init commit

Dockerfile

@@ -0,0 +1,30 @@
+# Use the specified base image
+FROM nvcr.io/nvidia/pytorch:23.12-py3
+
+# Set the working directory to your project directory
+WORKDIR ./
+
+# Copy the contents of your project into the Docker image
+COPY . .
+
+# Create and activate Conda environment
+# RUN conda create --name plm python=3.10
+# SHELL ["conda", "run", "-n", "plm", "/bin/bash", "-c"]
+# RUN conda activate plm
+
+# Install Miniconda
+# RUN wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh && \
+#     /bin/bash miniconda.sh -b -p /opt/conda && \
+#     rm miniconda.sh
+# ENV PATH="/opt/conda/bin:${PATH}"
+
+# Install dependencies
+# RUN cd protein_lm/modeling/models/libs/ && pip install -e causal-conv1d && pip install -e mamba && cd ../../../../
+# RUN pip install transformers datasets accelerate evaluate pytest fair-esm biopython deepspeed
+# RUN pip install -e .
+# RUN pip install hydra-core --upgrade
+# RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+# source "$HOME/.cargo/env"
+# RUN pip install -e protein_lm/tokenizer/rust_trie
+
+

LICENSE.md

@@ -0,0 +1,13 @@
+Copyright 2023 the authors
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.

README.md

@@ -1,18 +1,101 @@
----
-license: cc-by-nc-nd-4.0
-extra_gated_prompt: "You agree to share any computational and experimental results (along with any generated sequences) from this model's use with the authors prior to publication or dissemination (pranam.[email protected])."
-extra_gated_fields:
-  Name: text
-  Company: text
-  Country: country
-  Specific date: date_picker
-  I want to use this model for:
-    type: select
-    options: 
-      - Research
-      - Education
-      - label: Other
-        value: other
-  I agree not to file patents on any sequences generated by this model: checkbox
-  I agree to use this model for non-commercial use ONLY: checkbox
----
+# PTM-Mamba
+
+A PTM-Aware Protein Language Model with Bidirectional Gated Mamba Blocks.
+
+## Install Enviroment
+
+## Docker
+
+Setting up env for mamba could be a pain, alternatively we suggest using docker containers.
+
+#### Run container in interactive and detach mode, and mounte project dir to the container workspace.
+
+```
+docker run --gpus all -v $(pwd):/workspace -d -it --name plm_benji nvcr.io/nvidia/pytorch:23.12-py3 /bin/bash && docker attach plm_benji
+```
+
+#### Install pkgs in container
+
+```
+mkdir /root/.cache/torch/hub/checkpoints/ -p; wget -O /root/.cache/torch/hub/checkpoints/esm2_t33_650M_UR50D.pt https://dl.fbaipublicfiles.com/fair-esm/models/esm2_t33_650M_UR50D.pt
+cd protein_lm/modeling/models/libs/ && pip install -e causal-conv1d && pip install -e mamba && cd ../../../../
+pip install transformers datasets accelerate evaluate pytest fair-esm biopython deepspeed wandb
+pip install torch_geometric
+pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.0.0+cu117.html
+pip install -e .
+pip install hydra-core --upgrade
+curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
+source "$HOME/.cargo/env"
+pip install -e protein_lm/tokenizer/rust_trie
+
+
+```
+
+## Data
+
+We collect protein sequences and their PTM annotations from Uniprot-Swissprot. The PTM annotations are represented as tokens and used to replaced the corresponding amino acids. The data can be downloaded from [here](https://drive.google.com/file/d/151KUp79tgBxphoIky1-ohyuvzIS1gtNS/view?usp=drive_link). Please place the data on  `protein_lm/dataset/`.
+
+## Configs
+
+The training and testing configs are `protein_lm/configs`. We provide a basic training config at `protein_lm/configs/train/base.yaml`.
+
+## Training
+
+##### Single-GPU Training
+
+```
+python ./protein_lm/modeling/scripts/train.py +train=base 
+```
+
+The commond will use the configs in `protein_lm/configs/train/base.yaml`.
+
+##### Multi-GPU Training
+
+We use [Distributed training with 🤗 Accelerate (huggingface.co)](https://huggingface.co/docs/transformers/main/accelerate).
+
+```
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch --num_processes=8 --multi_gpu protein_lm/modeling/scripts/train.py +train=base   train.report_to='wandb' train.training_arguments.per_device_train_batch_size=256 train.training_arguments.use_esm=True train.training_arguments.save_dir='ckpt/ptm_mamba' train.model.model_type='bidirectional_mamba' train.training_arguments.max_tokens_per_batch=40000 
+```
+
+- `report_to='wandb'`  tracks the training using wandb.
+- `training_arguments.per_device_train_batch_size=300` sets the max batch size per device when constructing a batch.
+- `training_arguments.max_tokens_per_batch=80000` sets the max num of tokens within a batch. If a batch exceeds the max token limit(depending on the seq len), we will trim the batch. Tune the `per_device_train_batch_size` and ``max_tokens_per_batch`` togather to maximize the memory usage during training. The rule of thumb is setting a large batch size (e.g., 300) while searching for the max num token that fits your GPU memory.
+- `training_arguments.use_esm=True` uses the ESM embedding. By default, we use ESM 650M, and set the `model.esm_embed_dim: 1280` in `base.yaml`.  If disabled, the model will use its own embeddings.
+- `training_arguments.save_dir='ckpt/bi_directional_mamba-esm'` where the model ckpts will be saved.
+- `training_arguments.sample_len_ascending=true` is enable by default, samples sequences from short to long during the training.
+
+##### Multi-GPU training with Deepspeed
+
+Setup deepspeed with
+
+```
+accelerate config
+```
+
+and answer the questions asked. It will ask whether you want to use a config file for DeepSpeed to which you should answer no. Then answer the following questions to generate a basic DeepSpeed config. Use ZeRo 2 and FP32, which are sufficient for training our ~300M model without introducing overhead. This will generate a config file that will be used automatically to properly set the default options when launching training.
+
+## Inference
+
+The inference example is at `protein_lm/modeling/scripts/infer.py.` The model checkpoints can be downloaded from [here](https://drive.google.com/file/d/1x_rKff0xswWU7_ixKYvZvWYZPR23cd8x/view?usp=sharing). The outputs are:
+
+```
+Output = namedtuple("output", ["logits", "hidden_states"])
+```
+
+```
+from protein_lm.modeling.scripts.infer import PTMMamba
+
+ckpt_path = "ckpt/bi_mamba-esm-ptm_token_input/best.ckpt"
+mamba = PTMMamba(ckpt_path,device='cuda:0')
+seq = "M<N-acetylalanine>K"
+output = mamba(seq)
+print(output.logits.shape)
+print(output.hidden_states.shape)
+```
+
+## Ackonwledgement
+
+This project is based on the  following codebase. Please give them a star if you like our code.
+
+- [OpenBioML/protein-lm-scaling (github.com)](https://github.com/OpenBioML/protein-lm-scaling)
+- [state-spaces/mamba (github.com)](https://github.com/state-spaces/mamba)

protein_lm/configs/train/base.yaml

@@ -0,0 +1,43 @@
+seed: 42
+report_to: "none"
+
+
+dataset:
+  dataset: "ptm"
+  dataset_type: "csv"
+  dataset_loc: "protein_lm/dataset/ptm_labels.csv"
+  subsample_size: null
+  split_seed: 2
+  val_size: 100
+  test_size: 0
+  sequence_column_name: "ori_seq"
+  max_sequence_length: 1024
+  cache_dir: "protein_lm/dataset/cache/ptm"
+
+
+training_arguments:
+  save_dir: "checkpoints/ptm-mamba"
+  num_train_epochs: 3000
+  lr: 2.0e-4
+  per_device_train_batch_size: 256 # 64 if 1 GPU
+  resume_from_checkpoint: null
+  use_esm: true
+  max_tokens_per_batch: 80000 # 80G GPU
+  sort_by_seq: true
+  sample_len_ascending: true
+  log_steps: 100
+  max_sequence_length: null
+  
+model: # default mamba 130M
+  model_type: "mamba" # ["mamba", "bidirectional_mamba"]
+  d_model: 768 
+  n_layer: 24
+  vocab_size: null # will be set by tokenizer
+  ssm_cfg: {}
+  rms_norm: true
+  residual_in_fp32: true
+  fused_add_norm: true
+  
+  esm_embed_dim: 1280
+  pretrained_checkpoint: null
+

protein_lm/modeling/getters/collate.py

@@ -0,0 +1,84 @@
+import torch
+from torch.nn.utils.rnn import pad_sequence
+from torch.utils.data import Sampler
+from omegaconf import DictConfig, OmegaConf
+from functools import partial
+
+
+def crop_seq(input_ids, max_seq_len):
+    """
+    randomly crop sequences to max_seq_len
+    Args:
+        input_ids: tensor of shape (seq_len)
+        max_seq_len: int
+    """
+    seq_len = len(input_ids)
+    if seq_len <= max_seq_len:
+        return input_ids
+    else:
+        start_idx = torch.randint(0, seq_len - max_seq_len + 1, (1,)).item()
+        return input_ids[start_idx : start_idx + max_seq_len]
+
+class DataCollatorWithPadding:
+    def __init__(
+        self,
+        max_tokens,
+        tokenizer,
+        batch_by_tokens=False,
+        max_seq_len=None,
+    ) -> None:
+        self.max_tokens = max_tokens
+        self.tokenizer = tokenizer
+        self.batch_by_tokens = batch_by_tokens
+        self.max_seq_len = max_seq_len
+        self.crop_fn = partial(crop_seq, max_seq_len=max_seq_len) if max_seq_len is not None else lambda x: x
+        
+    def __call__(self, batch):
+        """
+        generate a batch of data g
+        Args:
+        batch: list of dictionaries with keys 'input_ids' and 'labels'
+
+        """
+        input_ids = [self.crop_fn(i["input_ids"]) for i in batch]
+        input_ids = pad_sequence(
+            [torch.tensor(x) for x in input_ids],
+            batch_first=True,
+            padding_value=self.tokenizer.pad_token_id,
+        )
+        if self.batch_by_tokens:
+            # keep a few sequences to make the total number of tokens in the batch <= max_tokens
+            total_tokens = input_ids.numel()
+            if total_tokens > self.max_tokens:
+                max_num_seq = self.max_tokens // input_ids.shape[-1] + 1
+                # randomly select max_num_seq sequences from the batch to keep
+                indices = torch.randperm(len(input_ids))[:max_num_seq]
+                input_ids = input_ids[indices]
+        pad_mask = input_ids != self.tokenizer.pad_token_id
+
+        return {
+            "input_ids": input_ids,
+            "pad_mask": pad_mask,
+        }
+
+
+class SequenceLengthSampler(Sampler):
+    def __init__(self, dataset, sort=True, sample_len_ascending=True):
+        """
+        Args:
+            dataset: a dataset with keys 'input_ids' and 'labels'
+            sample_len_ascending: if True, sample shorter sequences first
+        """
+        self.dataset = dataset
+        self.indices = list(range(len(dataset)))
+        if sort is True:
+            self.indices.sort(
+                key=lambda x: len(dataset[x]["input_ids"]),
+                reverse=not sample_len_ascending
+            )
+
+    def __iter__(self):
+        return iter(self.indices)
+
+    def __len__(self):
+        return len(self.indices)

protein_lm/modeling/getters/dataset.py

@@ -0,0 +1,16 @@
+import pandas as pd
+from typing import Dict, Literal, Optional
+from datasets import Dataset, load_dataset
+from datasets.dataset_dict import DatasetDict
+from pydantic import BaseModel
+from protein_lm.modeling.getters.ptm_dataset import get_ptm_dataset
+from protein_lm.modeling.getters.uniref_dataset import get_uniref_dataset
+
+
+def get_dataset(config_dict: Dict, tokenizer) -> Dataset:
+    if config_dict["dataset"] == "ptm":
+        return get_ptm_dataset(config_dict, tokenizer)
+    elif config_dict["dataset"] == "uniref50":
+        return get_uniref_dataset(config_dict, tokenizer)
+    else:
+        raise ValueError(f"Invalid dataset {config_dict['dataset']}!")

protein_lm/modeling/getters/log.py

@@ -0,0 +1,12 @@
+import logging
+
+
+class TrainLogger:
+    def __init__(self, log_file="logfile.log"):
+        logging.basicConfig(
+            filename=log_file, level=logging.INFO, format="%(asctime)s - %(message)s"
+        )
+
+    def log(self, data):
+        data = {k: round(v, 4) if isinstance(v, float) else v for k, v in data.items()}
+        logging.info(data)

protein_lm/modeling/getters/mask.py

@@ -0,0 +1,54 @@
+import torch
+
+
+class Masker:
+    def __init__(self, tokenizer) -> None:
+        self.tokenizer = tokenizer
+        self.mask_token_id = self.tokenizer.mask_token_id
+
+    def random_mask(self, input_ids, mask_prob=0.15):
+        device = input_ids.device
+        mask = (torch.rand(input_ids.shape) < mask_prob).to(device)
+        mask = mask & (torch.logical_not(self.tokenizer.is_special_token(input_ids)))
+        masked_input_ids = input_ids.clone()
+        masked_input_ids[mask] = self.mask_token_id
+        return masked_input_ids, mask
+
+    def mask_ptm_tokens(
+        self,
+        input_ids,
+    ):
+        device = input_ids.device
+        is_ptm_mask = self.tokenizer.is_ptm_token(input_ids).to(device)
+        is_ptm_mask = is_ptm_mask & (
+            torch.logical_not(self.tokenizer.is_special_token(input_ids))
+        )
+        masked_input_ids = input_ids.clone()
+        masked_input_ids[is_ptm_mask] = self.mask_token_id
+        return masked_input_ids, is_ptm_mask
+
+    def random_and_ptm_mask(self, input_ids, mask_prob=0.15):
+        device = input_ids.device
+        mask = (torch.rand(input_ids.shape) < mask_prob).to(device)
+        mask = mask & (torch.logical_not(self.tokenizer.is_special_token(input_ids)))
+        is_ptm_mask = self.tokenizer.is_ptm_token(input_ids).to(device)
+        is_ptm_mask = is_ptm_mask & (
+            torch.logical_not(self.tokenizer.is_special_token(input_ids))
+        )
+        mask = mask | is_ptm_mask
+        masked_input_ids = input_ids.clone()
+        masked_input_ids[mask] = self.mask_token_id
+        return masked_input_ids, mask
+
+    def random_or_random_and_ptm_mask(
+        self, input_ids, ranom_mask_prob=0.15, alternate_prob=0.2
+    ):
+        """
+        alternate between [(1) random mask] and [(2) random mask & ptm mask] by probability alternate_prob
+        """
+        p = torch.rand(1).item()
+        if p < alternate_prob:
+            return self.random_mask(input_ids, ranom_mask_prob)
+        else:
+            return self.random_and_ptm_mask(input_ids, ranom_mask_prob)
+

protein_lm/modeling/getters/ptm_dataset.py

@@ -0,0 +1,212 @@
+import pandas as pd
+from typing import Dict, Literal, Optional
+from datasets import Dataset, load_dataset
+from datasets.dataset_dict import DatasetDict
+from pydantic import BaseModel
+
+
+
+
+class DatasetConfig(BaseModel):
+    dataset: Literal['ptm', 'uniref50']
+    dataset_type: Literal["csv", "huggingface", "fasta"]
+
+    # The path if local or the huggingface dataset name if huggingface
+    dataset_loc: str
+
+    # sample size to limit to, if any, usually for debugging
+    subsample_size: Optional[int] = None
+
+    """
+    Args for splitting into train, val, test
+    to be updated once we have more options
+    """
+    # split seed
+    split_seed: Optional[int] = None
+    # size of validation dataset
+    val_size: int
+    # size of test dataset
+    test_size: int
+
+    # name of the column that contains the sequence
+    sequence_column_name: str
+
+    max_sequence_length: Optional[int] = None
+    cache_dir: Optional[str] = None
+
+
+def set_labels(result):
+    result["labels"] = result["input_ids"].copy()
+    return result
+
+
+def train_val_test_split(
+    dataset_dict: DatasetDict,
+    config: DatasetConfig,
+) -> DatasetDict:
+    """
+    Given a dictionary of datasets that only contains the split "train",
+    optionally subsamples it, and then splits it
+    so that it has potentially 3 splits: "train", "val", "test", where
+    "val" and "test" splits do not exist if the specified sizes are 0
+    """
+    assert set(dataset_dict.keys()) == {
+        "train"
+    }, f"{train_val_test_split.__name__} expects its input to have the keys \
+        ['train'] but the input has keys {list(dataset_dict.keys())}"
+
+    dataset = dataset_dict["train"]
+
+    val_size = config.val_size
+    test_size = config.test_size
+
+    assert isinstance(
+        dataset, Dataset
+    ), f"Invalid dataset type {type(dataset)}, only datasets.Dataset allowed"
+
+    dataset = dataset.shuffle(seed=config.split_seed)
+
+    if config.subsample_size is not None:
+        dataset = dataset.select(range(config.subsample_size))
+
+    valtest_size = val_size + test_size
+
+    if valtest_size > 0:
+        train_valtest = dataset.train_test_split(
+            test_size=val_size + test_size,
+            shuffle=False,
+        )
+        split_dict = {
+            "train": train_valtest["train"],
+        }
+        if test_size > 0 and val_size > 0:
+            test_val = train_valtest["test"].train_test_split(
+                test_size=test_size,
+                shuffle=False,
+            )
+            split_dict["val"] = test_val["train"]
+            split_dict["test"] = test_val["test"]
+        elif val_size > 0:
+            split_dict["val"] = train_valtest["test"]
+        else:
+            split_dict["test"] = train_valtest["test"]
+    else:
+        split_dict = {
+            "train": dataset,
+        }
+
+    split_dataset_dict = DatasetDict(split_dict)
+    return split_dataset_dict
+
+
+
+
+def load_ptm_dataset(df: pd.DataFrame, config: DatasetConfig) -> DatasetDict:
+    ds = Dataset.from_pandas(df)
+    ds_dict = DatasetDict({"train": ds})
+    return train_val_test_split(ds_dict, config)
+
+
+def create_token_dict_from_dataframe(
+    df, seq_col="ori_seq", pos_col="pos", token_col="token"
+):
+    result_dict = {}
+
+    for index, row in df.iterrows():
+        ac_id = row[seq_col]
+        pos = row[pos_col]
+        token = row[token_col]
+
+        if ac_id not in result_dict:
+            result_dict[ac_id] = {}
+
+        result_dict[ac_id][pos] = token
+
+    return result_dict
+
+
+def subsitute_tokens(sequence_lst, token_dict):
+    if isinstance(sequence_lst, list):
+        return [
+            _substitute_token(sequence, token_dict[sequence])
+            for sequence in sequence_lst
+        ]
+    elif isinstance(sequence_lst, str):
+        return _substitute_token(sequence_lst, token_dict[sequence_lst])
+
+
+def _substitute_token(sequence, token_dict):
+    result = list(sequence)
+    for position, new_tokens in token_dict.items():
+        result[position] = new_tokens
+    return "".join(result)
+
+
+def construct_ptm_seq(
+    batch, tokenizer, ptm_token_dict, sequence_column_name, max_sequence_length
+):
+    """
+    apply transform to the batch to replace the tokens with the PTM tokens
+    """
+    batch['wt_seq'] = batch[sequence_column_name]
+    batch[sequence_column_name] = subsitute_tokens(
+        batch[sequence_column_name], ptm_token_dict
+    )
+    batch['ptm_seq'] = batch[sequence_column_name]
+    batch["input_ids"] = tokenizer(
+        batch[sequence_column_name],
+        add_special_tokens=True,
+        max_sequence_length=max_sequence_length,
+    )
+    # batch["labels"] = batch["input_ids"]
+    return batch
+
+
+def get_ptm_dataset(config_dict: Dict, tokenizer) -> Dataset:
+    config = DatasetConfig(**config_dict)
+
+    if config.dataset_type == "csv":
+        df = pd.read_csv(config.dataset_loc)
+        ptm_token_dict = create_token_dict_from_dataframe(df)
+        df.drop(df.filter(regex="Unnamed"), axis=1, inplace=True)
+        df.drop_duplicates(subset=config.sequence_column_name, inplace=True)
+        split_dict = load_ptm_dataset(df, config)
+    else:
+        raise ValueError(f"Invalid dataset_type {config.dataset_type}!")
+
+    split_dict = split_dict.map(
+        lambda e: construct_ptm_seq(
+            batch=e,
+            tokenizer=tokenizer,
+            ptm_token_dict=ptm_token_dict,
+            sequence_column_name=config.sequence_column_name,
+            max_sequence_length=config.max_sequence_length,
+        ),
+        batched=True,
+        keep_in_memory=True,
+    )
+
+    return split_dict
+
+
+
+if __name__ == "__main__":
+    from protein_lm.tokenizer.tokenizer import PTMTokenizer
+    from transformers import DataCollatorWithPadding, default_data_collator
+
+    tokenizer = PTMTokenizer()
+    data_collator = DataCollatorWithPadding(
+        tokenizer=tokenizer, padding="longest", max_length=1024, return_tensors="pt"
+    )
+    config_dict = {
+        "dataset_type": "csv",
+        "dataset_loc": "protein_lm/dataset/ptm_labels.csv",
+        "subsample_size": None,
+        "val_size": 0,
+        "test_size": 0,
+        "sequence_column_name": "ori_seq",
+        "max_sequence_length": None,
+        "dataset": "ptm",
+    }
+    dataset = get_ptm_dataset(config_dict, tokenizer)
+    samples = dataset["train"][:8]

protein_lm/modeling/getters/scheduler.py

@@ -0,0 +1,198 @@
+import math
+
+from torch.optim.lr_scheduler import _LRScheduler, CosineAnnealingLR
+
+
+class ConstantLRScheduler(_LRScheduler):
+    def __init__(
+        self,
+        optimizer,
+        last_epoch: int = -1,
+        verbose: bool = False,
+        init_lr: float = 0.0,
+    ):
+        """
+        This is an implementation of constant learning rate scheduler.
+        Args:
+            optimizer: Optimizer
+
+            last_epoch: The index of last epoch. Default: -1
+
+            verbose: If ``True``, prints a message to stdout for each update. Default: ``False``
+
+            init_lr: Initial learning rate
+        """
+
+        self.init_lr = init_lr
+        super().__init__(optimizer, last_epoch, verbose)
+
+    def state_dict(self):
+        state_dict = {k: v for k, v in self.__dict__.items() if k not in ["optimizer"]}
+        return state_dict
+
+    def load_state_dict(self, state_dict):
+        self.__dict__.update(state_dict)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            raise RuntimeError(
+                "To get the last learning rate computed by the scheduler, use "
+                "get_last_lr()"
+            )
+
+        return [self.init_lr for group in self.optimizer.param_groups]
+
+
+class CosineAnnealingLRScheduler(_LRScheduler):
+    def __init__(
+        self,
+        optimizer,
+        last_epoch: int = -1,
+        verbose: bool = False,
+        init_lr: float = 4.0e-5,
+        max_lr: float = 4e-4,
+        final_lr: float = 4e-5,
+        warmup_steps: int = 2000,
+        cosine_steps: int = 10000,
+    ):
+        """
+        This is an implementation of cosine annealing learning rate scheduler.
+        Args:
+            optimizer: Optimizer
+
+            last_epoch: The index of last epoch. Default: -1
+
+            verbose: If ``True``, prints a message to stdout for each update. Default: ``False``
+
+            init_lr: Initial learning rate
+
+            max_lr: Maximum learning rate after warmup
+
+            final_lr: Final learning rate after decay
+
+            warmup_steps: Number of steps for warmup
+
+            cosine_steps: Number of steps for cosine annealing
+        """
+
+        self.init_lr = init_lr
+        self.max_lr = max_lr
+        self.final_lr = final_lr
+        self.warmup_steps = warmup_steps
+        self.cosine_steps = cosine_steps
+        super(CosineAnnealingLRScheduler, self).__init__(optimizer, last_epoch, verbose)
+
+    def state_dict(self):
+        state_dict = {k: v for k, v in self.__dict__.items() if k not in ["optimizer"]}
+        return state_dict
+
+    def load_state_dict(self, state_dict):
+        self.__dict__.update(state_dict)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            raise RuntimeError(
+                "To get the last learning rate computed by the scheduler, use "
+                "get_last_lr()"
+            )
+
+        step_no = self.last_epoch
+
+        if step_no <= self.warmup_steps:
+            lr = self.init_lr + step_no / self.warmup_steps * (
+                self.max_lr - self.init_lr
+            )
+
+        else:
+            lr = self.final_lr + 0.5 * (self.max_lr - self.final_lr) * (
+                1
+                + math.cos(math.pi * (step_no - self.warmup_steps) / self.cosine_steps)
+            )
+
+        return [lr for group in self.optimizer.param_groups]
+
+
+class Esm2LRScheduler(_LRScheduler):
+    def __init__(
+        self,
+        optimizer,
+        last_epoch: int = -1,
+        verbose: bool = False,
+        init_lr: float = 4e-5,
+        max_lr: float = 4e-4,
+        final_lr: float = 4e-5,
+        warmup_steps: int = 2000,
+        start_decay_after_n_steps: int = 500000,
+        end_decay_after_n_steps: int = 5000000,
+        on_use: bool = True,
+    ):
+        """
+        An implementation of ESM2's learning rate scheduler.
+        Args:
+            optimizer: Optimizer
+
+            last_epoch: The index of last epoch. Default: -1
+
+            verbose: If ``True``, prints a message to stdout for each update. Default: ``False``
+
+            init_lr: Initial learning rate
+
+            max_lr: Maximum learning rate after warmup
+
+            final_lr: Final learning rate after decay
+
+            warmup_steps: Number of steps for warmup
+
+            start_decay_after_n_steps: Start decay after this number of steps
+
+            end_decay_after_n_steps: End decay after this number of steps
+
+            on_use: Whether to use this scheduler. If ``False``, the scheduler will not change the learning rate
+            and will only use the ``init_lr``. Default: ``True``
+        """
+
+        self.init_lr = init_lr
+        self.max_lr = max_lr
+        self.final_lr = final_lr
+        self.warmup_steps = warmup_steps
+        self.start_decay_after_n_steps = start_decay_after_n_steps
+        self.end_decay_after_n_steps = end_decay_after_n_steps
+        self.on_use = on_use
+        super(Esm2LRScheduler, self).__init__(optimizer, last_epoch, verbose)
+
+    def state_dict(self):
+        state_dict = {k: v for k, v in self.__dict__.items() if k not in ["optimizer"]}
+        return state_dict
+
+    def load_state_dict(self, state_dict):
+        self.__dict__.update(state_dict)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            raise RuntimeError(
+                "To get the last learning rate computed by the scheduler, use "
+                "get_last_lr()"
+            )
+
+        step_no = self.last_epoch
+        if not self.on_use:
+            return [base_lr for base_lr in self.base_lrs]
+
+        if step_no <= self.warmup_steps:
+            lr = self.init_lr + step_no / self.warmup_steps * (
+                self.max_lr - self.init_lr
+            )
+
+        elif step_no <= self.start_decay_after_n_steps:
+            lr = self.max_lr
+
+        elif step_no <= self.end_decay_after_n_steps:
+            portion = (step_no - self.start_decay_after_n_steps) / (
+                self.end_decay_after_n_steps - self.start_decay_after_n_steps
+            )
+            lr = self.max_lr - portion * (self.max_lr - self.final_lr)
+
+        else:
+            lr = self.final_lr
+
+        return [lr for group in self.optimizer.param_groups]

protein_lm/modeling/getters/training_args.py

@@ -0,0 +1,14 @@
+import os
+from typing import Dict
+
+from transformers import TrainingArguments
+
+
+def get_training_args(config_dict: Dict) -> TrainingArguments:
+    config = TrainingArguments(**config_dict)
+
+    if not os.path.isdir(config.output_dir):
+        print(f"creating checkpoint directory at {config.output_dir}")
+        os.makedirs(config.output_dir)
+
+    return config

protein_lm/modeling/getters/uniref_dataset.py

@@ -0,0 +1,59 @@
+import os
+import pandas as pd
+
+from Bio import SeqIO
+from typing import Dict, Literal, Optional
+from datasets import Dataset, load_dataset
+from datasets.dataset_dict import DatasetDict
+from typing import Dict, Literal, Optional
+from protein_lm.modeling.getters.ptm_dataset import DatasetConfig, train_val_test_split
+
+
+def read_fasta_file(fasta_file_path, subsample_size):
+    ids = []
+    seqs = []
+    with open(fasta_file_path, "r") as fasta_file:
+        for i, record in enumerate(SeqIO.parse(fasta_file, "fasta")):
+            if subsample_size and i >= subsample_size:
+                break
+            ids.append(record.id)
+            seqs.append(str(record.seq))
+
+    return {"id": ids, "seq": seqs}
+
+
+def load_uniref_dataset(seq_dict, config) -> DatasetDict:
+    ds = Dataset.from_dict(seq_dict)
+    ds_dict = DatasetDict({"train": ds})
+    return train_val_test_split(ds_dict, config)
+
+
+def seq2token(batch, tokenizer, sequence_column_name, max_sequence_length):
+    batch["input_ids"] = tokenizer(
+        batch[sequence_column_name],
+        add_special_tokens=True,
+        max_sequence_length=max_sequence_length,
+    )
+    return batch
+
+
+def get_uniref_dataset(config: Dict, tokenizer) -> Dataset:
+    # config = DatasetConfig(**config_dict)
+    if config.cache_dir is not None and os.path.exists(config.cache_dir):
+        split_dict = DatasetDict.load_from_disk(config.cache_dir)
+        return split_dict
+    seq_dict = read_fasta_file(config.dataset_loc, config.subsample_size)
+    split_dict = load_uniref_dataset(seq_dict, config)
+    split_dict = split_dict.map(
+        lambda e: seq2token(
+            batch=e,
+            tokenizer=tokenizer,
+            sequence_column_name="seq",
+            max_sequence_length=config.max_sequence_length,
+        ),
+        batched=True,
+    )
+    if config.cache_dir is not None:
+        os.makedirs(config.cache_dir, exist_ok=True)
+        split_dict.save_to_disk(config.cache_dir)
+    return split_dict

protein_lm/modeling/models/libs/causal-conv1d/AUTHORS

@@ -0,0 +1 @@
+Tri Dao, [email protected]

protein_lm/modeling/models/libs/causal-conv1d/LICENSE

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2022, the respective contributors, as shown by the AUTHORS file.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions and the following disclaimer in the documentation
+  and/or other materials provided with the distribution.
+
+* Neither the name of the copyright holder nor the names of its
+  contributors may be used to endorse or promote products derived from
+  this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

protein_lm/modeling/models/libs/causal-conv1d/README.md

@@ -0,0 +1 @@
+# Causal depthwise conv1d in CUDA with a PyTorch interface

protein_lm/modeling/models/libs/causal-conv1d/causal_conv1d/__init__.py

@@ -0,0 +1,3 @@
+__version__ = "1.0.0"
+
+from causal_conv1d.causal_conv1d_interface import causal_conv1d_fn, causal_conv1d_update

protein_lm/modeling/models/libs/causal-conv1d/causal_conv1d/causal_conv1d_interface.py

@@ -0,0 +1,104 @@
+# Copyright (c) 2023, Tri Dao.
+
+import torch
+import torch.nn.functional as F
+
+
+import causal_conv1d_cuda
+
+
+class CausalConv1dFn(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, weight, bias=None, activation=None):
+        if activation not in [None, "silu", "swish"]:
+            raise NotImplementedError("activation must be None, silu, or swish")
+        if x.stride(2) != 1 and x.stride(1) != 1:
+            x = x.contiguous()
+        bias = bias.contiguous() if bias is not None else None
+        ctx.save_for_backward(x, weight, bias)
+        ctx.activation = activation in ["silu", "swish"]
+        out = causal_conv1d_cuda.causal_conv1d_fwd(x, weight, bias, ctx.activation)
+        return out
+
+    @staticmethod
+    def backward(ctx, dout):
+        x, weight, bias = ctx.saved_tensors
+        if dout.stride(2) != 1 and dout.stride(1) != 1:
+            dout = dout.contiguous()
+        # The kernel supports passing in a pre-allocated dx (e.g., in case we want to fuse the
+        # backward of conv1d with the backward of chunk).
+        # Here we just pass in None and dx will be allocated in the C++ code.
+        dx, dweight, dbias = causal_conv1d_cuda.causal_conv1d_bwd(
+            x, weight, bias, dout, None, ctx.activation
+        )
+        return dx, dweight, dbias if bias is not None else None, None
+
+
+def causal_conv1d_fn(x, weight, bias=None, activation=None):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+    activation: either None or "silu" or "swish"
+
+    out: (batch, dim, seqlen)
+    """
+    return CausalConv1dFn.apply(x, weight, bias, activation)
+
+
+def causal_conv1d_ref(x, weight, bias=None, activation=None):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+
+    out: (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    x = x.to(weight.dtype)
+    seqlen = x.shape[-1]
+    dim, width = weight.shape
+    out = F.conv1d(x, weight.unsqueeze(1), bias, padding=width - 1, groups=dim)
+    out = out[..., :seqlen]
+    return (out if activation is None else F.silu(out)).to(dtype=dtype_in)
+
+
+def causal_conv1d_update(x, conv_state, weight, bias=None, activation=None):
+    """
+    x: (batch, dim)
+    conv_state: (batch, dim, width)
+    weight: (dim, width)
+    bias: (dim,)
+
+    out: (batch, dim)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    activation = activation in ["silu", "swish"]
+    return causal_conv1d_cuda.causal_conv1d_update(x, conv_state, weight, bias, activation)
+
+
+def causal_conv1d_update_ref(x, conv_state, weight, bias=None, activation=None):
+    """
+    x: (batch, dim)
+    conv_state: (batch, dim, width)
+    weight: (dim, width)
+    bias: (dim,)
+
+    out: (batch, dim)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    batch, dim = x.shape
+    width = weight.shape[1]
+    assert conv_state.shape == (batch, dim, width)
+    assert weight.shape == (dim, width)
+    conv_state.copy_(torch.roll(conv_state, shifts=-1, dims=-1)) # Update state (B D W)
+    conv_state[:, :, -1] = x
+    out = torch.sum(conv_state * weight, dim=-1) # (B D)
+    if bias is not None:
+        out += bias
+    return (out if activation is None else F.silu(out)).to(dtype=dtype_in)

protein_lm/modeling/models/libs/causal-conv1d/csrc/causal_conv1d.cpp

@@ -0,0 +1,333 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/extension.h>
+#include <vector>
+
+#include "causal_conv1d.h"
+
+#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")
+
+#define DISPATCH_ITYPE_FLOAT_AND_HALF_AND_BF16(ITYPE, NAME, ...)                    \
+    if (ITYPE == at::ScalarType::Half) {                                            \
+        using input_t = at::Half;                                                   \
+        __VA_ARGS__();                                                              \
+    } else if (ITYPE == at::ScalarType::BFloat16) {                                 \
+        using input_t = at::BFloat16;                                               \
+        __VA_ARGS__();                                                              \
+    } else if (ITYPE == at::ScalarType::Float)  {                                   \
+        using input_t = float;                                                      \
+        __VA_ARGS__();                                                              \
+    } else {                                                                        \
+        AT_ERROR(#NAME, " not implemented for input type '", toString(ITYPE), "'"); \
+    }
+
+#define DISPATCH_WTYPE_FLOAT_AND_HALF_AND_BF16(WTYPE, NAME, ...)                     \
+    if (WTYPE == at::ScalarType::Half) {                                             \
+        using weight_t = at::Half;                                                   \
+        __VA_ARGS__();                                                               \
+    } else if (WTYPE == at::ScalarType::BFloat16) {                                  \
+        using weight_t = at::BFloat16;                                               \
+        __VA_ARGS__();                                                               \
+    } else if (WTYPE == at::ScalarType::Float)  {                                    \
+        using weight_t = float;                                                      \
+        __VA_ARGS__();                                                               \
+    } else {                                                                         \
+        AT_ERROR(#NAME, " not implemented for weight type '", toString(WTYPE), "'"); \
+    }
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_fwd_cuda(ConvParamsBase &params, cudaStream_t stream);
+template <typename input_t, typename weight_t>
+void causal_conv1d_channellast_fwd_cuda(ConvParamsBase &params, cudaStream_t stream);
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_bwd_cuda(ConvParamsBwd &params, cudaStream_t stream);
+template<typename input_t, typename weight_t>
+void causal_conv1d_channellast_bwd_cuda(ConvParamsBwd &params, cudaStream_t stream);
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_update_cuda(ConvParamsBase &params, cudaStream_t stream);
+
+void set_conv_params_fwd(ConvParamsBase &params,
+                         // sizes
+                         const size_t batch,
+                         const size_t dim,
+                         const size_t seqlen,
+                         const size_t width,
+                         // device pointers
+                         const at::Tensor x,
+                         const at::Tensor weight,
+                         const at::Tensor out,
+                         void* bias_ptr,
+                         bool silu_activation) {
+
+    // Reset the parameters
+    memset(&params, 0, sizeof(params));
+
+    params.batch = batch;
+    params.dim = dim;
+    params.seqlen = seqlen;
+    params.width = width;
+
+    params.silu_activation = silu_activation;
+
+    // Set the pointers and strides.
+    params.x_ptr = x.data_ptr();
+    params.weight_ptr = weight.data_ptr();
+    params.bias_ptr = bias_ptr;
+    params.out_ptr = out.data_ptr();
+    // All stride are in elements, not bytes.
+    params.x_batch_stride = x.stride(0);
+    params.x_c_stride = x.stride(1);
+    params.x_l_stride = x.stride(-1);
+    params.weight_c_stride = weight.stride(0);
+    params.weight_width_stride = weight.stride(1);
+    params.out_batch_stride = out.stride(0);
+    params.out_c_stride = out.stride(1);
+    params.out_l_stride = out.stride(-1);
+}
+
+
+void set_conv_params_bwd(ConvParamsBwd &params,
+                         // sizes
+                         const size_t batch,
+                         const size_t dim,
+                         const size_t seqlen,
+                         const size_t width,
+                         // device pointers
+                         const at::Tensor x,
+                         const at::Tensor weight,
+                         void* bias_ptr,
+                         const at::Tensor dout,
+                         const at::Tensor dx,
+                         const at::Tensor dweight,
+                         void* dbias_ptr,
+                         bool silu_activation) {
+    // Pass in "dout" instead of "out", we're not gonna use "out" at all.
+    set_conv_params_fwd(params, batch, dim, seqlen, width,
+                        x, weight, dout, bias_ptr, silu_activation);
+
+    // Set the pointers and strides.
+    params.dout_ptr = dout.data_ptr();
+    params.dx_ptr = dx.data_ptr();
+    params.dweight_ptr = dweight.data_ptr();
+    params.dbias_ptr = dbias_ptr;
+    // All stride are in elements, not bytes.
+    params.dout_batch_stride = dout.stride(0);
+    params.dout_c_stride = dout.stride(1);
+    params.dout_l_stride = dout.stride(2);
+    params.dweight_c_stride = dweight.stride(0);
+    params.dweight_width_stride = dweight.stride(1);
+    params.dx_batch_stride = dx.stride(0);
+    params.dx_c_stride = dx.stride(1);
+    params.dx_l_stride = dx.stride(2);
+}
+
+at::Tensor
+causal_conv1d_fwd(const at::Tensor &x, const at::Tensor &weight,
+                  const c10::optional<at::Tensor> &bias_,
+                  bool silu_activation) {
+    auto input_type = x.scalar_type();
+    auto weight_type = weight.scalar_type();
+    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::Half || weight_type == at::ScalarType::BFloat16);
+
+    TORCH_CHECK(x.is_cuda());
+    TORCH_CHECK(weight.is_cuda());
+
+    const auto sizes = x.sizes();
+    const int batch_size = sizes[0];
+    const int dim = sizes[1];
+    const int seqlen = sizes[2];
+    const int width = weight.size(-1);
+
+    CHECK_SHAPE(x, batch_size, dim, seqlen);
+    CHECK_SHAPE(weight, dim, width);
+
+    TORCH_CHECK(x.stride(2) == 1 || x.stride(1) == 1);
+    const bool is_channel_last = x.stride(1) == 1 && x.stride(2) > 1;
+
+    if (is_channel_last) {
+        TORCH_CHECK(dim % 8 == 0, "causal_conv1d only supports channel dimension divisible by 8 for now");
+    }
+    TORCH_CHECK(width >= 2 && width <= 4, "causal_conv1d only supports width between 2 and 4");
+
+
+    if (bias_.has_value()) {
+        auto bias = bias_.value();
+        TORCH_CHECK(bias.scalar_type() == weight_type);
+        TORCH_CHECK(bias.is_cuda());
+        TORCH_CHECK(bias.stride(-1) == 1);
+        CHECK_SHAPE(bias, dim);
+    }
+
+    at::Tensor out = torch::empty_like(x);
+
+    ConvParamsBase params;
+    set_conv_params_fwd(params, batch_size, dim, seqlen, width, x, weight, out,
+                        bias_.has_value() ? bias_.value().data_ptr() : nullptr,
+                        silu_activation);
+
+    // Otherwise the kernel will be launched from cuda:0 device
+    // Cast to char to avoid compiler warning about narrowing
+    at::cuda::CUDAGuard device_guard{(char)x.get_device()};
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    DISPATCH_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_fwd", [&] {
+        DISPATCH_WTYPE_FLOAT_AND_HALF_AND_BF16(weight.scalar_type(), "causal_conv1d_fwd", [&] {
+            if (!is_channel_last) {
+                causal_conv1d_fwd_cuda<input_t, weight_t>(params, stream);
+            } else {
+                causal_conv1d_channellast_fwd_cuda<input_t, weight_t>(params, stream);
+            }
+        });
+    });
+    return out;
+}
+
+std::vector<at::Tensor>
+causal_conv1d_bwd(const at::Tensor &x, const at::Tensor &weight,
+                  const c10::optional<at::Tensor> &bias_,
+                  at::Tensor &dout,
+                  c10::optional<at::Tensor> &dx_,
+                  bool silu_activation) {
+    auto input_type = x.scalar_type();
+    auto weight_type = weight.scalar_type();
+    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::Half || weight_type == at::ScalarType::BFloat16);
+
+    TORCH_CHECK(x.is_cuda());
+    TORCH_CHECK(weight.is_cuda());
+    TORCH_CHECK(dout.is_cuda());
+
+    const auto sizes = x.sizes();
+    const int batch_size = sizes[0];
+    const int dim = sizes[1];
+    const int seqlen = sizes[2];
+    const int width = weight.size(-1);
+
+    TORCH_CHECK(width >= 2 && width <= 4, "causal_conv1d only supports width between 2 and 4");
+
+    CHECK_SHAPE(x, batch_size, dim, seqlen);
+    CHECK_SHAPE(weight, dim, width);
+    CHECK_SHAPE(dout, batch_size, dim, seqlen);
+
+    TORCH_CHECK(x.stride(2) == 1 || x.stride(1) == 1);
+    const bool is_channel_last = x.stride(1) == 1 && x.stride(2) > 1;
+    if (!is_channel_last && dout.stride(2) != 1) { dout = dout.contiguous(); }
+    if (is_channel_last && dout.stride(1) != 1) { dout = dout.transpose(-1, -2).contiguous().transpose(-1, -2); }
+
+    if (bias_.has_value()) {
+        auto bias = bias_.value();
+        TORCH_CHECK(bias.scalar_type() == weight_type);
+        TORCH_CHECK(bias.is_cuda());
+        TORCH_CHECK(bias.stride(-1) == 1);
+        CHECK_SHAPE(bias, dim);
+    }
+
+    at::Tensor dx;
+    if (dx_.has_value()) {
+        dx = dx_.value();
+        TORCH_CHECK(dx.scalar_type() == input_type);
+        TORCH_CHECK(dx.is_cuda());
+        CHECK_SHAPE(dx, batch_size, dim, seqlen);
+        if (!is_channel_last) { TORCH_CHECK(dx.stride(2) == 1); }
+        if (is_channel_last) { TORCH_CHECK(dx.stride(1) == 1); }
+    } else {
+        dx = torch::empty_like(x);
+    }
+
+    // Otherwise the kernel will be launched from cuda:0 device
+    // Cast to char to avoid compiler warning about narrowing
+    at::cuda::CUDAGuard device_guard{(char)x.get_device()};
+
+    at::Tensor dweight = torch::zeros_like(weight, weight.options().dtype(at::kFloat));
+    at::Tensor dbias;
+    if (bias_.has_value()) { dbias = torch::zeros_like(bias_.value(), bias_.value().options().dtype(at::kFloat)); }
+
+    ConvParamsBwd params;
+    set_conv_params_bwd(params, batch_size, dim, seqlen, width,
+                        x, weight, bias_.has_value() ? bias_.value().data_ptr() : nullptr,
+                        dout, dx, dweight, bias_.has_value() ? dbias.data_ptr() : nullptr,
+                        silu_activation);
+
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    DISPATCH_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_bwd", [&] {
+        DISPATCH_WTYPE_FLOAT_AND_HALF_AND_BF16(weight.scalar_type(), "causal_conv1d_bwd", [&] {
+            if (!is_channel_last) {
+                causal_conv1d_bwd_cuda<input_t, weight_t>(params, stream);
+            } else {
+                causal_conv1d_channellast_bwd_cuda<input_t, weight_t>(params, stream);
+            }
+        });
+    });
+    return {dx, dweight.to(weight.dtype()), bias_.has_value() ? dbias.to(bias_.value().dtype()) : dbias};
+}
+
+at::Tensor
+causal_conv1d_update(const at::Tensor &x,
+                     const at::Tensor &conv_state,
+                     const at::Tensor &weight,
+                     const c10::optional<at::Tensor> &bias_,
+                  bool silu_activation) {
+    auto input_type = x.scalar_type();
+    auto weight_type = weight.scalar_type();
+    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::Half || weight_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(conv_state.scalar_type() == input_type);
+
+    TORCH_CHECK(x.is_cuda());
+    TORCH_CHECK(conv_state.is_cuda());
+    TORCH_CHECK(weight.is_cuda());
+
+    const auto sizes = x.sizes();
+    const int batch_size = sizes[0];
+    const int dim = sizes[1];
+    const int width = weight.size(-1);
+
+    CHECK_SHAPE(x, batch_size, dim);
+    CHECK_SHAPE(conv_state, batch_size, dim, width);
+    CHECK_SHAPE(weight, dim, width);
+
+    TORCH_CHECK(width >= 2 && width <= 4, "causal_conv1d only supports width between 2 and 4");
+
+    if (bias_.has_value()) {
+        auto bias = bias_.value();
+        TORCH_CHECK(bias.scalar_type() == weight_type);
+        TORCH_CHECK(bias.is_cuda());
+        TORCH_CHECK(bias.stride(-1) == 1);
+        CHECK_SHAPE(bias, dim);
+    }
+
+    at::Tensor out = torch::empty_like(x);
+
+    ConvParamsBase params;
+    set_conv_params_fwd(params, batch_size, dim, /*seqlen=*/1, width, x, weight, out,
+                        bias_.has_value() ? bias_.value().data_ptr() : nullptr,
+                        silu_activation);
+    params.conv_state_ptr = conv_state.data_ptr();
+    // All stride are in elements, not bytes.
+    params.conv_state_batch_stride = conv_state.stride(0);
+    params.conv_state_c_stride = conv_state.stride(1);
+    params.conv_state_l_stride = conv_state.stride(2);
+
+    // Otherwise the kernel will be launched from cuda:0 device
+    // Cast to char to avoid compiler warning about narrowing
+    at::cuda::CUDAGuard device_guard{(char)x.get_device()};
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    DISPATCH_ITYPE_FLOAT_AND_HALF_AND_BF16(x.scalar_type(), "causal_conv1d_update", [&] {
+        DISPATCH_WTYPE_FLOAT_AND_HALF_AND_BF16(weight.scalar_type(), "causal_conv1d_update", [&] {
+            causal_conv1d_update_cuda<input_t, weight_t>(params, stream);
+        });
+    });
+    return out;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("causal_conv1d_fwd", &causal_conv1d_fwd, "Causal conv1d forward");
+    m.def("causal_conv1d_bwd", &causal_conv1d_bwd, "Causal conv1d backward");
+    m.def("causal_conv1d_update", &causal_conv1d_update, "Causal conv1d update");
+}

protein_lm/modeling/models/libs/causal-conv1d/csrc/causal_conv1d.h

@@ -0,0 +1,53 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+struct ConvParamsBase {
+    using index_t = uint32_t;
+
+    int batch, dim, seqlen, width;
+    bool silu_activation;
+
+    index_t x_batch_stride;
+    index_t x_c_stride;
+    index_t x_l_stride;
+    index_t weight_c_stride;
+    index_t weight_width_stride;
+    index_t out_batch_stride;
+    index_t out_c_stride;
+    index_t out_l_stride;
+
+    index_t conv_state_batch_stride;
+    index_t conv_state_c_stride;
+    index_t conv_state_l_stride;
+
+    // Common data pointers.
+    void *__restrict__ x_ptr;
+    void *__restrict__ weight_ptr;
+    void *__restrict__ bias_ptr;
+    void *__restrict__ out_ptr;
+
+    void *__restrict__ conv_state_ptr;
+};
+
+struct ConvParamsBwd: public ConvParamsBase {
+    index_t dx_batch_stride;
+    index_t dx_c_stride;
+    index_t dx_l_stride;
+    index_t dweight_c_stride;
+    index_t dweight_width_stride;
+    index_t dout_batch_stride;
+    index_t dout_c_stride;
+    index_t dout_l_stride;
+
+    // Common data pointers.
+    void *__restrict__ dx_ptr;
+    void *__restrict__ dweight_ptr;
+    void *__restrict__ dbias_ptr;
+    void *__restrict__ dout_ptr;
+};
+

protein_lm/modeling/models/libs/causal-conv1d/csrc/causal_conv1d_bwd.cu

@@ -0,0 +1,525 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#include <c10/util/BFloat16.h>
+#include <c10/util/Half.h>
+#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
+
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_store.cuh>
+#include <cub/block/block_reduce.cuh>
+
+#include "causal_conv1d.h"
+#include "causal_conv1d_common.h"
+#include "static_switch.h"
+
+template<int kNThreads_, int kWidth_, bool kSiluAct_, bool kIsVecLoad_, typename input_t_, typename weight_t_>
+struct Causal_conv1d_bwd_kernel_traits {
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr int kNThreads = kNThreads_;
+    static constexpr int kWidth = kWidth_;
+    static constexpr bool kSiluAct = kSiluAct_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+    static constexpr int kNElts = kNBytes == 4 ? 4 : 8;
+    static_assert(kWidth <= kNElts);
+    // It's possible that we need to do 2 rounds of exchange if input_t is 16 bits
+    // (since then we'd have 8 values of float, and each round we can exchange 4 floats).
+    static constexpr int kNExchangeRounds = sizeof(float) / sizeof(input_t);
+    static constexpr bool kIsVecLoad = kIsVecLoad_;
+    using vec_t = typename BytesToType<kNBytes * kNElts>::Type;
+    using BlockLoadT = cub::BlockLoad<input_t, kNThreads, kNElts, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockLoadVecT = cub::BlockLoad<vec_t, kNThreads, 1, cub::BLOCK_LOAD_DIRECT>;
+    using BlockStoreT = cub::BlockStore<input_t, kNThreads, kNElts, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+    using BlockStoreVecT = cub::BlockStore<vec_t, kNThreads, 1, cub::BLOCK_STORE_DIRECT>;
+    using BlockReduceFloatT = cub::BlockReduce<float, kNThreads>;
+    static constexpr int kSmemIOSize = kIsVecLoad
+        ? 0
+        : std::max({sizeof(typename BlockLoadT::TempStorage), sizeof(typename BlockStoreT::TempStorage)});
+    static constexpr int kSmemExchangeSize = kNThreads * kNBytes * kNElts * (!kSiluAct ? 1 : kNExchangeRounds + 1);
+    static constexpr int kSmemSize = std::max({kSmemExchangeSize,
+            int(sizeof(typename BlockReduceFloatT::TempStorage))}) + (kIsVecLoad ? 0 : kSmemIOSize);
+};
+
+template<typename Ktraits>
+__global__ __launch_bounds__(Ktraits::kNThreads)
+void causal_conv1d_bwd_kernel(ConvParamsBwd params) {
+    constexpr int kWidth = Ktraits::kWidth;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    constexpr bool kSiluAct = Ktraits::kSiluAct;
+    constexpr int kNElts = Ktraits::kNElts;
+    constexpr int kNExchangeRounds = Ktraits::kNExchangeRounds;
+    constexpr bool kIsVecLoad = Ktraits::kIsVecLoad;
+    using input_t = typename Ktraits::input_t;
+    using vec_t = typename Ktraits::vec_t;
+    using weight_t = typename Ktraits::weight_t;
+
+    // Shared memory.
+    extern __shared__ char smem_[];
+    auto& smem_load = reinterpret_cast<typename Ktraits::BlockLoadT::TempStorage&>(smem_);
+    auto& smem_load_vec = reinterpret_cast<typename Ktraits::BlockLoadVecT::TempStorage&>(smem_);
+    auto& smem_store = reinterpret_cast<typename Ktraits::BlockStoreT::TempStorage&>(smem_);
+    auto& smem_store_vec = reinterpret_cast<typename Ktraits::BlockStoreVecT::TempStorage&>(smem_);
+    vec_t *smem_exchange = reinterpret_cast<vec_t *>(smem_ + Ktraits::kSmemIOSize);
+    vec_t *smem_exchange_x = reinterpret_cast<vec_t *>(smem_ + Ktraits::kSmemIOSize) + kNThreads * kNExchangeRounds;
+    auto& smem_reduce_float = *reinterpret_cast<typename Ktraits::BlockReduceFloatT::TempStorage*>(smem_ + Ktraits::kSmemIOSize);
+
+    const int tidx = threadIdx.x;
+    const int batch_id = blockIdx.x;
+    const int dim_id = blockIdx.y;
+    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + batch_id * params.x_batch_stride
+        + dim_id * params.x_c_stride;
+    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr) + dim_id * params.weight_c_stride;
+    input_t *dout = reinterpret_cast<input_t *>(params.dout_ptr) + batch_id * params.dout_batch_stride
+        + dim_id * params.dout_c_stride;
+    input_t *dx = reinterpret_cast<input_t *>(params.dx_ptr) + batch_id * params.dx_batch_stride
+        + dim_id * params.dx_c_stride;
+    float *dweight = reinterpret_cast<float *>(params.dweight_ptr) + dim_id * params.dweight_c_stride;
+    float bias_val = params.bias_ptr == nullptr ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[dim_id]);
+
+    // Thread kNThreads - 1 will load the first elements of the next chunk so we initialize those to 0.
+    if (tidx == 0) {
+        if constexpr (!kSiluAct) {
+            input_t zeros[kNElts] = {0};
+            smem_exchange[0] = reinterpret_cast<vec_t *>(zeros)[0];
+        } else {
+            float zeros[kNElts] = {0};
+            #pragma unroll
+            for (int r = 0; r < kNExchangeRounds; ++r) {
+                smem_exchange[r * kNThreads] = reinterpret_cast<vec_t *>(zeros)[r];
+            }
+        }
+    }
+
+    float weight_vals[kWidth];
+    #pragma unroll
+    for (int i = 0; i < kWidth; ++i) { weight_vals[i] = weight[i * params.weight_width_stride]; }
+
+    float dweight_vals[kWidth] = {0};
+    float dbias_val = 0;
+
+    constexpr int kChunkSize = kNThreads * kNElts;
+    const int n_chunks = (params.seqlen + kChunkSize - 1) / kChunkSize;
+    x += (n_chunks - 1) * kChunkSize;
+    dout += (n_chunks - 1) * kChunkSize;
+    dx += (n_chunks - 1) * kChunkSize;
+    for (int chunk = n_chunks - 1; chunk >= 0; --chunk) {
+        input_t x_vals_load[2 * kNElts] = {0};
+        input_t dout_vals_load[2 * kNElts] = {0};
+        if constexpr(kIsVecLoad) {
+            Ktraits::BlockLoadVecT(smem_load_vec).Load(reinterpret_cast<vec_t*>(x), *reinterpret_cast<vec_t (*)[1]>(&x_vals_load[kNElts]), (params.seqlen - chunk * kChunkSize) / kNElts);
+            Ktraits::BlockLoadVecT(smem_load_vec).Load(reinterpret_cast<vec_t*>(dout), *reinterpret_cast<vec_t (*)[1]>(&dout_vals_load[0]), (params.seqlen - chunk * kChunkSize) / kNElts);
+        } else {
+            __syncthreads();
+            Ktraits::BlockLoadT(smem_load).Load(x, *reinterpret_cast<input_t (*)[kNElts]>(&x_vals_load[kNElts]), params.seqlen - chunk * kChunkSize);
+            __syncthreads();
+            Ktraits::BlockLoadT(smem_load).Load(dout, *reinterpret_cast<input_t (*)[kNElts]>(&dout_vals_load[0]), params.seqlen - chunk * kChunkSize);
+        }
+        float dout_vals[2 * kNElts], x_vals[2 * kNElts];
+        if constexpr (!kSiluAct) {
+            __syncthreads();
+            // Thread 0 don't write yet, so that thread kNThreads - 1 can read
+            // the first elements of the next chunk.
+            if (tidx > 0) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(dout_vals_load)[0]; }
+            __syncthreads();
+            reinterpret_cast<vec_t *>(dout_vals_load)[1] = smem_exchange[tidx < kNThreads - 1 ? tidx + 1 : 0];
+            __syncthreads();
+            // Now thread 0 can write the first elements of the current chunk.
+            if (tidx == 0) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(dout_vals_load)[0]; }
+            #pragma unroll
+            for (int i = 0; i < 2 * kNElts; ++i) {
+                dout_vals[i] = float(dout_vals_load[i]);
+                x_vals[i] = float(x_vals_load[i]);
+            }
+        } else {
+            if (tidx == 0 && chunk > 0) {
+                if constexpr(kIsVecLoad) {
+                    reinterpret_cast<vec_t *>(x_vals_load)[0] = reinterpret_cast<vec_t *>(x)[-1];
+                } else {
+                    #pragma unroll
+                    for (int i = 0; i < kNElts; ++i) {
+                        if (chunk * kChunkSize + i < params.seqlen) { x_vals_load[i] = x[-kNElts + i]; }
+                    }
+                }
+            }
+            __syncthreads();
+            smem_exchange_x[tidx] = reinterpret_cast<vec_t *>(x_vals_load)[1];
+            __syncthreads();
+            if (tidx > 0) { reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange_x[tidx - 1]; }
+            #pragma unroll
+            for (int i = 0; i < 2 * kNElts; ++i) { x_vals[i] = float(x_vals_load[i]); }
+            // Recompute the output
+            #pragma unroll
+            for (int i = 0; i < kNElts; ++i) {
+                float out_val = bias_val;
+                #pragma unroll
+                for (int w = 0; w < kWidth; ++w) {
+                    out_val += weight_vals[w] * x_vals[kNElts + i - (kWidth - w - 1)];
+                }
+                float out_sigmoid_val = 1.0f / (1.0f + expf(-out_val));
+                dout_vals[i] = float(dout_vals_load[i]) * out_sigmoid_val
+                               * (1.0f + out_val * (1.0f - out_sigmoid_val));
+            }
+            // Exchange the dout_vals. It's possible that we need to do 2 rounds of exchange
+            // if input_t is 16 bits (since then we'd have 8 values of float)
+            __syncthreads();
+            // Thread 0 don't write yet, so that thread kNThreads - 1 can read
+            // the first elements of the next chunk.
+            if (tidx > 0) {
+                #pragma unroll
+                for (int r = 0; r < kNExchangeRounds; ++r) {
+                    smem_exchange[r * kNThreads + tidx] = reinterpret_cast<vec_t *>(dout_vals)[r];
+                }
+            }
+            __syncthreads();
+            #pragma unroll
+            for (int r = 0; r < kNExchangeRounds; ++r) {
+                reinterpret_cast<vec_t *>(dout_vals)[kNExchangeRounds + r]
+                    = smem_exchange[r * kNThreads + (tidx < kNThreads - 1 ? tidx + 1 : 0)];
+            }
+            __syncthreads();
+            // Now thread 0 can write the first elements of the current chunk.
+            if (tidx == 0) {
+                #pragma unroll
+                for (int r = 0; r < kNExchangeRounds; ++r) {
+                    smem_exchange[r * kNThreads + tidx] = reinterpret_cast<vec_t *>(dout_vals)[r];
+                }
+            }
+        }
+        dout -= kChunkSize;
+        x -= kChunkSize;
+
+        #pragma unroll
+        for (int i = 0; i < kNElts; ++i) { dbias_val += dout_vals[i]; }
+
+        float dx_vals[kNElts] = {0};
+        #pragma unroll
+        for (int i = 0; i < kNElts; ++i) {
+            #pragma unroll
+            for (int w = 0; w < kWidth; ++w) {
+                dx_vals[i] += weight_vals[w] * dout_vals[i + kWidth - w - 1];
+            }
+        }
+
+        input_t dx_vals_store[kNElts];
+        #pragma unroll
+        for (int i = 0; i < kNElts; ++i) { dx_vals_store[i] = dx_vals[i]; }
+        if constexpr(kIsVecLoad) {
+            Ktraits::BlockStoreVecT(smem_store_vec).Store(reinterpret_cast<vec_t*>(dx), reinterpret_cast<vec_t (&)[1]>(dx_vals_store), (params.seqlen - chunk * kChunkSize) / kNElts);
+        } else {
+            Ktraits::BlockStoreT(smem_store).Store(dx, dx_vals_store, params.seqlen - chunk * kChunkSize);
+        }
+        dx -= kChunkSize;
+
+        #pragma unroll
+        for (int w = 0; w < kWidth; ++w) {
+            #pragma unroll
+            for (int i = 0; i < kNElts; ++i) {
+                dweight_vals[w] += x_vals[kNElts + i] * dout_vals[i + kWidth - w - 1];
+            }
+        }
+    }
+
+    #pragma unroll
+    for (int w = 0; w < kWidth; ++w) {
+        __syncthreads();
+        dweight_vals[w] = Ktraits::BlockReduceFloatT(smem_reduce_float).Sum(dweight_vals[w]);
+        if (tidx == 0) {
+            atomicAdd(&reinterpret_cast<float *>(dweight)[w * params.dweight_width_stride], dweight_vals[w]);
+        }
+    }
+    if (params.bias_ptr != nullptr) {
+        __syncthreads();
+        dbias_val = Ktraits::BlockReduceFloatT(smem_reduce_float).Sum(dbias_val);
+        if (tidx == 0) {
+            atomicAdd(&reinterpret_cast<float *>(params.dbias_ptr)[dim_id], dbias_val);
+        }
+    }
+}
+
+template<int kNThreads, int kWidth, typename input_t, typename weight_t>
+void causal_conv1d_bwd_launch(ConvParamsBwd &params, cudaStream_t stream) {
+    static constexpr int kNElts = sizeof(input_t) == 4 ? 4 : 8;
+    BOOL_SWITCH(params.seqlen % kNElts == 0, kIsVecLoad, [&] {
+        BOOL_SWITCH(params.silu_activation, kSiluAct, [&] {
+            using Ktraits = Causal_conv1d_bwd_kernel_traits<kNThreads, kWidth, kSiluAct, kIsVecLoad, input_t, weight_t>;
+            constexpr int kSmemSize = Ktraits::kSmemSize;
+            dim3 grid(params.batch, params.dim);
+            auto kernel = &causal_conv1d_bwd_kernel<Ktraits>;
+            if (kSmemSize >= 48 * 1024) {
+                C10_CUDA_CHECK(cudaFuncSetAttribute(
+                    kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+                }
+            kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+            C10_CUDA_KERNEL_LAUNCH_CHECK();
+        });
+    });
+}
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_bwd_cuda(ConvParamsBwd &params, cudaStream_t stream) {
+    if (params.width == 2) {
+        causal_conv1d_bwd_launch<128, 2, input_t, weight_t>(params, stream);
+    } else if (params.width == 3) {
+        causal_conv1d_bwd_launch<128, 3, input_t, weight_t>(params, stream);
+    } else if (params.width == 4) {
+        causal_conv1d_bwd_launch<128, 4, input_t, weight_t>(params, stream);
+    }
+}
+
+template<int kNThreads_, int kWidth_, int kChunkSizeL_, bool kSiluAct_, bool kIsVecLoad_, typename input_t_, typename weight_t_>
+struct Causal_conv1d_channellast_bwd_kernel_traits {
+    // The cache line is 128 bytes, and we try to read 16 bytes per thread.
+    // So we have 8 threads per "row", so 32 or 64 elements in the channel dimension.
+    // That leaves 4 columns per warp, and so 16 columns per block (assuming each block has 128
+    // threads). Each each load is 16 x 32|64 elements in the L x C dimensions.
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr bool kSiluAct = kSiluAct_;
+    static constexpr int kNThreads = kNThreads_;
+    static_assert(kNThreads % 32 == 0);
+    static constexpr int kNWarps = kNThreads / 32;
+    static constexpr int kWidth = kWidth_;
+    static constexpr int kChunkSizeL = kChunkSizeL_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+    static constexpr int kNElts = kNBytes == 4 ? 4 : 8;
+    static constexpr int kNEltsPerRow = 128 / kNBytes;
+    static constexpr int kNThreadsPerRow = kNEltsPerRow / kNElts;  // Always 8 for now
+    static_assert(kNThreadsPerRow * kNBytes * kNElts == 128);
+    static constexpr int kNColsPerWarp = 32 / kNThreadsPerRow;  // Always 4 for now
+    static_assert(kNColsPerWarp * kNThreadsPerRow == 32);
+    static constexpr int kNColsPerLoad = kNColsPerWarp * kNWarps;
+    static constexpr int kNLoads = kChunkSizeL / kNColsPerLoad;
+    static_assert(kNLoads * kNColsPerLoad == kChunkSizeL);
+    static constexpr bool kIsVecLoad = kIsVecLoad_;
+    using vec_t = typename BytesToType<kNBytes * kNElts>::Type;
+    // using BlockLoadT = cub::BlockLoad<input_t, kNThreads, kNItems, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    // using BlockStoreT = cub::BlockStore<input_t, kNThreads, kNItems, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+    // static constexpr int kSmemSize = std::max({sizeof(typename BlockLoadT::TempStorage),
+    //                                            sizeof(typename BlockStoreT::TempStorage)});
+    // static constexpr int kSmemSize = kChunkSizeL * kNEltsPerRow * kNBytes;
+};
+
+template<typename Ktraits>
+__global__ __launch_bounds__(Ktraits::kNThreads)
+void causal_conv1d_channellast_bwd_kernel(ConvParamsBwd params) {
+    constexpr int kWidth = Ktraits::kWidth;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    constexpr bool kSiluAct = Ktraits::kSiluAct;
+    constexpr int kNElts = Ktraits::kNElts;
+    constexpr int kNWarp = Ktraits::kNWarps;
+    constexpr int kNThreadsPerC = Ktraits::kNThreadsPerRow;
+    constexpr int kLPerLoad = Ktraits::kNColsPerLoad;
+    constexpr int kChunkSizeL = Ktraits::kChunkSizeL;
+    constexpr int kChunkSizeC = Ktraits::kNEltsPerRow;
+    using input_t = typename Ktraits::input_t;
+    using vec_t = typename Ktraits::vec_t;
+    using weight_t = typename Ktraits::weight_t;
+
+    // Shared memory.
+    __shared__ input_t dout_smem[kChunkSizeL + kWidth - 1][kChunkSizeC + kNElts];
+    __shared__ input_t x_smem[kWidth - 1 + kChunkSizeL + kWidth - 1][kChunkSizeC + kNElts];
+
+    const int tid = threadIdx.x;
+    const int l_idx = tid / kNThreadsPerC;
+    const int c_idx = tid % kNThreadsPerC;
+    const int batch_id = blockIdx.x;
+    const int chunk_l_id = blockIdx.y;
+    const int chunk_c_id = blockIdx.z;
+    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + batch_id * params.x_batch_stride
+        + (chunk_l_id * kChunkSizeL + l_idx) * params.x_l_stride + chunk_c_id * kChunkSizeC + c_idx * kNElts;
+    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr)
+        + chunk_c_id * kChunkSizeC * params.weight_c_stride;
+    input_t *dout = reinterpret_cast<input_t *>(params.dout_ptr) + batch_id * params.dout_batch_stride
+        + (chunk_l_id * kChunkSizeL + l_idx) * params.dout_l_stride + chunk_c_id * kChunkSizeC + c_idx * kNElts;
+    input_t *dx = reinterpret_cast<input_t *>(params.dx_ptr) + batch_id * params.dx_batch_stride
+        + (chunk_l_id * kChunkSizeL + l_idx) * params.dx_l_stride + chunk_c_id * kChunkSizeC + c_idx * kNElts;
+    float *dweight = reinterpret_cast<float *>(params.dweight_ptr)
+        + chunk_c_id * kChunkSizeC * params.dweight_c_stride;
+
+    #pragma unroll
+    for (int l = 0; l < Ktraits::kNLoads; ++l) {
+        input_t dout_vals_load[kNElts] = {0};
+        input_t x_vals_load[kNElts] = {0};
+        if (chunk_l_id * kChunkSizeL + l * kLPerLoad + l_idx < params.seqlen
+            && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+            reinterpret_cast<vec_t *>(dout_vals_load)[0] = *reinterpret_cast<vec_t *>(dout + l * kLPerLoad * params.dout_l_stride);
+            reinterpret_cast<vec_t *>(x_vals_load)[0] = *reinterpret_cast<vec_t *>(x + l * kLPerLoad * params.x_l_stride);
+        }
+        reinterpret_cast<vec_t *>(dout_smem[l * kLPerLoad + l_idx])[c_idx] = reinterpret_cast<vec_t *>(dout_vals_load)[0];
+        reinterpret_cast<vec_t *>(x_smem[kWidth - 1 + l * kLPerLoad + l_idx])[c_idx] = reinterpret_cast<vec_t *>(x_vals_load)[0];
+    }
+    // Load the elements from the previous chunk or next chunk that are needed for convolution.
+    if (l_idx < kWidth - 1) {
+        input_t dout_vals_load[kNElts] = {0};
+        input_t x_vals_load[kNElts] = {0};
+        if ((chunk_l_id + 1) * kChunkSizeL + l_idx < params.seqlen
+            && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+            reinterpret_cast<vec_t *>(dout_vals_load)[0] = *reinterpret_cast<vec_t *>(dout + kChunkSizeL * params.dout_l_stride);
+        }
+        if (chunk_l_id * kChunkSizeL + l_idx - (kWidth - 1) >= 0
+            && chunk_l_id * kChunkSizeL + l_idx - (kWidth - 1) < params.seqlen
+            && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+            reinterpret_cast<vec_t *>(x_vals_load)[0] = *reinterpret_cast<vec_t *>(x - (kWidth - 1) * params.x_l_stride);
+        }
+        reinterpret_cast<vec_t *>(dout_smem[kChunkSizeL + l_idx])[c_idx] = reinterpret_cast<vec_t *>(dout_vals_load)[0];
+        reinterpret_cast<vec_t *>(x_smem[l_idx])[c_idx] = reinterpret_cast<vec_t *>(x_vals_load)[0];
+    }
+    // Need to load (kWdith - 1) extra x's on the right to recompute the (kChunkSizeL + kWidth - 1) outputs
+    if constexpr (kSiluAct) {
+        if (l_idx < kWidth - 1) {
+            input_t x_vals_load[kNElts] = {0};
+            if ((chunk_l_id + 1) * kChunkSizeL + l_idx < params.seqlen
+                && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+                reinterpret_cast<vec_t *>(x_vals_load)[0] = *reinterpret_cast<vec_t *>(x + kChunkSizeL * params.x_l_stride);
+            }
+            reinterpret_cast<vec_t *>(x_smem[kWidth - 1 + kChunkSizeL + l_idx])[c_idx] = reinterpret_cast<vec_t *>(x_vals_load)[0];
+        }
+    }
+
+    __syncthreads();
+
+    constexpr int kLPerThread = std::min(kChunkSizeL * kChunkSizeC / kNThreads, kChunkSizeL);
+    static_assert(kLPerThread * kNThreads == kChunkSizeL * kChunkSizeC);
+    constexpr int kNThreadsPerRow = kChunkSizeL / kLPerThread;
+    static_assert(kNThreadsPerRow * kLPerThread == kChunkSizeL);
+    // kChunkSizeL, kLPerThread, kNThreadsPerRow should be powers of 2 for simplicity
+    static_assert((kChunkSizeL & (kChunkSizeL - 1)) == 0);
+    static_assert((kLPerThread & (kLPerThread - 1)) == 0);
+    static_assert((kNThreadsPerRow & (kNThreadsPerRow - 1)) == 0);
+    static_assert(kNThreadsPerRow <= 32);
+
+    const int row_idx = tid / kNThreadsPerRow;
+    const int col_idx = tid % kNThreadsPerRow;
+
+    float bias_val = params.bias_ptr == nullptr || chunk_c_id * kChunkSizeC + row_idx >= params.dim ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[chunk_c_id * kChunkSizeC + row_idx]);
+    float weight_vals[kWidth] = {0};
+    if (chunk_c_id * kChunkSizeC + row_idx < params.dim) {
+        #pragma unroll
+        for (int w = 0; w < kWidth; ++w) {
+            weight_vals[w] = weight[row_idx * params.weight_c_stride + w * params.weight_width_stride];
+        }
+    }
+    float dout_vals[kLPerThread + kWidth - 1];
+    float x_vals[kWidth - 1 + kLPerThread + kWidth - 1];
+    #pragma unroll
+    for (int i = 0; i < kWidth - 1 + kLPerThread; ++i) {
+        dout_vals[i] = float(dout_smem[col_idx * kLPerThread + i][row_idx]);
+        x_vals[i] = float(x_smem[col_idx * kLPerThread + i][row_idx]);
+    }
+
+    if constexpr (kSiluAct) {  // Recompute the output
+        #pragma unroll
+        for (int i = kWidth - 1 + kLPerThread; i < kWidth - 1 + kLPerThread + kWidth - 1; ++i) {
+            x_vals[i] = float(x_smem[col_idx * kLPerThread + i][row_idx]);
+        }
+        #pragma unroll
+        for (int i = 0; i < kLPerThread + kWidth - 1; ++i) {
+            float out_val = bias_val;
+            #pragma unroll
+            for (int w = 0; w < kWidth; ++w) { out_val += weight_vals[w] * x_vals[i + w]; }
+            float out_val_sigmoid = 1.f / (1.f + expf(-out_val));
+            dout_vals[i] *= out_val_sigmoid * (1 + out_val * (1 - out_val_sigmoid));
+        }
+    }
+
+    float dweight_vals[kWidth] = {0};
+    SumOp<float> sum_op;
+    #pragma unroll
+    for (int w = 0; w < kWidth; ++w) {
+        #pragma unroll
+        for (int i = 0; i < kLPerThread; ++i) { dweight_vals[w] += x_vals[i + w] * dout_vals[i]; }
+        dweight_vals[w] = Allreduce<kNThreadsPerRow>::run(dweight_vals[w], sum_op);
+        if (col_idx == 0 && chunk_c_id * kChunkSizeC + row_idx < params.dim) {
+            atomicAdd(&reinterpret_cast<float *>(dweight)[row_idx * params.dweight_c_stride + w * params.dweight_width_stride], dweight_vals[w]);
+        }
+    }
+
+    if (params.bias_ptr != nullptr) {
+        float dbias_val = 0.f;
+        for (int i = 0; i < kLPerThread; ++i) { dbias_val += dout_vals[i]; }
+        dbias_val = Allreduce<kNThreadsPerRow>::run(dbias_val, sum_op);
+        if (col_idx == 0 && chunk_c_id * kChunkSizeC + row_idx < params.dim) {
+            atomicAdd(&reinterpret_cast<float *>(params.dbias_ptr)[chunk_c_id * kChunkSizeC + row_idx], dbias_val);
+        }
+    }
+
+    float dx_vals[kLPerThread] = {0};
+    #pragma unroll
+    for (int i = 0; i < kLPerThread; ++i) {
+        #pragma unroll
+        for (int w = 0; w < kWidth; ++w) { dx_vals[i] += weight_vals[kWidth - 1 - w] * dout_vals[i + w]; }
+    }
+    // Since kNThreadsPerRow is a power of 2 and <= 32, we only need syncwarp and not syncthreads.
+    __syncwarp();
+    #pragma unroll
+    for (int i = 0; i < kLPerThread; ++i) { x_smem[col_idx * kLPerThread + i][row_idx] = dx_vals[i]; }
+    __syncthreads();
+
+    #pragma unroll
+    for (int l = 0; l < Ktraits::kNLoads; ++l) {
+        input_t dx_vals_store[kNElts];
+        reinterpret_cast<vec_t *>(dx_vals_store)[0] = reinterpret_cast<vec_t *>(x_smem[l * kLPerLoad + l_idx])[c_idx];
+        if (chunk_l_id * kChunkSizeL + l * kLPerLoad + l_idx < params.seqlen
+            && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+            *reinterpret_cast<vec_t *>(dx + l * kLPerLoad * params.dx_l_stride) = reinterpret_cast<vec_t *>(dx_vals_store)[0];
+        }
+    }
+
+}
+
+template<int kNThreads, int kWidth, typename input_t, typename weight_t>
+void causal_conv1d_channellast_bwd_launch(ConvParamsBwd &params, cudaStream_t stream) {
+    BOOL_SWITCH(params.silu_activation, kSiluAct, [&] {
+        using Ktraits = Causal_conv1d_channellast_bwd_kernel_traits<kNThreads, kWidth, 64, kSiluAct, true, input_t, weight_t>;
+        // constexpr int kSmemSize = Ktraits::kSmemSize;
+        constexpr int kChunkSizeL = Ktraits::kChunkSizeL;
+        constexpr int kChunkSizeC = Ktraits::kNEltsPerRow;
+        const int n_chunks_L = (params.seqlen + kChunkSizeL - 1) / kChunkSizeL;
+        const int n_chunks_C = (params.dim + kChunkSizeC - 1) / kChunkSizeC;
+        dim3 grid(params.batch, n_chunks_L, n_chunks_C);
+        dim3 block(Ktraits::kNThreads);
+        auto kernel = &causal_conv1d_channellast_bwd_kernel<Ktraits>;
+        // if (kSmemSize >= 48 * 1024) {
+        //     C10_CUDA_CHECK(cudaFuncSetAttribute(
+        //         kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+        //     }
+        // kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+        kernel<<<grid, Ktraits::kNThreads, 0, stream>>>(params);
+        C10_CUDA_KERNEL_LAUNCH_CHECK();
+    });
+}
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_channellast_bwd_cuda(ConvParamsBwd &params, cudaStream_t stream) {
+    if (params.width == 2) {
+        causal_conv1d_channellast_bwd_launch<128, 2, input_t, weight_t>(params, stream);
+    } else if (params.width == 3) {
+        causal_conv1d_channellast_bwd_launch<128, 3, input_t, weight_t>(params, stream);
+    } else if (params.width == 4) {
+        causal_conv1d_channellast_bwd_launch<128, 4, input_t, weight_t>(params, stream);
+    }
+}
+
+template void causal_conv1d_bwd_cuda<float, float>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<at::Half, float>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<at::BFloat16, float>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<float, at::Half>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<at::Half, at::Half>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<at::BFloat16, at::Half>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<float, at::BFloat16>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<at::Half, at::BFloat16>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_bwd_cuda<at::BFloat16, at::BFloat16>(ConvParamsBwd &params, cudaStream_t stream);
+
+template void causal_conv1d_channellast_bwd_cuda<float, float>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<at::Half, float>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<at::BFloat16, float>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<float, at::Half>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<at::Half, at::Half>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<at::BFloat16, at::Half>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<float, at::BFloat16>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<at::Half, at::BFloat16>(ConvParamsBwd &params, cudaStream_t stream);
+template void causal_conv1d_channellast_bwd_cuda<at::BFloat16, at::BFloat16>(ConvParamsBwd &params, cudaStream_t stream);

protein_lm/modeling/models/libs/causal-conv1d/csrc/causal_conv1d_common.h

@@ -0,0 +1,64 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template<int BYTES> struct BytesToType {};
+
+template<> struct BytesToType<16> {
+    using Type = uint4;
+    static_assert(sizeof(Type) == 16);
+};
+
+template<> struct BytesToType<8> {
+    using Type = uint64_t;
+    static_assert(sizeof(Type) == 8);
+};
+
+template<> struct BytesToType<4> {
+    using Type = uint32_t;
+    static_assert(sizeof(Type) == 4);
+};
+
+template<> struct BytesToType<2> {
+    using Type = uint16_t;
+    static_assert(sizeof(Type) == 2);
+};
+
+template<> struct BytesToType<1> {
+    using Type = uint8_t;
+    static_assert(sizeof(Type) == 1);
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template<typename T>
+struct SumOp {
+__device__ inline T operator()(T const & x, T const & y) { return x + y; }
+};
+
+template<int THREADS>
+struct Allreduce {
+    static_assert(THREADS == 32 || THREADS == 16 || THREADS == 8 || THREADS == 4);
+    template<typename T, typename Operator>
+    static __device__ inline T run(T x, Operator &op) {
+        constexpr int OFFSET = THREADS / 2;
+        x = op(x, __shfl_xor_sync(uint32_t(-1), x, OFFSET));
+        return Allreduce<OFFSET>::run(x, op);
+    }
+};
+
+template<>
+struct Allreduce<2> {
+template<typename T, typename Operator>
+static __device__ inline T run(T x, Operator &op) {
+    x = op(x, __shfl_xor_sync(uint32_t(-1), x, 1));
+    return x;
+}
+};

protein_lm/modeling/models/libs/causal-conv1d/csrc/causal_conv1d_fwd.cu

@@ -0,0 +1,350 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#include <c10/util/BFloat16.h>
+#include <c10/util/Half.h>
+#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
+
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_store.cuh>
+
+#include "causal_conv1d.h"
+#include "causal_conv1d_common.h"
+#include "static_switch.h"
+
+template<int kNThreads_, int kWidth_, bool kIsVecLoad_, typename input_t_, typename weight_t_>
+struct Causal_conv1d_fwd_kernel_traits {
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr int kNThreads = kNThreads_;
+    static constexpr int kWidth = kWidth_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+    static constexpr int kNElts = kNBytes == 4 ? 4 : 8;
+    static_assert(kWidth <= kNElts);
+    static constexpr bool kIsVecLoad = kIsVecLoad_;
+    using vec_t = typename BytesToType<kNBytes * kNElts>::Type;
+    using BlockLoadT = cub::BlockLoad<input_t, kNThreads, kNElts, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockLoadVecT = cub::BlockLoad<vec_t, kNThreads, 1, cub::BLOCK_LOAD_DIRECT>;
+    using BlockStoreT = cub::BlockStore<input_t, kNThreads, kNElts, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+    using BlockStoreVecT = cub::BlockStore<vec_t, kNThreads, 1, cub::BLOCK_STORE_DIRECT>;
+    static constexpr int kSmemIOSize = kIsVecLoad
+        ? 0
+        : std::max({sizeof(typename BlockLoadT::TempStorage), sizeof(typename BlockStoreT::TempStorage)});
+    static constexpr int kSmemExchangeSize = kNThreads * kNBytes * kNElts;
+    static constexpr int kSmemSize = kSmemIOSize + kSmemExchangeSize;
+};
+
+template<typename Ktraits>
+__global__ __launch_bounds__(Ktraits::kNThreads)
+void causal_conv1d_fwd_kernel(ConvParamsBase params) {
+    constexpr int kWidth = Ktraits::kWidth;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    constexpr int kNElts = Ktraits::kNElts;
+    constexpr bool kIsVecLoad = Ktraits::kIsVecLoad;
+    using input_t = typename Ktraits::input_t;
+    using vec_t = typename Ktraits::vec_t;
+    using weight_t = typename Ktraits::weight_t;
+
+    // Shared memory.
+    extern __shared__ char smem_[];
+    auto& smem_load = reinterpret_cast<typename Ktraits::BlockLoadT::TempStorage&>(smem_);
+    auto& smem_load_vec = reinterpret_cast<typename Ktraits::BlockLoadVecT::TempStorage&>(smem_);
+    auto& smem_store = reinterpret_cast<typename Ktraits::BlockStoreT::TempStorage&>(smem_);
+    auto& smem_store_vec = reinterpret_cast<typename Ktraits::BlockStoreVecT::TempStorage&>(smem_);
+    vec_t *smem_exchange = reinterpret_cast<vec_t *>(smem_ + Ktraits::kSmemIOSize);
+
+    const int tidx = threadIdx.x;
+    const int batch_id = blockIdx.x;
+    const int channel_id = blockIdx.y;
+    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + batch_id * params.x_batch_stride
+        + channel_id * params.x_c_stride;
+    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr) + channel_id * params.weight_c_stride;
+    input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + batch_id * params.out_batch_stride
+        + channel_id * params.out_c_stride;
+    float bias_val = params.bias_ptr == nullptr ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[channel_id]);
+
+    // Thread 0 will load the last elements of the previous chunk, so we initialize those to 0.
+    if (tidx == 0) {
+        input_t zeros[kNElts] = {0};
+        smem_exchange[kNThreads - 1] = reinterpret_cast<vec_t *>(zeros)[0];
+    }
+
+    float weight_vals[kWidth];
+    #pragma unroll
+    for (int i = 0; i < kWidth; ++i) { weight_vals[i] = float(weight[i * params.weight_width_stride]); }
+
+    constexpr int kChunkSize = kNThreads * kNElts;
+    const int n_chunks = (params.seqlen + kChunkSize - 1) / kChunkSize;
+    for (int chunk = 0; chunk < n_chunks; ++chunk) {
+        input_t x_vals_load[2 * kNElts] = {0};
+        if constexpr(kIsVecLoad) {
+            Ktraits::BlockLoadVecT(smem_load_vec).Load(reinterpret_cast<vec_t*>(x), *reinterpret_cast<vec_t (*)[1]>(&x_vals_load[kNElts]), (params.seqlen - chunk * kChunkSize) / kNElts);
+        } else {
+            __syncthreads();
+            Ktraits::BlockLoadT(smem_load).Load(x, *reinterpret_cast<input_t (*)[kNElts]>(&x_vals_load[kNElts]), params.seqlen - chunk * kChunkSize);
+        }
+        x += kChunkSize;
+        __syncthreads();
+        // Thread kNThreads - 1 don't write yet, so that thread 0 can read
+        // the last elements of the previous chunk.
+        if (tidx < kNThreads - 1) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(x_vals_load)[1]; }
+        __syncthreads();
+        reinterpret_cast<vec_t *>(x_vals_load)[0] = smem_exchange[tidx > 0 ? tidx - 1 : kNThreads - 1];
+        __syncthreads();
+        // Now thread kNThreads - 1 can write the last elements of the current chunk.
+        if (tidx == kNThreads - 1) { smem_exchange[tidx] = reinterpret_cast<vec_t *>(x_vals_load)[1]; }
+
+        float x_vals[2 * kNElts];
+        #pragma unroll
+        for (int i = 0; i < 2 * kNElts; ++i) { x_vals[i] = float(x_vals_load[i]); }
+
+        float out_vals[kNElts];
+        #pragma unroll
+        for (int i = 0; i < kNElts; ++i) {
+            out_vals[i] = bias_val;
+            #pragma unroll
+            for (int w = 0; w < kWidth; ++w) {
+                out_vals[i] += weight_vals[w] * x_vals[kNElts + i - (kWidth - w - 1)];
+            }
+        }
+
+        if (params.silu_activation) {
+            #pragma unroll
+            for (int i = 0; i < kNElts; ++i) {
+                out_vals[i] = out_vals[i] / (1 + expf(-out_vals[i]));
+            }
+        }
+
+        input_t out_vals_store[kNElts];
+        #pragma unroll
+        for (int i = 0; i < kNElts; ++i) { out_vals_store[i] = out_vals[i]; }
+        if constexpr(kIsVecLoad) {
+            Ktraits::BlockStoreVecT(smem_store_vec).Store(reinterpret_cast<vec_t*>(out), reinterpret_cast<vec_t (&)[1]>(out_vals_store), (params.seqlen - chunk * kChunkSize) / kNElts);
+        } else {
+            Ktraits::BlockStoreT(smem_store).Store(out, out_vals_store, params.seqlen - chunk * kChunkSize);
+        }
+        out += kChunkSize;
+    }
+}
+
+template<int kNThreads, int kWidth, typename input_t, typename weight_t>
+void causal_conv1d_fwd_launch(ConvParamsBase &params, cudaStream_t stream) {
+    static constexpr int kNElts = sizeof(input_t) == 4 ? 4 : 8;
+    BOOL_SWITCH(params.seqlen % kNElts == 0, kIsVecLoad, [&] {
+        using Ktraits = Causal_conv1d_fwd_kernel_traits<kNThreads, kWidth, kIsVecLoad, input_t, weight_t>;
+        constexpr int kSmemSize = Ktraits::kSmemSize;
+        dim3 grid(params.batch, params.dim);
+        auto kernel = &causal_conv1d_fwd_kernel<Ktraits>;
+        if (kSmemSize >= 48 * 1024) {
+            C10_CUDA_CHECK(cudaFuncSetAttribute(
+                kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+            }
+        kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+        C10_CUDA_KERNEL_LAUNCH_CHECK();
+    });
+}
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_fwd_cuda(ConvParamsBase &params, cudaStream_t stream) {
+    if (params.width == 2) {
+        causal_conv1d_fwd_launch<128, 2, input_t, weight_t>(params, stream);
+    } else if (params.width == 3) {
+        causal_conv1d_fwd_launch<128, 3, input_t, weight_t>(params, stream);
+    } else if (params.width == 4) {
+        causal_conv1d_fwd_launch<128, 4, input_t, weight_t>(params, stream);
+    }
+}
+
+template<int kNThreads_, int kWidth_, int kChunkSizeL_, bool kIsVecLoad_, typename input_t_, typename weight_t_>
+struct Causal_conv1d_channellast_fwd_kernel_traits {
+    // The cache line is 128 bytes, and we try to read 16 bytes per thread.
+    // So we have 8 threads per "row", so 32 or 64 elements in the channel dimension.
+    // That leaves 4 columns per warp, and so 16 columns per block (assuming each block has 128
+    // threads). Each each load is 16 x 32|64 elements in the L x C dimensions.
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr int kNThreads = kNThreads_;
+    static_assert(kNThreads % 32 == 0);
+    static constexpr int kNWarps = kNThreads / 32;
+    static constexpr int kWidth = kWidth_;
+    static constexpr int kChunkSizeL = kChunkSizeL_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+    static constexpr int kNElts = kNBytes == 4 ? 4 : 8;
+    static constexpr int kNEltsPerRow = 128 / kNBytes;
+    static constexpr int kNThreadsPerRow = kNEltsPerRow / kNElts;  // Always 8 for now
+    static_assert(kNThreadsPerRow * kNBytes * kNElts == 128);
+    static constexpr int kNColsPerWarp = 32 / kNThreadsPerRow;  // Always 4 for now
+    static_assert(kNColsPerWarp * kNThreadsPerRow == 32);
+    static constexpr int kNColsPerLoad = kNColsPerWarp * kNWarps;
+    static constexpr int kNLoads = kChunkSizeL / kNColsPerLoad;
+    static_assert(kNLoads * kNColsPerLoad == kChunkSizeL);
+    static constexpr bool kIsVecLoad = kIsVecLoad_;
+    using vec_t = typename BytesToType<kNBytes * kNElts>::Type;
+    // using BlockLoadT = cub::BlockLoad<input_t, kNThreads, kNItems, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    // using BlockStoreT = cub::BlockStore<input_t, kNThreads, kNItems, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+    // static constexpr int kSmemSize = std::max({sizeof(typename BlockLoadT::TempStorage),
+    //                                            sizeof(typename BlockStoreT::TempStorage)});
+    // static constexpr int kSmemSize = kChunkSizeL * kNEltsPerRow * kNBytes;
+};
+
+template<typename Ktraits>
+__global__ __launch_bounds__(Ktraits::kNThreads)
+void causal_conv1d_channellast_fwd_kernel(ConvParamsBase params) {
+    constexpr int kWidth = Ktraits::kWidth;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    constexpr int kNElts = Ktraits::kNElts;
+    constexpr int kNWarp = Ktraits::kNWarps;
+    constexpr int kNThreadsPerC = Ktraits::kNThreadsPerRow;
+    constexpr int kLPerLoad = Ktraits::kNColsPerLoad;
+    constexpr int kChunkSizeL = Ktraits::kChunkSizeL;
+    constexpr int kChunkSizeC = Ktraits::kNEltsPerRow;
+    using input_t = typename Ktraits::input_t;
+    using vec_t = typename Ktraits::vec_t;
+    using weight_t = typename Ktraits::weight_t;
+
+    // Shared memory.
+    __shared__ input_t x_smem[kWidth - 1 + kChunkSizeL][kChunkSizeC + kNElts];
+
+    const int tid = threadIdx.x;
+    const int l_idx = tid / kNThreadsPerC;
+    const int c_idx = tid % kNThreadsPerC;
+    const int batch_id = blockIdx.x;
+    const int chunk_l_id = blockIdx.y;
+    const int chunk_c_id = blockIdx.z;
+    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + batch_id * params.x_batch_stride
+        + (chunk_l_id * kChunkSizeL + l_idx) * params.x_l_stride + chunk_c_id * kChunkSizeC + c_idx * kNElts;
+    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr)
+        + chunk_c_id * kChunkSizeC * params.weight_c_stride;
+    input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + batch_id * params.out_batch_stride
+        + (chunk_l_id * kChunkSizeL + l_idx) * params.out_l_stride + chunk_c_id * kChunkSizeC + c_idx * kNElts;
+
+    #pragma unroll
+    for (int l = 0; l < Ktraits::kNLoads; ++l) {
+        input_t x_vals_load[kNElts] = {0};
+        if (chunk_l_id * kChunkSizeL + l * kLPerLoad + l_idx < params.seqlen
+            && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+            reinterpret_cast<vec_t *>(x_vals_load)[0] = *reinterpret_cast<vec_t *>(x + l * kLPerLoad * params.x_l_stride);
+        }
+        reinterpret_cast<vec_t *>(x_smem[kWidth - 1 + l * kLPerLoad + l_idx])[c_idx] = reinterpret_cast<vec_t *>(x_vals_load)[0];
+    }
+    // Load the elements from the previous chunk that are needed for convolution.
+    if (l_idx < kWidth - 1) {
+        input_t x_vals_load[kNElts] = {0};
+        if (chunk_l_id * kChunkSizeL + l_idx - (kWidth - 1) >= 0
+            && chunk_l_id * kChunkSizeL + l_idx - (kWidth - 1) < params.seqlen
+            && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+            reinterpret_cast<vec_t *>(x_vals_load)[0] = *reinterpret_cast<vec_t *>(x - (kWidth - 1) * params.x_l_stride);
+        }
+        reinterpret_cast<vec_t *>(x_smem[l_idx])[c_idx] = reinterpret_cast<vec_t *>(x_vals_load)[0];
+    }
+
+    __syncthreads();
+
+    constexpr int kLPerThread = std::min(kChunkSizeL * kChunkSizeC / kNThreads, kChunkSizeL);
+    static_assert(kLPerThread * kNThreads == kChunkSizeL * kChunkSizeC);
+    constexpr int kNThreadsPerRow = kChunkSizeL / kLPerThread;
+    static_assert(kNThreadsPerRow * kLPerThread == kChunkSizeL);
+    // kChunkSizeL, kLPerThread, kNThreadsPerRow should be powers of 2 for simplicity
+    static_assert((kChunkSizeL & (kChunkSizeL - 1)) == 0);
+    static_assert((kLPerThread & (kLPerThread - 1)) == 0);
+    static_assert((kNThreadsPerRow & (kNThreadsPerRow - 1)) == 0);
+    static_assert(kNThreadsPerRow <= 32);
+
+    const int row_idx = tid / kNThreadsPerRow;
+    const int col_idx = tid % kNThreadsPerRow;
+
+    float bias_val = params.bias_ptr == nullptr || chunk_c_id * kChunkSizeC + row_idx >= params.dim ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[chunk_c_id * kChunkSizeC + row_idx]);
+    float weight_vals[kWidth] = {0};
+    if (chunk_c_id + kChunkSizeC + row_idx < params.dim) {
+        #pragma unroll
+        for (int w = 0; w < kWidth; ++w) {
+            weight_vals[w] = weight[row_idx * params.weight_c_stride + w * params.weight_width_stride];
+        }
+    }
+    float x_vals[kWidth - 1 + kLPerThread];
+    #pragma unroll
+    for (int i = 0; i < kWidth - 1 + kLPerThread; ++i) {
+        x_vals[i] = float(x_smem[col_idx * kLPerThread + i][row_idx]);
+    }
+
+    float out_vals[kLPerThread];
+    #pragma unroll
+    for (int i = 0; i < kLPerThread; ++i) {
+        out_vals[i] = bias_val;
+        #pragma unroll
+        for (int w = 0; w < kWidth; ++w) { out_vals[i] += weight_vals[w] * x_vals[i + w]; }
+        if (params.silu_activation) {out_vals[i] = out_vals[i] / (1 + expf(-out_vals[i])); }
+    }
+
+    // Since kNThreadsPerRow is a power of 2 and <= 32, we only need syncwarp and not syncthreads.
+    __syncwarp();
+    #pragma unroll
+    for (int i = 0; i < kLPerThread; ++i) { x_smem[col_idx * kLPerThread + i][row_idx] = out_vals[i]; }
+    __syncthreads();
+
+    #pragma unroll
+    for (int l = 0; l < Ktraits::kNLoads; ++l) {
+        input_t out_vals_store[kNElts];
+        reinterpret_cast<vec_t *>(out_vals_store)[0] = reinterpret_cast<vec_t *>(x_smem[l * kLPerLoad + l_idx])[c_idx];
+        if (chunk_l_id * kChunkSizeL + l * kLPerLoad + l_idx < params.seqlen
+            && chunk_c_id * kChunkSizeC + c_idx * kNElts < params.dim) {
+            *reinterpret_cast<vec_t *>(out + l * kLPerLoad * params.out_l_stride) = reinterpret_cast<vec_t *>(out_vals_store)[0];
+        }
+    }
+
+}
+
+template<int kNThreads, int kWidth, typename input_t, typename weight_t>
+void causal_conv1d_channellast_fwd_launch(ConvParamsBase &params, cudaStream_t stream) {
+    using Ktraits = Causal_conv1d_channellast_fwd_kernel_traits<kNThreads, kWidth, 64, true, input_t, weight_t>;
+    // constexpr int kSmemSize = Ktraits::kSmemSize;
+    constexpr int kChunkSizeL = Ktraits::kChunkSizeL;
+    constexpr int kChunkSizeC = Ktraits::kNEltsPerRow;
+    const int n_chunks_L = (params.seqlen + kChunkSizeL - 1) / kChunkSizeL;
+    const int n_chunks_C = (params.dim + kChunkSizeC - 1) / kChunkSizeC;
+    // printf("n_chunks_L: %d, n_chunks_C: %d\n", n_chunks_L, n_chunks_C);
+    dim3 grid(params.batch, n_chunks_L, n_chunks_C);
+    dim3 block(Ktraits::kNThreads);
+    auto kernel = &causal_conv1d_channellast_fwd_kernel<Ktraits>;
+    // if (kSmemSize >= 48 * 1024) {
+    //     C10_CUDA_CHECK(cudaFuncSetAttribute(
+    //         kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+    //     }
+    // kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+    kernel<<<grid, Ktraits::kNThreads, 0, stream>>>(params);
+    C10_CUDA_KERNEL_LAUNCH_CHECK();
+}
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_channellast_fwd_cuda(ConvParamsBase &params, cudaStream_t stream) {
+    if (params.width == 2) {
+        causal_conv1d_channellast_fwd_launch<128, 2, input_t, weight_t>(params, stream);
+    } else if (params.width == 3) {
+        causal_conv1d_channellast_fwd_launch<128, 3, input_t, weight_t>(params, stream);
+    } else if (params.width == 4) {
+        causal_conv1d_channellast_fwd_launch<128, 4, input_t, weight_t>(params, stream);
+    }
+}
+
+template void causal_conv1d_fwd_cuda<float, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::Half, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::BFloat16, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<float, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::Half, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::BFloat16, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<float, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::Half, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_fwd_cuda<at::BFloat16, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+
+template void causal_conv1d_channellast_fwd_cuda<float, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<at::Half, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<at::BFloat16, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<float, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<at::Half, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<at::BFloat16, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<float, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<at::Half, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_channellast_fwd_cuda<at::BFloat16, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);

protein_lm/modeling/models/libs/causal-conv1d/csrc/causal_conv1d_update.cu

@@ -0,0 +1,96 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#include <c10/util/BFloat16.h>
+#include <c10/util/Half.h>
+#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
+
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_store.cuh>
+
+#include "causal_conv1d.h"
+#include "causal_conv1d_common.h"
+#include "static_switch.h"
+
+template<int kNThreads_, int kWidth_, typename input_t_, typename weight_t_>
+struct Causal_conv1d_update_kernel_traits {
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr int kNThreads = kNThreads_;
+    static constexpr int kWidth = kWidth_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+};
+
+template<typename Ktraits>
+__global__ __launch_bounds__(Ktraits::kNThreads)
+void causal_conv1d_update_kernel(ConvParamsBase params) {
+    constexpr int kWidth = Ktraits::kWidth;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    using input_t = typename Ktraits::input_t;
+    using weight_t = typename Ktraits::weight_t;
+
+    const int tidx = threadIdx.x;
+    const int batch_id = blockIdx.x;
+    const int channel_id = blockIdx.y * kNThreads + tidx;
+    input_t *x = reinterpret_cast<input_t *>(params.x_ptr) + batch_id * params.x_batch_stride
+        + channel_id * params.x_c_stride;
+    input_t *conv_state = reinterpret_cast<input_t *>(params.conv_state_ptr) + batch_id * params.conv_state_batch_stride
+        + channel_id * params.conv_state_c_stride;
+    weight_t *weight = reinterpret_cast<weight_t *>(params.weight_ptr) + channel_id * params.weight_c_stride;
+    input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + batch_id * params.out_batch_stride
+        + channel_id * params.out_c_stride;
+    float bias_val = params.bias_ptr == nullptr || channel_id >= params.dim ? 0.f : float(reinterpret_cast<weight_t *>(params.bias_ptr)[channel_id]);
+
+    float weight_vals[kWidth] = {0};
+    if (channel_id < params.dim) {
+        #pragma unroll
+        for (int i = 0; i < kWidth; ++i) { weight_vals[i] = float(weight[i * params.weight_width_stride]); }
+    }
+
+    float x_vals[kWidth] = {0};
+    if (channel_id < params.dim) {
+        #pragma unroll
+        for (int i = 0; i < kWidth - 1; ++i) { x_vals[i] = float(conv_state[(i + 1) * params.conv_state_l_stride]); }
+        x_vals[kWidth - 1] = float(x[0]);
+        #pragma unroll
+        for (int i = 0; i < kWidth; ++i) { conv_state[i * params.conv_state_l_stride] = input_t(x_vals[i]); }
+    }
+
+    float out_val = bias_val;
+    #pragma unroll
+    for (int i = 0; i < kWidth; ++i) { out_val += weight_vals[i] * x_vals[i]; }
+    if (params.silu_activation) { out_val = out_val / (1 + expf(-out_val)); }
+    if (channel_id < params.dim) { out[0] = input_t(out_val); }
+}
+
+template<int kNThreads, int kWidth, typename input_t, typename weight_t>
+void causal_conv1d_update_launch(ConvParamsBase &params, cudaStream_t stream) {
+    using Ktraits = Causal_conv1d_update_kernel_traits<kNThreads, kWidth, input_t, weight_t>;
+    dim3 grid(params.batch, (params.dim + kNThreads - 1) / kNThreads);
+    auto kernel = &causal_conv1d_update_kernel<Ktraits>;
+    kernel<<<grid, Ktraits::kNThreads, 0, stream>>>(params);
+    C10_CUDA_KERNEL_LAUNCH_CHECK();
+}
+
+template<typename input_t, typename weight_t>
+void causal_conv1d_update_cuda(ConvParamsBase &params, cudaStream_t stream) {
+    if (params.width == 2) {
+        causal_conv1d_update_launch<64, 2, input_t, weight_t>(params, stream);
+    } else if (params.width == 3) {
+        causal_conv1d_update_launch<64, 3, input_t, weight_t>(params, stream);
+    } else if (params.width == 4) {
+        causal_conv1d_update_launch<64, 4, input_t, weight_t>(params, stream);
+    }
+}
+
+template void causal_conv1d_update_cuda<float, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::Half, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::BFloat16, float>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<float, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::Half, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::BFloat16, at::Half>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<float, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::Half, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);
+template void causal_conv1d_update_cuda<at::BFloat16, at::BFloat16>(ConvParamsBase &params, cudaStream_t stream);

protein_lm/modeling/models/libs/causal-conv1d/csrc/static_switch.h

@@ -0,0 +1,25 @@
+// Inspired by https://github.com/NVIDIA/DALI/blob/main/include/dali/core/static_switch.h
+// and https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/Dispatch.h
+
+#pragma once
+
+/// @param COND       - a boolean expression to switch by
+/// @param CONST_NAME - a name given for the constexpr bool variable.
+/// @param ...       - code to execute for true and false
+///
+/// Usage:
+/// ```
+/// BOOL_SWITCH(flag, BoolConst, [&] {
+///     some_function<BoolConst>(...);
+/// });
+/// ```
+#define BOOL_SWITCH(COND, CONST_NAME, ...)                                           \
+    [&] {                                                                            \
+        if (COND) {                                                                  \
+            static constexpr bool CONST_NAME = true;                                 \
+            return __VA_ARGS__();                                                    \
+        } else {                                                                     \
+            static constexpr bool CONST_NAME = false;                                \
+            return __VA_ARGS__();                                                    \
+        }                                                                            \
+    }()

protein_lm/modeling/models/libs/causal-conv1d/setup.py

@@ -0,0 +1,264 @@
+# Copyright (c) 2023, Tri Dao.
+import sys
+import warnings
+import os
+import re
+import ast
+from pathlib import Path
+from packaging.version import parse, Version
+import platform
+
+from setuptools import setup, find_packages
+import subprocess
+
+import urllib.request
+import urllib.error
+from wheel.bdist_wheel import bdist_wheel as _bdist_wheel
+
+import torch
+from torch.utils.cpp_extension import (
+    BuildExtension,
+    CppExtension,
+    CUDAExtension,
+    CUDA_HOME,
+)
+
+
+with open("README.md", "r", encoding="utf-8") as fh:
+    long_description = fh.read()
+
+
+# ninja build does not work unless include_dirs are abs path
+this_dir = os.path.dirname(os.path.abspath(__file__))
+
+PACKAGE_NAME = "causal_conv1d"
+
+BASE_WHEEL_URL = "https://github.com/Dao-AILab/causal-conv1d/releases/download/{tag_name}/{wheel_name}"
+
+# FORCE_BUILD: Force a fresh build locally, instead of attempting to find prebuilt wheels
+# SKIP_CUDA_BUILD: Intended to allow CI to use a simple `python setup.py sdist` run to copy over raw files, without any cuda compilation
+FORCE_BUILD = os.getenv("CAUSAL_CONV1D_FORCE_BUILD", "FALSE") == "TRUE"
+SKIP_CUDA_BUILD = os.getenv("CAUSAL_CONV1D_SKIP_CUDA_BUILD", "FALSE") == "TRUE"
+# For CI, we want the option to build with C++11 ABI since the nvcr images use C++11 ABI
+FORCE_CXX11_ABI = os.getenv("CAUSAL_CONV1D_FORCE_CXX11_ABI", "FALSE") == "TRUE"
+
+
+def get_platform():
+    """
+    Returns the platform name as used in wheel filenames.
+    """
+    if sys.platform.startswith("linux"):
+        return "linux_x86_64"
+    elif sys.platform == "darwin":
+        mac_version = ".".join(platform.mac_ver()[0].split(".")[:2])
+        return f"macosx_{mac_version}_x86_64"
+    elif sys.platform == "win32":
+        return "win_amd64"
+    else:
+        raise ValueError("Unsupported platform: {}".format(sys.platform))
+
+
+def get_cuda_bare_metal_version(cuda_dir):
+    raw_output = subprocess.check_output(
+        [cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True
+    )
+    output = raw_output.split()
+    release_idx = output.index("release") + 1
+    bare_metal_version = parse(output[release_idx].split(",")[0])
+
+    return raw_output, bare_metal_version
+
+
+def check_if_cuda_home_none(global_option: str) -> None:
+    if CUDA_HOME is not None:
+        return
+    # warn instead of error because user could be downloading prebuilt wheels, so nvcc won't be necessary
+    # in that case.
+    warnings.warn(
+        f"{global_option} was requested, but nvcc was not found.  Are you sure your environment has nvcc available?  "
+        "If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, "
+        "only images whose names contain 'devel' will provide nvcc."
+    )
+
+
+def append_nvcc_threads(nvcc_extra_args):
+    return nvcc_extra_args + ["--threads", "4"]
+
+
+cmdclass = {}
+ext_modules = []
+
+if not SKIP_CUDA_BUILD:
+    print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
+    TORCH_MAJOR = int(torch.__version__.split(".")[0])
+    TORCH_MINOR = int(torch.__version__.split(".")[1])
+
+    check_if_cuda_home_none("causal_conv1d")
+    # Check, if CUDA11 is installed for compute capability 8.0
+    cc_flag = []
+    if CUDA_HOME is not None:
+        _, bare_metal_version = get_cuda_bare_metal_version(CUDA_HOME)
+        if bare_metal_version < Version("11.6"):
+            raise RuntimeError(
+                "causal_conv1d is only supported on CUDA 11.6 and above.  "
+                "Note: make sure nvcc has a supported version by running nvcc -V."
+            )
+
+    cc_flag.append("-gencode")
+    cc_flag.append("arch=compute_70,code=sm_70")
+    cc_flag.append("-gencode")
+    cc_flag.append("arch=compute_80,code=sm_80")
+    if bare_metal_version >= Version("11.8"):
+        cc_flag.append("-gencode")
+        cc_flag.append("arch=compute_90,code=sm_90")
+
+    # HACK: The compiler flag -D_GLIBCXX_USE_CXX11_ABI is set to be the same as
+    # torch._C._GLIBCXX_USE_CXX11_ABI
+    # https://github.com/pytorch/pytorch/blob/8472c24e3b5b60150096486616d98b7bea01500b/torch/utils/cpp_extension.py#L920
+    if FORCE_CXX11_ABI:
+        torch._C._GLIBCXX_USE_CXX11_ABI = True
+
+    ext_modules.append(
+        CUDAExtension(
+            name="causal_conv1d_cuda",
+            sources=[
+                "csrc/causal_conv1d.cpp",
+                "csrc/causal_conv1d_fwd.cu",
+                "csrc/causal_conv1d_bwd.cu",
+                "csrc/causal_conv1d_update.cu",
+            ],
+            extra_compile_args={
+                "cxx": ["-O3"],
+                "nvcc": append_nvcc_threads(
+                    [
+                        "-O3",
+                        "-U__CUDA_NO_HALF_OPERATORS__",
+                        "-U__CUDA_NO_HALF_CONVERSIONS__",
+                        "-U__CUDA_NO_BFLOAT16_OPERATORS__",
+                        "-U__CUDA_NO_BFLOAT16_CONVERSIONS__",
+                        "-U__CUDA_NO_BFLOAT162_OPERATORS__",
+                        "-U__CUDA_NO_BFLOAT162_CONVERSIONS__",
+                        "--expt-relaxed-constexpr",
+                        "--expt-extended-lambda",
+                        "--use_fast_math",
+                        "--ptxas-options=-v",
+                        "-lineinfo",
+                    ]
+                    + cc_flag
+                ),
+            },
+            include_dirs=[this_dir],
+        )
+    )
+
+
+def get_package_version():
+    with open(Path(this_dir) / "causal_conv1d" / "__init__.py", "r") as f:
+        version_match = re.search(r"^__version__\s*=\s*(.*)$", f.read(), re.MULTILINE)
+    public_version = ast.literal_eval(version_match.group(1))
+    local_version = os.environ.get("CAUSAL_CONV1D_LOCAL_VERSION")
+    if local_version:
+        return f"{public_version}+{local_version}"
+    else:
+        return str(public_version)
+
+
+def get_wheel_url():
+    # Determine the version numbers that will be used to determine the correct wheel
+    # We're using the CUDA version used to build torch, not the one currently installed
+    # _, cuda_version_raw = get_cuda_bare_metal_version(CUDA_HOME)
+    torch_cuda_version = parse(torch.version.cuda)
+    torch_version_raw = parse(torch.__version__)
+    # For CUDA 11, we only compile for CUDA 11.8, and for CUDA 12 we only compile for CUDA 12.2
+    # to save CI time. Minor versions should be compatible.
+    torch_cuda_version = parse("11.8") if torch_cuda_version.major == 11 else parse("12.2")
+    python_version = f"cp{sys.version_info.major}{sys.version_info.minor}"
+    platform_name = get_platform()
+    causal_conv1d_version = get_package_version()
+    # cuda_version = f"{cuda_version_raw.major}{cuda_version_raw.minor}"
+    cuda_version = f"{torch_cuda_version.major}{torch_cuda_version.minor}"
+    torch_version = f"{torch_version_raw.major}.{torch_version_raw.minor}"
+    cxx11_abi = str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()
+
+    # Determine wheel URL based on CUDA version, torch version, python version and OS
+    wheel_filename = f"{PACKAGE_NAME}-{causal_conv1d_version}+cu{cuda_version}torch{torch_version}cxx11abi{cxx11_abi}-{python_version}-{python_version}-{platform_name}.whl"
+    wheel_url = BASE_WHEEL_URL.format(
+        tag_name=f"v{causal_conv1d_version}", wheel_name=wheel_filename
+    )
+    return wheel_url, wheel_filename
+
+
+class CachedWheelsCommand(_bdist_wheel):
+    """
+    The CachedWheelsCommand plugs into the default bdist wheel, which is ran by pip when it cannot
+    find an existing wheel (which is currently the case for all installs). We use
+    the environment parameters to detect whether there is already a pre-built version of a compatible
+    wheel available and short-circuits the standard full build pipeline.
+    """
+
+    def run(self):
+        if FORCE_BUILD:
+            return super().run()
+
+        wheel_url, wheel_filename = get_wheel_url()
+        print("Guessing wheel URL: ", wheel_url)
+        try:
+            urllib.request.urlretrieve(wheel_url, wheel_filename)
+
+            # Make the archive
+            # Lifted from the root wheel processing command
+            # https://github.com/pypa/wheel/blob/cf71108ff9f6ffc36978069acb28824b44ae028e/src/wheel/bdist_wheel.py#LL381C9-L381C85
+            if not os.path.exists(self.dist_dir):
+                os.makedirs(self.dist_dir)
+
+            impl_tag, abi_tag, plat_tag = self.get_tag()
+            archive_basename = f"{self.wheel_dist_name}-{impl_tag}-{abi_tag}-{plat_tag}"
+
+            wheel_path = os.path.join(self.dist_dir, archive_basename + ".whl")
+            print("Raw wheel path", wheel_path)
+            os.rename(wheel_filename, wheel_path)
+        except urllib.error.HTTPError:
+            print("Precompiled wheel not found. Building from source...")
+            # If the wheel could not be downloaded, build from source
+            super().run()
+
+
+setup(
+    name=PACKAGE_NAME,
+    version=get_package_version(),
+    packages=find_packages(
+        exclude=(
+            "build",
+            "csrc",
+            "include",
+            "tests",
+            "dist",
+            "docs",
+            "benchmarks",
+            "causal_conv1d.egg-info",
+        )
+    ),
+    author="Tri Dao",
+    author_email="[email protected]",
+    description="Causal depthwise conv1d in CUDA, with a PyTorch interface",
+    long_description=long_description,
+    long_description_content_type="text/markdown",
+    url="https://github.com/Dao-AILab/causal-conv1d",
+    classifiers=[
+        "Programming Language :: Python :: 3",
+        "License :: OSI Approved :: BSD License",
+        "Operating System :: Unix",
+    ],
+    ext_modules=ext_modules,
+    cmdclass={"bdist_wheel": CachedWheelsCommand, "build_ext": BuildExtension}
+    if ext_modules
+    else {
+        "bdist_wheel": CachedWheelsCommand,
+    },
+    python_requires=">=3.7",
+    install_requires=[
+        "torch",
+        "packaging",
+        "ninja",
+    ],
+)

protein_lm/modeling/models/libs/causal-conv1d/tests/test_causal_conv1d.py

@@ -0,0 +1,173 @@
+# Copyright (C) 2023, Tri Dao.
+
+import math
+
+import torch
+import pytest
+
+from einops import rearrange
+
+from causal_conv1d.causal_conv1d_interface import causal_conv1d_fn, causal_conv1d_ref
+from causal_conv1d.causal_conv1d_interface import causal_conv1d_update, causal_conv1d_update_ref
+
+
+@pytest.mark.parametrize("channel_last", [False, True])
+# @pytest.mark.parametrize('channel_last', [True])
+@pytest.mark.parametrize("itype", [torch.float32, torch.float16, torch.bfloat16])
+# @pytest.mark.parametrize('itype', [torch.float16])
+@pytest.mark.parametrize("silu_activation", [False, True])
+# @pytest.mark.parametrize('silu_activation', [True])
+@pytest.mark.parametrize("has_bias", [False, True])
+# @pytest.mark.parametrize('has_bias', [True])
+@pytest.mark.parametrize("width", [2, 3, 4])
+# @pytest.mark.parametrize('width', [2])
+@pytest.mark.parametrize(
+    "seqlen", [8, 16, 32, 64, 128, 151, 256, 372, 512, 784, 1024, 1134, 2048, 4096]
+)
+# @pytest.mark.parametrize('seqlen', [8, 16, 32, 64, 128, 256, 512, 784, 1024, 2048, 4096])
+# @pytest.mark.parametrize('seqlen', [128])
+def test_causal_conv1d(seqlen, width, has_bias, silu_activation, itype, channel_last):
+    device = "cuda"
+    rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (3e-3, 5e-3)
+    if itype == torch.bfloat16:
+        rtol, atol = 1e-2, 5e-2
+    rtolw, atolw = (1e-3, 1e-3)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 2
+    # batch_size = 1
+    dim = 4096 + 32  # Try dim not divisible by 64
+    # dim = 64
+    if not channel_last:
+        x = torch.randn(batch_size, 4096 + dim + 64, seqlen, device=device, dtype=itype)[:, 4096:4096 + dim, :].requires_grad_()
+    else:
+        x = rearrange(
+            torch.randn(batch_size, seqlen, 4096 + dim + 64, device=device, dtype=itype)[:, :, 4096:4096 + dim], "b s d -> b d s"
+        ).requires_grad_()
+    weight = torch.randn(dim, width, device=device, dtype=torch.float32, requires_grad=True)
+    if has_bias:
+        bias = torch.randn(dim, device=device, dtype=torch.float32, requires_grad=True)
+    else:
+        bias = None
+    x_ref = x.detach().clone().requires_grad_()
+    weight_ref = weight.detach().clone().requires_grad_()
+    bias_ref = bias.detach().clone().requires_grad_() if bias is not None else None
+    activation = None if not silu_activation else "silu"
+    out = causal_conv1d_fn(x, weight, bias, activation=activation)
+    out_ref = causal_conv1d_ref(x_ref, weight_ref, bias_ref, activation=activation)
+
+    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
+    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
+    assert torch.allclose(out, out_ref, rtol=rtol, atol=atol)
+
+    g = torch.randn_like(out)
+    out_ref.backward(g)
+    out.backward(g)
+
+    print(f"dx max diff: {(x.grad - x_ref.grad).abs().max().item()}")
+    print(f"dweight max diff: {(weight.grad - weight_ref.grad).abs().max().item()}")
+    if has_bias:
+        print(f"dbias max diff: {(bias.grad - bias_ref.grad).abs().max().item()}")
+
+    assert torch.allclose(x.grad, x_ref.grad.to(dtype=itype), rtol=rtol, atol=atol)
+    assert torch.allclose(weight.grad, weight_ref.grad, rtol=rtolw, atol=atolw)
+    if has_bias:
+        assert torch.allclose(bias.grad, bias_ref.grad, rtol=rtolw, atol=atolw)
+
+
+@pytest.mark.parametrize("itype", [torch.float32, torch.float16, torch.bfloat16])
+# @pytest.mark.parametrize('itype', [torch.float16])
+@pytest.mark.parametrize("silu_activation", [False, True])
+# @pytest.mark.parametrize('silu_activation', [False])
+@pytest.mark.parametrize("has_bias", [False, True])
+# @pytest.mark.parametrize('has_bias', [True])
+@pytest.mark.parametrize("width", [2, 3, 4])
+# @pytest.mark.parametrize('width', [2])
+@pytest.mark.parametrize("dim", [2048, 2048 + 16, 4096])
+# @pytest.mark.parametrize("dim", [2048])
+def test_causal_conv1d_update(dim, width, has_bias, silu_activation, itype):
+    device = "cuda"
+    rtol, atol = (3e-4, 1e-3) if itype == torch.float32 else (3e-3, 5e-3)
+    if itype == torch.bfloat16:
+        rtol, atol = 1e-2, 5e-2
+    rtolw, atolw = (1e-3, 1e-3)
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 2
+    # batch_size = 1
+    # dim = 64
+    x = torch.randn(batch_size, dim, device=device, dtype=itype)
+    conv_state = torch.randn(batch_size, dim, width, device=device, dtype=itype)
+    weight = torch.randn(dim, width, device=device, dtype=torch.float32, requires_grad=True)
+    if has_bias:
+        bias = torch.randn(dim, device=device, dtype=torch.float32, requires_grad=True)
+    else:
+        bias = None
+    conv_state_ref = conv_state.detach().clone()
+    activation = None if not silu_activation else "silu"
+    out = causal_conv1d_update(x, conv_state, weight, bias, activation=activation)
+    out_ref = causal_conv1d_update_ref(x, conv_state_ref, weight, bias, activation=activation)
+
+    print(f"Output max diff: {(out - out_ref).abs().max().item()}")
+    print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
+    assert torch.equal(conv_state, conv_state_ref)
+    assert torch.allclose(out, out_ref, rtol=rtol, atol=atol)
+
+
+# @pytest.mark.parametrize("channel_last", [False, True])
+@pytest.mark.parametrize('channel_last', [True])
+# @pytest.mark.parametrize("itype", [torch.float32, torch.float16, torch.bfloat16])
+@pytest.mark.parametrize('itype', [torch.bfloat16])
+# @pytest.mark.parametrize("silu_activation", [False, True])
+@pytest.mark.parametrize('silu_activation', [True])
+# @pytest.mark.parametrize("has_bias", [False, True])
+@pytest.mark.parametrize('has_bias', [True])
+# @pytest.mark.parametrize("width", [2, 3, 4])
+@pytest.mark.parametrize('width', [4])
+@pytest.mark.parametrize(
+    # "seqlen", [8, 16, 32, 64, 128, 151, 256, 372, 512, 784, 1024, 1134, 2048, 4096]
+    "seqlen", [2048]
+)
+# @pytest.mark.parametrize('seqlen', [8, 16, 32, 64, 128, 256, 512, 784, 1024, 2048, 4096])
+# @pytest.mark.parametrize('seqlen', [128])
+def test_causal_conv1d_race_condition(seqlen, width, has_bias, silu_activation, itype, channel_last):
+    device = "cuda"
+    # set seed
+    torch.random.manual_seed(0)
+    batch_size = 2
+    # batch_size = 1
+    dim = 4096 + 32  # Try dim not divisible by 64
+    # dim = 64
+    if not channel_last:
+        x = torch.randn(batch_size, 4096 + dim + 64, seqlen, device=device, dtype=itype)[:, 4096:4096 + dim, :].requires_grad_()
+    else:
+        x = rearrange(
+            torch.randn(batch_size, seqlen, 4096 + dim + 64, device=device, dtype=itype)[:, :, 4096:4096 + dim], "b s d -> b d s"
+        ).requires_grad_()
+    weight = torch.randn(dim, width, device=device, dtype=torch.float32, requires_grad=True)
+    if has_bias:
+        bias = torch.randn(dim, device=device, dtype=torch.float32, requires_grad=True)
+    else:
+        bias = None
+    activation = None if not silu_activation else "silu"
+    out0 = causal_conv1d_fn(x, weight, bias, activation=activation)
+    g = torch.randn_like(out0)
+    dx0, dw0, db0 = torch.autograd.grad(out0, (x, weight, bias), g)
+    dw_atol = 1e-4
+    db_atol = 1e-4
+
+    for i in range(10000):
+        out = causal_conv1d_fn(x, weight, bias, activation=activation)
+        dx, dw, db = torch.autograd.grad(out, (x, weight, bias), g)
+        dw_equal = torch.allclose(dw, dw0, atol=dw_atol)
+        # if not dw_equal:
+        #     breakpoint()
+        if has_bias:
+            db_equal = torch.allclose(db, db0, atol=db_atol)
+            # if not db_equal:
+            #     breakpoint()
+        assert torch.equal(out, out0)
+        assert torch.equal(dx, dx0)
+        assert dw_equal
+        if has_bias:
+            assert dw_equal

protein_lm/modeling/models/libs/mamba/.gitmodules

@@ -0,0 +1,3 @@
+[submodule "3rdparty/lm-evaluation-harness"]
+	path = 3rdparty/lm-evaluation-harness
+	url = https://github.com/EleutherAI/lm-evaluation-harness/

protein_lm/modeling/models/libs/mamba/AUTHORS

@@ -0,0 +1,2 @@
+Tri Dao, [email protected]
+Albert Gu, [email protected]

protein_lm/modeling/models/libs/mamba/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2023 Tri Dao, Albert Gu
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

protein_lm/modeling/models/libs/mamba/README.md

@@ -0,0 +1,149 @@
+# Mamba
+
+![Mamba](assets/selection.png "Selective State Space")
+> **Mamba: Linear-Time Sequence Modeling with Selective State Spaces**\
+> Albert Gu*, Tri Dao*\
+> Paper: https://arxiv.org/abs/2312.00752
+
+## About
+
+Mamba is a new state space model architecture showing promising performance on information-dense data such as language modeling, where previous subquadratic models fall short of Transformers.
+It is based on the line of progress on [structured state space models](https://github.com/state-spaces/s4),
+with an efficient hardware-aware design and implementation in the spirit of [FlashAttention](https://github.com/Dao-AILab/flash-attention).
+
+## Installation
+
+- `pip install causal-conv1d`: an efficient implementation of a simple causal Conv1d layer used inside the Mamba block.
+- `pip install mamba-ssm`: the core Mamba package.
+
+It can also be built from source with `pip install .` from this repository.
+
+If `pip` complains about PyTorch versions, try passing `--no-build-isolation` to `pip`.
+
+Other requirements:
+- Linux
+- NVIDIA GPU
+- PyTorch 1.12+
+- CUDA 11.6+
+
+## Usage
+
+We expose several levels of interface with the Mamba model.
+
+### Selective SSM
+
+Mamba is based on a selective SSM layer, which is the focus of the paper (Section 3; Algorithm 2).
+
+Source: [ops/selective_scan_interface.py](mamba_ssm/ops/selective_scan_interface.py).
+
+### Mamba Block
+
+The main module of this repository is the Mamba architecture block wrapping the selective SSM.
+
+Source: [modules/mamba_simple.py](mamba_ssm/modules/mamba_simple.py).
+
+Usage:
+```
+from mamba_ssm import Mamba
+
+batch, length, dim = 2, 64, 16
+x = torch.randn(batch, length, dim).to("cuda")
+model = Mamba(
+    # This module uses roughly 3 * expand * d_model^2 parameters
+    d_model=dim, # Model dimension d_model
+    d_state=16,  # SSM state expansion factor
+    d_conv=4,    # Local convolution width
+    expand=2,    # Block expansion factor
+).to("cuda")
+y = model(x)
+assert y.shape == x.shape
+```
+
+### Mamba Language Model
+
+Finally, we provide an example of a complete language model: a deep sequence model backbone (with repeating Mamba blocks) + language model head.
+
+Source: [models/mixer_seq_simple.py](mamba_ssm/models/mixer_seq_simple.py).
+
+This is an example of how to integrate Mamba into an end-to-end neural network.
+This example is used in the generation scripts below.
+
+
+
+## Pretrained Models
+
+Pretrained models are uploaded to
+[HuggingFace](https://huggingface.co/state-spaces): `mamba-130m`, `mamba-370m`,
+`mamba-790m`, `mamba-1.4b`, `mamba-2.8b`.
+
+The models will be autodownloaded by the generation script below.
+
+These models were trained on the [Pile](https://huggingface.co/datasets/EleutherAI/pile), and follow the standard model dimensions described by GPT-3 and followed by many open source models:
+
+| Parameters | Layers | Model dim. | 
+|------------|--------|------------|
+| 130M       | 12     | 768        |
+| 370M       | 24     | 1024       |
+| 790M       | 24     | 1536       |
+| 1.4B       | 24     | 2048       |
+| 2.8B       | 32     | 2560       |
+
+(The layer count of Mamba should be doubled, as two Mamba blocks are needed for each "layer" (MHA block + MLP block) of a Transformer.)
+
+Note: these are base models trained only for 300B tokens, without any form of downstream modification (instruction tuning, etc.).
+Performance is expected to be comparable or better than other architectures trained on similar data, but not to match larger or fine-tuned models.
+
+
+## Evaluations
+
+To run zero-shot evaluations of models (corresponding to Table 3 of the paper),
+we use the
+[lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/big-refactor)
+library.
+
+1. Pull the `lm-evaluation-harness` repo by `git submodule update --init
+   --recursive`. We use the `big-refactor` branch.
+2. Install `lm-evaluation-harness`: `pip install -e 3rdparty/lm-evaluation-harness`
+3. Run evaluation with (more documentation at the [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness/tree/big-refactor) repo):
+```
+python evals/lm_harness_eval.py --model mamba --model_args pretrained=state-spaces/mamba-130m --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande --device cuda --batch_size 64
+python evals/lm_harness_eval.py --model hf --model_args pretrained=EleutherAI/pythia-160m --tasks lambada_openai,hellaswag,piqa,arc_easy,arc_challenge,winogrande --device cuda --batch_size 64
+```
+
+Note that the result of each task might differ from reported values by 0.1-0.3 due to noise in the evaluation process.
+
+## Inference
+
+The script [benchmarks/benchmark_generation_mamba_simple.py](benchmarks/benchmark_generation_mamba_simple.py)
+1. autoloads a model from the HuggingFace Hub,
+2. generates completions of a user-specified prompt,
+3. benchmarks the inference speed of this generation.
+
+Other configurable options include the top-p (nucleus sampling) probability, and the softmax temperature.
+
+### Examples
+
+To test generation latency (e.g. batch size = 1) with different sampling strategies:
+
+```
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba-2.8b" --prompt "My cat wrote all this CUDA code for a new language model and" --topp 0.9 --temperature 0.5
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "EleutherAI/pythia-2.8b" --prompt "My cat wrote all this CUDA code for a new language model and" --topp 0.9 --temperature 0.5
+```
+
+To test generation throughput with random prompts (e.g. large batch size):
+```
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "state-spaces/mamba-2.8b" --batch 128
+python benchmarks/benchmark_generation_mamba_simple.py --model-name "EleutherAI/pythia-2.8b" --batch 128
+```
+
+## Citation
+
+If you use this codebase, or otherwise found our work valuable, please cite Mamba:
+```
+@article{mamba,
+  title={Mamba: Linear-Time Sequence Modeling with Selective State Spaces},
+  author={Gu, Albert and Dao, Tri},
+  journal={arXiv preprint arXiv:2312.00752},
+  year={2023}
+}
+```

protein_lm/modeling/models/libs/mamba/benchmarks/benchmark_generation_mamba_simple.py

@@ -0,0 +1,88 @@
+# Copyright (c) 2023, Tri Dao, Albert Gu.
+
+import argparse
+import time
+import json
+
+import torch
+import torch.nn.functional as F
+
+from einops import rearrange
+
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
+
+
+parser = argparse.ArgumentParser(description="Generation benchmarking")
+parser.add_argument("--model-name", type=str, default="state-spaces/mamba-130m")
+parser.add_argument("--prompt", type=str, default=None)
+parser.add_argument("--promptlen", type=int, default=100)
+parser.add_argument("--genlen", type=int, default=100)
+parser.add_argument("--temperature", type=float, default=1.0)
+parser.add_argument("--topk", type=int, default=1)
+parser.add_argument("--topp", type=float, default=1.0)
+parser.add_argument("--batch", type=int, default=1)
+args = parser.parse_args()
+
+repeats = 3
+device = "cuda"
+dtype = torch.float16
+
+print(f"Loading model {args.model_name}")
+is_mamba = args.model_name.startswith("state-spaces/mamba-") or "mamba" in args.model_name
+
+if is_mamba:
+    tokenizer = AutoTokenizer.from_pretrained("/home/zhulianghui/VisionProjects/mamba/ckpts/gpt-neox-20b-tokenizer")
+    model = MambaLMHeadModel.from_pretrained(args.model_name, device=device, dtype=dtype)
+else:
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+    model = AutoModelForCausalLM.from_pretrained(args.model_name, device_map={"": device}, torch_dtype=dtype)
+model.eval()
+print(f"Number of parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
+
+torch.random.manual_seed(0)
+if args.prompt is None:
+    input_ids = torch.randint(1, 1000, (args.batch, args.promptlen), dtype=torch.long, device="cuda")
+    attn_mask = torch.ones_like(input_ids, dtype=torch.long, device="cuda")
+else:
+    tokens = tokenizer(args.prompt, return_tensors="pt")
+    input_ids = tokens.input_ids.to(device=device)
+    attn_mask = tokens.attention_mask.to(device=device)
+max_length = input_ids.shape[1] + args.genlen
+
+if is_mamba:
+    fn = lambda: model.generate(
+        input_ids=input_ids,
+        max_length=max_length,
+        cg=True,
+        return_dict_in_generate=True,
+        output_scores=True,
+        enable_timing=False,
+        temperature=args.temperature,
+        top_k=args.topk,
+        top_p=args.topp,
+    )
+else:
+    fn = lambda: model.generate(
+        input_ids=input_ids,
+        attention_mask=attn_mask,
+        max_length=max_length,
+        return_dict_in_generate=True,
+        pad_token_id=tokenizer.eos_token_id,
+        do_sample=True,
+        temperature=args.temperature,
+        top_k=args.topk,
+        top_p=args.topp,
+    )
+out = fn()
+if args.prompt is not None:
+    print(tokenizer.batch_decode(out.sequences.tolist()))
+
+torch.cuda.synchronize()
+start = time.time()
+for _ in range(repeats):
+    fn()
+torch.cuda.synchronize()
+print(f"Prompt length: {len(input_ids[0])}, generation length: {len(out.sequences[0]) - len(input_ids[0])}")
+print(f"{args.model_name} prompt processing + decoding time: {(time.time() - start) / repeats * 1000:.0f}ms")

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/reverse_scan.cuh

@@ -0,0 +1,401 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include <cub/config.cuh>
+
+#include <cub/util_ptx.cuh>
+#include <cub/util_type.cuh>
+#include <cub/block/block_raking_layout.cuh>
+// #include <cub/detail/uninitialized_copy.cuh>
+#include "uninitialized_copy.cuh"
+
+/**
+ * Perform a reverse sequential reduction over \p LENGTH elements of the \p input array.  The aggregate is returned.
+ */
+template <
+    int         LENGTH,
+    typename    T,
+    typename    ReductionOp>
+__device__ __forceinline__ T ThreadReverseReduce(const T (&input)[LENGTH], ReductionOp reduction_op) {
+    static_assert(LENGTH > 0);
+    T retval = input[LENGTH - 1];
+    #pragma unroll
+    for (int i = LENGTH - 2; i >= 0; --i) { retval = reduction_op(retval, input[i]); }
+    return retval;
+}
+
+/**
+ * Perform a sequential inclusive postfix reverse scan over the statically-sized \p input array, seeded with the specified \p postfix.  The aggregate is returned.
+ */
+template <
+    int         LENGTH,
+    typename    T,
+    typename    ScanOp>
+__device__ __forceinline__ T ThreadReverseScanInclusive(
+    const T (&input)[LENGTH],
+    T (&output)[LENGTH],
+    ScanOp scan_op,
+    const T postfix)
+{
+    T inclusive = postfix;
+    #pragma unroll
+    for (int i = LENGTH - 1; i >= 0; --i) {
+        inclusive = scan_op(inclusive, input[i]);
+        output[i] = inclusive;
+    }
+}
+
+/**
+ * Perform a sequential exclusive postfix reverse scan over the statically-sized \p input array, seeded with the specified \p postfix.  The aggregate is returned.
+ */
+template <
+    int         LENGTH,
+    typename    T,
+    typename    ScanOp>
+__device__ __forceinline__ T ThreadReverseScanExclusive(
+    const T (&input)[LENGTH],
+    T (&output)[LENGTH],
+    ScanOp scan_op,
+    const T postfix)
+{
+    // Careful, output maybe be aliased to input
+    T exclusive = postfix;
+    T inclusive;
+    #pragma unroll
+    for (int i = LENGTH - 1; i >= 0; --i) {
+        inclusive = scan_op(exclusive, input[i]);
+        output[i] = exclusive;
+        exclusive = inclusive;
+    }
+    return inclusive;
+}
+
+
+/**
+ * \brief WarpReverseScan provides SHFL-based variants of parallel postfix scan of items partitioned across a CUDA thread warp.
+ *
+ * LOGICAL_WARP_THREADS must be a power-of-two
+ */
+template <
+    typename    T,                      ///< Data type being scanned
+    int         LOGICAL_WARP_THREADS    ///< Number of threads per logical warp
+    >
+struct WarpReverseScan {
+    //---------------------------------------------------------------------
+    // Constants and type definitions
+    //---------------------------------------------------------------------
+
+    /// Whether the logical warp size and the PTX warp size coincide
+    static constexpr bool IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(0));
+    /// The number of warp scan steps
+    static constexpr int STEPS = cub::Log2<LOGICAL_WARP_THREADS>::VALUE;
+    static_assert(LOGICAL_WARP_THREADS == 1 << STEPS);
+
+
+    //---------------------------------------------------------------------
+    // Thread fields
+    //---------------------------------------------------------------------
+
+    /// Lane index in logical warp
+    unsigned int lane_id;
+
+    /// Logical warp index in 32-thread physical warp
+    unsigned int warp_id;
+
+    /// 32-thread physical warp member mask of logical warp
+    unsigned int member_mask;
+
+    //---------------------------------------------------------------------
+    // Construction
+    //---------------------------------------------------------------------
+
+    /// Constructor
+    explicit __device__ __forceinline__
+    WarpReverseScan()
+        : lane_id(cub::LaneId())
+        , warp_id(IS_ARCH_WARP ? 0 : (lane_id / LOGICAL_WARP_THREADS))
+        , member_mask(cub::WarpMask<LOGICAL_WARP_THREADS>(warp_id))
+    {
+        if (!IS_ARCH_WARP) {
+            lane_id = lane_id % LOGICAL_WARP_THREADS;
+        }
+    }
+
+
+    /// Broadcast
+    __device__ __forceinline__ T Broadcast(
+        T               input,              ///< [in] The value to broadcast
+        int             src_lane)           ///< [in] Which warp lane is to do the broadcasting
+    {
+        return cub::ShuffleIndex<LOGICAL_WARP_THREADS>(input, src_lane, member_mask);
+    }
+
+
+    /// Inclusive scan
+    template <typename ScanOpT>
+    __device__ __forceinline__ void InclusiveReverseScan(
+        T               input,              ///< [in] Calling thread's input item.
+        T               &inclusive_output,  ///< [out] Calling thread's output item.  May be aliased with \p input.
+        ScanOpT         scan_op)            ///< [in] Binary scan operator
+    {
+        inclusive_output = input;
+        #pragma unroll
+        for (int STEP = 0; STEP < STEPS; STEP++) {
+            int offset = 1 << STEP;
+            T temp = cub::ShuffleDown<LOGICAL_WARP_THREADS>(
+                inclusive_output, offset, LOGICAL_WARP_THREADS - 1, member_mask
+            );
+            // Perform scan op if from a valid peer
+            inclusive_output = static_cast<int>(lane_id) >= LOGICAL_WARP_THREADS - offset
+                ? inclusive_output : scan_op(temp, inclusive_output);
+        }
+    }
+
+    /// Exclusive scan
+    // Get exclusive from inclusive
+    template <typename ScanOpT>
+    __device__ __forceinline__ void ExclusiveReverseScan(
+        T              input,              ///< [in] Calling thread's input item.
+        T              &exclusive_output,  ///< [out] Calling thread's output item.  May be aliased with \p input.
+        ScanOpT        scan_op,            ///< [in] Binary scan operator
+        T              &warp_aggregate)    ///< [out] Warp-wide aggregate reduction of input items.
+    {
+        T inclusive_output;
+        InclusiveReverseScan(input, inclusive_output, scan_op);
+        warp_aggregate = cub::ShuffleIndex<LOGICAL_WARP_THREADS>(inclusive_output, 0, member_mask);
+        // initial value unknown
+        exclusive_output = cub::ShuffleDown<LOGICAL_WARP_THREADS>(
+            inclusive_output, 1, LOGICAL_WARP_THREADS - 1, member_mask
+        );
+    }
+
+    /**
+     * \brief Computes both inclusive and exclusive reverse scans using the specified binary scan functor across the calling warp.  Because no initial value is supplied, the \p exclusive_output computed for the last <em>warp-lane</em> is undefined.
+     */
+    template <typename ScanOpT>
+    __device__ __forceinline__ void ReverseScan(
+        T               input,              ///< [in] Calling thread's input item.
+        T               &inclusive_output,  ///< [out] Calling thread's inclusive-scan output item.
+        T               &exclusive_output,  ///< [out] Calling thread's exclusive-scan output item.
+        ScanOpT         scan_op)            ///< [in] Binary scan operator
+    {
+        InclusiveReverseScan(input, inclusive_output, scan_op);
+        // initial value unknown
+        exclusive_output = cub::ShuffleDown<LOGICAL_WARP_THREADS>(
+            inclusive_output, 1, LOGICAL_WARP_THREADS - 1, member_mask
+        );
+    }
+
+};
+
+/**
+ * \brief BlockReverseScan provides variants of raking-based parallel postfix scan across a CUDA thread block.
+ */
+template <
+    typename    T,              ///< Data type being scanned
+    int         BLOCK_DIM_X,    ///< The thread block length in threads along the X dimension
+    bool        MEMOIZE=false   ///< Whether or not to buffer outer raking scan partials to incur fewer shared memory reads at the expense of higher register pressure
+    >
+struct BlockReverseScan {
+    //---------------------------------------------------------------------
+    // Types and constants
+    //---------------------------------------------------------------------
+
+    /// Constants
+    /// The thread block size in threads
+    static constexpr int BLOCK_THREADS = BLOCK_DIM_X;
+
+    /// Layout type for padded thread block raking grid
+    using BlockRakingLayout = cub::BlockRakingLayout<T, BLOCK_THREADS>;
+    // The number of reduction elements is not a multiple of the number of raking threads for now
+    static_assert(BlockRakingLayout::UNGUARDED);
+
+    /// Number of raking threads
+    static constexpr int RAKING_THREADS = BlockRakingLayout::RAKING_THREADS;
+    /// Number of raking elements per warp synchronous raking thread
+    static constexpr int SEGMENT_LENGTH = BlockRakingLayout::SEGMENT_LENGTH;
+    /// Cooperative work can be entirely warp synchronous
+    static constexpr bool WARP_SYNCHRONOUS = (int(BLOCK_THREADS) == int(RAKING_THREADS));
+
+    ///  WarpReverseScan utility type
+    using WarpReverseScan = WarpReverseScan<T, RAKING_THREADS>;
+
+    /// Shared memory storage layout type
+    struct _TempStorage {
+        typename BlockRakingLayout::TempStorage raking_grid;     ///< Padded thread block raking grid
+    };
+
+
+    /// Alias wrapper allowing storage to be unioned
+    struct TempStorage : cub::Uninitialized<_TempStorage> {};
+
+
+    //---------------------------------------------------------------------
+    // Per-thread fields
+    //---------------------------------------------------------------------
+
+    // Thread fields
+    _TempStorage    &temp_storage;
+    unsigned int    linear_tid;
+    T               cached_segment[SEGMENT_LENGTH];
+
+
+    //---------------------------------------------------------------------
+    // Utility methods
+    //---------------------------------------------------------------------
+
+    /// Performs upsweep raking reduction, returning the aggregate
+    template <typename ScanOp>
+    __device__ __forceinline__ T Upsweep(ScanOp scan_op) {
+        T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);
+        // Read data into registers
+        #pragma unroll
+        for (int i = 0; i < SEGMENT_LENGTH; ++i) { cached_segment[i] = smem_raking_ptr[i]; }
+        T raking_partial = cached_segment[SEGMENT_LENGTH - 1];
+        #pragma unroll
+        for (int i = SEGMENT_LENGTH - 2; i >= 0; --i) {
+            raking_partial = scan_op(raking_partial, cached_segment[i]);
+        }
+        return raking_partial;
+    }
+
+
+    /// Performs exclusive downsweep raking scan
+    template <typename ScanOp>
+    __device__ __forceinline__ void ExclusiveDownsweep(
+        ScanOp          scan_op,
+        T               raking_partial)
+    {
+        T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);
+        // Read data back into registers
+        if (!MEMOIZE) {
+            #pragma unroll
+            for (int i = 0; i < SEGMENT_LENGTH; ++i) { cached_segment[i] = smem_raking_ptr[i]; }
+        }
+        ThreadReverseScanExclusive(cached_segment, cached_segment, scan_op, raking_partial);
+        // Write data back to smem
+        #pragma unroll
+        for (int i = 0; i < SEGMENT_LENGTH; ++i) { smem_raking_ptr[i] = cached_segment[i]; }
+    }
+
+
+    //---------------------------------------------------------------------
+    // Constructors
+    //---------------------------------------------------------------------
+
+    /// Constructor
+    __device__ __forceinline__ BlockReverseScan(
+        TempStorage &temp_storage)
+    :
+        temp_storage(temp_storage.Alias()),
+        linear_tid(cub::RowMajorTid(BLOCK_DIM_X, 1, 1))
+    {}
+
+
+    /// Computes an exclusive thread block-wide postfix scan using the specified binary \p scan_op functor.  Each thread contributes one input element.  the call-back functor \p block_postfix_callback_op is invoked by the first warp in the block, and the value returned by <em>lane</em><sub>0</sub> in that warp is used as the "seed" value that logically postfixes the thread block's scan inputs.  Also provides every thread with the block-wide \p block_aggregate of all inputs.
+    template <
+        typename ScanOp,
+        typename BlockPostfixCallbackOp>
+    __device__ __forceinline__ void ExclusiveReverseScan(
+        T                       input,                          ///< [in] Calling thread's input item
+        T                       &exclusive_output,              ///< [out] Calling thread's output item (may be aliased to \p input)
+        ScanOp                  scan_op,                        ///< [in] Binary scan operator
+        BlockPostfixCallbackOp  &block_postfix_callback_op)     ///< [in-out] <b>[<em>warp</em><sub>0</sub> only]</b> Call-back functor for specifying a thread block-wide postfix to be applied to all inputs.
+    {
+        if (WARP_SYNCHRONOUS) {
+            // Short-circuit directly to warp-synchronous scan
+            T block_aggregate;
+            WarpReverseScan warp_scan;
+            warp_scan.ExclusiveReverseScan(input, exclusive_output, scan_op, block_aggregate);
+            // Obtain warp-wide postfix in lane0, then broadcast to other lanes
+            T block_postfix = block_postfix_callback_op(block_aggregate);
+            block_postfix = warp_scan.Broadcast(block_postfix, 0);
+            exclusive_output = linear_tid == BLOCK_THREADS - 1 ? block_postfix : scan_op(block_postfix, exclusive_output);
+        } else {
+            // Place thread partial into shared memory raking grid
+            T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);
+            detail::uninitialized_copy(placement_ptr, input);
+            cub::CTA_SYNC();
+            // Reduce parallelism down to just raking threads
+            if (linear_tid < RAKING_THREADS) {
+                WarpReverseScan warp_scan;
+                // Raking upsweep reduction across shared partials
+                T upsweep_partial = Upsweep(scan_op);
+                // Warp-synchronous scan
+                T exclusive_partial, block_aggregate;
+                warp_scan.ExclusiveReverseScan(upsweep_partial, exclusive_partial, scan_op, block_aggregate);
+                // Obtain block-wide postfix in lane0, then broadcast to other lanes
+                T block_postfix = block_postfix_callback_op(block_aggregate);
+                block_postfix = warp_scan.Broadcast(block_postfix, 0);
+                // Update postfix with warpscan exclusive partial
+                T downsweep_postfix = linear_tid == RAKING_THREADS - 1
+                    ? block_postfix : scan_op(block_postfix, exclusive_partial);
+                // Exclusive raking downsweep scan
+                ExclusiveDownsweep(scan_op, downsweep_postfix);
+            }
+            cub::CTA_SYNC();
+            // Grab thread postfix from shared memory
+            exclusive_output = *placement_ptr;
+
+            // // Compute warp scan in each warp.
+            // // The exclusive output from the last lane in each warp is invalid.
+            // T inclusive_output;
+            // WarpReverseScan warp_scan;
+            // warp_scan.ReverseScan(input, inclusive_output, exclusive_output, scan_op);
+
+            // // Compute the warp-wide postfix and block-wide aggregate for each warp.  Warp postfix for the last warp is invalid.
+            // T block_aggregate;
+            // T warp_postfix = ComputeWarpPostfix(scan_op, inclusive_output, block_aggregate);
+
+            // // Apply warp postfix to our lane's partial
+            // if (warp_id != 0) {
+            //     exclusive_output = scan_op(warp_postfix, exclusive_output);
+            //     if (lane_id == 0) { exclusive_output = warp_postfix; }
+            // }
+
+            // // Use the first warp to determine the thread block postfix, returning the result in lane0
+            // if (warp_id == 0) {
+            //     T block_postfix = block_postfix_callback_op(block_aggregate);
+            //     if (lane_id == 0) {
+            //         // Share the postfix with all threads
+            //         detail::uninitialized_copy(&temp_storage.block_postfix,
+            //                                   block_postfix);
+
+            //         exclusive_output = block_postfix; // The block postfix is the exclusive output for tid0
+            //     }
+            // }
+
+            // cub::CTA_SYNC();
+
+            // // Incorporate thread block postfix into outputs
+            // T block_postfix = temp_storage.block_postfix;
+            // if (linear_tid > 0) { exclusive_output = scan_op(block_postfix, exclusive_output); }
+        }
+    }
+
+
+    /**
+     * \brief Computes an inclusive block-wide postfix scan using the specified binary \p scan_op functor.  Each thread contributes an array of consecutive input elements.  the call-back functor \p block_postfix_callback_op is invoked by the first warp in the block, and the value returned by <em>lane</em><sub>0</sub> in that warp is used as the "seed" value that logically postfixes the thread block's scan inputs.  Also provides every thread with the block-wide \p block_aggregate of all inputs.
+     */
+    template <
+        int             ITEMS_PER_THREAD,
+        typename        ScanOp,
+        typename        BlockPostfixCallbackOp>
+    __device__ __forceinline__ void InclusiveReverseScan(
+        T                       (&input)[ITEMS_PER_THREAD],     ///< [in] Calling thread's input items
+        T                       (&output)[ITEMS_PER_THREAD],    ///< [out] Calling thread's output items (may be aliased to \p input)
+        ScanOp                  scan_op,                        ///< [in] Binary scan functor
+        BlockPostfixCallbackOp   &block_postfix_callback_op)    ///< [in-out] <b>[<em>warp</em><sub>0</sub> only]</b> Call-back functor for specifying a block-wide postfix to be applied to the logical input sequence.
+    {
+        // Reduce consecutive thread items in registers
+        T thread_postfix = ThreadReverseReduce(input, scan_op);
+        // Exclusive thread block-scan
+        ExclusiveReverseScan(thread_postfix, thread_postfix, scan_op, block_postfix_callback_op);
+        // Inclusive scan in registers with postfix as seed
+        ThreadReverseScanInclusive(input, output, scan_op, thread_postfix);
+    }
+
+};

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan.cpp

@@ -0,0 +1,497 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <torch/extension.h>
+#include <vector>
+
+#include "selective_scan.h"
+
+#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")
+
+#define DISPATCH_ITYPE_FLOAT_AND_HALF_AND_BF16(ITYPE, NAME, ...)                    \
+    if (ITYPE == at::ScalarType::Half) {                                            \
+        using input_t = at::Half;                                                   \
+        __VA_ARGS__();                                                              \
+    } else if (ITYPE == at::ScalarType::BFloat16) {                                 \
+        using input_t = at::BFloat16;                                               \
+        __VA_ARGS__();                                                              \
+    } else if (ITYPE == at::ScalarType::Float)  {                                   \
+        using input_t = float;                                                      \
+        __VA_ARGS__();                                                              \
+    } else {                                                                        \
+        AT_ERROR(#NAME, " not implemented for input type '", toString(ITYPE), "'"); \
+    }
+
+#define DISPATCH_WTYPE_FLOAT_AND_HALF_AND_BF16(WTYPE, NAME, ...)                     \
+    if (WTYPE == at::ScalarType::Half) {                                             \
+        using weight_t = at::Half;                                                   \
+        __VA_ARGS__();                                                               \
+    } else if (WTYPE == at::ScalarType::BFloat16) {                                  \
+        using weight_t = at::BFloat16;                                               \
+        __VA_ARGS__();                                                               \
+    } else if (WTYPE == at::ScalarType::Float)  {                                    \
+        using weight_t = float;                                                      \
+        __VA_ARGS__();                                                               \
+    } else {                                                                         \
+        AT_ERROR(#NAME, " not implemented for weight type '", toString(WTYPE), "'"); \
+    }
+
+#define DISPATCH_WTYPE_FLOAT_AND_COMPLEX(WTYPE, NAME, ...)                           \
+    if (WTYPE == at::ScalarType::Float) {                                            \
+       using weight_t = float;                                                       \
+        __VA_ARGS__();                                                               \
+    } else if (WTYPE == at::ScalarType::ComplexFloat) {                              \
+        using weight_t = c10::complex<float>;                                        \
+        __VA_ARGS__();                                                               \
+    } else {                                                                         \
+        AT_ERROR(#NAME, " not implemented for weight type '", toString(WTYPE), "'"); \
+    }
+
+template<typename input_t, typename weight_t>
+void selective_scan_fwd_cuda(SSMParamsBase &params, cudaStream_t stream);
+
+template <typename input_t, typename weight_t>
+void selective_scan_bwd_cuda(SSMParamsBwd &params, cudaStream_t stream);
+
+void set_ssm_params_fwd(SSMParamsBase &params,
+                        // sizes
+                        const size_t batch,
+                        const size_t dim,
+                        const size_t seqlen,
+                        const size_t dstate,
+                        const size_t n_groups,
+                        const size_t n_chunks,
+                        const bool is_variable_B,
+                        const bool is_variable_C,
+                        // device pointers
+                        const at::Tensor u,
+                        const at::Tensor delta,
+                        const at::Tensor A,
+                        const at::Tensor B,
+                        const at::Tensor C,
+                        const at::Tensor out,
+                        const at::Tensor z,
+                        const at::Tensor out_z,
+                        void* D_ptr,
+                        void* delta_bias_ptr,
+                        void* x_ptr,
+                        bool has_z,
+                        bool delta_softplus) {
+
+    // Reset the parameters
+    memset(&params, 0, sizeof(params));
+
+    params.batch = batch;
+    params.dim = dim;
+    params.seqlen = seqlen;
+    params.dstate = dstate;
+    params.n_groups = n_groups;
+    params.n_chunks = n_chunks;
+    params.dim_ngroups_ratio = dim / n_groups;
+
+    params.delta_softplus = delta_softplus;
+
+    params.is_variable_B = is_variable_B;
+    params.is_variable_C = is_variable_C;
+
+    // Set the pointers and strides.
+    params.u_ptr = u.data_ptr();
+    params.delta_ptr = delta.data_ptr();
+    params.A_ptr = A.data_ptr();
+    params.B_ptr = B.data_ptr();
+    params.C_ptr = C.data_ptr();
+    params.D_ptr = D_ptr;
+    params.delta_bias_ptr = delta_bias_ptr;
+    params.out_ptr = out.data_ptr();
+    params.x_ptr = x_ptr;
+    params.z_ptr = has_z ? z.data_ptr() : nullptr;
+    params.out_z_ptr = has_z ? out_z.data_ptr() : nullptr;
+    // All stride are in elements, not bytes.
+    params.A_d_stride = A.stride(0);
+    params.A_dstate_stride = A.stride(1);
+    if (!is_variable_B) {
+        params.B_d_stride = B.stride(0);
+    } else {
+        params.B_batch_stride = B.stride(0);
+        params.B_group_stride = B.stride(1);
+    }
+    params.B_dstate_stride = !is_variable_B ? B.stride(1) : B.stride(2);
+    if (!is_variable_C) {
+        params.C_d_stride = C.stride(0);
+    } else {
+        params.C_batch_stride = C.stride(0);
+        params.C_group_stride = C.stride(1);
+    }
+    params.C_dstate_stride = !is_variable_C ? C.stride(1) : C.stride(2);
+    params.u_batch_stride = u.stride(0);
+    params.u_d_stride = u.stride(1);
+    params.delta_batch_stride = delta.stride(0);
+    params.delta_d_stride = delta.stride(1);
+    if (has_z) {
+        params.z_batch_stride = z.stride(0);
+        params.z_d_stride = z.stride(1);
+        params.out_z_batch_stride = out_z.stride(0);
+        params.out_z_d_stride = out_z.stride(1);
+    }
+    params.out_batch_stride = out.stride(0);
+    params.out_d_stride = out.stride(1);
+}
+
+void set_ssm_params_bwd(SSMParamsBwd &params,
+                        // sizes
+                        const size_t batch,
+                        const size_t dim,
+                        const size_t seqlen,
+                        const size_t dstate,
+                        const size_t n_groups,
+                        const size_t n_chunks,
+                        const bool is_variable_B,
+                        const bool is_variable_C,
+                        // device pointers
+                        const at::Tensor u,
+                        const at::Tensor delta,
+                        const at::Tensor A,
+                        const at::Tensor B,
+                        const at::Tensor C,
+                        const at::Tensor z,
+                        const at::Tensor out,
+                        const at::Tensor out_z,
+                        void* D_ptr,
+                        void* delta_bias_ptr,
+                        void* x_ptr,
+                        const at::Tensor dout,
+                        const at::Tensor du,
+                        const at::Tensor ddelta,
+                        const at::Tensor dA,
+                        const at::Tensor dB,
+                        const at::Tensor dC,
+                        const at::Tensor dz,
+                        void* dD_ptr,
+                        void* ddelta_bias_ptr,
+                        bool has_z,
+                        bool delta_softplus,
+                        bool recompute_out_z) {
+    // Pass in "dout" instead of "out", we're not gonna use "out" unless we have z
+    set_ssm_params_fwd(params, batch, dim, seqlen, dstate, n_groups, n_chunks, is_variable_B, is_variable_C,
+                       u, delta, A, B, C, has_z ? out : dout,
+                       has_z ? z : dout,
+                       // If not recompute_out_z, pass dout instead of out_z.
+                       // This won't be used by the bwd kernel
+                       recompute_out_z ? out_z : dout,
+                       D_ptr, delta_bias_ptr, x_ptr, has_z, delta_softplus);
+    if (!recompute_out_z) { params.out_z_ptr = nullptr; }
+
+    // Set the pointers and strides.
+    params.dout_ptr = dout.data_ptr();
+    params.du_ptr = du.data_ptr();
+    params.dA_ptr = dA.data_ptr();
+    params.dB_ptr = dB.data_ptr();
+    params.dC_ptr = dC.data_ptr();
+    params.dD_ptr = dD_ptr;
+    params.ddelta_ptr = ddelta.data_ptr();
+    params.ddelta_bias_ptr = ddelta_bias_ptr;
+    params.dz_ptr = has_z ? dz.data_ptr() : nullptr;
+    // All stride are in elements, not bytes.
+    params.dout_batch_stride = dout.stride(0);
+    params.dout_d_stride = dout.stride(1);
+    params.dA_d_stride = dA.stride(0);
+    params.dA_dstate_stride = dA.stride(1);
+    if (!is_variable_B) {
+        params.dB_d_stride = dB.stride(0);
+    } else {
+        params.dB_batch_stride = dB.stride(0);
+        params.dB_group_stride = dB.stride(1);
+    }
+    params.dB_dstate_stride = !is_variable_B ? dB.stride(1) : dB.stride(2);
+    if (!is_variable_C) {
+        params.dC_d_stride = dC.stride(0);
+    } else {
+        params.dC_batch_stride = dC.stride(0);
+        params.dC_group_stride = dC.stride(1);
+    }
+    params.dC_dstate_stride = !is_variable_C ? dC.stride(1) : dC.stride(2);
+    params.du_batch_stride = du.stride(0);
+    params.du_d_stride = du.stride(1);
+    params.ddelta_batch_stride = ddelta.stride(0);
+    params.ddelta_d_stride = ddelta.stride(1);
+    if (has_z) {
+        params.dz_batch_stride = dz.stride(0);
+        params.dz_d_stride = dz.stride(1);
+    }
+}
+
+std::vector<at::Tensor>
+selective_scan_fwd(const at::Tensor &u, const at::Tensor &delta,
+                  const at::Tensor &A, const at::Tensor &B, const at::Tensor &C,
+                  const c10::optional<at::Tensor> &D_,
+                  const c10::optional<at::Tensor> &z_,
+                  const c10::optional<at::Tensor> &delta_bias_,
+                  bool delta_softplus) {
+    auto input_type = u.scalar_type();
+    auto weight_type = A.scalar_type();
+    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::ComplexFloat);
+
+    const bool is_variable_B = B.dim() >= 3;
+    const bool is_variable_C = C.dim() >= 3;
+    const bool is_complex = weight_type == at::ScalarType::ComplexFloat;
+
+    TORCH_CHECK(delta.scalar_type() == input_type);
+    TORCH_CHECK(B.scalar_type() == (!is_variable_B ? weight_type : input_type));
+    TORCH_CHECK(C.scalar_type() == (!is_variable_C ? weight_type : input_type));
+
+    TORCH_CHECK(u.is_cuda());
+    TORCH_CHECK(delta.is_cuda());
+    TORCH_CHECK(A.is_cuda());
+    TORCH_CHECK(B.is_cuda());
+    TORCH_CHECK(C.is_cuda());
+
+    TORCH_CHECK(u.stride(-1) == 1);
+    TORCH_CHECK(delta.stride(-1) == 1);
+
+    const auto sizes = u.sizes();
+    const int batch_size = sizes[0];
+    const int dim = sizes[1];
+    const int seqlen = sizes[2];
+    const int dstate = A.size(1);
+    const int n_groups = is_variable_B ? B.size(1) : 1;
+
+    TORCH_CHECK(dstate <= 256, "selective_scan only supports state dimension <= 256");
+
+    CHECK_SHAPE(u, batch_size, dim, seqlen);
+    CHECK_SHAPE(delta, batch_size, dim, seqlen);
+    CHECK_SHAPE(A, dim, dstate);
+    if (!is_variable_B) {
+        CHECK_SHAPE(B, dim, dstate);
+    } else {
+        CHECK_SHAPE(B, batch_size, n_groups, dstate, !is_complex ? seqlen : seqlen * 2);
+        TORCH_CHECK(B.stride(-1) == 1);
+    }
+    if (!is_variable_C) {
+        CHECK_SHAPE(C, dim, dstate);
+    } else {
+        CHECK_SHAPE(C, batch_size, n_groups, dstate, !is_complex ? seqlen: seqlen * 2);
+        TORCH_CHECK(C.stride(-1) == 1);
+    }
+
+    if (D_.has_value()) {
+        auto D = D_.value();
+        TORCH_CHECK(D.scalar_type() == at::ScalarType::Float);
+        TORCH_CHECK(D.is_cuda());
+        TORCH_CHECK(D.stride(-1) == 1);
+        CHECK_SHAPE(D, dim);
+    }
+
+    if (delta_bias_.has_value()) {
+        auto delta_bias = delta_bias_.value();
+        TORCH_CHECK(delta_bias.scalar_type() == at::ScalarType::Float);
+        TORCH_CHECK(delta_bias.is_cuda());
+        TORCH_CHECK(delta_bias.stride(-1) == 1);
+        CHECK_SHAPE(delta_bias, dim);
+    }
+
+    at::Tensor z, out_z;
+    const bool has_z = z_.has_value();
+    if (has_z) {
+        z = z_.value();
+        TORCH_CHECK(z.scalar_type() == input_type);
+        TORCH_CHECK(z.is_cuda());
+        TORCH_CHECK(z.stride(-1) == 1);
+        CHECK_SHAPE(z, batch_size, dim, seqlen);
+        out_z = torch::empty_like(z);
+    }
+
+    const int n_chunks = (seqlen + 2048 - 1) / 2048;
+    // const int n_chunks = (seqlen + 1024 - 1) / 1024;
+    // at::Tensor out = torch::empty_like(u);
+    // Right now u has BHL layout and delta has HBL layout, and we want out to have HBL layout
+    at::Tensor out = torch::empty_like(delta);
+    at::Tensor x;
+    x = torch::empty({batch_size, dim, n_chunks, dstate * 2}, u.options().dtype(weight_type));
+
+    SSMParamsBase params;
+    set_ssm_params_fwd(params, batch_size, dim, seqlen, dstate, n_groups, n_chunks, is_variable_B, is_variable_C,
+                       u, delta, A, B, C, out, z, out_z,
+                       D_.has_value() ? D_.value().data_ptr() : nullptr,
+                       delta_bias_.has_value() ? delta_bias_.value().data_ptr() : nullptr,
+                       x.data_ptr(),
+                       has_z,
+                       delta_softplus);
+
+    // Otherwise the kernel will be launched from cuda:0 device
+    // Cast to char to avoid compiler warning about narrowing
+    at::cuda::CUDAGuard device_guard{(char)u.get_device()};
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    DISPATCH_ITYPE_FLOAT_AND_HALF_AND_BF16(u.scalar_type(), "selective_scan_fwd", [&] {
+        DISPATCH_WTYPE_FLOAT_AND_COMPLEX(A.scalar_type(), "selective_scan_fwd", [&] {
+            selective_scan_fwd_cuda<input_t, weight_t>(params, stream);
+        });
+    });
+    std::vector<at::Tensor> result = {out, x};
+    if (has_z) { result.push_back(out_z); }
+    return result;
+}
+
+std::vector<at::Tensor>
+selective_scan_bwd(const at::Tensor &u, const at::Tensor &delta,
+                  const at::Tensor &A, const at::Tensor &B, const at::Tensor &C,
+                  const c10::optional<at::Tensor> &D_,
+                  const c10::optional<at::Tensor> &z_,
+                  const c10::optional<at::Tensor> &delta_bias_,
+                  const at::Tensor &dout,
+                  const c10::optional<at::Tensor> &x_,
+                  const c10::optional<at::Tensor> &out_,
+                  c10::optional<at::Tensor> &dz_,
+                  bool delta_softplus,
+                  bool recompute_out_z) {
+    auto input_type = u.scalar_type();
+    auto weight_type = A.scalar_type();
+    TORCH_CHECK(input_type == at::ScalarType::Float || input_type == at::ScalarType::Half || input_type == at::ScalarType::BFloat16);
+    TORCH_CHECK(weight_type == at::ScalarType::Float || weight_type == at::ScalarType::ComplexFloat);
+
+    const bool is_variable_B = B.dim() >= 3;
+    const bool is_variable_C = C.dim() >= 3;
+    const bool is_complex = weight_type == at::ScalarType::ComplexFloat;
+
+    TORCH_CHECK(delta.scalar_type() == input_type);
+    TORCH_CHECK(B.scalar_type() == (!is_variable_B ? weight_type : input_type));
+    TORCH_CHECK(C.scalar_type() == (!is_variable_C ? weight_type : input_type));
+    TORCH_CHECK(dout.scalar_type() == input_type);
+
+    TORCH_CHECK(u.is_cuda());
+    TORCH_CHECK(delta.is_cuda());
+    TORCH_CHECK(A.is_cuda());
+    TORCH_CHECK(B.is_cuda());
+    TORCH_CHECK(C.is_cuda());
+    TORCH_CHECK(dout.is_cuda());
+
+    TORCH_CHECK(u.stride(-1) == 1);
+    TORCH_CHECK(delta.stride(-1) == 1);
+    TORCH_CHECK(dout.stride(-1) == 1);
+
+    const auto sizes = u.sizes();
+    const int batch_size = sizes[0];
+    const int dim = sizes[1];
+    const int seqlen = sizes[2];
+    const int dstate = A.size(1);
+    const int n_groups = is_variable_B ? B.size(1) : 1;
+
+    TORCH_CHECK(dstate <= 256, "selective_scan only supports state dimension <= 256");
+
+    CHECK_SHAPE(u, batch_size, dim, seqlen);
+    CHECK_SHAPE(delta, batch_size, dim, seqlen);
+    CHECK_SHAPE(A, dim, dstate);
+    if (!is_variable_B) {
+        CHECK_SHAPE(B, dim, dstate);
+    } else {
+        CHECK_SHAPE(B, batch_size, n_groups, dstate, !is_complex ? seqlen : seqlen * 2);
+        TORCH_CHECK(B.stride(-1) == 1);
+    }
+    if (!is_variable_C) {
+        CHECK_SHAPE(C, dim, dstate);
+    } else {
+        CHECK_SHAPE(C, batch_size, n_groups, dstate, !is_complex ? seqlen: seqlen * 2);
+        TORCH_CHECK(C.stride(-1) == 1);
+    }
+    CHECK_SHAPE(dout, batch_size, dim, seqlen);
+
+    if (D_.has_value()) {
+        auto D = D_.value();
+        TORCH_CHECK(D.scalar_type() == at::ScalarType::Float);
+        TORCH_CHECK(D.is_cuda());
+        TORCH_CHECK(D.stride(-1) == 1);
+        CHECK_SHAPE(D, dim);
+    }
+
+    if (delta_bias_.has_value()) {
+        auto delta_bias = delta_bias_.value();
+        TORCH_CHECK(delta_bias.scalar_type() == at::ScalarType::Float);
+        TORCH_CHECK(delta_bias.is_cuda());
+        TORCH_CHECK(delta_bias.stride(-1) == 1);
+        CHECK_SHAPE(delta_bias, dim);
+    }
+
+    at::Tensor z, out, dz, out_z;
+    const bool has_z = z_.has_value();
+    if (has_z) {
+        z = z_.value();
+        TORCH_CHECK(z.scalar_type() == input_type);
+        TORCH_CHECK(z.is_cuda());
+        TORCH_CHECK(z.stride(-1) == 1);
+        CHECK_SHAPE(z, batch_size, dim, seqlen);
+
+        TORCH_CHECK(out_.has_value());
+        out = out_.value();
+        TORCH_CHECK(out.scalar_type() == input_type);
+        TORCH_CHECK(out.is_cuda());
+        TORCH_CHECK(out.stride(-1) == 1);
+        CHECK_SHAPE(out, batch_size, dim, seqlen);
+
+        if (dz_.has_value()) {
+            dz = dz_.value();
+            TORCH_CHECK(dz.scalar_type() == input_type);
+            TORCH_CHECK(dz.is_cuda());
+            TORCH_CHECK(dz.stride(-1) == 1);
+            CHECK_SHAPE(dz, batch_size, dim, seqlen);
+        } else {
+            dz = torch::empty_like(z);
+        }
+        if (recompute_out_z) {
+            out_z = torch::empty_like(out);
+        }
+    }
+
+    const int n_chunks = (seqlen + 2048 - 1) / 2048;
+    // const int n_chunks = (seqlen + 1024 - 1) / 1024;
+    if (n_chunks > 1) { TORCH_CHECK(x_.has_value()); }
+    if (x_.has_value()) {
+        auto x = x_.value();
+        TORCH_CHECK(x.scalar_type() == weight_type);
+        TORCH_CHECK(x.is_cuda());
+        TORCH_CHECK(x.is_contiguous());
+        CHECK_SHAPE(x, batch_size, dim, n_chunks, 2 * dstate);
+    }
+
+    at::Tensor du = torch::empty_like(u);
+    at::Tensor ddelta = torch::empty_like(delta);
+    at::Tensor dA = torch::zeros_like(A);
+    at::Tensor dB = !is_variable_B ? torch::zeros_like(B) : torch::zeros_like(B, B.options().dtype(torch::kFloat32));
+    at::Tensor dC = !is_variable_C ? torch::zeros_like(C) : torch::zeros_like(C, C.options().dtype(torch::kFloat32));
+    at::Tensor dD;
+    if (D_.has_value()) { dD = torch::zeros_like(D_.value()); }
+    at::Tensor ddelta_bias;
+    if (delta_bias_.has_value()) { ddelta_bias = torch::zeros_like(delta_bias_.value()); }
+
+    SSMParamsBwd params;
+    set_ssm_params_bwd(params, batch_size, dim, seqlen, dstate, n_groups, n_chunks, is_variable_B, is_variable_C,
+                       u, delta, A, B, C, z, out, out_z,
+                       D_.has_value() ? D_.value().data_ptr() : nullptr,
+                       delta_bias_.has_value() ? delta_bias_.value().data_ptr() : nullptr,
+                       x_.has_value() ? x_.value().data_ptr() : nullptr,
+                       dout, du, ddelta, dA, dB, dC, dz,
+                       D_.has_value() ? dD.data_ptr() : nullptr,
+                       delta_bias_.has_value() ? ddelta_bias.data_ptr() : nullptr,
+                       has_z, delta_softplus, recompute_out_z);
+
+    // Otherwise the kernel will be launched from cuda:0 device
+    // Cast to char to avoid compiler warning about narrowing
+    at::cuda::CUDAGuard device_guard{(char)u.get_device()};
+    auto stream = at::cuda::getCurrentCUDAStream().stream();
+    DISPATCH_ITYPE_FLOAT_AND_HALF_AND_BF16(u.scalar_type(), "selective_scan_bwd", [&] {
+        DISPATCH_WTYPE_FLOAT_AND_COMPLEX(A.scalar_type(), "selective_scan_bwd", [&] {
+            selective_scan_bwd_cuda<input_t, weight_t>(params, stream);
+        });
+    });
+    std::vector<at::Tensor> result = {du, ddelta, dA, dB.to(B.dtype()), dC.to(C.dtype()), dD, ddelta_bias};
+    if (has_z) { result.push_back(dz); }
+    if (recompute_out_z) { result.push_back(out_z); }
+    return result;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("fwd", &selective_scan_fwd, "Selective scan forward");
+    m.def("bwd", &selective_scan_bwd, "Selective scan backward");
+}

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan.h

@@ -0,0 +1,101 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+struct SSMScanParamsBase {
+    using index_t = uint32_t;
+
+    int batch, seqlen, n_chunks;
+    index_t a_batch_stride;
+    index_t b_batch_stride;
+    index_t out_batch_stride;
+
+    // Common data pointers.
+    void *__restrict__ a_ptr;
+    void *__restrict__ b_ptr;
+    void *__restrict__ out_ptr;
+    void *__restrict__ x_ptr;
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+struct SSMParamsBase {
+    using index_t = uint32_t;
+
+    int batch, dim, seqlen, dstate, n_groups, n_chunks;
+    int dim_ngroups_ratio;
+    bool is_variable_B;
+    bool is_variable_C;
+
+    bool delta_softplus;
+
+    index_t A_d_stride;
+    index_t A_dstate_stride;
+    index_t B_batch_stride;
+    index_t B_d_stride;
+    index_t B_dstate_stride;
+    index_t B_group_stride;
+    index_t C_batch_stride;
+    index_t C_d_stride;
+    index_t C_dstate_stride;
+    index_t C_group_stride;
+    index_t u_batch_stride;
+    index_t u_d_stride;
+    index_t delta_batch_stride;
+    index_t delta_d_stride;
+    index_t z_batch_stride;
+    index_t z_d_stride;
+    index_t out_batch_stride;
+    index_t out_d_stride;
+    index_t out_z_batch_stride;
+    index_t out_z_d_stride;
+
+    // Common data pointers.
+    void *__restrict__ A_ptr;
+    void *__restrict__ B_ptr;
+    void *__restrict__ C_ptr;
+    void *__restrict__ D_ptr;
+    void *__restrict__ u_ptr;
+    void *__restrict__ delta_ptr;
+    void *__restrict__ delta_bias_ptr;
+    void *__restrict__ out_ptr;
+    void *__restrict__ x_ptr;
+    void *__restrict__ z_ptr;
+    void *__restrict__ out_z_ptr;
+};
+
+struct SSMParamsBwd: public SSMParamsBase {
+    index_t dout_batch_stride;
+    index_t dout_d_stride;
+    index_t dA_d_stride;
+    index_t dA_dstate_stride;
+    index_t dB_batch_stride;
+    index_t dB_group_stride;
+    index_t dB_d_stride;
+    index_t dB_dstate_stride;
+    index_t dC_batch_stride;
+    index_t dC_group_stride;
+    index_t dC_d_stride;
+    index_t dC_dstate_stride;
+    index_t du_batch_stride;
+    index_t du_d_stride;
+    index_t dz_batch_stride;
+    index_t dz_d_stride;
+    index_t ddelta_batch_stride;
+    index_t ddelta_d_stride;
+
+    // Common data pointers.
+    void *__restrict__ dout_ptr;
+    void *__restrict__ dA_ptr;
+    void *__restrict__ dB_ptr;
+    void *__restrict__ dC_ptr;
+    void *__restrict__ dD_ptr;
+    void *__restrict__ du_ptr;
+    void *__restrict__ dz_ptr;
+    void *__restrict__ ddelta_ptr;
+    void *__restrict__ ddelta_bias_ptr;
+};

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_bwd_bf16_complex.cu

@@ -0,0 +1,9 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_bwd_kernel.cuh"
+
+template void selective_scan_bwd_cuda<at::BFloat16, complex_t>(SSMParamsBwd &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_bwd_bf16_real.cu

@@ -0,0 +1,9 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_bwd_kernel.cuh"
+
+template void selective_scan_bwd_cuda<at::BFloat16, float>(SSMParamsBwd &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_bwd_fp16_complex.cu

@@ -0,0 +1,9 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_bwd_kernel.cuh"
+
+template void selective_scan_bwd_cuda<at::Half, complex_t>(SSMParamsBwd &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_bwd_fp16_real.cu

@@ -0,0 +1,9 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_bwd_kernel.cuh"
+
+template void selective_scan_bwd_cuda<at::Half, float>(SSMParamsBwd &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_bwd_fp32_complex.cu

@@ -0,0 +1,9 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_bwd_kernel.cuh"
+
+template void selective_scan_bwd_cuda<float, complex_t>(SSMParamsBwd &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_bwd_fp32_real.cu

@@ -0,0 +1,9 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_bwd_kernel.cuh"
+
+template void selective_scan_bwd_cuda<float, float>(SSMParamsBwd &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_bwd_kernel.cuh

@@ -0,0 +1,531 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include <c10/util/BFloat16.h>
+#include <c10/util/Half.h>
+#include <c10/cuda/CUDAException.h>  // For C10_CUDA_CHECK and C10_CUDA_KERNEL_LAUNCH_CHECK
+#include <ATen/cuda/Atomic.cuh>  // For atomicAdd on complex
+
+#include <cub/block/block_load.cuh>
+#include <cub/block/block_store.cuh>
+#include <cub/block/block_scan.cuh>
+#include <cub/block/block_reduce.cuh>
+
+#include "selective_scan.h"
+#include "selective_scan_common.h"
+#include "reverse_scan.cuh"
+#include "static_switch.h"
+
+template<typename scalar_t> __device__ __forceinline__ scalar_t conj(scalar_t x);
+template<> __device__ __forceinline__ float conj<float>(float x) { return x; }
+template<> __device__ __forceinline__ complex_t conj<complex_t>(complex_t x) { return std::conj(x); }
+
+template<int kNThreads_, int kNItems_, bool kIsEvenLen_, bool kIsVariableB_, bool kIsVariableC_,
+         bool kDeltaSoftplus_, bool kHasZ_, typename input_t_, typename weight_t_>
+struct Selective_Scan_bwd_kernel_traits {
+    static_assert(kNItems_ % 4 == 0);
+    using input_t = input_t_;
+    using weight_t = weight_t_;
+    static constexpr int kNThreads = kNThreads_;
+    static constexpr int kNItems = kNItems_;
+    static constexpr int kNBytes = sizeof(input_t);
+    static_assert(kNBytes == 2 || kNBytes == 4);
+    static constexpr int kNElts = kNBytes == 4 ? 4 : std::min(8, kNItems);
+    static_assert(kNItems % kNElts == 0);
+    static constexpr int kNLoads = kNItems / kNElts;
+    static constexpr bool kIsComplex = std::is_same_v<weight_t, complex_t>;
+    static constexpr bool kIsEvenLen = kIsEvenLen_;
+    static constexpr bool kIsVariableB = kIsVariableB_;
+    static constexpr bool kIsVariableC = kIsVariableC_;
+    static constexpr bool kDeltaSoftplus = kDeltaSoftplus_;
+    static constexpr bool kHasZ = kHasZ_;
+    // Setting MinBlocksPerMP to be 3 (instead of 2) for 128 threads with float improves occupancy.
+    // For complex this would lead to massive register spilling, so we keep it at 2.
+    static constexpr int kMinBlocks = kNThreads == 128 && !kIsComplex ? 3 : 2;
+    using vec_t = typename BytesToType<kNBytes * kNElts>::Type;
+    using scan_t = std::conditional_t<!kIsComplex, float2, float4>;
+    using BlockLoadT = cub::BlockLoad<input_t, kNThreads, kNItems, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockLoadVecT = cub::BlockLoad<vec_t, kNThreads, kNLoads, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockLoadWeightT = cub::BlockLoad<input_t, kNThreads, !kIsComplex ? kNItems : kNItems * 2, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockLoadWeightVecT = cub::BlockLoad<vec_t, kNThreads, !kIsComplex ? kNLoads : kNLoads * 2, cub::BLOCK_LOAD_WARP_TRANSPOSE>;
+    using BlockStoreT = cub::BlockStore<input_t, kNThreads, kNItems, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+    using BlockStoreVecT = cub::BlockStore<vec_t, kNThreads, kNLoads, cub::BLOCK_STORE_WARP_TRANSPOSE>;
+    // using BlockScanT = cub::BlockScan<scan_t, kNThreads, cub::BLOCK_SCAN_RAKING_MEMOIZE>;
+    using BlockScanT = cub::BlockScan<scan_t, kNThreads, cub::BLOCK_SCAN_RAKING>;
+    // using BlockScanT = cub::BlockScan<scan_t, kNThreads, cub::BLOCK_SCAN_WARP_SCANS>;
+    using BlockReverseScanT = BlockReverseScan<scan_t, kNThreads>;
+    using BlockReduceT = cub::BlockReduce<scan_t, kNThreads>;
+    using BlockReduceFloatT = cub::BlockReduce<float, kNThreads>;
+    using BlockReduceComplexT = cub::BlockReduce<complex_t, kNThreads>;
+    using BlockExchangeT = cub::BlockExchange<float, kNThreads, !kIsComplex ? kNItems : kNItems * 2>;
+    static constexpr int kSmemIOSize = std::max({sizeof(typename BlockLoadT::TempStorage),
+                                                 sizeof(typename BlockLoadVecT::TempStorage),
+                                                 (int(kIsVariableB) + int(kIsVariableC)) * sizeof(typename BlockLoadWeightT::TempStorage),
+                                                 (int(kIsVariableB) + int(kIsVariableC)) * sizeof(typename BlockLoadWeightVecT::TempStorage),
+                                                 sizeof(typename BlockStoreT::TempStorage),
+                                                 sizeof(typename BlockStoreVecT::TempStorage)});
+    static constexpr int kSmemExchangeSize = (int(kIsVariableB) + int(kIsVariableC)) * sizeof(typename BlockExchangeT::TempStorage);
+    static constexpr int kSmemReduceSize = sizeof(typename BlockReduceT::TempStorage);
+    static constexpr int kSmemSize = kSmemIOSize + kSmemExchangeSize + kSmemReduceSize + sizeof(typename BlockScanT::TempStorage) + sizeof(typename BlockReverseScanT::TempStorage);
+};
+
+template<typename Ktraits>
+__global__ __launch_bounds__(Ktraits::kNThreads, Ktraits::kMinBlocks)
+void selective_scan_bwd_kernel(SSMParamsBwd params) {
+    constexpr bool kIsComplex = Ktraits::kIsComplex;
+    constexpr bool kIsVariableB = Ktraits::kIsVariableB;
+    constexpr bool kIsVariableC = Ktraits::kIsVariableC;
+    constexpr bool kDeltaSoftplus = Ktraits::kDeltaSoftplus;
+    constexpr bool kHasZ = Ktraits::kHasZ;
+    constexpr int kNThreads = Ktraits::kNThreads;
+    constexpr int kNItems = Ktraits::kNItems;
+    using input_t = typename Ktraits::input_t;
+    using weight_t = typename Ktraits::weight_t;
+    using scan_t = typename Ktraits::scan_t;
+
+    // Shared memory.
+    extern __shared__ char smem_[];
+    // cast to lvalue reference of expected type
+    // char *smem_loadstorescan = smem_ + 2 * MAX_DSTATE * sizeof(weight_t);
+    // auto& smem_load = reinterpret_cast<typename BlockLoadT::TempStorage&>(smem_ + 2 * MAX_DSTATE * sizeof(weight_t));
+    // auto& smem_load = reinterpret_cast<typename BlockLoadT::TempStorage&>(smem_loadstorescan);
+    auto& smem_load = reinterpret_cast<typename Ktraits::BlockLoadT::TempStorage&>(smem_);
+    auto& smem_load_weight = reinterpret_cast<typename Ktraits::BlockLoadWeightT::TempStorage&>(smem_);
+    auto& smem_load_weight1 = *reinterpret_cast<typename Ktraits::BlockLoadWeightT::TempStorage*>(smem_ + sizeof(typename Ktraits::BlockLoadWeightT::TempStorage));
+    auto& smem_store = reinterpret_cast<typename Ktraits::BlockStoreT::TempStorage&>(smem_);
+    auto& smem_exchange = *reinterpret_cast<typename Ktraits::BlockExchangeT::TempStorage*>(smem_ + Ktraits::kSmemIOSize);
+    auto& smem_exchange1 = *reinterpret_cast<typename Ktraits::BlockExchangeT::TempStorage*>(smem_ + Ktraits::kSmemIOSize + sizeof(typename Ktraits::BlockExchangeT::TempStorage));
+    auto& smem_reduce = *reinterpret_cast<typename Ktraits::BlockReduceT::TempStorage*>(reinterpret_cast<char *>(&smem_exchange) + Ktraits::kSmemExchangeSize);
+    auto& smem_reduce_float = *reinterpret_cast<typename Ktraits::BlockReduceFloatT::TempStorage*>(&smem_reduce);
+    auto& smem_reduce_complex = *reinterpret_cast<typename Ktraits::BlockReduceComplexT::TempStorage*>(&smem_reduce);
+    auto& smem_scan = *reinterpret_cast<typename Ktraits::BlockScanT::TempStorage*>(reinterpret_cast<char *>(&smem_reduce) + Ktraits::kSmemReduceSize);
+    auto& smem_reverse_scan = *reinterpret_cast<typename Ktraits::BlockReverseScanT::TempStorage*>(reinterpret_cast<char *>(&smem_scan) + sizeof(typename Ktraits::BlockScanT::TempStorage));
+    weight_t *smem_delta_a = reinterpret_cast<weight_t *>(smem_ + Ktraits::kSmemSize);
+    scan_t *smem_running_postfix = reinterpret_cast<scan_t *>(smem_delta_a + 2 * MAX_DSTATE + kNThreads);
+    weight_t *smem_da = reinterpret_cast<weight_t *>(smem_running_postfix + MAX_DSTATE);
+    weight_t *smem_dbc = reinterpret_cast<weight_t *>(smem_da + MAX_DSTATE);
+
+    const int batch_id = blockIdx.x;
+    const int dim_id = blockIdx.y;
+    const int group_id = dim_id / (params.dim_ngroups_ratio);
+    input_t *u = reinterpret_cast<input_t *>(params.u_ptr) + batch_id * params.u_batch_stride
+        + dim_id * params.u_d_stride;
+    input_t *delta = reinterpret_cast<input_t *>(params.delta_ptr) + batch_id * params.delta_batch_stride
+        + dim_id * params.delta_d_stride;
+    input_t *dout = reinterpret_cast<input_t *>(params.dout_ptr) + batch_id * params.dout_batch_stride
+        + dim_id * params.dout_d_stride;
+    weight_t *A = reinterpret_cast<weight_t *>(params.A_ptr) + dim_id * params.A_d_stride;
+    weight_t *B = reinterpret_cast<weight_t *>(params.B_ptr) + dim_id * params.B_d_stride;
+    input_t *Bvar = reinterpret_cast<input_t *>(params.B_ptr) + batch_id * params.B_batch_stride + group_id * params.B_group_stride;
+    weight_t *C = reinterpret_cast<weight_t *>(params.C_ptr) + dim_id * params.C_d_stride;
+    input_t *Cvar = reinterpret_cast<input_t *>(params.C_ptr) + batch_id * params.C_batch_stride + group_id * params.C_group_stride;
+    weight_t *dA = reinterpret_cast<weight_t *>(params.dA_ptr) + dim_id * params.dA_d_stride;
+    weight_t *dB = reinterpret_cast<weight_t *>(params.dB_ptr)
+        + (!kIsVariableB ? dim_id * params.dB_d_stride : batch_id * (!kIsComplex ? params.dB_batch_stride : params.dB_batch_stride / 2) + group_id * params.dB_group_stride);
+    weight_t *dC = reinterpret_cast<weight_t *>(params.dC_ptr)
+        + (!kIsVariableC ? dim_id * params.dC_d_stride : batch_id * (!kIsComplex ? params.dC_batch_stride : params.dC_batch_stride / 2) + group_id * params.dC_group_stride);
+    float *dD = params.dD_ptr == nullptr ? nullptr : reinterpret_cast<float *>(params.dD_ptr) + dim_id;
+    float D_val = params.D_ptr == nullptr ? 0 : reinterpret_cast<float *>(params.D_ptr)[dim_id];
+    float *ddelta_bias = params.ddelta_bias_ptr == nullptr ? nullptr : reinterpret_cast<float *>(params.ddelta_bias_ptr) + dim_id;
+    float delta_bias = params.delta_bias_ptr == nullptr ? 0 : reinterpret_cast<float *>(params.delta_bias_ptr)[dim_id];
+    scan_t *x = params.x_ptr == nullptr
+        ? nullptr
+        : reinterpret_cast<scan_t *>(params.x_ptr) + (batch_id * params.dim + dim_id) * (params.n_chunks) * params.dstate;
+    float dD_val = 0;
+    float ddelta_bias_val = 0;
+
+    constexpr int kChunkSize = kNThreads * kNItems;
+    u += (params.n_chunks - 1) * kChunkSize;
+    delta += (params.n_chunks - 1) * kChunkSize;
+    dout += (params.n_chunks - 1) * kChunkSize;
+    Bvar += (params.n_chunks - 1) * kChunkSize * (!kIsComplex ? 1 : 2);
+    Cvar += (params.n_chunks - 1) * kChunkSize * (!kIsComplex ? 1 : 2);
+    for (int chunk = params.n_chunks - 1; chunk >= 0; --chunk) {
+        input_t u_vals[kNItems];
+        input_t delta_vals_load[kNItems];
+        input_t dout_vals_load[kNItems];
+        __syncthreads();
+        load_input<Ktraits>(u, u_vals, smem_load, params.seqlen - chunk * kChunkSize);
+        u -= kChunkSize;
+        __syncthreads();
+        load_input<Ktraits>(delta, delta_vals_load, smem_load, params.seqlen - chunk * kChunkSize);
+        // Will reload delta at the same location if kDeltaSoftplus
+        if constexpr (!kDeltaSoftplus) { delta -= kChunkSize; }
+        __syncthreads();
+        load_input<Ktraits>(dout, dout_vals_load, smem_load, params.seqlen - chunk * kChunkSize);
+        dout -= kChunkSize;
+
+        float dout_vals[kNItems], delta_vals[kNItems];
+        #pragma unroll
+        for (int i = 0; i < kNItems; ++i) {
+            dout_vals[i] = float(dout_vals_load[i]);
+            delta_vals[i] = float(delta_vals_load[i]) + delta_bias;
+            if constexpr (kDeltaSoftplus) {
+                delta_vals[i] = delta_vals[i] <= 20.f ? log1pf(expf(delta_vals[i])) : delta_vals[i];
+            }
+        }
+
+        if constexpr (kHasZ) {
+            input_t *z = reinterpret_cast<input_t *>(params.z_ptr) + batch_id * params.z_batch_stride
+                + dim_id * params.z_d_stride + chunk * kChunkSize;
+            input_t *out = reinterpret_cast<input_t *>(params.out_ptr) + batch_id * params.out_batch_stride
+                + dim_id * params.out_d_stride + chunk * kChunkSize;
+            input_t *dz = reinterpret_cast<input_t *>(params.dz_ptr) + batch_id * params.dz_batch_stride
+                + dim_id * params.dz_d_stride + chunk * kChunkSize;
+            input_t z_vals[kNItems], out_vals[kNItems];
+            __syncthreads();
+            load_input<Ktraits>(z, z_vals, smem_load, params.seqlen - chunk * kChunkSize);
+            __syncthreads();
+            load_input<Ktraits>(out, out_vals, smem_load, params.seqlen - chunk * kChunkSize);
+            float dz_vals[kNItems], z_silu_vals[kNItems];
+            #pragma unroll
+            for (int i = 0; i < kNItems; ++i) {
+                float z_val = z_vals[i];
+                float z_sigmoid_val = 1.0f / (1.0f + expf(-z_val));
+                z_silu_vals[i] = z_val * z_sigmoid_val;
+                dz_vals[i] = dout_vals[i] * float(out_vals[i]) * z_sigmoid_val
+                             * (1.0f + z_val * (1.0f - z_sigmoid_val));
+                dout_vals[i] *= z_silu_vals[i];
+            }
+            __syncthreads();
+            store_output<Ktraits>(dz, dz_vals, smem_store, params.seqlen - chunk * kChunkSize);
+            if (params.out_z_ptr != nullptr) {  // Recompute and store out_z
+                float out_z_vals[kNItems];
+                #pragma unroll
+                for (int i = 0; i < kNItems; ++i) { out_z_vals[i] = float(out_vals[i]) * z_silu_vals[i]; }
+                // if (blockIdx.x == 0 && blockIdx.y == 0 && threadIdx.x == 0) {
+                    // printf("out_val=%f, z_silu_val = %f, out_z_val = %f\n", float(out_vals[0]), z_silu_vals[0], out_z_vals[0]);
+                // }
+                input_t *out_z = reinterpret_cast<input_t *>(params.out_z_ptr) + batch_id * params.out_z_batch_stride
+                    + dim_id * params.out_z_d_stride + chunk * kChunkSize;
+                __syncthreads();
+                store_output<Ktraits>(out_z, out_z_vals, smem_store, params.seqlen - chunk * kChunkSize);
+            }
+        }
+
+        float du_vals[kNItems];
+        #pragma unroll
+        for (int i = 0; i < kNItems; ++i) { du_vals[i] = D_val * dout_vals[i]; }
+        #pragma unroll
+        for (int i = 0; i < kNItems; ++i) { dD_val += dout_vals[i] * float(u_vals[i]); }
+
+        float ddelta_vals[kNItems] = {0};
+        __syncthreads();
+        for (int state_idx = 0; state_idx < params.dstate; ++state_idx) {
+            const weight_t A_val = A[state_idx * params.A_dstate_stride];
+            // Multiply the real part of A with LOG2E so we can use exp2f instead of expf.
+            weight_t A_scaled;
+            constexpr float kLog2e = M_LOG2E;
+            if constexpr (!kIsComplex) {
+                A_scaled = A_val * kLog2e;
+            } else {
+                A_scaled = complex_t(A_val.real_ * kLog2e, A_val.imag_);
+            }
+            weight_t B_val, C_val;
+            weight_t B_vals[kNItems], C_vals[kNItems];
+            if constexpr (!kIsVariableB) {
+                B_val = B[state_idx * params.B_dstate_stride];
+            } else {
+                load_weight<Ktraits>(Bvar + state_idx * params.B_dstate_stride, B_vals,
+                    smem_load_weight, (params.seqlen - chunk * kChunkSize) * (!kIsComplex ? 1 : 2));
+            }
+            if constexpr (!kIsVariableC) {
+                C_val = C[state_idx * params.C_dstate_stride];
+            } else {
+                auto &smem_load_weight_C = !kIsVariableB ? smem_load_weight : smem_load_weight1;
+                load_weight<Ktraits>(Cvar + state_idx * params.C_dstate_stride, C_vals,
+                    smem_load_weight_C, (params.seqlen - chunk * kChunkSize) * (!kIsComplex ? 1 : 2));
+            }
+            // const weight_t A_val = smem_a[state_idx];
+            scan_t thread_data[kNItems], thread_reverse_data[kNItems];
+            if constexpr (!kIsComplex) {
+                #pragma unroll
+                for (int i = 0; i < kNItems; ++i) {
+                    const float delta_a_exp = exp2f(delta_vals[i] * A_scaled);
+                    thread_data[i] = make_float2(delta_a_exp, !kIsVariableB ? delta_vals[i] * float(u_vals[i]) : delta_vals[i] * float(u_vals[i]) * B_vals[i]);
+                    if (i == 0) {
+                        smem_delta_a[threadIdx.x == 0 ? state_idx + (chunk % 2) * MAX_DSTATE : threadIdx.x + 2 * MAX_DSTATE] = delta_a_exp;
+                    } else {
+                        thread_reverse_data[i - 1].x = delta_a_exp;
+                    }
+                    thread_reverse_data[i].y = dout_vals[i] *
+                        (!kIsVariableC
+                         ? (!kIsVariableB ? B_val * C_val : C_val)
+                         : (!kIsVariableB ? B_val * C_vals[i] : C_vals[i]));
+                }
+                __syncthreads();
+                thread_reverse_data[kNItems - 1].x = threadIdx.x == kNThreads - 1
+                    ? (chunk == params.n_chunks - 1 ? 1.f : smem_delta_a[state_idx + ((chunk + 1) % 2) * MAX_DSTATE])
+                    : smem_delta_a[threadIdx.x + 1 + 2 * MAX_DSTATE];
+                // Initialize running total
+                scan_t running_prefix = chunk > 0 && threadIdx.x % 32 == 0 ? x[(chunk - 1) * params.dstate + state_idx] : make_float2(1.f, 0.f);
+                SSMScanPrefixCallbackOp<weight_t> prefix_op(running_prefix);
+                Ktraits::BlockScanT(smem_scan).InclusiveScan(
+                    thread_data, thread_data, SSMScanOp<weight_t>(), prefix_op
+                );
+                scan_t running_postfix = chunk < params.n_chunks - 1 && threadIdx.x % 32 == 0 ? smem_running_postfix[state_idx] : make_float2(1.f, 0.f);
+                SSMScanPrefixCallbackOp<weight_t> postfix_op(running_postfix);
+                Ktraits::BlockReverseScanT(smem_reverse_scan).InclusiveReverseScan(
+                    thread_reverse_data, thread_reverse_data, SSMScanOp<weight_t>(), postfix_op
+                );
+                if (threadIdx.x == 0) { smem_running_postfix[state_idx] = postfix_op.running_prefix; }
+                weight_t dA_val = 0, dBC_val = 0;
+                weight_t dB_vals[kNItems], dC_vals[kNItems];
+                #pragma unroll
+                for (int i = 0; i < kNItems; ++i) {
+                    const float dx = thread_reverse_data[i].y;
+                    const float ddelta_u = !kIsVariableB ? dx : dx * B_vals[i];
+                    du_vals[i] += ddelta_u * delta_vals[i];
+                    const float a = thread_data[i].y - (!kIsVariableB ? delta_vals[i] * float(u_vals[i]) : delta_vals[i] * float(u_vals[i]) * B_vals[i]);
+                    ddelta_vals[i] += ddelta_u * float(u_vals[i]) + dx * A_val * a;
+                    dA_val += dx * delta_vals[i] * a;
+                    if constexpr (!kIsVariableB || !kIsVariableC) {
+                        if constexpr (!kIsVariableB) {  // dBC_val is dB_val
+                            dBC_val += dout_vals[i] * (!kIsVariableC ? thread_data[i].y : thread_data[i].y * C_vals[i]);
+                        } else {  // dBC_val is dC_val
+                            dBC_val += dout_vals[i] * thread_data[i].y;
+                        }
+                    }
+                    if constexpr (kIsVariableB) { dB_vals[i] = dx * delta_vals[i] * float(u_vals[i]); }
+                    if constexpr (kIsVariableC) {
+                        dC_vals[i] = dout_vals[i] * (!kIsVariableB ? thread_data[i].y * B_val : thread_data[i].y);
+                    }
+                }
+                // Block-exchange to make the atomicAdd's coalesced, otherwise they're much slower
+                if constexpr (kIsVariableB || kIsVariableC) {
+                    if constexpr (kIsVariableB) {
+                        Ktraits::BlockExchangeT(smem_exchange).BlockedToStriped(dB_vals, dB_vals);
+                    }
+                    if constexpr (kIsVariableC) {
+                        auto &smem_exchange_C = !kIsVariableB ? smem_exchange : smem_exchange1;
+                        Ktraits::BlockExchangeT(smem_exchange_C).BlockedToStriped(dC_vals, dC_vals);
+                    }
+                    const int seqlen_remaining = params.seqlen - chunk * kChunkSize - threadIdx.x;
+                    weight_t *dB_cur = dB + state_idx * params.dB_dstate_stride + chunk * kChunkSize + threadIdx.x;
+                    weight_t *dC_cur = dC + state_idx * params.dC_dstate_stride + chunk * kChunkSize + threadIdx.x;
+                    #pragma unroll
+                    for (int i = 0; i < kNItems; ++i) {
+                        if (i * kNThreads < seqlen_remaining) {
+                            if constexpr (kIsVariableB) { gpuAtomicAdd(dB_cur + i * kNThreads, dB_vals[i]); }
+                            if constexpr (kIsVariableC) { gpuAtomicAdd(dC_cur + i * kNThreads, dC_vals[i]); }
+                        }
+                    }
+                }
+                if constexpr (!kIsVariableB || !kIsVariableC) {
+                    float2 dA_dBC_val = make_float2(dA_val, dBC_val);
+                    dA_dBC_val = Ktraits::BlockReduceT(smem_reduce).Sum(dA_dBC_val);
+                    dA_val = dA_dBC_val.x;
+                    if (threadIdx.x == 0) {
+                        smem_dbc[state_idx] = chunk == params.n_chunks - 1 ? dA_dBC_val.y : dA_dBC_val.y + smem_dbc[state_idx];
+                    }
+                } else {
+                    dA_val = Ktraits::BlockReduceFloatT(smem_reduce_float).Sum(dA_val);
+                }
+                if (threadIdx.x == 0) {
+                    smem_da[state_idx] = chunk == params.n_chunks - 1 ? dA_val : dA_val + smem_da[state_idx];
+                }
+            } else {
+                #pragma unroll
+                for (int i = 0; i < kNItems; ++i) {
+                    // Pytorch's implementation of complex exp (which calls thrust) is very slow
+                    complex_t delta_a_exp = cexp2f(delta_vals[i] * A_scaled);
+                    weight_t B_delta_u_val = !kIsVariableB ? delta_vals[i] * float(u_vals[i]) : B_vals[i] * delta_vals[i] * float(u_vals[i]);
+                    thread_data[i] = make_float4(delta_a_exp.real_, delta_a_exp.imag_, B_delta_u_val.real_, B_delta_u_val.imag_);
+                    if (i == 0) {
+                        smem_delta_a[threadIdx.x == 0 ? state_idx + (chunk % 2) * MAX_DSTATE : threadIdx.x + 2 * MAX_DSTATE] = delta_a_exp;
+                    } else {
+                        thread_reverse_data[i - 1].x = delta_a_exp.real_;
+                        thread_reverse_data[i - 1].y = -delta_a_exp.imag_;
+                    }
+                    complex_t dout_BC = 2 * dout_vals[i]
+                        * conj(!kIsVariableC
+                                ? (!kIsVariableB ? B_val * C_val : C_val)
+                                : (!kIsVariableB ? B_val * C_vals[i] : C_vals[i]));
+                    thread_reverse_data[i].z = dout_BC.real_;
+                    thread_reverse_data[i].w = dout_BC.imag_;
+                }
+                __syncthreads();
+                complex_t delta_a_exp = threadIdx.x == kNThreads - 1
+                    ? (chunk == params.n_chunks - 1 ? 1.f : smem_delta_a[state_idx + ((chunk + 1) % 2) * MAX_DSTATE])
+                    : smem_delta_a[threadIdx.x + 1 + 2 * MAX_DSTATE];
+                thread_reverse_data[kNItems - 1].x = delta_a_exp.real_;
+                thread_reverse_data[kNItems - 1].y = -delta_a_exp.imag_;
+                // Initialize running total
+                scan_t running_prefix = chunk > 0 && threadIdx.x % 32 == 0 ? x[(chunk - 1) * params.dstate + state_idx] : make_float4(1.f, 0.f, 0.f, 0.f);
+                SSMScanPrefixCallbackOp<weight_t> prefix_op(running_prefix);
+                Ktraits::BlockScanT(smem_scan).InclusiveScan(
+                    thread_data, thread_data, SSMScanOp<weight_t>(), prefix_op
+                );
+                scan_t running_postfix = chunk < params.n_chunks - 1 && threadIdx.x % 32 == 0 ? smem_running_postfix[state_idx] : make_float4(1.f, 0.f, 0.f, 0.f);
+                SSMScanPrefixCallbackOp<weight_t> postfix_op(running_postfix);
+                Ktraits::BlockReverseScanT(smem_reverse_scan).InclusiveReverseScan(
+                    thread_reverse_data, thread_reverse_data, SSMScanOp<weight_t>(), postfix_op
+                );
+                if (threadIdx.x == 0) { smem_running_postfix[state_idx] = postfix_op.running_prefix; }
+                weight_t dA_val = 0, dBC_val = 0;
+                weight_t dB_vals[kNItems], dC_vals[kNItems];
+                #pragma unroll
+                for (int i = 0; i < kNItems; ++i) {
+                    complex_t x = complex_t(thread_data[i].z, thread_data[i].w);
+                    complex_t dx = complex_t(thread_reverse_data[i].z, thread_reverse_data[i].w);
+                    float ddelta_u = !kIsVariableB ? dx.real_ : (dx * conj(B_vals[i])).real_;
+                    if constexpr (!kIsVariableB || !kIsVariableC) {
+                        if constexpr (!kIsVariableB) {  // dBC_val is dB_val
+                            dBC_val += (2 * dout_vals[i]) * conj(!kIsVariableC ? x : x * C_vals[i]);
+                        } else {  // dBC_val is dC_val
+                            dBC_val += (2 * dout_vals[i]) * conj(x);
+                        }
+                    }
+                    const complex_t a_conj = conj(x - (!kIsVariableB ? delta_vals[i] * float(u_vals[i]) : delta_vals[i] * float(u_vals[i]) * B_vals[i]));
+                    du_vals[i] += ddelta_u * delta_vals[i];
+                    ddelta_vals[i] += ddelta_u * float(u_vals[i]) + (dx * conj(A_val) * a_conj).real_;
+                    dA_val += delta_vals[i] * dx * a_conj;
+                    if constexpr (kIsVariableB) { dB_vals[i] = dx * delta_vals[i] * float(u_vals[i]); }
+                    if constexpr (kIsVariableC) {
+                        dC_vals[i] = (2 * dout_vals[i]) * conj(!kIsVariableB ? x * B_val : x);
+                    }
+                }
+                // Block-exchange to make the atomicAdd's coalesced, otherwise they're much slower
+                if constexpr (kIsVariableB || kIsVariableC) {
+                    float dB_vals_f[kNItems * 2], dC_vals_f[kNItems * 2];
+                    if constexpr (kIsVariableB) {
+                        #pragma unroll
+                        for (int i = 0; i < kNItems; ++i) {
+                            dB_vals_f[i * 2] = dB_vals[i].real_;
+                            dB_vals_f[i * 2 + 1] = dB_vals[i].imag_;
+                        }
+                        Ktraits::BlockExchangeT(smem_exchange).BlockedToStriped(dB_vals_f, dB_vals_f);
+                    }
+                    if constexpr (kIsVariableC) {
+                        #pragma unroll
+                        for (int i = 0; i < kNItems; ++i) {
+                            dC_vals_f[i * 2] = dC_vals[i].real_;
+                            dC_vals_f[i * 2 + 1] = dC_vals[i].imag_;
+                        }
+                        auto &smem_exchange_C = !kIsVariableB ? smem_exchange : smem_exchange1;
+                        Ktraits::BlockExchangeT(smem_exchange_C).BlockedToStriped(dC_vals_f, dC_vals_f);
+                    }
+                    const int seqlen_remaining = (params.seqlen - chunk * kChunkSize) * 2 - threadIdx.x;
+                    float *dB_cur = reinterpret_cast<float *>(dB) + state_idx * params.dB_dstate_stride + chunk * kChunkSize * 2 + threadIdx.x;
+                    float *dC_cur = reinterpret_cast<float *>(dC) + state_idx * params.dC_dstate_stride + chunk * kChunkSize * 2 + threadIdx.x;
+                    #pragma unroll
+                    for (int i = 0; i < kNItems * 2; ++i) {
+                        if (i * kNThreads < seqlen_remaining) {
+                            if constexpr (kIsVariableB) { gpuAtomicAdd(dB_cur + i * kNThreads, dB_vals_f[i]); }
+                            if constexpr (kIsVariableC) { gpuAtomicAdd(dC_cur + i * kNThreads, dC_vals_f[i]); }
+                        }
+                    }
+                }
+                if constexpr (!kIsVariableB || !kIsVariableC) {
+                    float4 dA_dBC_val = make_float4(dA_val.real_, dA_val.imag_, dBC_val.real_, dBC_val.imag_);
+                    dA_dBC_val = Ktraits::BlockReduceT(smem_reduce).Sum(dA_dBC_val);
+                    dA_val = complex_t(dA_dBC_val.x, dA_dBC_val.y);
+                    dBC_val = complex_t(dA_dBC_val.z, dA_dBC_val.w);
+                    if (threadIdx.x == 0) {
+                        smem_dbc[state_idx] = chunk == params.n_chunks - 1 ? dBC_val : dBC_val + smem_dbc[state_idx];
+                    }
+                } else {
+                    dA_val = Ktraits::BlockReduceComplexT(smem_reduce_complex).Sum(dA_val);
+                }
+                if (threadIdx.x == 0) {
+                    smem_da[state_idx] = chunk == params.n_chunks - 1 ? dA_val : dA_val + smem_da[state_idx];
+                }
+            }
+        }
+
+        if constexpr (kDeltaSoftplus) {
+            __syncthreads();
+            input_t delta_vals_load[kNItems];
+            load_input<Ktraits>(delta, delta_vals_load, smem_load, params.seqlen - chunk * kChunkSize);
+            delta -= kChunkSize;
+            #pragma unroll
+            for (int i = 0; i < kNItems; ++i) {
+                float delta_val = float(delta_vals_load[i]) + delta_bias;
+                float delta_val_neg_exp = expf(-delta_val);
+                ddelta_vals[i] = delta_val <= 20.f
+                    ? ddelta_vals[i] / (1.f + delta_val_neg_exp)
+                    : ddelta_vals[i];
+            }
+        }
+        for (int i = 0; i < kNItems; ++i) { ddelta_bias_val += ddelta_vals[i]; }
+
+        input_t *du = reinterpret_cast<input_t *>(params.du_ptr) + batch_id * params.du_batch_stride
+            + dim_id * params.du_d_stride + chunk * kChunkSize;
+        input_t *ddelta = reinterpret_cast<input_t *>(params.ddelta_ptr) + batch_id * params.ddelta_batch_stride
+            + dim_id * params.ddelta_d_stride + chunk * kChunkSize;
+        __syncthreads();
+        store_output<Ktraits>(du, du_vals, smem_store, params.seqlen - chunk * kChunkSize);
+        __syncthreads();
+        store_output<Ktraits>(ddelta, ddelta_vals, smem_store, params.seqlen - chunk * kChunkSize);
+
+        Bvar -= kChunkSize * (!kIsComplex ? 1 : 2);
+        Cvar -= kChunkSize * (!kIsComplex ? 1 : 2);
+    }
+    if (params.dD_ptr != nullptr) {
+        dD_val = Ktraits::BlockReduceFloatT(smem_reduce_float).Sum(dD_val);
+        if (threadIdx.x == 0) { gpuAtomicAdd(dD, dD_val); }
+    }
+    if (params.ddelta_bias_ptr != nullptr) {
+        __syncthreads();
+        ddelta_bias_val = Ktraits::BlockReduceFloatT(smem_reduce_float).Sum(ddelta_bias_val);
+        if (threadIdx.x == 0) { gpuAtomicAdd(ddelta_bias, ddelta_bias_val); }
+    }
+    for (int state_idx = threadIdx.x; state_idx < params.dstate; state_idx += blockDim.x) {
+        gpuAtomicAdd(&(dA[state_idx * params.dA_dstate_stride]), smem_da[state_idx]);
+        weight_t dBC_val;
+        if (!kIsVariableB || !kIsVariableC) { dBC_val = smem_dbc[state_idx]; }
+        if constexpr (!kIsVariableB) {
+            gpuAtomicAdd(&(dB[state_idx * params.dB_dstate_stride]),
+                         !kIsVariableC ? dBC_val * conj(C[state_idx * params.C_dstate_stride]) : dBC_val);
+        }
+        if constexpr (!kIsVariableC) {
+            gpuAtomicAdd(&(dC[state_idx * params.dC_dstate_stride]),
+                        !kIsVariableB ? dBC_val * conj(B[state_idx * params.B_dstate_stride]) : dBC_val);
+        }
+    }
+}
+
+template<int kNThreads, int kNItems, typename input_t, typename weight_t>
+void selective_scan_bwd_launch(SSMParamsBwd &params, cudaStream_t stream) {
+    BOOL_SWITCH(params.seqlen % (kNThreads * kNItems) == 0, kIsEvenLen, [&] {
+        BOOL_SWITCH(params.is_variable_B, kIsVariableB, [&] {
+            BOOL_SWITCH(params.is_variable_C, kIsVariableC, [&] {
+                BOOL_SWITCH(params.delta_softplus, kDeltaSoftplus, [&] {
+                    BOOL_SWITCH(params.z_ptr != nullptr , kHasZ, [&] {
+                        using Ktraits = Selective_Scan_bwd_kernel_traits<kNThreads, kNItems, kIsEvenLen, kIsVariableB, kIsVariableC, kDeltaSoftplus, kHasZ, input_t, weight_t>;
+                        // using Ktraits = Selective_Scan_bwd_kernel_traits<kNThreads, kNItems, true, kIsVariableB, kIsVariableC, kDeltaSoftplus, kHasZ, input_t, weight_t>;
+                        // TODO: check this
+                        constexpr int kSmemSize = Ktraits::kSmemSize + MAX_DSTATE * sizeof(typename Ktraits::scan_t) + (kNThreads + 4 * MAX_DSTATE) * sizeof(typename Ktraits::weight_t);
+                        // printf("smem_size = %d\n", kSmemSize);
+                        dim3 grid(params.batch, params.dim);
+                        auto kernel = &selective_scan_bwd_kernel<Ktraits>;
+                        if (kSmemSize >= 48 * 1024) {
+                            C10_CUDA_CHECK(cudaFuncSetAttribute(
+                                kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, kSmemSize));
+                        }
+                        kernel<<<grid, Ktraits::kNThreads, kSmemSize, stream>>>(params);
+                        C10_CUDA_KERNEL_LAUNCH_CHECK();
+                    });
+                });
+            });
+        });
+    });
+}
+
+template<typename input_t, typename weight_t>
+void selective_scan_bwd_cuda(SSMParamsBwd &params, cudaStream_t stream) {
+    if (params.seqlen <= 128) {
+        selective_scan_bwd_launch<32, 4, input_t, weight_t>(params, stream);
+    } else if (params.seqlen <= 256) {
+        selective_scan_bwd_launch<32, 8, input_t, weight_t>(params, stream);
+    } else if (params.seqlen <= 512) {
+        selective_scan_bwd_launch<32, 16, input_t, weight_t>(params, stream);
+    } else if (params.seqlen <= 1024) {
+        selective_scan_bwd_launch<64, 16, input_t, weight_t>(params, stream);
+    } else {
+        selective_scan_bwd_launch<128, 16, input_t, weight_t>(params, stream);
+    }
+}

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_common.h

@@ -0,0 +1,221 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+#pragma once
+
+#include <cuda_bf16.h>
+#include <cuda_fp16.h>
+#include <c10/util/complex.h>  // For scalar_value_type
+
+#define MAX_DSTATE 256
+
+using complex_t = c10::complex<float>;
+
+inline __device__ float2 operator+(const float2 & a, const float2 & b){
+    return {a.x + b.x, a.y + b.y};
+}
+
+inline __device__ float3 operator+(const float3 &a, const float3 &b) {
+  return {a.x + b.x, a.y + b.y, a.z + b.z};
+}
+
+inline __device__ float4 operator+(const float4 & a, const float4 & b){
+    return {a.x + b.x, a.y + b.y, a.z + b.z, a.w + b.w};
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template<int BYTES> struct BytesToType {};
+
+template<> struct BytesToType<16> {
+    using Type = uint4;
+    static_assert(sizeof(Type) == 16);
+};
+
+template<> struct BytesToType<8> {
+    using Type = uint64_t;
+    static_assert(sizeof(Type) == 8);
+};
+
+template<> struct BytesToType<4> {
+    using Type = uint32_t;
+    static_assert(sizeof(Type) == 4);
+};
+
+template<> struct BytesToType<2> {
+    using Type = uint16_t;
+    static_assert(sizeof(Type) == 2);
+};
+
+template<> struct BytesToType<1> {
+    using Type = uint8_t;
+    static_assert(sizeof(Type) == 1);
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template<typename scalar_t, int N>
+struct Converter{
+    static inline __device__ void to_float(const scalar_t (&src)[N], float (&dst)[N]) {
+        #pragma unroll
+        for (int i = 0; i < N; ++i) { dst[i] = src[i]; }
+    }
+};
+
+template<int N>
+struct Converter<at::Half, N>{
+    static inline __device__ void to_float(const at::Half (&src)[N], float (&dst)[N]) {
+        static_assert(N % 2 == 0);
+        auto &src2 = reinterpret_cast<const half2 (&)[N / 2]>(src);
+        auto &dst2 = reinterpret_cast<float2 (&)[N / 2]>(dst);
+        #pragma unroll
+        for (int i = 0; i < N / 2; ++i) { dst2[i] = __half22float2(src2[i]); }
+    }
+};
+
+#if __CUDA_ARCH__ >= 800
+template<int N>
+struct Converter<at::BFloat16, N>{
+    static inline __device__ void to_float(const at::BFloat16 (&src)[N], float (&dst)[N]) {
+        static_assert(N % 2 == 0);
+        auto &src2 = reinterpret_cast<const nv_bfloat162 (&)[N / 2]>(src);
+        auto &dst2 = reinterpret_cast<float2 (&)[N / 2]>(dst);
+        #pragma unroll
+        for (int i = 0; i < N / 2; ++i) { dst2[i] = __bfloat1622float2(src2[i]); }
+    }
+};
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// From https://stackoverflow.com/questions/9860711/cucomplex-h-and-exp
+// and https://forums.developer.nvidia.com/t/complex-number-exponential-function/24696
+__device__ __forceinline__ complex_t cexp2f(complex_t z) {
+    float t = exp2f(z.real_);
+    float c, s;
+    sincosf(z.imag_, &s, &c);
+    return complex_t(c * t, s * t);
+}
+
+__device__ __forceinline__ complex_t cexpf(complex_t z) {
+    float t = expf(z.real_);
+    float c, s;
+    sincosf(z.imag_, &s, &c);
+    return complex_t(c * t, s * t);
+}
+
+template<typename scalar_t> struct SSMScanOp;
+
+template<>
+struct SSMScanOp<float> {
+    __device__ __forceinline__ float2 operator()(const float2 &ab0, const float2 &ab1) const {
+        return make_float2(ab1.x * ab0.x, ab1.x * ab0.y + ab1.y);
+    }
+};
+
+template<>
+struct SSMScanOp<complex_t> {
+    __device__ __forceinline__ float4 operator()(const float4 &ab0, const float4 &ab1) const {
+        complex_t a0 = complex_t(ab0.x, ab0.y);
+        complex_t b0 = complex_t(ab0.z, ab0.w);
+        complex_t a1 = complex_t(ab1.x, ab1.y);
+        complex_t b1 = complex_t(ab1.z, ab1.w);
+        complex_t out_a = a1 * a0;
+        complex_t out_b = a1 * b0 + b1;
+        return make_float4(out_a.real_, out_a.imag_, out_b.real_, out_b.imag_);
+    }
+};
+
+// A stateful callback functor that maintains a running prefix to be applied
+// during consecutive scan operations.
+template <typename scalar_t> struct SSMScanPrefixCallbackOp {
+    using scan_t = std::conditional_t<std::is_same_v<scalar_t, float>, float2, float4>;
+    scan_t running_prefix;
+    // Constructor
+    __device__ SSMScanPrefixCallbackOp(scan_t running_prefix_) : running_prefix(running_prefix_) {}
+    // Callback operator to be entered by the first warp of threads in the block.
+    // Thread-0 is responsible for returning a value for seeding the block-wide scan.
+    __device__ scan_t operator()(scan_t block_aggregate) {
+        scan_t old_prefix = running_prefix;
+        running_prefix = SSMScanOp<scalar_t>()(running_prefix, block_aggregate);
+        return old_prefix;
+    }
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+
+template<typename Ktraits>
+inline __device__ void load_input(typename Ktraits::input_t *u,
+                                  typename Ktraits::input_t (&u_vals)[Ktraits::kNItems],
+                                  typename Ktraits::BlockLoadT::TempStorage &smem_load,
+                                  int seqlen) {
+    if constexpr (Ktraits::kIsEvenLen) {
+        auto& smem_load_vec = reinterpret_cast<typename Ktraits::BlockLoadVecT::TempStorage&>(smem_load);
+        using vec_t = typename Ktraits::vec_t;
+        Ktraits::BlockLoadVecT(smem_load_vec).Load(
+            reinterpret_cast<vec_t*>(u),
+            reinterpret_cast<vec_t(&)[Ktraits::kNLoads]>(u_vals)
+       );
+    } else {
+        Ktraits::BlockLoadT(smem_load).Load(u, u_vals, seqlen, 0.f);
+    }
+}
+
+template<typename Ktraits>
+inline __device__ void load_weight(typename Ktraits::input_t *Bvar,
+                                   typename Ktraits::weight_t (&B_vals)[Ktraits::kNItems],
+                                   typename Ktraits::BlockLoadWeightT::TempStorage &smem_load_weight,
+                                   int seqlen) {
+    constexpr int kNItems = Ktraits::kNItems;
+    if constexpr (!Ktraits::kIsComplex) {
+        typename Ktraits::input_t B_vals_load[kNItems];
+        if constexpr (Ktraits::kIsEvenLen) {
+            auto& smem_load_weight_vec = reinterpret_cast<typename Ktraits::BlockLoadWeightVecT::TempStorage&>(smem_load_weight);
+            using vec_t = typename Ktraits::vec_t;
+            Ktraits::BlockLoadWeightVecT(smem_load_weight_vec).Load(
+                reinterpret_cast<vec_t*>(Bvar),
+                reinterpret_cast<vec_t(&)[Ktraits::kNLoads]>(B_vals_load)
+          );
+        } else {
+            Ktraits::BlockLoadWeightT(smem_load_weight).Load(Bvar, B_vals_load, seqlen, 0.f);
+        }
+        // #pragma unroll
+        // for (int i = 0; i < kNItems; ++i) { B_vals[i] = B_vals_load[i]; }
+        Converter<typename Ktraits::input_t, kNItems>::to_float(B_vals_load, B_vals);
+    } else {
+        typename Ktraits::input_t B_vals_load[kNItems * 2];
+        if constexpr (Ktraits::kIsEvenLen) {
+            auto& smem_load_weight_vec = reinterpret_cast<typename Ktraits::BlockLoadWeightVecT::TempStorage&>(smem_load_weight);
+            using vec_t = typename Ktraits::vec_t;
+            Ktraits::BlockLoadWeightVecT(smem_load_weight_vec).Load(
+                reinterpret_cast<vec_t*>(Bvar),
+                reinterpret_cast<vec_t(&)[Ktraits::kNLoads * 2]>(B_vals_load)
+          );
+        } else {
+            Ktraits::BlockLoadWeightT(smem_load_weight).Load(Bvar, B_vals_load, seqlen, 0.f);
+        }
+        #pragma unroll
+        for (int i = 0; i < kNItems; ++i) { B_vals[i] = complex_t(B_vals_load[i * 2], B_vals_load[i * 2 + 1]); }
+    }
+}
+
+template<typename Ktraits>
+inline __device__ void store_output(typename Ktraits::input_t *out,
+                                    const float (&out_vals)[Ktraits::kNItems],
+                                    typename Ktraits::BlockStoreT::TempStorage &smem_store,
+                                    int seqlen) {
+    typename Ktraits::input_t write_vals[Ktraits::kNItems];
+    #pragma unroll
+    for (int i = 0; i < Ktraits::kNItems; ++i) { write_vals[i] = out_vals[i]; }
+    if constexpr (Ktraits::kIsEvenLen) {
+        auto& smem_store_vec = reinterpret_cast<typename Ktraits::BlockStoreVecT::TempStorage&>(smem_store);
+        using vec_t = typename Ktraits::vec_t;
+        Ktraits::BlockStoreVecT(smem_store_vec).Store(
+            reinterpret_cast<vec_t*>(out),
+            reinterpret_cast<vec_t(&)[Ktraits::kNLoads]>(write_vals)
+       );
+    } else {
+        Ktraits::BlockStoreT(smem_store).Store(out, write_vals, seqlen);
+    }
+}

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_fwd_bf16.cu

@@ -0,0 +1,10 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_fwd_kernel.cuh"
+
+template void selective_scan_fwd_cuda<at::BFloat16, float>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<at::BFloat16, complex_t>(SSMParamsBase &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_fwd_fp16.cu

@@ -0,0 +1,10 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_fwd_kernel.cuh"
+
+template void selective_scan_fwd_cuda<at::Half, float>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<at::Half, complex_t>(SSMParamsBase &params, cudaStream_t stream);

protein_lm/modeling/models/libs/mamba/csrc/selective_scan/selective_scan_fwd_fp32.cu

@@ -0,0 +1,10 @@
+/******************************************************************************
+ * Copyright (c) 2023, Tri Dao.
+ ******************************************************************************/
+
+// Split into multiple files to compile in paralell
+
+#include "selective_scan_fwd_kernel.cuh"
+
+template void selective_scan_fwd_cuda<float, float>(SSMParamsBase &params, cudaStream_t stream);
+template void selective_scan_fwd_cuda<float, complex_t>(SSMParamsBase &params, cudaStream_t stream);