_params=None,
+        **kwargs,
+    ):
+        """
+        Arguments:
+            x: (batch, seqlen, hidden_dim) (where hidden_dim = num heads * head dim) if
+                cu_seqlens is None and max_seqlen is None, else (total, hidden_dim) where total
+                is the is the sum of the sequence lengths in the batch.
+            x_kv: (batch, seqlen, hidden_dim), only applicable for cross-attention. If None, use x.
+            cu_seqlens: (batch_size + 1,), dtype torch.int32. The cumulative sequence lengths
+                of the sequences in the batch, used to index into x. Only applicable when using
+                FlashAttention.
+            max_seqlen: int. Maximum sequence length in the batch.
+            key_padding_mask: boolean mask, True means to keep, False means to mask out.
+                (batch, seqlen). Only applicable when not using FlashAttention.
+            mixer_subset: for cross-attention only. If not None, will take a subset of x
+                before applying the query projection. Useful for e.g., ViT where we only care
+                about the CLS token in the last layer.
+            inference_params: for generation. Adapted from Megatron-LM (and Apex)
+            https://github.com/NVIDIA/apex/blob/3ff1a10f72ec07067c4e44759442329804ac5162/apex/transformer/testing/standalone_transformer_lm.py#L470
+        """
+        if cu_seqlens is not None:
+            assert max_seqlen is not None
+            assert key_padding_mask is None
+            assert self.use_flash_attn
+            assert not self.dwconv
+        if key_padding_mask is not None:
+            assert cu_seqlens is None
+            assert max_seqlen is None
+            assert not self.use_flash_attn
+        if inference_params is not None:
+            assert key_padding_mask is None
+            assert cu_seqlens is None and max_seqlen is None
+            assert not self.dwconv
+
+        kwargs = (
+            {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen, **kwargs}
+            if self.use_flash_attn
+            else {"key_padding_mask": key_padding_mask, **kwargs}
+        )
+        seqlen_offset = (
+            0
+            if inference_params is None
+            else (
+                inference_params.lengths_per_sample
+                if inference_params.lengths_per_sample is not None
+                else inference_params.seqlen_offset
+            )
+        )
+        rotary_max_seqlen = (
+            inference_params.max_sequence_len if inference_params is not None else max_seqlen
+        )
+        batch, seqlen = x.shape[:2]
+        if not self.cross_attn and self.num_heads_kv == self.num_heads:
+            assert x_kv is None and mixer_subset is None
+            if not self.return_residual:
+                qkv = self.Wqkv(x)
+            else:
+                qkv, x = self.Wqkv(x)
+            if self.dwconv:
+                qkv = rearrange(
+                    self.dwconv_qkv(rearrange(qkv, "b s d -> b d s"))[..., :-2], "b d s -> b s d"
+                ).contiguous()
+            qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
+            if (
+                inference_params is None
+                or inference_params.seqlen_offset == 0
+                or (self.rotary_emb_dim == 0 or self.rotary_emb_dim % 16 != 0)
+                or not self.use_flash_attn
+            ):
+                if inference_params is None:
+                    if not self.checkpointing:
+                        context = self.inner_attn(qkv, **kwargs)
+                    else:
+                        context = torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, **kwargs)
+                else:
+                    context = self._update_kvcache_attention(
+                        qkv[:, :, 0], qkv[:, :, 1:], inference_params
+                    )
+            else:
+                context = self._apply_rotary_update_kvcache_attention(
+                    qkv[:, :, 0], qkv[:, :, 1:], inference_params
+                )
+        else:
+            if self.cross_attn:
+                if not self.return_residual:
+                    q = self.Wq(x if mixer_subset is None else x[:, mixer_subset])
+                    kv = self.Wkv(x_kv if x_kv is not None else x)
+                else:
+                    if x_kv is not None:
+                        kv, x_kv = self.Wkv(x_kv)
+                    else:
+                        kv, x = self.Wkv(x)
+                    q = self.Wq(x if mixer_subset is None else x[:, mixer_subset])
+            else:
+                assert self.num_heads_kv != self.num_heads
+                if not self.return_residual:
+                    qkv = self.Wqkv(x)
+                else:
+                    qkv, x = self.Wqkv(x)
+                q = qkv[..., : self.num_heads * self.head_dim]
+                kv = qkv[..., self.num_heads * self.head_dim :]
+            q = rearrange(q, "... (h d) -> ... h d", d=self.head_dim)
+            kv = rearrange(kv, "... (two hkv d) -> ... two hkv d", two=2, d=self.head_dim)
+            if self.dwconv:
+                q = rearrange(
+                    self.dwconv_q(rearrange(q, "b s d -> b d s"))[..., :-2], "b d s -> b s d"
+                ).contiguous()
+                kv = rearrange(
+                    self.dwconv_kv(rearrange(kv, "b s d -> b d s"))[..., :-2], "b d s -> b s d"
+                ).contiguous()
+            if (
+                inference_params is None
+                or inference_params.seqlen_offset == 0
+                or (self.rotary_emb_dim == 0 or self.rotary_emb_dim % 16 != 0)
+                or not self.use_flash_attn
+            ):
+                if inference_params is None:
+                    if not self.checkpointing:
+                        context = self.inner_cross_attn(q, kv, **kwargs)
+                    else:
+                        context = torch.utils.checkpoint.checkpoint(
+                            self.inner_cross_attn, q, kv, **kwargs
+                        )
+                else:
+                    context = self._update_kvcache_attention(q, kv, inference_params)
+            else:
+                context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
+        out = self.out_proj(rearrange(context, "... h d -> ... (h d)"))
+        return out if not self.return_residual else (out, x)

mlp.py

@@ -0,0 +1,194 @@
+# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mlp.py
+# Commit id: c3b219665292c61a51153d0ded4473c494296382
+
+# Copyright (c) 2023, Tri Dao.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.distributed import ProcessGroup
+
+
+try:
+    from flash_attn.ops.activations import swiglu
+except ImportError:
+    swiglu = None
+
+try:
+    from flash_attn.ops.fused_dense import ColumnParallelLinear, RowParallelLinear
+except ImportError:
+    ColumnParallelLinear, RowParallelLinear = None, None
+
+try:
+    from flash_attn.ops.fused_dense import FusedMLP, ParallelFusedMLP
+except ImportError:
+    FusedMLP, ParallelFusedMLP = None, None
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        activation=F.gelu,
+        bias1=True,
+        bias2=True,
+        return_residual=False,
+        device=None,
+        dtype=None,
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        out_features = out_features if out_features is not None else in_features
+        hidden_features = hidden_features if hidden_features is not None else in_features * 4
+        self.return_residual = return_residual
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias1, **factory_kwargs)
+        self.activation = activation
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)
+
+    def forward(self, x):
+        y = self.fc1(x)
+        y = self.activation(y)
+        y = self.fc2(y)
+        return y if not self.return_residual else (y, x)
+
+
+class ParallelMLP(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        activation=F.gelu,
+        process_group: ProcessGroup = None,
+        sequence_parallel=True,
+        bias1=True,
+        bias2=True,
+        device=None,
+        dtype=None,
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        assert ColumnParallelLinear is not None, "Need to install fused_dense"
+        assert RowParallelLinear is not None, "Need to install fused_dense"
+        out_features = out_features if out_features is not None else in_features
+        hidden_features = hidden_features if hidden_features is not None else in_features * 4
+        self.fc1 = ColumnParallelLinear(
+            in_features,
+            hidden_features,
+            process_group,
+            bias=bias1,
+            sequence_parallel=sequence_parallel,
+            **factory_kwargs,
+        )
+        self.activation = activation
+        self.fc2 = RowParallelLinear(
+            hidden_features,
+            out_features,
+            process_group,
+            bias=bias2,
+            sequence_parallel=sequence_parallel,
+            **factory_kwargs,
+        )
+
+    def forward(self, x):
+        y = self.fc1(x)
+        y = self.activation(y)
+        y = self.fc2(y)
+        return y
+
+
+class GatedMlp(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        activation=F.sigmoid,
+        bias1=True,
+        bias2=True,
+        multiple_of=128,
+        return_residual=False,
+        device=None,
+        dtype=None,
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        out_features = out_features if out_features is not None else in_features
+        hidden_features = (
+            hidden_features if hidden_features is not None else int(8 * in_features / 3)
+        )
+        hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
+        self.return_residual = return_residual
+        self.fc1 = nn.Linear(in_features, 2 * hidden_features, bias=bias1, **factory_kwargs)
+        self.activation = activation
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)
+
+    def forward(self, x):
+        y = self.fc1(x)
+        if self.activation == F.sigmoid:  # Special case for GLU
+            y = F.glu(y, dim=-1)
+        elif self.activation == F.silu and swiglu is not None:  # Special case for SwiGLU
+            y, gate = y.chunk(2, dim=-1)
+            y = swiglu(gate, y)
+        else:
+            y, gate = y.chunk(2, dim=-1)
+            y = y * self.activation(gate)
+        y = self.fc2(y)
+        return y if not self.return_residual else (y, x)
+
+
+class ParallelGatedMlp(nn.Module):
+    """Parallel GatedMlp"""
+
+    def __init__(
+        self,
+        in_features,
+        process_group,
+        hidden_features=None,
+        out_features=None,
+        activation=F.sigmoid,
+        bias1=True,
+        bias2=True,
+        multiple_of=128,
+        sequence_parallel=True,
+        device=None,
+        dtype=None,
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        out_features = out_features if out_features is not None else in_features
+        hidden_features = (
+            hidden_features if hidden_features is not None else int(8 * in_features / 3)
+        )
+        hidden_features = (hidden_features + multiple_of - 1) // multiple_of * multiple_of
+        if ColumnParallelLinear is None or RowParallelLinear is None:
+            raise ImportError("fused_dense is not installed")
+        self.fc1 = ColumnParallelLinear(
+            in_features,
+            2 * hidden_features,
+            process_group,
+            bias=bias1,
+            sequence_parallel=sequence_parallel,
+            **factory_kwargs,
+        )
+        self.activation = activation
+        self.fc2 = RowParallelLinear(
+            hidden_features,
+            out_features,
+            process_group,
+            bias=bias2,
+            sequence_parallel=sequence_parallel,
+            **factory_kwargs,
+        )
+
+    def forward(self, x):
+        y = self.fc1(x)
+        if self.activation == F.sigmoid:  # Special case for GLU
+            y = F.glu(y, dim=-1)
+        else:
+            y, gate = y.chunk(2, dim=-1)
+            y = y * self.activation(gate)
+        y = self.fc2(y)
+        return y

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b416a3cd3ba537175d19829b09324707b9ee80bb603518dce2f3a2284feea76
+size 556892306

modeling_xlm_roberta.py

@@ -0,0 +1,1119 @@
+# This implementation was adopted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/models/bert.py
+# Commit id: abbc1311731867310635f9edc2a9ec18317c8c48
+# Copyright (c) 2022, Tri Dao.
+# This BERT implementation is based on our MLPerf 2.0 and MLPerf 2.1 BERT implementation.
+# https://github.com/mlcommons/training_results_v2.0/blob/main/HazyResearch/benchmarks/bert/implementations/pytorch/modeling.py
+# https://github.com/mlcommons/training_results_v2.1/blob/main/Azure-HazyResearch/benchmarks/bert/implementations/ND96amsr_A100_v4/modeling.py
+
+# Inspired by https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py
+
+import importlib.util
+import logging
+import re
+from collections import OrderedDict
+from collections.abc import Sequence
+from functools import partial
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+from einops import rearrange
+from transformers import PretrainedConfig
+from transformers.modeling_utils import PreTrainedModel
+from transformers.modeling_outputs import MaskedLMOutput,SequenceClassifierOutput
+from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead
+
+from transformers.models.bert.modeling_bert import (
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    BertForPreTrainingOutput,
+)
+
+from typing import List, Optional, Tuple, Union
+
+from .xlm_padding import (
+    index_first_axis,
+    index_first_axis_residual,
+    pad_input,
+    unpad_input,
+)
+from .configuration_xlm_roberta import XLMRobertaFlashConfig
+from .block import Block
+from .embedding import XLMRobertaEmbeddings
+from .mha import MHA
+from .mlp import FusedMLP, Mlp
+
+try:
+    from flash_attn.ops.fused_dense import FusedDense
+except ImportError:
+    FusedDense = None
+
+try:
+    from flash_attn.ops.triton.layer_norm import layer_norm_fn
+except ImportError:
+    layer_norm_fn = None
+
+
+try:
+    from flash_attn.losses.cross_entropy import CrossEntropyLoss
+except ImportError:
+    CrossEntropyLoss = torch.nn.CrossEntropyLoss
+
+try:
+    from tqdm.autonotebook import trange
+except ImportError:
+    trange = None
+
+
+logger = logging.getLogger(__name__)
+
+
+def get_use_flash_attn(config: XLMRobertaFlashConfig):
+    if not getattr(config, "use_flash_attn", False):
+        return False
+    if not torch.cuda.is_available():
+        return False
+    if importlib.util.find_spec("flash_attn") is None:
+        logger.warning(
+            'flash_attn is not installed. Using PyTorch native attention implementation.'
+        )
+        return False
+    return True
+
+
+def create_mixer_cls(config, cross_attn=False, return_residual=False):
+    use_flash_attn = get_use_flash_attn(config)
+    fused_bias_fc = getattr(config, "fused_bias_fc", False)
+
+    mixer_cls = partial(
+        MHA,
+        num_heads=config.num_attention_heads,
+        cross_attn=cross_attn,
+        dropout=config.attention_probs_dropout_prob,
+        causal=False,
+        fused_bias_fc=fused_bias_fc,
+        use_flash_attn=use_flash_attn,
+        return_residual=return_residual,
+    )
+    return mixer_cls
+
+
+def create_mlp_cls(config, layer_idx=None, return_residual=False):
+    inner_dim = config.intermediate_size
+    fused_mlp = getattr(config, "fused_mlp", False)
+    if fused_mlp:
+        assert config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"], (
+            "fused_mlp only " "supports approximate gelu"
+        )
+    if not fused_mlp:
+        approximate = (
+            "tanh"
+            if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
+            else "none"
+        )
+        mlp_cls = partial(
+            Mlp,
+            hidden_features=inner_dim,
+            activation=partial(F.gelu, approximate=approximate),
+            return_residual=return_residual,
+        )
+    else:
+        if FusedMLP is None:
+            raise ImportError("fused_dense is not installed")
+        mlp_checkpoint_lvl = getattr(config, "mlp_checkpoint_lvl", 0)
+        # mlp_checkpoint_lvl could be a list, which contains the checkpoint_lvl for each layer
+        if isinstance(mlp_checkpoint_lvl, Sequence):
+            assert layer_idx is not None
+            mlp_checkpoint_lvl = mlp_checkpoint_lvl[layer_idx]
+        mlp_cls = partial(
+            FusedMLP,
+            hidden_features=inner_dim,
+            checkpoint_lvl=mlp_checkpoint_lvl,
+            return_residual=return_residual,
+        )
+    return mlp_cls
+
+
+def create_block(config, layer_idx=None):
+    last_layer_subset = getattr(config, "last_layer_subset", False)
+    cross_attn = last_layer_subset and layer_idx == config.num_hidden_layers - 1
+    # TD [2022-12-19]: For cross attention (last layer), we actually want to return the
+    # residual x_kv, not residual x. But it's annoying to change the API (and it only affects
+    # one layer) so we just choose not to return residual in this case.
+    return_residual = not cross_attn
+    mixer_cls = create_mixer_cls(config, cross_attn, return_residual=return_residual)
+    mlp_cls = create_mlp_cls(config, layer_idx, return_residual=return_residual)
+    norm_cls = partial(nn.LayerNorm, eps=config.layer_norm_eps)
+    block = Block(
+        config.hidden_size,
+        mixer_cls,
+        mlp_cls,
+        norm_cls=norm_cls,
+        prenorm=False,
+        resid_dropout1=config.hidden_dropout_prob,
+        resid_dropout2=config.hidden_dropout_prob,
+        fused_dropout_add_ln=getattr(config, "fused_dropout_add_ln", False),
+        return_residual=return_residual,
+    )
+    return block
+
+
+# https://github.com/huggingface/transformers/blob/7032e0203262ebb2ebf55da8d2e01f873973e835/src/transformers/models/bert/modeling_bert.py#L748
+def _init_weights(module, initializer_range=0.02):
+    if isinstance(module, nn.Linear):
+        nn.init.normal_(module.weight, std=initializer_range)
+        if module.bias is not None:
+            nn.init.zeros_(module.bias)
+    elif isinstance(module, nn.Embedding):
+        nn.init.normal_(module.weight, std=initializer_range)
+        if module.padding_idx is not None:
+            nn.init.zeros_(module.weight[module.padding_idx])
+
+
+class XLMRobertaEncoder(nn.Module):
+    def __init__(self, config: XLMRobertaFlashConfig):
+        super().__init__()
+        self.use_flash_attn = get_use_flash_attn(config)
+        self.layers = nn.ModuleList(
+            [create_block(config, layer_idx=i) for i in range(config.num_hidden_layers)]
+        )
+        self._grad_checkpointing = False
+
+    @property
+    def gradient_checkpointing(self):
+        return self._grad_checkpointing
+
+    @gradient_checkpointing.setter
+    def gradient_checkpointing(self, value):
+        self._grad_checkpointing = value
+
+    def forward(self, hidden_states, key_padding_mask=None, subset_mask=None):
+        """If subset_mask is not None, we only want output for the subset of the sequence.
+        This means that we only compute the last layer output for these tokens.
+        subset_mask: (batch, seqlen), dtype=torch.bool
+        """
+        if key_padding_mask is None or not self.use_flash_attn:
+            mixer_kwargs = (
+                {"key_padding_mask": key_padding_mask.bool()}
+                if key_padding_mask is not None
+                else None
+            )
+            for layer in self.layers:
+                if self._grad_checkpointing:
+                    hidden_states = torch.utils.checkpoint.checkpoint(
+                        layer,
+                        hidden_states,
+                        use_reentrant=False,
+                        mixer_kwargs=mixer_kwargs,
+                    )
+                else:
+                    hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
+            if subset_mask is not None:
+                hidden_states = hidden_states[subset_mask]
+        else:
+            batch, seqlen = hidden_states.shape[:2]
+            hidden_states, indices, cu_seqlens, max_seqlen_in_batch = unpad_input(
+                hidden_states, key_padding_mask
+            )
+            mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch}
+            if subset_mask is None:
+                for layer in self.layers:
+                    if self._grad_checkpointing:
+                        hidden_states = torch.utils.checkpoint.checkpoint(
+                            layer,
+                            hidden_states,
+                            use_reentrant=False,
+                            mixer_kwargs=mixer_kwargs,
+                        )
+                    else:
+                        hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
+                hidden_states = pad_input(hidden_states, indices, batch, seqlen)
+            else:
+                for layer in self.layers[:-1]:
+                    if self._grad_checkpointing:
+                        hidden_states = torch.utils.checkpoint.checkpoint(
+                            layer,
+                            hidden_states,
+                            use_reentrant=False,
+                            mixer_kwargs=mixer_kwargs,
+                        )
+                    else:
+                        hidden_states = layer(hidden_states, mixer_kwargs=mixer_kwargs)
+                if key_padding_mask is not None:
+                    subset_idx = torch.nonzero(
+                        subset_mask[key_padding_mask], as_tuple=False
+                    ).flatten()
+                    subset_seqlens = (subset_mask & key_padding_mask).sum(
+                        dim=-1, dtype=torch.int32
+                    )
+                    subset_cu_seqlens = F.pad(
+                        torch.cumsum(subset_seqlens, dim=0, dtype=torch.torch.int32),
+                        (1, 0),
+                    )
+                else:
+                    subset_idx = torch.nonzero(subset_mask, as_tuple=False).flatten()
+                    subset_seqlens = subset_mask.sum(dim=-1, dtype=torch.int32)
+                    subset_cu_seqlens = F.pad(
+                        torch.cumsum(subset_seqlens, dim=0, dtype=torch.torch.int32),
+                        (1, 0),
+                    )
+                hidden_states_subset, hidden_states = index_first_axis_residual(
+                    hidden_states, subset_idx
+                )
+                # It's ok to set max_seqlen_q to be much larger
+                mixer_kwargs = {
+                    "x_kv": hidden_states,
+                    "cu_seqlens": subset_cu_seqlens,
+                    "max_seqlen": max_seqlen_in_batch,
+                    "cu_seqlens_k": cu_seqlens,
+                    "max_seqlen_k": max_seqlen_in_batch,
+                }
+                if self._grad_checkpointing:
+                    torch.utils.checkpoint.checkpoint(
+                        self.layers[-1],
+                        hidden_states_subset,
+                        use_reentrant=False,
+                        mixer_kwargs=mixer_kwargs,
+                    )
+                else:
+                    hidden_states = self.layers[-1](
+                        hidden_states_subset, mixer_kwargs=mixer_kwargs
+                    )
+        return hidden_states
+
+
+class XLMRobertaPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        fused_bias_fc = getattr(config, "fused_bias_fc", False)
+        if fused_bias_fc and FusedDense is None:
+            raise ImportError("fused_dense is not installed")
+        linear_cls = nn.Linear if not fused_bias_fc else FusedDense
+        self.dense = linear_cls(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states, pool=True):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0] if pool else hidden_states
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class XLMRobertaPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        fused_bias_fc = getattr(config, "fused_bias_fc", False)
+        if fused_bias_fc and FusedDense is None:
+            raise ImportError("fused_dense is not installed")
+        self.fused_dropout_add_ln = getattr(config, "fused_dropout_add_ln", False)
+        if self.fused_dropout_add_ln and layer_norm_fn is None:
+            raise ImportError("Triton is not installed")
+        linear_cls = nn.Linear if not fused_bias_fc else FusedDense
+        self.dense = linear_cls(config.hidden_size, config.hidden_size)
+        approximate = (
+            "tanh"
+            if config.hidden_act in ["gelu_new", "gelu_fast", "gelu_pytorch_tanh"]
+            else "none"
+        )
+        self.transform_act_fn = nn.GELU(approximate=approximate)
+        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        if not self.fused_dropout_add_ln:
+            hidden_states = self.layer_norm(hidden_states)
+        else:
+            hidden_states = layer_norm_fn(
+                hidden_states,
+                self.layer_norm.weight,
+                self.layer_norm.bias,
+                eps=self.layer_norm.eps,
+            )
+        return hidden_states
+
+
+class XLMRobertaLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        fused_bias_fc = getattr(config, "fused_bias_fc", False)
+        if fused_bias_fc and FusedDense is None:
+            raise ImportError("fused_dense is not installed")
+        linear_cls = nn.Linear if not fused_bias_fc else FusedDense
+
+        self.transform = XLMRobertaPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = linear_cls(config.hidden_size, config.vocab_size, bias=True)
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class XLMRobertaPreTrainingHeads(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = XLMRobertaLMPredictionHead(config)
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, sequence_output, pooled_output):
+        prediction_scores = self.predictions(sequence_output)
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return prediction_scores, seq_relationship_score
+
+
+class XLMRobertaPreTrainedModel(PreTrainedModel):
+    """An abstract class to handle weights initialization and
+    a simple interface for dowloading and loading pretrained models.
+    """
+
+    config_class = XLMRobertaFlashConfig
+    base_model_prefix = "roberta"
+    supports_gradient_checkpointing = True
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, XLMRobertaEncoder):
+            module.gradient_checkpointing = value
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        *args,
+        **kwargs,
+    ):
+        if not 'torch_dtype' in kwargs:
+            kwargs['torch_dtype'] = 'auto'
+        return super().from_pretrained(*args, **kwargs)
+
+
+
+class XLMRobertaModel(XLMRobertaPreTrainedModel):
+    def __init__(self, config: XLMRobertaFlashConfig, add_pooling_layer=True):
+        super().__init__(config)
+        self.pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
+        if config.vocab_size % self.pad_vocab_size_multiple != 0:
+            config.vocab_size += self.pad_vocab_size_multiple - (
+                config.vocab_size % self.pad_vocab_size_multiple
+            )
+        self.fused_dropout_add_ln = getattr(config, "fused_dropout_add_ln", False)
+        if self.fused_dropout_add_ln and layer_norm_fn is None:
+            raise ImportError("Triton is not installed")
+        assert config.hidden_act in [
+            "gelu",
+            "gelu_new",
+            "gelu_fast",
+            "gelu_pytorch_tanh",
+        ]
+
+        self.embeddings = XLMRobertaEmbeddings(
+            config.hidden_size,
+            config.vocab_size,
+            config.max_position_embeddings if config.position_embedding_type == 'absolute' else -1,
+            config.type_vocab_size,
+            padding_idx=config.pad_token_id,
+        )
+        self.emb_drop = nn.Dropout(config.hidden_dropout_prob)
+        self.emb_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.encoder = XLMRobertaEncoder(config)
+        self.pooler = XLMRobertaPooler(config) if add_pooling_layer else None
+
+        self.apply(partial(_init_weights, initializer_range=config.initializer_range))
+
+
+    @torch.inference_mode()
+    def encode(
+        self: 'XLMRobertaModel',
+        sentences: Union[str, List[str]],
+        batch_size: int = 32,
+        show_progress_bar: Optional[bool] = None,
+        output_value: str = 'sentence_embedding',
+        convert_to_numpy: bool = True,
+        convert_to_tensor: bool = False,
+        device: Optional[torch.device] = None,
+        normalize_embeddings: bool = False,
+        truncate_dim: Optional[int] = None,
+        **tokenizer_kwargs,
+    ) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
+        """
+        Computes sentence embeddings
+        Args:
+            sentences(`str` or `List[str]`):
+                Sentence or sentences to be encoded
+            batch_size(`int`, *optional*, defaults to 32):
+                Batch size for the computation
+            show_progress_bar(`bool`, *optional*, defaults to None):
+                Show a progress bar when encoding sentences.
+                If set to None, progress bar is only shown when
+                `logger.level == logging.INFO` or `logger.level == logging.DEBUG`.
+            output_value(`str`, *optional*, defaults to 'sentence_embedding'):
+                Default sentence_embedding, to get sentence embeddings.
+                Can be set to token_embeddings to get wordpiece token embeddings.
+                Set to None, to get all output values
+            convert_to_numpy(`bool`, *optional*, defaults to True):
+                If true, the output is a list of numpy vectors.
+                Else, it is a list of pytorch tensors.
+            convert_to_tensor(`bool`, *optional*, defaults to False):
+                If true, you get one large tensor as return.
+                Overwrites any setting from convert_to_numpy
+            device(`torch.device`, *optional*, defaults to None):
+                Which torch.device to use for the computation
+            normalize_embeddings(`bool`, *optional*, defaults to False):
+                If set to true, returned vectors will have length 1. In that case, the
+                faster dot-product (util.dot_score) instead of cosine similarity can
+                be used.
+            truncate_dim(`int`, *optional*, defaults to None):
+                The dimension to truncate sentence embeddings to. `None` does no truncation.
+            tokenizer_kwargs(`Dict[str, Any]`, *optional*, defaults to {}):
+                Keyword arguments for the tokenizer
+        Returns:
+            By default, a list of tensors is returned.
+            If convert_to_tensor, a stacked tensor is returned.
+            If convert_to_numpy, a numpy matrix is returned.
+        """
+        from transformers import AutoTokenizer
+
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            self.name_or_path, trust_remote_code=True
+        )
+
+        is_training = self.training
+        self.eval()
+
+        if show_progress_bar is None:
+            show_progress_bar = (
+                logger.getEffectiveLevel() == logging.INFO
+                or logger.getEffectiveLevel() == logging.DEBUG
+            )
+
+        if convert_to_tensor:
+            convert_to_numpy = False
+
+        if output_value != 'sentence_embedding':
+            convert_to_tensor = False
+            convert_to_numpy = False
+
+        input_was_string = False
+        if isinstance(sentences, str) or not hasattr(sentences, '__len__'):
+            sentences = [sentences]
+            input_was_string = True
+
+        if device is not None:
+            self.to(device)
+
+        permutation = np.argsort([-len(i) for i in sentences])
+        inverse_permutation = np.argsort(permutation)
+        sentences = [sentences[idx] for idx in permutation]
+
+        tokenizer_kwargs['padding'] = tokenizer_kwargs.get('padding', True)
+        tokenizer_kwargs['max_length'] = tokenizer_kwargs.get(
+            'max_length', self.tokenizer.init_kwargs.get('model_max_length', 8192)
+        )
+        tokenizer_kwargs['truncation'] = tokenizer_kwargs.get('truncation', True)
+
+        all_embeddings = []
+
+        if trange is not None:
+            range_iter = trange(
+                0,
+                len(sentences),
+                batch_size,
+                desc="Encoding",
+                disable=not show_progress_bar,
+            )
+        else:
+            range_iter = range(0, len(sentences), batch_size)
+
+        for i in range_iter:
+            encoded_input = self.tokenizer(
+                sentences[i : i + batch_size],
+                return_tensors='pt',
+                **tokenizer_kwargs,
+            ).to(self.device)
+            token_embs = self.forward(**encoded_input)[0]
+
+            # Accumulate in fp32 to avoid overflow
+            token_embs = token_embs.float()
+
+            if output_value == 'token_embeddings':
+                raise NotImplementedError
+            elif output_value is None:
+                raise NotImplementedError
+            else:
+                if self.config.emb_pooler == 'cls':
+                    embeddings = self.cls_pooling(
+                        token_embs, encoded_input['attention_mask']
+                    )
+                else:
+                    embeddings = self.mean_pooling(
+                        token_embs, encoded_input['attention_mask']
+                    )
+
+                if normalize_embeddings:
+                    embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)
+
+                if convert_to_numpy:
+                    embeddings = embeddings.cpu()
+            all_embeddings.extend(embeddings)
+
+        all_embeddings = [all_embeddings[idx] for idx in inverse_permutation]
+
+        truncate_dim = truncate_dim or self.config.truncate_dim
+        if truncate_dim:
+            all_embeddings = self.truncate_embeddings(all_embeddings, truncate_dim)
+
+        if convert_to_tensor:
+            all_embeddings = torch.stack(all_embeddings)
+        elif convert_to_numpy:
+            all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])
+
+        if input_was_string:
+            all_embeddings = all_embeddings[0]
+
+        self.train(is_training)
+        return all_embeddings
+
+
+    def truncate_embeddings(self, embeddings, truncate_dim):
+        if not self.config.matryoshka_dimensions:
+            logger.warning(
+                'Matryoshka embeddings are not supported, so dimension truncation will not be performed.'
+            )
+            return embeddings
+        elif truncate_dim in self.config.matryoshka_dimensions:
+            return [tensor[:truncate_dim] for tensor in embeddings]
+        else:
+            raise ValueError(f'The provided `truncate_dim` value of {truncate_dim} is not supported. '
+                             f'Supported dimensions are {self.config.matryoshka_dimensions}.')
+
+    def mean_pooling(
+        self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
+    ):
+        input_mask_expanded = (
+            attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
+        )
+        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
+            input_mask_expanded.sum(1), min=1e-9
+        )
+
+
+    def cls_pooling(
+        self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
+    ):
+        return token_embeddings[:,0]
+
+
+    def forward(
+        self,
+        input_ids,
+        position_ids=None,
+        token_type_ids=None,
+        attention_mask=None,
+        masked_tokens_mask=None,
+        return_dict=None,
+        **kwargs,
+    ):
+        """If masked_tokens_mask is not None (i.e. last_layer_subset == True in XLMForPreTraining),
+        we only want the output for the masked tokens. This means that we only compute the last
+        layer output for these tokens.
+        masked_tokens_mask: (batch, seqlen), dtype=torch.bool
+        """
+
+        if kwargs:
+            for key, value in kwargs.items():
+                if value is not None:
+                    logger.warning(
+                        'Flash attention implementation does not support kwargs: %s',
+                        key,
+                    )
+
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        hidden_states = self.embeddings(
+            input_ids, position_ids=position_ids, token_type_ids=token_type_ids
+        )
+        # TD [2022-12:18]: Don't need to force residual in fp32
+        # BERT puts embedding LayerNorm before embedding dropout.
+        if not self.fused_dropout_add_ln:
+            hidden_states = self.emb_ln(hidden_states)
+        else:
+            hidden_states = layer_norm_fn(
+                hidden_states, self.emb_ln.weight, self.emb_ln.bias, eps=self.emb_ln.eps
+            )
+        hidden_states = self.emb_drop(hidden_states)
+
+        if masked_tokens_mask is not None:
+            batch_size, seqlen = input_ids.shape[:2]
+            # We also need the first column for the CLS token
+            first_col_mask = torch.zeros(
+                batch_size, seqlen, dtype=torch.bool, device=input_ids.device
+            )
+            first_col_mask[:, 0] = True
+            subset_mask = masked_tokens_mask | first_col_mask
+        else:
+            subset_mask = None
+
+        sequence_output = self.encoder(
+            hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask
+        )
+
+        if masked_tokens_mask is None:
+            pooled_output = (
+                self.pooler(sequence_output) if self.pooler is not None else None
+            )
+        else:
+            # TD [2022-03-01]: the indexing here is very tricky.
+            if attention_mask is not None:
+                subset_idx = subset_mask[attention_mask]
+                pool_input = sequence_output[first_col_mask[attention_mask][subset_idx]]
+                sequence_output = sequence_output[
+                    masked_tokens_mask[attention_mask][subset_idx]
+                ]
+            else:
+                pool_input = sequence_output[first_col_mask[subset_mask]]
+                sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
+            pooled_output = (
+                self.pooler(pool_input, pool=False) if self.pooler is not None else None
+            )
+
+        if not return_dict:
+            return sequence_output, pooled_output
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+        )
+
+
+class XLMRobertaForMaskedLM(XLMRobertaPreTrainedModel):
+    _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `XLMRobertaForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        self.roberta = XLMRobertaModel(config, add_pooling_layer=False)
+        self.lm_head = XLMRobertaLMHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.roberta.embeddings.word_embeddings
+
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(prediction_scores.device)
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(
+                prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)
+            )
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return (
+                ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+            )
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaClassificationHead with Roberta->XLMRoberta
+class XLMRobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        fused_bias_fc = getattr(config, "fused_bias_fc", False)
+        if fused_bias_fc and FusedDense is None:
+            raise ImportError("fused_dense is not installed")
+        linear_cls = nn.Linear if not fused_bias_fc else FusedDense
+        self.dense = linear_cls(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout
+            if config.classifier_dropout is not None
+            else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = linear_cls(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaForSequenceClassification with Roberta->XLMRoberta, ROBERTA->XLM_ROBERTA
+class XLMRobertaForSequenceClassification(XLMRobertaPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = XLMRobertaModel(config, add_pooling_layer=False)
+        self.classifier = XLMRobertaClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        outputs = self.roberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (
+                    labels.dtype == torch.long or labels.dtype == torch.int
+                ):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+    @torch.inference_mode()
+    def compute_score(
+        self,
+        sentence_pairs: Union[List[Tuple[str, str]], Tuple[str, str]],
+        batch_size: int = 32,
+        max_length: Optional[int] = None,
+    ) -> List[float]:
+
+        if not hasattr(self, "_tokenizer"):
+            from transformers import AutoTokenizer
+
+            self._tokenizer = AutoTokenizer.from_pretrained(
+                self.name_or_path, trust_remote_code=True
+            )
+
+        assert isinstance(sentence_pairs, list)
+        if isinstance(sentence_pairs[0], str):
+            sentence_pairs = [sentence_pairs]
+
+        all_scores = []
+        for start_index in range(
+            0, len(sentence_pairs), batch_size
+        ):
+            sentences_batch = sentence_pairs[
+                start_index : start_index + batch_size
+            ]
+            inputs = self._tokenizer(
+                sentences_batch,
+                padding=True,
+                truncation=True,
+                return_tensors='pt',
+                max_length=max_length,
+            ).to(self.device)
+            scores = (
+                self.forward(**inputs, return_dict=True)
+                .logits.view(
+                    -1,
+                )
+                .float()
+            )
+            scores = torch.sigmoid(scores)
+            all_scores.extend(scores.cpu().numpy().tolist())
+
+        if len(all_scores) == 1:
+            return all_scores[0]
+        return all_scores
+
+    def predict(
+        self,
+        sentence_pairs: Union[List[Tuple[str, str]], Tuple[str, str]],
+        batch_size: int = 32,
+        max_length: Optional[int] = None,
+    ) -> List[float]:
+        # used for beir evaluation
+        return self.compute_score(sentence_pairs, batch_size=batch_size, max_length=max_length)
+
+    def rerank(
+        self,
+        query: str,
+        documents: List[str],
+        batch_size: int = 32,
+        max_length: int = 1024,
+        max_query_length: int = 512,
+        overlap_tokens: int = 80,
+        top_n: Optional[int] = None,
+        **kwargs,
+    ):
+        assert max_length >= max_query_length * 2, (
+            f'max_length ({max_length}) must be greater than or equal to '
+            f'max_query_length ({max_query_length}) * 2'
+        )
+
+        if not hasattr(self, "_tokenizer"):
+            from transformers import AutoTokenizer
+
+            self._tokenizer = AutoTokenizer.from_pretrained(
+                self.name_or_path, trust_remote_code=True
+            )
+
+        # preproc of tokenization
+        sentence_pairs, sentence_pairs_pids = reranker_tokenize_preproc(
+            query,
+            documents,
+            tokenizer=self._tokenizer,
+            max_length=max_length,
+            max_query_length=max_query_length,
+            overlap_tokens=overlap_tokens,
+        )
+
+        tot_scores = []
+        with torch.no_grad():
+            for k in range(0, len(sentence_pairs), batch_size):
+                batch = self._tokenizer.pad(
+                    sentence_pairs[k : k + batch_size],
+                    padding=True,
+                    max_length=max_length,
+                    pad_to_multiple_of=None,
+                    return_tensors="pt",
+                )
+                batch_on_device = {k: v.to(self.device) for k, v in batch.items()}
+                scores = (
+                    self.forward(**batch_on_device, return_dict=True)
+                    .logits.view(
+                        -1,
+                    )
+                    .float()
+                )
+                scores = torch.sigmoid(scores)
+                tot_scores.extend(scores.cpu().numpy().tolist())
+
+        # ranking
+        merge_scores = [0 for _ in range(len(documents))]
+        for pid, score in zip(sentence_pairs_pids, tot_scores):
+            merge_scores[pid] = max(merge_scores[pid], score)
+
+        merge_scores_argsort = np.argsort(merge_scores)[::-1]
+        sorted_documents = []
+        sorted_scores = []
+        for mid in merge_scores_argsort:
+            sorted_scores.append(merge_scores[mid])
+            sorted_documents.append(documents[mid])
+
+        top_n = min(top_n or len(sorted_documents), len(sorted_documents))
+
+        return [
+            {
+                'document': sorted_documents[i],
+                'relevance_score': sorted_scores[i],
+                'index': merge_scores_argsort[i],
+            }
+            for i in range(top_n)
+        ]
+
+
+def reranker_tokenize_preproc(
+    query: str,
+    passages: List[str],
+    tokenizer=None,
+    max_length: int = 1024,
+    max_query_length: int = 512,
+    overlap_tokens: int = 80,
+):
+    from copy import deepcopy
+
+    assert tokenizer is not None, "Please provide a valid tokenizer for tokenization!"
+    sep_id = tokenizer.sep_token_id
+
+    def _merge_inputs(chunk1_raw, chunk2):
+        chunk1 = deepcopy(chunk1_raw)
+        chunk1['input_ids'].append(sep_id)
+        chunk1['input_ids'].extend(chunk2['input_ids'])
+        chunk1['input_ids'].append(sep_id)
+        chunk1['attention_mask'].append(chunk2['attention_mask'][0])
+        chunk1['attention_mask'].extend(chunk2['attention_mask'])
+        chunk1['attention_mask'].append(chunk2['attention_mask'][-1])
+        if 'token_type_ids' in chunk1:
+            token_type_ids = [1 for _ in range(len(chunk2['token_type_ids']) + 2)]
+            chunk1['token_type_ids'].extend(token_type_ids)
+        return chunk1
+
+    # Note: the long query will be truncated to 256 tokens by default
+    query_inputs = tokenizer.encode_plus(
+        query, truncation=True, padding=False, max_length=max_query_length
+    )
+
+    max_passage_inputs_length = max_length - len(query_inputs['input_ids']) - 2
+    # assert (
+    #     max_passage_inputs_length > 100
+    # ), "Your query is too long! Please make sure your query less than 500 tokens!"
+
+    overlap_tokens_implt = min(overlap_tokens, max_passage_inputs_length // 4)
+
+    res_merge_inputs = []
+    res_merge_inputs_pids = []
+    for pid, passage in enumerate(passages):
+        passage_inputs = tokenizer.encode_plus(
+            passage,
+            truncation=False,
+            padding=False,
+            add_special_tokens=False,
+            max_length=0,
+        )
+        passage_inputs_length = len(passage_inputs['input_ids'])
+
+        if passage_inputs_length <= max_passage_inputs_length:
+            qp_merge_inputs = _merge_inputs(query_inputs, passage_inputs)
+            res_merge_inputs.append(qp_merge_inputs)
+            res_merge_inputs_pids.append(pid)
+        else:
+            start_id = 0
+            while start_id < passage_inputs_length:
+                end_id = start_id + max_passage_inputs_length
+                # make sure the length of the last chunk is `max_passage_inputs_length`
+                if end_id >= passage_inputs_length:
+                    sub_passage_inputs = {
+                        k: v[-max_passage_inputs_length:]
+                        for k, v in passage_inputs.items()
+                    }
+                else:
+                    sub_passage_inputs = {
+                        k: v[start_id:end_id] for k, v in passage_inputs.items()
+                    }
+                start_id = (
+                    end_id - overlap_tokens_implt
+                    if end_id < passage_inputs_length
+                    else end_id
+                )
+
+                qp_merge_inputs = _merge_inputs(query_inputs, sub_passage_inputs)
+                res_merge_inputs.append(qp_merge_inputs)
+                res_merge_inputs_pids.append(pid)
+
+    return res_merge_inputs, res_merge_inputs_pids

special_tokens_map.json

@@ -0,0 +1,51 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e802fe5337779428818439760a1e6161ed36ceed72d4ebcbda9c139a2108fc99
+size 17082988

tokenizer_config.json

@@ -0,0 +1,55 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "250001": {
+      "content": "<mask>",
+      "lstrip": true,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "extra_special_tokens": {},
+  "mask_token": "<mask>",
+  "model_max_length": 1024,
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "tokenizer_class": "XLMRobertaTokenizerFast",
+  "unk_token": "<unk>"
+}

xlm_padding.py

@@ -0,0 +1,218 @@
+# This implementation was adapted from https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/block.py
+# Commit id: c94cd09744d20f0ac587a351ff6ff2e8ad11ae1b
+
+# Previously adapted from https://github.com/mlcommons/training_results_v1.1/blob/main/NVIDIA/benchmarks/bert/implementations/pytorch/padding.py
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange, repeat
+
+
+class IndexFirstAxis(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input, indices):
+        ctx.save_for_backward(indices)
+        assert input.ndim >= 2
+        ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]
+        second_dim = other_shape.numel()
+        # TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
+        # return input[indices]
+        return torch.gather(
+            rearrange(input, "b ... -> b (...)"), 0, repeat(indices, "z -> z d", d=second_dim)
+        ).reshape(-1, *other_shape)
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        (indices,) = ctx.saved_tensors
+        assert grad_output.ndim >= 2
+        other_shape = grad_output.shape[1:]
+        grad_output = rearrange(grad_output, "b ... -> b (...)")
+        grad_input = torch.zeros(
+            [ctx.first_axis_dim, grad_output.shape[1]],
+            device=grad_output.device,
+            dtype=grad_output.dtype,
+        )
+        # TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
+        # grad_input[indices] = grad_output
+        grad_input.scatter_(0, repeat(indices, "z -> z d", d=grad_output.shape[1]), grad_output)
+        return grad_input.reshape(ctx.first_axis_dim, *other_shape), None
+
+
+index_first_axis = IndexFirstAxis.apply
+
+
+class IndexPutFirstAxis(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, values, indices, first_axis_dim):
+        ctx.save_for_backward(indices)
+        assert indices.ndim == 1
+        assert values.ndim >= 2
+        output = torch.zeros(
+            first_axis_dim, *values.shape[1:], device=values.device, dtype=values.dtype
+        )
+        # TD [2022-03-04] For some reason torch.scatter is a bit faster than indexing.
+        output[indices] = values
+        # output.scatter_(0, repeat(indices, 'z -> z d', d=values.shape[1]), values)
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        (indices,) = ctx.saved_tensors
+        # TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
+        grad_values = grad_output[indices]
+        # grad_values = torch.gather(grad_output, 0, repeat(indices, 'z -> z d', d=grad_output.shape[1]))
+        return grad_values, None, None
+
+
+index_put_first_axis = IndexPutFirstAxis.apply
+
+
+class IndexFirstAxisResidual(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input, indices):
+        ctx.save_for_backward(indices)
+        assert input.ndim >= 2
+        ctx.first_axis_dim, other_shape = input.shape[0], input.shape[1:]
+        second_dim = other_shape.numel()
+        # TD [2022-03-04] For some reason torch.gather is a bit faster than indexing.
+        output = input[indices]
+        # We don't want to reshape input (b ... -> b (...)) since it could change the channel_last
+        # memory format to channel_first. In other words, input might not be contiguous.
+        # If we don't detach, Pytorch complains about output being a view and is being modified inplace
+        return output, input.detach()
+
+    @staticmethod
+    def backward(ctx, grad_output, grad_residual):
+        (indices,) = ctx.saved_tensors
+        assert grad_output.ndim >= 2
+        other_shape = grad_output.shape[1:]
+        assert grad_residual.shape[1:] == other_shape
+        grad_input = grad_residual
+        # grad_input[indices] += grad_output
+        indices = indices.reshape(indices.shape[0], *((1,) * (grad_output.ndim - 1)))
+        indices = indices.expand_as(grad_output)
+        grad_input.scatter_add_(0, indices, grad_output)
+        return grad_input.reshape(ctx.first_axis_dim, *other_shape), None
+
+
+index_first_axis_residual = IndexFirstAxisResidual.apply
+
+
+def unpad_input(hidden_states, attention_mask):
+    """
+    Arguments:
+        hidden_states: (batch, seqlen, ...)
+        attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
+    Return:
+        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
+        indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.
+        cu_seqlens: (batch + 1), the cumulative sequence lengths, used to index into hidden_states.
+        max_seqlen_in_batch: int
+    """
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+    # TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the
+    # bool mask, then call nonzero to get the indices, then index with those. The indices is @dim
+    # times larger than it needs to be, wasting memory. It's faster and more memory-efficient to
+    # index with integer indices. Moreover, torch's index is a bit slower than it needs to be,
+    # so we write custom forward and backward to make it a bit faster.
+    return (
+        index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices),
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def unpad_input_for_concatenated_sequences(hidden_states, attention_mask_in_length):
+    """
+    Supports concatenating short samples in one sequence. The attention_mask_in_length is utilized to mask other short samples. It helps efficient training of variant lengths-based samples (e.g., the supervised fine-tuning task in large language model).
+    The motivation for this function is explained [here](https://github.com/Dao-AILab/flash-attention/issues/432#issuecomment-1668822286).
+
+    For example, if batch = 3 and seqlen = 6, the attention_mask_in_length is:
+        ```
+        [
+          [2, 3, 0, 0, 0, 0],
+          [3, 2, 0, 0, 0, 0],
+          [6, 0, 0, 0, 0, 0]
+        ]
+        ```
+    , which refers to the 3D-attention mask:
+        ```
+        [
+          [
+            [1, 0, 0, 0, 0, 0],
+            [1, 1, 0, 0, 0, 0],
+            [0, 0, 1, 0, 0, 0],
+            [0, 0, 1, 1, 0, 0],
+            [0, 0, 1, 1, 1, 0],
+            [0, 0, 0, 0, 0, 1]
+          ],
+          [
+            [1, 0, 0, 0, 0, 0],
+            [1, 1, 0, 0, 0, 0],
+            [1, 1, 1, 0, 0, 0],
+            [0, 0, 0, 1, 0, 0],
+            [0, 0, 0, 1, 1, 0],
+            [0, 0, 0, 0, 0, 1]
+          ],
+          [
+            [1, 0, 0, 0, 0, 0],
+            [1, 1, 0, 0, 0, 0],
+            [1, 1, 1, 0, 0, 0],
+            [1, 1, 1, 1, 0, 0],
+            [1, 1, 1, 1, 1, 0],
+            [1, 1, 1, 1, 1, 1]
+          ]
+        ]
+        ```.
+
+    Arguments:
+        hidden_states: (batch, seqlen, ...)
+        attention_mask_in_length: (batch, seqlen), int, a nonzero number (e.g., 1, 2, 3, etc.) means length of concatenated sequence in b-th batch, and 0 means none.
+    Return:
+        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
+        indices: (total_nnz), the indices of non-masked tokens from the flattened input sequence.
+        cu_seqlens: (batch + 1), the cumulative sequence lengths, used to index into hidden_states.
+        max_seqlen_in_batch: int
+    """
+    length = attention_mask_in_length.sum(dim=-1)
+    seqlen = attention_mask_in_length.size(-1)
+    attention_mask_2d = torch.arange(seqlen, device=length.device, dtype=length.dtype).expand(len(length),
+                                                                                              seqlen) < length.unsqueeze(
+        1)
+    real_indices_idx = torch.nonzero(attention_mask_in_length.flatten(), as_tuple=False).flatten()
+    seqlens_in_batch = attention_mask_in_length.flatten()[real_indices_idx]
+    indices = torch.nonzero(attention_mask_2d.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+    # TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the
+    # bool mask, then call nonzero to get the indices, then index with those. The indices is @dim
+    # times larger than it needs to be, wasting memory. It's faster and more memory-efficient to
+    # index with integer indices. Moreover, torch's index is a bit slower than it needs to be,
+    # so we write custom forward and backward to make it a bit faster.
+    return (
+        index_first_axis(rearrange(hidden_states, "b s ... -> (b s) ..."), indices),
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def pad_input(hidden_states, indices, batch, seqlen):
+    """
+    Arguments:
+        hidden_states: (total_nnz, ...), where total_nnz = number of tokens in selected in attention_mask.
+        indices: (total_nnz), the indices that represent the non-masked tokens of the original padded input sequence.
+        batch: int, batch size for the padded sequence.
+        seqlen: int, maximum sequence length for the padded sequence.
+    Return:
+        hidden_states: (batch, seqlen, ...)
+    """
+    dim = hidden_states.shape[-1]
+    # output = torch.zeros((batch * seqlen), dim, device=hidden_states.device, dtype=hidden_states.dtype)
+    # output[indices] = hidden_states
+    output = index_put_first_axis(hidden_states, indices, batch * seqlen)
+    return rearrange(output, "(b s) ... -> b s ...", b=batch)