init

README.md

@@ -0,0 +1,82 @@
+---
+license: cc-by-sa-3.0
+datasets:
+- togethercomputer/RedPajama-Data-1T
+---
+
+# MPT-1b-RedPajama-200b-dolly
+
+MPT-1b-RedPajama-200b-dolly is a 1.3 billion parameter decoder-only transformer pre-trained on the [RedPajama dataset](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T) and subsequently fine-tuned on the [Databricks Dolly](https://github.com/databrickslabs/dolly/tree/master/data) instruction dataset.
+The model was pre-trained for 200B tokens by sampling from the subsets of the RedPajama dataset in the same proportions as were used by the [Llama series of models](https://arxiv.org/abs/2302.13971).
+This model was trained by [MosaicML](https://www.mosaicml.com) and follows a modified decoder-only transformer architecture.
+
+This model is an instruction fine-tuned version of [mpt-1b-redpajama-200b](https://huggingface.co/mosaicml/mpt-1b-redpajama-200b). In other words, the pre-trained version of this model is [mpt-1b-redpajama-200b](https://huggingface.co/mosaicml/mpt-1b-redpajama-200b).
+
+## Model Date
+
+April 20, 2023
+
+## How to Use
+
+Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method. 
+This is because we use a custom model architecture `MosaicGPT` that is not yet part of the `transformers` package.
+`MosaicGPT` includes options for many training efficiency features such as [FlashAttention (Dao et al. 2022)](https://arxiv.org/pdf/2205.14135.pdf), [ALIBI](https://arxiv.org/abs/2108.12409), QK LayerNorm, and more.
+
+```python
+import transformers
+model = transformers.AutoModelForCausalLM.from_pretrained('mosaicml/mpt-1b-redpajama-200b-dolly', trust_remote_code=True)
+```
+
+To use the optimized triton implementation of FlashAttention, you can load with `attn_impl='triton'` and move the model to `bfloat16` like so:
+```python
+model = transformers.AutoModelForCausalLM.from_pretrained('mosaicml/mpt-1b-redpajama-200b-dolly', trust_remote_code=True, attn_impl='triton')
+model.to(device='cuda:0', dtype=torch.bfloat16)
+```
+
+## Model Description
+
+This model uses the MosaicML LLM codebase, which can be found in the [MosaicML Examples Repository](https://github.com/mosaicml/examples/tree/v0.0.4/examples/llm).
+The architecture is a modification of a standard decoder-only transformer.
+The transformer has 24 layers, 16 attention heads, and width 2048.
+The model has been modified from a standard transformer in the following ways:
+* It uses ALiBi and does not use positional embeddings.
+* It uses QK LayerNorm.
+* It does not use biases.
+
+## Training Data
+
+### Pre-Training
+
+The model was pre-trained for 200B tokens (batch size 2200, sequence length 2048). It was trained on the following data mix:
+* 67% RedPajama Common Crawl
+* 15% [C4](https://huggingface.co/datasets/c4)
+* 4.5% RedPajama GitHub
+* 4.5% RedPajama Wikipedia
+* 4.5% RedPajama Books
+* 2.5% RedPajama Arxiv
+* 2% RedPajama StackExchange
+
+This is the same mix of data as was used in the Llama series of models](https://arxiv.org/abs/2302.13971).
+
+Each sample was chosen from one of the datasets, with the dataset selected with the probability specified above.
+The examples were shuffled within each dataset.
+Each example was constructed from as many sequences from that dataset as were necessary to fill the 2048 sequence length.
+
+The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer.
+
+### Fine-Tuning
+
+We fine tuned this model on the [databricks-dolly-15k dataset](https://github.com/databrickslabs/dolly/tree/master/data) released by Databricks, following the same hyperparameters found in their [train_dolly.py](https://github.com/databrickslabs/dolly/blob/master/train_dolly.py) script.
+
+## Training Configuration
+
+This model was pre-trained on 440 A100-40GBs for about half a day using the [MosaicML Platform](https://www.mosaicml.com/platform). The model was pre-trained with sharded data parallelism using FSDP.
+
+## Acknowledgements
+
+This model builds on the work of [Together](https://www.together.xyz), which created the RedPajama dataset with the goal of mimicking the training data used to create the Llama series of models.
+We gratefully acknowledge the hard work of the team that put together this dataset, and we hope this model serves as a useful companion to that work.
+
+This model also builds on the work of [Databricks](https://www.databricks.com/), which created the Dolly instruction fine-tuning dataset.
+
+We also gratefully acknowledge the work of the researchers who created the Llama series of models, which was the impetus for our efforts and those who worked on the RedPajama project.

attention.py

@@ -0,0 +1,408 @@
+# Copyright 2022 MosaicML Examples authors
+# SPDX-License-Identifier: Apache-2.0
+
+"""Attention layers."""
+
+import math
+import warnings
+from typing import Optional
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from torch import nn
+
+from .low_precision_layernorm import LPLayerNorm
+
+
+def _reset_is_causal(num_query_tokens: int, num_key_tokens: int,
+                     original_is_causal: bool):
+    if original_is_causal and num_query_tokens != num_key_tokens:
+        if num_query_tokens != 1:
+            raise NotImplementedError(
+                'MosaicGPT does not support query and key with different number of tokens, unless number of query tokens is 1.'
+            )
+        else:
+            return False
+    return original_is_causal
+
+
+def scaled_multihead_dot_product_attention(
+    query,
+    key,
+    value,
+    n_heads,
+    softmax_scale=None,
+    attn_bias=None,
+    key_padding_mask=None,
+    is_causal=False,
+    dropout_p=0.0,
+    training=False,
+    needs_weights=False,
+):
+
+    q = rearrange(query, 'b s (h d) -> b h s d', h=n_heads)
+    k = rearrange(key, 'b s (h d) -> b h d s', h=n_heads)  # includes key.t()
+    v = rearrange(value, 'b s (h d) -> b h s d', h=n_heads)
+
+    min_val = torch.finfo(q.dtype).min
+
+    b, _, s_q, d = q.shape
+    s_k = k.size(-1)
+
+    if softmax_scale is None:
+        softmax_scale = 1 / math.sqrt(d)
+
+    attn_weight = q.matmul(k) * softmax_scale
+
+    if attn_bias is not None:
+        if (attn_bias.size(-1) != 1 and
+                attn_bias.size(-1) != s_k) or (attn_bias.size(-2) != 1 and
+                                               attn_bias.size(-2) != s_q):
+            raise RuntimeError(
+                f'attn_bias (shape: {attn_bias.shape}) is expected to broadcast to shape: {attn_weight.shape}.'
+            )
+        attn_weight = attn_weight + attn_bias
+
+    if key_padding_mask is not None:
+        if attn_bias is not None:
+            warnings.warn(
+                'Propogating key_padding_mask to the attention module ' +\
+                'and applying it within the attention module can cause ' +\
+                'unneccessary computation/memory usage. Consider integrating ' +\
+                'into attn_bias once and passing that to each attention ' +\
+                'module instead.'
+            )
+        attn_weight = attn_weight.masked_fill(
+            ~key_padding_mask.view((b, 1, 1, s_k)), min_val)
+
+    if is_causal:
+        s = max(s_q, s_k)
+        causal_mask = attn_weight.new_ones(s, s, dtype=torch.float16)
+        causal_mask = causal_mask.tril()
+        causal_mask = causal_mask.to(torch.bool)
+        causal_mask = ~causal_mask
+        causal_mask = causal_mask[-s_q:, -s_k:]
+        attn_weight = attn_weight.masked_fill(causal_mask.view(1, 1, s_q, s_k),
+                                              min_val)
+
+    attn_weight = torch.softmax(attn_weight, dim=-1)
+
+    if dropout_p:
+        attn_weight = torch.nn.functional.dropout(attn_weight,
+                                                  p=dropout_p,
+                                                  training=training,
+                                                  inplace=True)
+
+    out = attn_weight.matmul(v)
+    out = rearrange(out, 'b h s d -> b s (h d)')
+
+    if needs_weights:
+        return out, attn_weight
+    return out, None
+
+
+def check_valid_inputs(*tensors, valid_dtypes=[torch.float16, torch.bfloat16]):
+    for tensor in tensors:
+        if tensor.dtype not in valid_dtypes:
+            raise TypeError(f'{tensor.dtype=} must be in {valid_dtypes=}.')
+        if not tensor.is_cuda:
+            raise TypeError(f'Inputs must be cuda tensors ({tensor.is_cuda=}).')
+
+
+def flash_attn_fn(
+    query,
+    key,
+    value,
+    n_heads,
+    softmax_scale=None,
+    attn_bias=None,
+    key_padding_mask=None,
+    is_causal=False,
+    dropout_p=0.0,
+    training=False,
+    needs_weights=False,
+):
+    try:
+        from flash_attn import bert_padding, flash_attn_interface
+    except:
+        raise RuntimeError('Please install flash_attn==0.2.8')
+
+    check_valid_inputs(query, key, value)
+
+    if attn_bias is not None:
+        raise NotImplementedError(f'attn_bias not implemented for flash attn.')
+
+    batch_size, seqlen = query.shape[:2]
+
+    if key_padding_mask is None:
+        key_padding_mask = torch.ones_like(key[:, :, 0], dtype=torch.bool)
+    query_padding_mask = key_padding_mask[:, -query.size(1):]
+
+    query_unpad, indices_q, cu_seqlens_q, max_seqlen_q = bert_padding.unpad_input(
+        query, query_padding_mask)
+    query_unpad = rearrange(query_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
+
+    key_unpad, _, cu_seqlens_k, max_seqlen_k = bert_padding.unpad_input(
+        key, key_padding_mask)
+    key_unpad = rearrange(key_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
+
+    value_unpad, _, _, _ = bert_padding.unpad_input(value, key_padding_mask)
+    value_unpad = rearrange(value_unpad, 'nnz (h d) -> nnz h d', h=n_heads)
+
+    dropout_p = dropout_p if training else 0.0
+
+    reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
+
+    output_unpad = flash_attn_interface.flash_attn_unpadded_func(
+        query_unpad,
+        key_unpad,
+        value_unpad,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlen_q,
+        max_seqlen_k,
+        dropout_p,
+        softmax_scale=softmax_scale,
+        causal=reset_is_causal,
+        return_attn_probs=needs_weights)
+
+    output = bert_padding.pad_input(
+        rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size,
+        seqlen)
+    return output, None
+
+
+def triton_flash_attn_fn(
+    query,
+    key,
+    value,
+    n_heads,
+    softmax_scale=None,
+    attn_bias=None,
+    key_padding_mask=None,
+    is_causal=False,
+    dropout_p=0.0,
+    training=False,
+    needs_weights=False,
+):
+    try:
+        from flash_attn import flash_attn_triton  # type: ignore
+    except:
+        raise RuntimeError('Please install flash_attn==0.2.8 and triton==2.0.0.dev20221202.')
+
+    check_valid_inputs(query, key, value)
+
+    if dropout_p:
+        raise NotImplementedError(
+            f'Dropout not implemented for attn_impl: triton.')
+
+    if needs_weights:
+        raise NotImplementedError(
+            f'attn_impl: triton cannot return attn weights.')
+
+    if key_padding_mask is not None:
+        warnings.warn(
+            'Propagating key_padding_mask to the attention module ' +\
+            'and applying it within the attention module can cause ' +\
+            'unnecessary computation/memory usage. Consider integrating ' +\
+            'into attn_bias once and passing that to each attention ' +\
+            'module instead.'
+        )
+        b_size, s_k = key_padding_mask.shape[:2]
+
+        if attn_bias is None:
+            attn_bias = query.new_zeros(b_size, 1, 1, s_k)
+
+        attn_bias = attn_bias.masked_fill(
+            ~key_padding_mask.view((b_size, 1, 1, s_k)),
+            torch.finfo(query.dtype).min)
+
+    query = rearrange(query, 'b s (h d) -> b s h d', h=n_heads)
+    key = rearrange(key, 'b s (h d) -> b s h d', h=n_heads)
+    value = rearrange(value, 'b s (h d) -> b s h d', h=n_heads)
+
+    reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
+    attn_output = flash_attn_triton.flash_attn_func(query, key, value,
+                                                    attn_bias, reset_is_causal,
+                                                    softmax_scale)
+
+    output = attn_output.view(*attn_output.shape[:2], -1)
+
+    return output, None
+
+
+class MultiheadAttention(nn.Module):
+    """Multi-head self attention.
+
+    Using torch or triton attention implemetation enables user to also use
+    additive bias.
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        n_heads: int,
+        attn_impl: str = 'triton',
+        attn_clip_qkv: Optional[float] = None,
+        attn_qk_ln: bool = False,
+        softmax_scale: Optional[float] = None,
+        attn_pdrop: float = 0.0,
+        low_precision_layernorm: bool = False,
+        device: Optional[str] = None,
+    ):
+        super().__init__()
+
+        self.attn_impl = attn_impl
+        self.clip_qkv = attn_clip_qkv
+        self.attn_qk_ln = attn_qk_ln
+
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.softmax_scale = softmax_scale
+        if self.softmax_scale is None:
+            self.softmax_scale = 1 / math.sqrt(self.d_model / self.n_heads)
+        self.attn_dropout_p = attn_pdrop
+
+        self.Wqkv = nn.Linear(self.d_model, 3 * self.d_model, device=device)
+        # for param init fn; enables shape based init of fused layers
+        fuse_splits = (d_model, 2 * d_model)
+        self.Wqkv._fused = (0, fuse_splits)  # type: ignore
+
+        if self.attn_qk_ln:
+            layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
+            self.q_ln = layernorm_class(self.d_model, device=device)
+            self.k_ln = layernorm_class(self.d_model, device=device)
+
+        if self.attn_impl == 'flash':
+            self.attn_fn = flash_attn_fn
+        elif self.attn_impl == 'triton':
+            self.attn_fn = triton_flash_attn_fn
+            warnings.warn(
+                'While `attn_impl: triton` can be faster than `attn_impl: flash` ' +\
+                'it uses more memory. When training larger models this can trigger '  +\
+                'alloc retries which hurts performance. If encountered, we recommend ' +\
+                'using `attn_impl: flash` if your model does not use `alibi` or `prefix_lm`.')
+        elif self.attn_impl == 'torch':
+            self.attn_fn = scaled_multihead_dot_product_attention
+            if torch.cuda.is_available():
+                warnings.warn(
+                    'Using `attn_impl: torch`. If your model does not use `alibi` or ' +\
+                    '`prefix_lm` we recommend using `attn_impl: flash` otherwise ' +\
+                    'we recommend using `attn_impl: triton`.'
+                )
+        else:
+            raise ValueError(f'{attn_impl=} is an invalid setting.')
+
+        self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
+        self.out_proj._is_residual = True  # type: ignore
+
+    def forward(self,
+                x,
+                past_key_value=None,
+                attn_bias=None,
+                attention_mask=None,
+                is_causal=True,
+                needs_weights=False):
+        qkv = self.Wqkv(x)
+
+        if self.clip_qkv:
+            qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
+
+        query, key, value = qkv.chunk(3, dim=2)
+
+        key_padding_mask = attention_mask
+
+        if self.attn_qk_ln:
+            # Applying layernorm to qk
+            dtype = query.dtype
+            query = self.q_ln(query).to(dtype)
+            key = self.k_ln(key).to(dtype)
+
+        if past_key_value is not None:
+            if len(past_key_value) != 0:
+                key = torch.cat([past_key_value[0], key], dim=1)
+                value = torch.cat([past_key_value[1], value], dim=1)
+
+            past_key_value = (key, value)
+
+        if attn_bias is not None:
+            attn_bias = attn_bias[:, :, -query.size(1):, -key.size(1):]
+
+        context, attn_weights = self.attn_fn(
+            query,
+            key,
+            value,
+            self.n_heads,
+            softmax_scale=self.softmax_scale,
+            attn_bias=attn_bias,
+            key_padding_mask=key_padding_mask,
+            is_causal=is_causal,
+            dropout_p=self.attn_dropout_p,
+            training=self.training,
+            needs_weights=needs_weights,
+        )
+
+        return self.out_proj(context), attn_weights, past_key_value
+
+
+def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal,
+                    use_sequence_id):
+    if attn_impl == 'flash':
+        return None
+    elif attn_impl in ['torch', 'triton']:
+        if alibi:
+            if (prefix_lm or not causal) or use_sequence_id:
+                return (1, n_heads, seq_len, seq_len)
+            return (1, n_heads, 1, seq_len)
+        elif prefix_lm or use_sequence_id:
+            return (1, 1, seq_len, seq_len)
+        return None
+    else:
+        raise ValueError(f'{attn_impl=} is an invalid setting.')
+
+
+def attn_bias(attn_impl,
+              attn_bias,
+              n_heads,
+              seq_len,
+              causal=False,
+              alibi=False,
+              alibi_bias_max=8):
+    if attn_impl == 'flash':
+        return None
+    elif attn_impl in ['torch', 'triton']:
+        if alibi:
+            # in place add alibi to attn bias
+            device, dtype = attn_bias.device, attn_bias.dtype
+            attn_bias = attn_bias.add(
+                alibi_bias(n_heads,
+                           seq_len,
+                           full=not causal,
+                           alibi_bias_max=alibi_bias_max,
+                           device=device,
+                           dtype=dtype))
+        return attn_bias
+    else:
+        raise ValueError(f'{attn_impl=} is an invalid setting.')
+
+
+def alibi_bias(n_heads,
+               seq_len,
+               full=False,
+               alibi_bias_max=8,
+               device=None,
+               dtype=None):
+    alibi_bias = torch.arange(1 - seq_len, 1, dtype=dtype,
+                              device=device).view(1, 1, 1, seq_len)
+    if full:
+        # generate 1 x Heads x SeqLen x SeqLen alibi bias mask
+        # otherwise the mask is 1 x Heads x 1 x SeqLen (which is broadcast to the appropriate size)
+        alibi_bias = alibi_bias - torch.arange(
+            1 - seq_len, 1, dtype=dtype, device=device).view(1, 1, seq_len, 1)
+        alibi_bias = alibi_bias.abs().mul(-1)
+
+    m = torch.arange(1, n_heads + 1, dtype=dtype, device=device)
+    m = m.mul(alibi_bias_max / n_heads)
+    alibi_bias = alibi_bias * (1. / (2**m.view(1, n_heads, 1, 1)))
+    return alibi_bias

config.json

@@ -0,0 +1,46 @@
+{
+  "_name_or_path": "mosaicml/mosaic-llama-redpajama-final-candidate",
+  "alibi": true,
+  "alibi_bias_max": 8,
+  "architectures": [
+    "MosaicGPT"
+  ],
+  "attn_clip_qkv": null,
+  "attn_impl": "torch",
+  "attn_pdrop": 0,
+  "attn_qk_ln": true,
+  "attn_uses_sequence_id": false,
+  "auto_map": {
+    "AutoConfig": "configuration_mosaic_gpt.MosaicGPTConfig",
+    "AutoModelForCausalLM": "mosaic_gpt.MosaicGPT"
+  },
+  "d_model": 2048,
+  "emb_init_std": null,
+  "emb_init_uniform_lim": null,
+  "emb_pdrop": 0,
+  "embedding_fraction": 1.0,
+  "fan_mode": "fan_in",
+  "init_device": "cpu",
+  "init_div_is_residual": true,
+  "init_gain": 0,
+  "init_nonlinearity": "relu",
+  "init_std": 0.02,
+  "logit_scale": null,
+  "low_precision_layernorm": true,
+  "max_seq_len": 2048,
+  "mlp_ratio": 4,
+  "model_type": "mosaic_gpt",
+  "n_heads": 16,
+  "n_layers": 24,
+  "no_bias": true,
+  "param_init_fn": "kaiming_normal_",
+  "prefix_lm": false,
+  "resid_pdrop": 0,
+  "softmax_scale": null,
+  "tokenizer_name": "EleutherAI/gpt-neox-20b",
+  "torch_dtype": "float32",
+  "transformers_version": "4.27.4",
+  "use_cache": false,
+  "verbose": 0,
+  "vocab_size": 50432
+}

configuration_mosaic_gpt.py

@@ -0,0 +1,168 @@
+# Copyright 2022 MosaicML Examples authors
+# SPDX-License-Identifier: Apache-2.0
+
+"""A HuggingFace-style model configuration."""
+
+from typing import Optional, Tuple, Union
+
+from transformers import PretrainedConfig
+
+
+class MosaicGPTConfig(PretrainedConfig):
+    model_type = 'mosaic_gpt'
+
+    def __init__(
+        self,
+        d_model: int = 2048,
+        n_heads: int = 16,
+        n_layers: int = 24,
+        mlp_ratio: int = 4,
+        max_seq_len: int = 2048,
+        vocab_size: int = 50368,
+        attn_pdrop: float = 0.0,
+        resid_pdrop: float = 0.0,
+        emb_pdrop: float = 0.0,
+        attn_impl: str = 'triton',
+        attn_qk_ln: bool = False,
+        attn_clip_qkv: Optional[float] = None,
+        softmax_scale: Optional[float] = None,
+        prefix_lm: Optional[bool] = False,
+        attn_uses_sequence_id: Optional[bool] = False,
+        alibi: bool = False,
+        alibi_bias_max: int = 8,
+        init_device: str = 'cpu',
+        logit_scale: Optional[Union[float, str]] = None,
+        no_bias: bool = False,
+        verbose: int = 0,
+        param_init_fn: str = 'kaiming_normal_',
+        init_div_is_residual: Union[int, float, str, bool] = True,
+        init_std: float = 0.02,
+        emb_init_std: Optional[float] = None,
+        emb_init_uniform_lim: Optional[Union[Tuple[float, float],
+                                             float]] = None,
+        init_gain: float = 0,
+        fan_mode: str = 'fan_in',
+        init_nonlinearity: str = 'relu',
+        embedding_fraction: float = 1.0,
+        low_precision_layernorm: bool = True,
+        use_cache: bool = False,
+        **kwargs,
+    ):
+        """The MosaicGPT configuration class.
+
+        Args:
+            d_model (int): The size of the embedding dimension of the model.
+            n_heads (int): The number of attention heads.
+            n_layers (int): The number of layers in the model.
+            mlp_ratio (int): The ratio of the up/down scale in the MLP.
+            max_seq_len (int): The maximum sequence length of the model.
+            vocab_size (int): The size of the vocabulary.
+            attn_pdrop (float): The dropout probability for the attention layers.
+            resid_pdrop (float): The dropout probability applied to the attention output before combining with residual.
+            emb_pdrop (float): The dropout probability for the embedding layer.
+            attn_impl (str): The attention implementation to use. One of 'torch', 'flash', or 'triton'.
+            attn_qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
+            attn_clip_qkv (Optional[float]): If not None, clip the queries, keys, and values in the attention layer to
+                this value.
+            softmax_scale (Optional[float]): If not None, scale the softmax in the attention layer by this value. If None,
+                use the default scale of ``1/sqrt(d_keys)``.
+            prefix_lm (Optional[bool]): Whether the model should operate as a Prefix LM. This requires passing an
+                extra `prefix_mask` argument which indicates which tokens belong to the prefix. Tokens in the prefix
+                can attend to one another bi-directionally. Tokens outside the prefix use causal attention.
+            attn_uses_sequence_id (Optional[bool]): Whether to restrict attention to tokens that have the same sequence_id.
+                When the model is in `train` mode, this requires passing an extra `sequence_id` argument which indicates
+                which sub-sequence each token belongs to.
+                Defaults to ``False`` meaning any provided `sequence_id` will be ignored.
+            alibi (bool): Whether to use the alibi bias instead of position embeddings.
+            alibi_bias_max (int): The maximum value of the alibi bias.
+            init_device (str): The device to use for parameter initialization.
+            logit_scale (Optional[Union[float, str]]): If not None, scale the logits by this value.
+            no_bias (bool): Whether to use bias in all layers.
+            verbose (int): The verbosity level. 0 is silent.
+            param_init_fn (str): The parameter initialization scheme to use. One of 'default_', 'baseline_', 'kaiming_uniform_',
+                'kaiming_normal_', 'neox_init_', 'small_init_', 'xavier_uniform_', or 'xavier_normal_'.
+            init_div_is_residual (Union[int, float, str, bool]): Value to divide initial weights by if ``module._is_residual`` is True.
+            init_std (float): The standard deviation of the normal distribution used to initialize the model,
+                if using the baseline_ parameter initialization scheme.
+            emb_init_std (Optional[float]): The standard deviation of the normal distribution used to initialize the embedding layer.
+            emb_init_uniform_lim (Optional[Union[Tuple[float, float], float]]): The lower and upper limits of the uniform distribution
+                used to initialize the embedding layer. Mutually exclusive with ``emb_init_std``.
+            init_gain (float): The gain to use for parameter initialization with kaiming or xavier initialization schemes.
+            fan_mode (str): The fan mode to use for parameter initialization with kaiming initialization schemes.
+            init_nonlinearity (str): The nonlinearity to use for parameter initialization with kaiming initialization schemes.
+            embedding_fraction (float): The fraction to scale the gradients of the embedding layer by.
+            low_precision_layernorm (bool): Whether to use low precision layer normalization.
+            use_cache (bool): Whether or not the model should return the last key/values attentions
+        """
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.mlp_ratio = mlp_ratio
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.attn_pdrop = attn_pdrop
+        self.resid_pdrop = resid_pdrop
+        self.emb_pdrop = emb_pdrop
+        self.attn_impl = attn_impl
+        self.attn_qk_ln = attn_qk_ln
+        self.attn_clip_qkv = attn_clip_qkv
+        self.softmax_scale = softmax_scale
+        self.prefix_lm = prefix_lm
+        self.attn_uses_sequence_id = attn_uses_sequence_id
+        self.alibi = alibi
+        self.alibi_bias_max = alibi_bias_max
+        self.init_device = init_device
+        self.logit_scale = logit_scale
+        self.no_bias = no_bias
+        self.verbose = verbose
+        self.param_init_fn = param_init_fn
+        self.init_div_is_residual = init_div_is_residual
+        self.init_std = init_std
+        self.emb_init_std = emb_init_std
+        self.emb_init_uniform_lim = emb_init_uniform_lim
+        self.init_std = init_std
+        self.init_gain = init_gain
+        self.fan_mode = fan_mode
+        self.init_nonlinearity = init_nonlinearity
+        self.embedding_fraction = embedding_fraction
+        self.low_precision_layernorm = low_precision_layernorm
+        self.use_cache = use_cache
+        if 'name' in kwargs:
+            del kwargs['name']
+        if 'loss_fn' in kwargs:
+            del kwargs['loss_fn']
+        super().__init__(**kwargs)
+
+        self._validate_config()
+
+    def _validate_config(self):
+        if self.d_model % self.n_heads != 0:
+            raise ValueError('d_model must be divisible by n_heads')
+        if any(prob < 0 or prob > 1
+               for prob in [self.attn_pdrop, self.resid_pdrop, self.emb_pdrop]):
+            raise ValueError(
+                'attn_pdrop, resid_pdrop, emb_pdrop are probabilities and must be between 0 and 1'
+            )
+        if self.attn_impl not in ['torch', 'flash', 'triton']:
+            raise ValueError(f'Unknown attn_impl={self.attn_impl}')
+        if self.prefix_lm and self.attn_impl not in ['torch', 'triton']:
+            raise NotImplementedError(
+                'prefix_lm only implemented with torch and triton attention.')
+        if self.alibi and self.attn_impl not in ['torch', 'triton']:
+            raise NotImplementedError(
+                'alibi only implemented with torch and triton attention.')
+        if self.attn_uses_sequence_id and self.attn_impl not in [
+                'torch', 'triton'
+        ]:
+            raise NotImplementedError(
+                'attn_uses_sequence_id only implemented with torch and triton attention.'
+            )
+        if self.embedding_fraction > 1 or self.embedding_fraction <= 0:
+            raise ValueError(
+                'model.embedding_fraction must be between 0 (exclusive) and 1 (inclusive)!'
+            )
+        if isinstance(self.logit_scale,
+                      str) and self.logit_scale != 'inv_sqrt_d_model':
+            raise ValueError(
+                f"{self.logit_scale=} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'."
+            )

generation_config.json

@@ -0,0 +1,5 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "4.27.4",
+  "use_cache": false
+}

gpt_blocks.py

@@ -0,0 +1,90 @@
+# Copyright 2022 MosaicML Examples authors
+# SPDX-License-Identifier: Apache-2.0
+
+"""GPT Blocks used for the GPT Model."""
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+from .attention import MultiheadAttention
+from .low_precision_layernorm import LPLayerNorm
+
+
+class GPTMLP(nn.Module):
+
+    def __init__(self,
+                 d_model: int,
+                 mlp_ratio: int,
+                 device: Optional[str] = None):
+        super().__init__()
+        self.mlp_up = nn.Linear(d_model, mlp_ratio * d_model, device=device)
+        self.mlp_act = nn.GELU(approximate='none')
+        self.mlp_down = nn.Linear(mlp_ratio * d_model, d_model, device=device)
+        self.mlp_down._is_residual = True  # type: ignore
+
+    def forward(self, x):
+        return self.mlp_down(self.mlp_act(self.mlp_up(x)))
+
+
+class GPTBlock(nn.Module):
+
+    def __init__(self,
+                 attn_impl: str,
+                 d_model: int,
+                 n_heads: int,
+                 mlp_ratio: int,
+                 attn_clip_qkv: Optional[float] = None,
+                 attn_qk_ln: bool = False,
+                 softmax_scale: Optional[float] = None,
+                 attn_pdrop: float = 0.0,
+                 alibi: bool = False,
+                 resid_pdrop: float = 0.0,
+                 low_precision_layernorm: bool = False,
+                 device: Optional[str] = None,
+                 **kwargs):
+        del kwargs  # unused, just to capture any extra args from the config
+        super().__init__()
+
+        layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
+
+        self.ln_1 = layernorm_class(d_model, device=device)
+        self.attn = MultiheadAttention(
+            attn_impl=attn_impl,
+            attn_clip_qkv=attn_clip_qkv,
+            attn_qk_ln=attn_qk_ln,
+            softmax_scale=softmax_scale,
+            attn_pdrop=attn_pdrop,
+            d_model=d_model,
+            n_heads=n_heads,
+            device=device,
+        )
+        self.ln_2 = layernorm_class(d_model, device=device)
+        self.mlp = GPTMLP(
+            d_model=d_model,
+            mlp_ratio=mlp_ratio,
+            device=device,
+        )
+        self.resid_attn_dropout = nn.Dropout(resid_pdrop)
+        self.resid_mlp_dropout = nn.Dropout(resid_pdrop)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attn_bias: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.ByteTensor] = None,
+        is_causal: bool = True,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
+        a = self.ln_1(x)
+        b, _, past_key_value = self.attn(a,
+                                         past_key_value=past_key_value,
+                                         attn_bias=attn_bias,
+                                         attention_mask=attention_mask,
+                                         is_causal=is_causal)
+        x = x + self.resid_attn_dropout(b)
+        m = self.ln_2(x)
+        n = self.mlp(m)
+        x = x + self.resid_mlp_dropout(n)
+        return x, past_key_value

low_precision_layernorm.py

@@ -0,0 +1,31 @@
+import torch
+import torch.nn.functional as F
+
+class LPLayerNorm(torch.nn.LayerNorm):
+    def __init__(self, normalized_shape, eps=1e-05, elementwise_affine=True, device=None, dtype=None):
+        super().__init__(
+            normalized_shape=normalized_shape,
+            eps=eps,
+            elementwise_affine=elementwise_affine,
+            device=device,
+            dtype=dtype,
+        )
+
+    def forward(self, x):
+        module_device = x.device
+        downcast_x = _cast_if_autocast_enabled(x)
+        downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
+        downcast_bias = _cast_if_autocast_enabled(self.bias) if self.bias is not None else self.bias
+        with torch.autocast(enabled=False, device_type=module_device.type):
+            return F.layer_norm(downcast_x, self.normalized_shape, downcast_weight, downcast_bias, self.eps)
+
+def _cast_if_autocast_enabled(tensor):
+    if torch.is_autocast_enabled():
+        if tensor.device.type == 'cuda':
+            dtype = torch.get_autocast_gpu_dtype()
+        elif tensor.device.type == 'cpu':
+            dtype = torch.get_autocast_cpu_dtype()
+        else:
+            raise NotImplementedError()
+        return tensor.to(dtype=dtype)
+    return tensor

mosaic_gpt.py

@@ -0,0 +1,449 @@
+# Copyright 2022 MosaicML Examples authors
+# SPDX-License-Identifier: Apache-2.0
+
+"""A simple, flexible implementation of a GPT model.
+
+Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
+"""
+
+import math
+import warnings
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoTokenizer, PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from .attention import attn_bias as module_attn_bias, attn_bias_shape as module_attn_bias_shape
+from .gpt_blocks import GPTBlock
+from .configuration_mosaic_gpt import \
+    MosaicGPTConfig
+from .param_init_fns import MODEL_INIT_REGISTRY
+from .low_precision_layernorm import LPLayerNorm
+
+
+class MosaicGPT(PreTrainedModel):
+    config_class = MosaicGPTConfig
+    base_model_prefix = 'mosaic_gpt'
+
+    def __init__(self, config: MosaicGPTConfig):
+        super().__init__(config)
+
+        if config.attn_impl == 'flash' and config.alibi:
+            raise RuntimeError("ALiBi is not supported with flash attention. Please use triton or torch.")
+
+        self.attn_impl = config.attn_impl
+        self.prefix_lm = config.prefix_lm
+        self.attn_uses_sequence_id = config.attn_uses_sequence_id
+        self.alibi = config.alibi
+        self.alibi_bias_max = config.alibi_bias_max
+
+        layernorm_class = LPLayerNorm if config.low_precision_layernorm else nn.LayerNorm
+
+        # CogView (https://arxiv.org/abs/2105.13290) and GLM-130B (https://arxiv.org/abs/2210.02414)
+        # both report this helping with stabilizing training
+        self.embedding_fraction = config.embedding_fraction
+
+        self.transformer = nn.ModuleDict({
+            'wte':
+                nn.Embedding(config.vocab_size,
+                             config.d_model,
+                             device=config.init_device)
+        })
+        if not self.alibi:
+            self.transformer.update({
+                'wpe':
+                    nn.Embedding(config.max_seq_len,
+                                 config.d_model,
+                                 device=config.init_device)
+            })
+        self.transformer.update({'emb_drop': nn.Dropout(config.emb_pdrop)})
+        self.transformer.update({
+            'blocks':
+                nn.ModuleList([
+                    GPTBlock(device=config.init_device,
+                                        **config.to_dict())
+                    for _ in range(config.n_layers)
+                ])
+        })
+        self.transformer.update({
+            'ln_f': layernorm_class(config.d_model, device=config.init_device)
+        })
+
+        # enables scaling output logits; similar to a softmax "temperature"
+        # PaLM paper uses scale 1/sqrt(config.d_model)
+        self.logit_scale = None
+        if config.logit_scale is not None:
+            logit_scale = config.logit_scale
+            if isinstance(logit_scale, str):
+                if logit_scale == 'inv_sqrt_d_model':
+                    logit_scale = 1 / math.sqrt(config.d_model)
+                else:
+                    raise ValueError(
+                        f"{logit_scale=} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'."
+                    )
+            self.logit_scale = logit_scale
+
+        if config.init_device != 'meta':
+            print(
+                f'You are using {config.init_device=}, but you can also use config.init_device="meta" with Composer + FSDP for fast initialization.'
+            )
+            self.apply(self.param_init_fn)
+
+        self.is_causal = not self.prefix_lm
+
+        # define attn mask
+        self._attn_bias_initialized = False
+        self.attn_bias = None
+        self.attn_bias_shape = module_attn_bias_shape(
+            self.attn_impl,
+            config.n_heads,
+            config.max_seq_len,
+            self.alibi,
+            prefix_lm=self.prefix_lm,
+            causal=self.is_causal,
+            use_sequence_id=self.attn_uses_sequence_id)
+
+        if config.no_bias:
+            for module in self.modules():
+                if hasattr(module, 'bias') and isinstance(
+                        module.bias, nn.Parameter):
+                    if config.verbose:
+                        print(f'Removing bias ({module.bias}) from {module}.')
+                    module.register_parameter('bias', None)
+
+        if config.verbose and config.verbose > 2:
+            print(self)
+
+    @torch.no_grad()
+    def _attn_bias(self,
+                   device,
+                   dtype,
+                   attention_mask: Optional[torch.ByteTensor] = None,
+                   prefix_mask: Optional[torch.ByteTensor] = None,
+                   sequence_id: Optional[torch.LongTensor] = None):
+        if not self._attn_bias_initialized:
+            if self.attn_bias_shape:
+                self.attn_bias = torch.zeros(self.attn_bias_shape,
+                                             device=device,
+                                             dtype=dtype)
+                self.attn_bias = module_attn_bias(
+                    self.attn_impl,
+                    self.attn_bias,
+                    self.config.n_heads,
+                    self.config.max_seq_len,
+                    causal=self.is_causal,
+                    alibi=self.alibi,
+                    alibi_bias_max=self.alibi_bias_max)
+            self._attn_bias_initialized = True
+
+        # flash does not support prefix_lm and will incorporate any
+        # attention_mask inside the attention module
+        if self.attn_impl == 'flash':
+            return self.attn_bias, attention_mask
+
+        attn_bias = self.attn_bias
+
+        # If using torch or triton, we incorporate the prefix_mask (if appropriate)
+        if self.prefix_lm:
+            assert isinstance(attn_bias, torch.Tensor)  # pyright
+            assert isinstance(prefix_mask, torch.Tensor)  # pyright
+            attn_bias = self._apply_prefix_mask(attn_bias, prefix_mask)
+
+        # If using torch or triton, we incorporate sequence_id (if appropriate)
+        if self.attn_uses_sequence_id and sequence_id is not None:
+            assert isinstance(attn_bias, torch.Tensor)  # pyright
+            attn_bias = self._apply_sequence_id(attn_bias, sequence_id)
+
+        # If using torch or triton, we incorporate attention_mask. This will output
+        # None in place of attention_mask since it will not be further needed in the
+        # attention modules.
+        if attention_mask is not None:
+            s_k = attention_mask.shape[-1]
+            if attn_bias is None:
+                attn_bias = torch.zeros((1, 1, 1, s_k),
+                                        device=device,
+                                        dtype=dtype)
+            else:
+                attn_bias = attn_bias[:, :, :, -s_k:]
+            if prefix_mask is not None and (attention_mask.shape !=
+                                            prefix_mask.shape):
+                raise ValueError(
+                    f'attention_mask shape={attention_mask.shape} ' +\
+                    f'and prefix_mask shape={prefix_mask.shape} are not equal.'
+                )
+            min_val = torch.finfo(attn_bias.dtype).min
+            attn_bias = attn_bias.masked_fill(
+                ~attention_mask.view(-1, 1, 1, s_k), min_val)
+
+        return attn_bias, None
+
+    def _apply_prefix_mask(self, attn_bias: torch.Tensor,
+                           prefix_mask: torch.Tensor):
+        s_k, s_q = attn_bias.shape[-2:]
+        if (s_k != self.config.max_seq_len) or (s_q != self.config.max_seq_len):
+            raise ValueError(
+                'attn_bias does not match the expected shape. ' +\
+                f'The last two dimensions should both be {self.config.max_length} ' +\
+                f'but are {s_k} and {s_q}.'
+            )
+        seq_len = prefix_mask.shape[-1]
+        if seq_len > self.config.max_seq_len:
+            raise ValueError(
+                f'prefix_mask sequence length cannot exceed max_seq_len={self.config.max_seq_len}'
+            )
+
+        # select seq_len subset of attn mask
+        attn_bias = attn_bias[..., :seq_len, :seq_len]
+
+        # Mix the causal max and the bidirectional mask to get the full
+        # allowable attention (i.e. full = not accounting for padding yet)
+        causal = torch.tril(
+            torch.ones((seq_len, seq_len),
+                       dtype=torch.bool,
+                       device=prefix_mask.device)).view(1, 1, seq_len, seq_len)
+        prefix = prefix_mask.view(-1, 1, 1, seq_len)
+        cannot_attend = ~torch.logical_or(causal, prefix.bool())
+
+        min_val = torch.finfo(attn_bias.dtype).min
+        attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
+
+        return attn_bias
+
+    def _apply_sequence_id(self, attn_bias: torch.Tensor,
+                           sequence_id: torch.LongTensor):
+        seq_len = sequence_id.shape[-1]
+        if seq_len > self.config.max_seq_len:
+            raise ValueError(
+                f'sequence_id sequence length cannot exceed max_seq_len={self.config.max_seq_len}'
+            )
+
+        # select seq_len subset of attn mask
+        attn_bias = attn_bias[..., :seq_len, :seq_len]
+
+        # Restrict attention to tokens that share the same value
+        # in sequence_id
+        cannot_attend = torch.logical_not(
+            torch.eq(sequence_id.view(-1, seq_len, 1),
+                     sequence_id.view(-1, 1, seq_len))).unsqueeze(1)
+        min_val = torch.finfo(attn_bias.dtype).min
+        attn_bias = attn_bias.masked_fill(cannot_attend, min_val)
+
+        return attn_bias
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor,
+            past_key_values: Optional[List[Tuple[torch.FloatTensor]]] = None,
+            attention_mask: Optional[torch.ByteTensor] = None,
+            prefix_mask: Optional[torch.ByteTensor] = None,
+            sequence_id: Optional[torch.LongTensor] = None,
+            return_dict: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            use_cache: Optional[bool] = None):
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        # These args are passed in by keyword in huggingface's generate function
+        # https://github.com/huggingface/transformers/blob/68287689f2f0d8b7063c400230b3766987abf18d/src/transformers/generation/utils.py#L2201-L2206
+        # but have not yet been fully implemented in MosaicGPT
+        if not return_dict:
+            raise NotImplementedError(
+                'return_dict False is not implemented yet for MosaicGPT')
+        if output_attentions:
+            raise NotImplementedError(
+                'output_attentions is not implemented yet for MosaicGPT')
+
+        if attention_mask is not None and attention_mask[:, 0].sum(
+        ) != attention_mask.shape[0] and self.training:
+            raise NotImplementedError(
+                'MosaicGPT does not support training with left padding.')
+
+        if self.prefix_lm and prefix_mask is None:
+            raise ValueError(
+                'prefix_mask is a required argument when MosaicGPT is configured with prefix_lm=True.'
+            )
+
+        if self.training:
+            if self.attn_uses_sequence_id and sequence_id is None:
+                raise ValueError(
+                    'sequence_id is a required argument when MosaicGPT is configured with attn_uses_sequence_id=True ' +\
+                    'and the model is in train mode.'
+                )
+            elif (self.attn_uses_sequence_id is False) and (sequence_id
+                                                            is not None):
+                warnings.warn(
+                    'MosaicGPT received non-None input for `sequence_id` but is configured with attn_uses_sequence_id=False. ' +\
+                    'This input will be ignored. If you want the model to use `sequence_id`, set attn_uses_sequence_id to True.'
+                )
+
+        S = input_ids.size(1)
+
+        assert (
+            S <= self.config.max_seq_len
+        ), f'Cannot forward input with seq_len={S}, this model only supports seq_len<={self.config.max_seq_len}'
+
+        tok_emb = self.transformer.wte(input_ids)  # type: ignore
+        if self.alibi:
+            x = tok_emb
+        else:
+            past_position = 0
+            if past_key_values is not None:
+                if len(past_key_values) != self.config.n_layers:
+                    raise ValueError(
+                        f'past_key_values must provide a past_key_value for each attention ' +\
+                        f'layer in the network ({len(past_key_values)=}; {self.config.n_layers=}).'
+                    )
+                # get the key tensor whose spec should be (batch, seq, dim), and
+                # collect the `seq`, so that the position embedding is shifted
+                past_position = past_key_values[0][0].size(1)
+
+            if S + past_position > self.config.max_seq_len:
+                raise ValueError(
+                    f'Cannot forward input with past sequence length {past_position} and current sequence length '
+                    f'{S + 1}, this model only supports total sequence length <= {self.config.max_seq_len}.'
+                )
+            pos = torch.arange(past_position,
+                               S + past_position,
+                               dtype=torch.long,
+                               device=input_ids.device).unsqueeze(0)
+            if attention_mask is not None:
+                # adjust the position indices to account for padding tokens
+                pos = torch.clamp(pos - torch.cumsum(
+                    (~attention_mask).to(torch.int32), dim=1)[:,
+                                                              past_position:],
+                                  min=0)
+
+            pos_emb = self.transformer.wpe(pos)  # type: ignore
+            x = tok_emb + pos_emb
+
+        if self.embedding_fraction == 1:
+            x = self.transformer.emb_drop(x)  # type: ignore
+        else:
+            # this implementation is proposed on page 7 of the GLM-130B paper https://arxiv.org/abs/2210.02414
+            x_shrunk = (x * self.embedding_fraction) + (
+                x.detach() * (1 - self.embedding_fraction))
+            assert isinstance(self.transformer.emb_drop, nn.Module)  # pyright
+            x = self.transformer.emb_drop(x_shrunk)
+
+        attn_bias, attention_mask = self._attn_bias(
+            device=x.device,
+            dtype=x.dtype,
+            attention_mask=attention_mask,
+            prefix_mask=prefix_mask,
+            sequence_id=sequence_id)
+
+        # initialize the past key values cache if it should be used
+        if use_cache and past_key_values is None:
+            past_key_values = [() for _ in range(self.config.n_layers)
+                              ]  # type: ignore
+
+        all_hidden_states = () if output_hidden_states else None
+        for b_idx, block in enumerate(self.transformer.blocks):  # type: ignore
+            if output_hidden_states:
+                assert all_hidden_states is not None  # pyright
+                all_hidden_states = all_hidden_states + (x,)
+            past_key_value = past_key_values[
+                b_idx] if past_key_values is not None else None
+            x, past_key_value = block(x,
+                                      past_key_value=past_key_value,
+                                      attn_bias=attn_bias,
+                                      attention_mask=attention_mask,
+                                      is_causal=self.is_causal)
+            if past_key_values is not None:
+                past_key_values[b_idx] = past_key_value
+
+        x = self.transformer.ln_f(x)  # type: ignore
+
+        # output embedding weight tied to input embedding
+        assert isinstance(self.transformer.wte, nn.Module)  # pyright
+        assert isinstance(self.transformer.wte.weight, torch.Tensor)  # pyright
+        logits = F.linear(x, self.transformer.wte.weight, None)
+
+        if self.logit_scale is not None:
+            if self.logit_scale == 0:
+                warnings.warn(
+                    f'Multiplying logits by {self.logit_scale=}. This will produce uniform (uninformative) outputs.'
+                )
+            logits *= self.logit_scale
+
+        return CausalLMOutputWithPast(logits=logits,
+                                      past_key_values=past_key_values,
+                                      hidden_states=all_hidden_states)
+
+    # Param Initialization, needed for device='meta' fast initialization
+    def param_init_fn(self, module):
+        init_fn_name = self.config.param_init_fn
+        if self.config.verbose > 1:
+            warnings.warn(f'Using {init_fn_name} initialization.')
+        MODEL_INIT_REGISTRY[init_fn_name](module=module,
+                                          **self.config.to_dict())
+
+    # FSDP Wrap function
+    def fsdp_wrap_fn(self, module):
+        return isinstance(module, GPTBlock)
+
+    # Activation Checkpointing
+    def activation_checkpointing_fn(self, module):
+        return isinstance(module, GPTBlock)
+
+    def prepare_inputs_for_generation(self,
+                                      input_ids,
+                                      past_key_values=None,
+                                      inputs_embeds=None,
+                                      **kwargs):
+        if inputs_embeds is not None:
+            raise NotImplementedError(
+                'inputs_embeds is not implemented for MosaicGPT yet')
+
+        attention_mask = kwargs['attention_mask'].bool()
+        if attention_mask[:, -1].sum() != attention_mask.shape[0]:
+            raise NotImplementedError(
+                'MosaicGPT does not support generation with right padding.')
+
+        if self.attn_uses_sequence_id and self.training:
+            sequence_id = torch.zeros_like(input_ids[:1])
+        else:
+            sequence_id = None
+
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+
+        if self.prefix_lm:
+            # Leverage a convenience of sequential generation!
+            prefix_mask = torch.ones_like(attention_mask)
+            # This requires that we're using the cache
+            if kwargs.get('use_cache') == False:
+                raise NotImplementedError(
+                    'MosaicGPT with prefix_lm=True does not support use_cache=False.'
+                )
+        else:
+            prefix_mask = None
+
+        return {
+            'input_ids': input_ids,
+            'attention_mask': attention_mask,
+            'prefix_mask': prefix_mask,
+            'sequence_id': sequence_id,
+            'past_key_values': past_key_values,
+            'use_cache': kwargs.get('use_cache', True),
+        }
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        """Used by HuggingFace generate when using beam search with kv-caching.
+
+        See https://github.com/huggingface/transformers/blob/3ec7a47664ebe40c40f4b722f6bb1cd30c3821ec/src/transformers/models/gpt2/modeling_gpt2.py#L1122-L1133
+        for an example in transformers.
+        """
+        reordered_past = []
+        for layer_past in past_key_values:
+            reordered_past += [
+                tuple(
+                    past_state.index_select(0, beam_idx)
+                    for past_state in layer_past)
+            ]
+        return reordered_past

param_init_fns.py

@@ -0,0 +1,464 @@
+# Copyright 2022 MosaicML Examples authors
+# SPDX-License-Identifier: Apache-2.0
+import math
+import warnings
+from collections.abc import Sequence
+from functools import partial
+from typing import Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+
+def torch_default_param_init_fn_(
+    module: nn.Module,
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+    if verbose > 1:
+        warnings.warn(
+            f"Initializing network using module's reset_parameters attribute")
+
+    if hasattr(module, 'reset_parameters'):
+        module.reset_parameters()  # type: ignore
+
+
+def fused_init_helper_(module: nn.Module, init_fn_):
+    # parameter initialization is often based on the parameters shape.
+    # If a layer is fused, initialization should be based on the shapes
+    # of the original tensor instead of the shape of the fused tensor.
+    # Layers which are fused should have the _fused attibute defined.
+    # The first element of _fused is the dimension along which the tensor is fused.
+    # This is followed by an iterable of split indices."
+
+    _fused = getattr(module, '_fused', None)
+
+    if _fused is None:
+        raise RuntimeError(f'Internal logic error')
+
+    dim, splits = _fused
+    splits = (0, *splits, module.weight.size(dim))  # type: ignore
+    for s, e in zip(splits[:-1], splits[1:]):
+        slice_indices = [slice(None)] * module.weight.ndim  # type: ignore
+        slice_indices[dim] = slice(s, e)
+        init_fn_(module.weight[slice_indices])  # type: ignore
+
+
+def generic_param_init_fn_(
+    module: nn.Module,
+    init_fn_,
+    n_layers: int,
+    d_model: Optional[int] = None,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+    if verbose > 1:
+        warnings.warn(
+            f'If model has bias parameters they are initialized to 0.')
+
+    # enable user to divide _is_residual weights by
+    # a value which defaults to math.sqrt(2 * cfg.n_layers)
+    init_div_is_residual = init_div_is_residual
+
+    if init_div_is_residual is False:
+        # not used, for pyright
+        div_is_residual = 1.0
+    elif init_div_is_residual is True:
+        div_is_residual = math.sqrt(2 * n_layers)
+    elif isinstance(init_div_is_residual, float) or isinstance(
+            init_div_is_residual, int):
+        div_is_residual = init_div_is_residual
+    elif isinstance(init_div_is_residual,
+                    str) and init_div_is_residual.isnumeric():
+        # do not trust YAML parsing to always convert numbers to numbers
+        div_is_residual = float(init_div_is_residual)
+    else:
+        # not used, for pyright
+        div_is_residual = 1.0
+        raise ValueError(
+            f'Expected init_div_is_residual to be boolean or numeric, got {init_div_is_residual}'
+        )
+
+    if init_div_is_residual is not False:
+        if verbose > 1:
+            warnings.warn(
+                f'Initializing _is_residual layers then dividing them by {div_is_residual}.' +\
+                f'set `init_div_is_residual: false` in model config to disable this.'
+            )
+
+    if isinstance(module, nn.Linear):
+        # Linear
+        if hasattr(module, '_fused'):
+            fused_init_helper_(module, init_fn_)
+        else:
+            init_fn_(module.weight)
+        if module.bias is not None:
+            torch.nn.init.zeros_(module.bias)
+
+        if init_div_is_residual is not False and getattr(
+                module, '_is_residual', False):
+            with torch.no_grad():
+                module.weight.div_(div_is_residual)
+
+    elif isinstance(module, nn.Embedding):
+        # Embedding
+        if emb_init_std is not None:
+            std = emb_init_std
+            if std == 0:
+                warnings.warn(f'Embedding layer initialized to 0.')
+            emb_init_fn_ = partial(torch.nn.init.normal_, mean=0.0, std=std)
+            if verbose > 1:
+                warnings.warn(
+                    f'Embedding layer initialized using normal distribution with mean=0 and {std=}.'
+                )
+        elif emb_init_uniform_lim is not None:
+            lim = emb_init_uniform_lim
+            if isinstance(lim, Sequence):
+                if len(lim) > 2:
+                    raise ValueError(
+                        f'Uniform init requires a min and a max limit. User input: {lim}.'
+                    )
+                if lim[0] == lim[1]:
+                    warnings.warn(f'Embedding layer initialized to {lim[0]}.')
+            else:
+                if lim == 0:
+                    warnings.warn(f'Embedding layer initialized to 0.')
+                lim = [-lim, lim]
+            a, b = lim
+            emb_init_fn_ = partial(torch.nn.init.uniform_, a=a, b=b)
+            if verbose > 1:
+                warnings.warn(
+                    f'Embedding layer initialized using uniform distribution in range {lim}.'
+                )
+        else:
+            emb_init_fn_ = init_fn_
+
+        emb_init_fn_(module.weight)
+
+    elif isinstance(module, nn.LayerNorm):
+        # LayerNorm
+        if verbose > 1:
+            warnings.warn(
+                f'LayerNorm gamma weights are set to 1. If the layer has a bias it is initialized to 0.'
+            )
+        torch.nn.init.ones_(module.weight)
+        if module.bias is not None:
+            torch.nn.init.zeros_(module.bias)
+
+    elif isinstance(module, nn.MultiheadAttention):
+        # torch's MultiheadAttention
+        if module._qkv_same_embed_dim:
+            assert module.in_proj_weight is not None
+            assert module.q_proj_weight is None and module.k_proj_weight is None and module.v_proj_weight is None
+            assert d_model is not None
+            # in_proj_weight is actually 3 layers and should be split up for width based init
+            _d = d_model
+            splits = (0, _d, 2 * _d, 3 * _d)
+            for s, e in zip(splits[:-1], splits[1:]):
+                init_fn_(module.in_proj_weight[s:e])
+        else:
+            assert module.q_proj_weight is not None and module.k_proj_weight is not None and module.v_proj_weight is not None
+            assert module.in_proj_weight is None
+            init_fn_(module.q_proj_weight)
+            init_fn_(module.k_proj_weight)
+            init_fn_(module.v_proj_weight)
+
+        # bias
+        if module.in_proj_bias is not None:
+            torch.nn.init.zeros_(module.in_proj_bias)
+        if module.bias_k is not None:
+            torch.nn.init.zeros_(module.bias_k)
+        if module.bias_v is not None:
+            torch.nn.init.zeros_(module.bias_v)
+
+        # out proj
+        init_fn_(module.out_proj.weight)
+        if init_div_is_residual is not False and getattr(
+                module.out_proj, '_is_residual', False):
+            with torch.no_grad():
+                module.out_proj.weight.div_(div_is_residual)
+        if module.out_proj.bias is not None:
+            torch.nn.init.zeros_(module.out_proj.bias)
+
+    else:
+        for _ in module.parameters(recurse=False):
+            # raise error if uninitialized module has any parameters
+            raise NotImplementedError(
+                f'{module.__class__.__name__} parameters are not initialized by param_init_fn.'
+            )
+
+
+def _normal_init_(std, mean=0.0):
+    return partial(torch.nn.init.normal_, mean=mean, std=std)
+
+
+def _normal_param_init_fn_(
+    module: nn.Module,
+    std: float,
+    n_layers: int,
+    d_model: Optional[int] = None,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+    init_fn_ = _normal_init_(std=std)
+
+    if verbose > 1:
+        warnings.warn(
+            f'Using torch.nn.init.normal_ init fn mean=0.0, std={std}')
+
+    generic_param_init_fn_(
+        module=module,
+        init_fn_=init_fn_,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=init_div_is_residual,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+def baseline_param_init_fn_(
+    module: nn.Module,
+    init_std: float,
+    n_layers: int,
+    d_model: Optional[int] = None,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+    if init_std is None:
+        raise ValueError(
+            'You must set model.init_std to a float value to use the default initialization scheme.'
+        )
+    _normal_param_init_fn_(
+        module=module,
+        std=init_std,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=init_div_is_residual,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+def small_param_init_fn_(
+    module: nn.Module,
+    n_layers: int,
+    d_model: int,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+    # very close to kaiming normal
+    # from Transformers without Tears (2019) - Nguyen & Salazar
+    std = math.sqrt(2 / (5 * d_model))
+    _normal_param_init_fn_(
+        module=module,
+        std=std,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=init_div_is_residual,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+def neox_param_init_fn_(
+    module: nn.Module,
+    n_layers: int,
+    d_model: int,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    verbose: int = 0,
+    **kwargs,
+):
+    """From section 2.3.1 of GPT-NeoX-20B:
+
+    An Open-Source AutoregressiveLanguage Model — Black et. al. (2022)
+    see https://github.com/EleutherAI/gpt-neox/blob/9610391ab319403cef079b438edd016a2443af54/megatron/model/init_functions.py#L151
+    and https://github.com/EleutherAI/gpt-neox/blob/main/megatron/model/transformer.py
+    """
+    del kwargs  # unused, just to capture any extra args from the config
+    residual_div = n_layers / math.sqrt(10)  # small std / wang std
+
+    if verbose > 1:
+        warnings.warn(f'setting init_div_is_residual to {residual_div}')
+
+    small_param_init_fn_(
+        module=module,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=residual_div,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+def kaiming_uniform_param_init_fn_(
+    module: nn.Module,
+    n_layers: int,
+    d_model: Optional[int] = None,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    init_gain: float = 0,
+    fan_mode: str = 'fan_in',
+    init_nonlinearity: str = 'leaky_relu',
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+
+    if verbose > 1:
+        warnings.warn(
+            f'Using nn.init.kaiming_uniform_ init fn with parameters: ' +\
+            f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}'
+        )
+
+    kaiming_uniform_ = partial(nn.init.kaiming_uniform_,
+                               a=init_gain,
+                               mode=fan_mode,
+                               nonlinearity=init_nonlinearity)
+
+    generic_param_init_fn_(
+        module=module,
+        init_fn_=kaiming_uniform_,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=init_div_is_residual,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+def kaiming_normal_param_init_fn_(
+    module: nn.Module,
+    n_layers: int,
+    d_model: Optional[int] = None,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    init_gain: float = 0,
+    fan_mode: str = 'fan_in',
+    init_nonlinearity: str = 'leaky_relu',
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+
+    if verbose > 1:
+        warnings.warn(
+            f'Using nn.init.kaiming_normal_ init fn with parameters: ' +\
+            f'a={init_gain}, mode={fan_mode}, nonlinearity={init_nonlinearity}'
+        )
+
+    kaiming_normal_ = partial(torch.nn.init.kaiming_normal_,
+                              a=init_gain,
+                              mode=fan_mode,
+                              nonlinearity=init_nonlinearity)
+
+    generic_param_init_fn_(
+        module=module,
+        init_fn_=kaiming_normal_,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=init_div_is_residual,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+def xavier_uniform_param_init_fn_(
+    module: nn.Module,
+    n_layers: int,
+    d_model: Optional[int] = None,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    init_gain: float = 0,
+    verbose: int = 0,
+    **kwargs,
+):
+    del kwargs  # unused, just to capture any extra args from the config
+    xavier_uniform_ = partial(torch.nn.init.xavier_uniform_, gain=init_gain)
+
+    if verbose > 1:
+        warnings.warn(
+            f'Using torch.nn.init.xavier_uniform_ init fn with parameters: ' +\
+            f'gain={init_gain}'
+        )
+
+    generic_param_init_fn_(
+        module=module,
+        init_fn_=xavier_uniform_,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=init_div_is_residual,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+def xavier_normal_param_init_fn_(
+    module: nn.Module,
+    n_layers: int,
+    d_model: Optional[int] = None,
+    init_div_is_residual: Union[int, float, str, bool] = True,
+    emb_init_std: Optional[float] = None,
+    emb_init_uniform_lim: Optional[Union[Tuple[float, float], float]] = None,
+    init_gain: float = 0,
+    verbose: int = 0,
+    **kwargs,
+):
+    xavier_normal_ = partial(torch.nn.init.xavier_normal_, gain=init_gain)
+
+    if verbose > 1:
+        warnings.warn(
+            f'Using torch.nn.init.xavier_normal_ init fn with parameters: ' +\
+            f'gain={init_gain}'
+        )
+
+    generic_param_init_fn_(
+        module=module,
+        init_fn_=xavier_normal_,
+        d_model=d_model,
+        n_layers=n_layers,
+        init_div_is_residual=init_div_is_residual,
+        emb_init_std=emb_init_std,
+        emb_init_uniform_lim=emb_init_uniform_lim,
+        verbose=verbose,
+    )
+
+
+MODEL_INIT_REGISTRY = {
+    'default_': torch_default_param_init_fn_,
+    'baseline_': baseline_param_init_fn_,
+    'kaiming_uniform_': kaiming_uniform_param_init_fn_,
+    'kaiming_normal_': kaiming_normal_param_init_fn_,
+    'neox_init_': neox_param_init_fn_,
+    'small_init_': small_param_init_fn_,
+    'xavier_uniform_': xavier_uniform_param_init_fn_,
+    'xavier_normal_': xavier_normal_param_init_fn_,
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f64cae13311155342c37179f2563fe046e627146bc51282f30b9261dcf4bbefd
+size 5245893681

special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>"
+}

tokenizer_config.json

@@ -0,0 +1,9 @@
+{
+  "add_prefix_space": false,
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 2048,
+  "special_tokens_map_file": "/root/.cache/huggingface/hub/models--EleutherAI--gpt-neox-20b/snapshots/4e49eadb5d14bd22f314ec3f45b69a87b88c7691/special_tokens_map.json",
+  "tokenizer_class": "GPTNeoXTokenizer",
+  "unk_token": "<|endoftext|>"
+}