Fast forward to latest mosaic version

README.md

@@ -14,23 +14,17 @@ datasets:
 inference: false
 ---
 
-### Attribution
-
-This model is derived from [MosaicML's MPT-7B model](https://huggingface.co/mosaicml/mpt-7b/tree/main), with changes from
-[cekal/mpt-7b-peft-compatible](https://huggingface.co/cekal/mpt-7b-peft-compatible) applied; each licensed under the
-Apache License, version 2.0.
-
 # MPT-7B
 
 MPT-7B is a decoder-style transformer pretrained from scratch on 1T tokens of English text and code.
 This model was trained by [MosaicML](https://www.mosaicml.com).
 
-MPT-7B is part of the family of MosaicPretrainedTransformer (MPT) models, which use a modified transformer architecture optimized for efficient training and inference. 
+MPT-7B is part of the family of MosaicPretrainedTransformer (MPT) models, which use a modified transformer architecture optimized for efficient training and inference.
 
-These architectural changes include performance-optimized layer implementations and the elimination of context length limits by replacing 
-positional embeddings with Attention with Linear Biases ([ALiBi](https://arxiv.org/abs/2108.12409)). 
-Thanks to these modifications, MPT models can be trained with high throughput efficiency and stable convergence. 
-MPT models can also be served efficiently with both standard HuggingFace pipelines and NVIDIA's [FasterTransformer](https://github.com/NVIDIA/FasterTransformer). 
+These architectural changes include performance-optimized layer implementations and the elimination of context length limits by replacing
+positional embeddings with Attention with Linear Biases ([ALiBi](https://arxiv.org/abs/2108.12409)).
+Thanks to these modifications, MPT models can be trained with high throughput efficiency and stable convergence.
+MPT models can also be served efficiently with both standard HuggingFace pipelines and NVIDIA's [FasterTransformer](https://github.com/NVIDIA/FasterTransformer).
 
 This model uses the MosaicML LLM codebase, which can be found in the [llm-foundry repository](https://github.com/mosaicml/llm-foundry). It was trained by MosaicML’s NLP team on the [MosaicML platform](https://www.mosaicml.com/training) for LLM pretraining, finetuning, and inference.
 
@@ -55,15 +49,13 @@ We demonstrate generations as long as 80k tokens on a single A100-80GB GPU in ou
   * License: Apache 2.0
 
 * [MPT-7B-Instruct](https://huggingface.co/mosaicml/mpt-7b-instruct): a model for short-form instruction following.
-Built by finetuning MPT-7B on a [dataset](https://huggingface.co/datasets/mosaicml/dolly_hhrlhf) we also release, derived from the [Databricks Dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) and the [Anthropic Helpful and Harmless (HH-RLHF)](https://huggingface.co/datasets/Anthropic/hh-rlhf) datasets. 
-  * License: _CC-By-SA-3.0_
-  * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-7b-instruct)
+Built by finetuning MPT-7B on a [dataset](https://huggingface.co/datasets/mosaicml/dolly_hhrlhf) we also release, derived from the [Databricks Dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) and the [Anthropic Helpful and Harmless (HH-RLHF)](https://huggingface.co/datasets/Anthropic/hh-rlhf) datasets.
+  * License: Apache 2.0
 
 * [MPT-7B-Chat](https://huggingface.co/mosaicml/mpt-7b-chat): a chatbot-like model for dialogue generation.
 Built by finetuning MPT-7B on the [ShareGPT-Vicuna](https://huggingface.co/datasets/jeffwan/sharegpt_vicuna), [HC3](https://huggingface.co/datasets/Hello-SimpleAI/HC3),
  [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca), [HH-RLHF](https://huggingface.co/datasets/Anthropic/hh-rlhf), and [Evol-Instruct](https://huggingface.co/datasets/victor123/evol_instruct_70k) datasets.
   * License: _CC-By-NC-SA-4.0_
-  * [Demo on Hugging Face Spaces](https://huggingface.co/spaces/mosaicml/mpt-7b-chat)
 
 ## Model Date
 
@@ -77,7 +69,7 @@ Apache-2.0
 
 * [Blog post: Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs](https://www.mosaicml.com/blog/mpt-7b)
 * [Codebase (mosaicml/llm-foundry repo)](https://github.com/mosaicml/llm-foundry/)
-* Questions: Feel free to contact us via the [MosaicML Community Slack](https://join.slack.com/t/mosaicml-community/shared_invite/zt-1btms90mc-GipE2ufuPkKY0QBrmF3LSA)!
+* Questions: Feel free to contact us via the [MosaicML Community Slack](https://mosaicml.me/slack)!
 
 
 ## How to Use
@@ -91,37 +83,41 @@ model = transformers.AutoModelForCausalLM.from_pretrained(
   trust_remote_code=True
 )
 ```
-Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method. 
+Note: This model requires that `trust_remote_code=True` be passed to the `from_pretrained` method.
 This is because we use a custom `MPT` model architecture that is not yet part of the Hugging Face `transformers` package.
 `MPT` includes options for many training efficiency features such as [FlashAttention](https://arxiv.org/pdf/2205.14135.pdf), [ALiBi](https://arxiv.org/abs/2108.12409), [QK LayerNorm](https://arxiv.org/abs/2010.04245), and more.
 
-To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model with `attn_impl='triton'` and move the model to `bfloat16`:
+To use the optimized [triton implementation](https://github.com/openai/triton) of FlashAttention, you can load the model on GPU (`cuda:0`) with `attn_impl='triton'` and with `bfloat16` precision:
 ```python
-config = transformers.AutoConfig.from_pretrained(
-  'mosaicml/mpt-7b',
-  trust_remote_code=True
-)
+import torch
+import transformers
+
+name = 'mosaicml/mpt-7b'
+
+config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
 config.attn_config['attn_impl'] = 'triton'
+config.init_device = 'cuda:0' # For fast initialization directly on GPU!
 
 model = transformers.AutoModelForCausalLM.from_pretrained(
-  'mosaicml/mpt-7b',
+  name,
   config=config,
-  torch_dtype=torch.bfloat16,
+  torch_dtype=torch.bfloat16, # Load model weights in bfloat16
   trust_remote_code=True
 )
-model.to(device='cuda:0')
 ```
 
 Although the model was trained with a sequence length of 2048, ALiBi enables users to increase the maximum sequence length during finetuning and/or inference. For example:
 
 ```python
-config = transformers.AutoConfig.from_pretrained(
-  'mosaicml/mpt-7b',
-  trust_remote_code=True
-)
-config.update({"max_seq_len": 4096})
+import transformers
+
+name = 'mosaicml/mpt-7b'
+
+config = transformers.AutoConfig.from_pretrained(name, trust_remote_code=True)
+config.max_seq_len = 4096 # (input + output) tokens can now be up to 4096
+
 model = transformers.AutoModelForCausalLM.from_pretrained(
-  'mosaicml/mpt-7b',
+  name,
   config=config,
   trust_remote_code=True
 )
@@ -131,7 +127,23 @@ This model was trained with the [EleutherAI/gpt-neox-20b](https://huggingface.co
 
 ```python
 from transformers import AutoTokenizer
-tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
+tokenizer = AutoTokenizer.from_pretrained('EleutherAI/gpt-neox-20b')
+```
+
+The model can then be used, for example, within a text-generation pipeline.  
+Note: when running Torch modules in lower precision, it is best practice to use the [torch.autocast context manager](https://pytorch.org/docs/stable/amp.html).
+
+```python
+from transformers import pipeline
+
+pipe = pipeline('text-generation', model=model, tokenizer=tokenizer, device='cuda:0')
+
+with torch.autocast('cuda', dtype=torch.bfloat16):
+    print(
+        pipe('Here is a recipe for vegan banana bread:\n',
+            max_new_tokens=100,
+            do_sample=True,
+            use_cache=True))
 ```
 
 ## Model Description
@@ -159,7 +171,7 @@ The model has been modified from a standard transformer in the following ways:
 
 ### Streaming Datasets
 
-Data was formatted using the MosaicML [StreamingDataset](https://github.com/mosaicml/streaming) library to host our data in object storage and efficiently stream it to our compute cluster during training. 
+Data was formatted using the MosaicML [StreamingDataset](https://github.com/mosaicml/streaming) library to host our data in object storage and efficiently stream it to our compute cluster during training.
 StreamingDataset obviates the need to download the whole dataset before starting training, and allows instant resumption of training from any point in the dataset.
 
 
@@ -184,24 +196,24 @@ The model was trained for 1T tokens (with batch size 1760 and sequence length 20
 Samples for each batch were selected from one of the datasets with the probability specified above.
 The examples were shuffled within each dataset, and 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. This BPE tokenizer has a number of desirable characteristics, 
-most of which are relevant for tokenizing code: 
-(1) It was trained on a diverse mix of data that includes code (The Pile) 
-(2) It applies consistent space delimitation, unlike the GPT2 tokenizer which tokenizes inconsistently depending on the presence of prefix spaces 
-(3) It contains tokens for repeated space characters, which allows superior compression of text with large amounts of repeated space characters. 
+The data was tokenized using the [EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b) tokenizer. This BPE tokenizer has a number of desirable characteristics,
+most of which are relevant for tokenizing code:
+(1) It was trained on a diverse mix of data that includes code (The Pile)
+(2) It applies consistent space delimitation, unlike the GPT2 tokenizer which tokenizes inconsistently depending on the presence of prefix spaces
+(3) It contains tokens for repeated space characters, which allows superior compression of text with large amounts of repeated space characters.
 
 The model vocabulary size of 50432 was set to be a multiple of 128 (as in [MEGATRON-LM](https://arxiv.org/abs/1909.08053)), model flop utilization (MFU) increased by up to four percentage points.
 
 ### Training Configuration
 
-This model was trained on 440 A100-40GBs for about 9.5 days using the [MosaicML Platform](https://www.mosaicml.com/platform). 
-The model was trained with sharded data parallelism using [FSDP](https://pytorch.org/docs/stable/fsdp.html) and used the [LION](https://arxiv.org/abs/2302.06675) optimizer. 
+This model was trained on 440 A100-40GBs for about 9.5 days using the [MosaicML Platform](https://www.mosaicml.com/platform).
+The model was trained with sharded data parallelism using [FSDP](https://pytorch.org/docs/stable/fsdp.html) and used the [LION](https://arxiv.org/abs/2302.06675) optimizer.
 
 ## Limitations and Biases
 
 _The following language is modified from [EleutherAI's GPT-NeoX-20B](https://huggingface.co/EleutherAI/gpt-neox-20b)_
 
-MPT-7B (Base) is **not** intended for deployment without finetuning. 
+MPT-7B (Base) is **not** intended for deployment without finetuning.
 It should not be used for human-facing interactions without further guardrails and user consent.
 
 MPT-7B can produce factually incorrect output, and should not be relied on to produce factually accurate information.
@@ -224,11 +236,11 @@ Please cite this model using the following format:
 ```
 @online{MosaicML2023Introducing,
     author    = {MosaicML NLP Team},
-    title     = {Introducing MPT-7B: A New Standard for Open-Source, 
-    ly Usable LLMs},
+    title     = {Introducing MPT-7B: A New Standard for Open-Source,
+    Commercially Usable LLMs},
     year      = {2023},
     url       = {www.mosaicml.com/blog/mpt-7b},
-    note      = {Accessed: 2023-03-28}, % change this date
-    urldate   = {2023-03-28} % change this date
+    note      = {Accessed: 2023-05-05},
+    urldate   = {2023-05-05}
 }
 ```

adapt_tokenizer.py

@@ -1,9 +1,8 @@
-from typing import Union
-from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast
-Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
+from typing import Any
+from transformers import AutoTokenizer, PreTrainedTokenizerBase
 NUM_SENTINEL_TOKENS: int = 100
 
-def adapt_tokenizer_for_denoising(tokenizer: Tokenizer):
+def adapt_tokenizer_for_denoising(tokenizer: PreTrainedTokenizerBase) -> None:
     """Adds sentinel tokens and padding token (if missing).
 
     Expands the tokenizer vocabulary to include sentinel tokens
@@ -34,7 +33,7 @@ class AutoTokenizerForMOD(AutoTokenizer):
     """
 
     @classmethod
-    def from_pretrained(cls, *args, **kwargs):
+    def from_pretrained(cls, *args: Any, **kwargs: Any) -> PreTrainedTokenizerBase:
         """See `AutoTokenizer.from_pretrained` docstring."""
         tokenizer = super().from_pretrained(*args, **kwargs)
         adapt_tokenizer_for_denoising(tokenizer)

attention.py

@@ -1,14 +1,30 @@
 """Attention layers."""
 import math
 import warnings
-from typing import Optional
+from typing import Any, List, Optional, Tuple
 import torch
 import torch.nn as nn
 from einops import rearrange
+from packaging import version
 from torch import nn
-from .norm import LPLayerNorm
+from .fc import FC_CLASS_REGISTRY
+from .norm import NORM_CLASS_REGISTRY
 
-def _reset_is_causal(num_query_tokens: int, num_key_tokens: int, original_is_causal: bool):
+def is_flash_v2_installed():
+    try:
+        import flash_attn as flash_attn
+    except:
+        return False
+    return version.parse(flash_attn.__version__) >= version.parse('2.0.0')
+
+def is_flash_v1_installed():
+    try:
+        import flash_attn as flash_attn
+    except:
+        return False
+    return version.parse(flash_attn.__version__) < version.parse('2.0.0')
+
+def _reset_is_causal(num_query_tokens: int, num_key_tokens: int, original_is_causal: bool) -> bool:
     if original_is_causal and num_query_tokens != num_key_tokens:
         if num_query_tokens != 1:
             raise NotImplementedError('MPT does not support query and key with different number of tokens, unless number of query tokens is 1.')
@@ -16,27 +32,57 @@ def _reset_is_causal(num_query_tokens: int, num_key_tokens: int, original_is_cau
             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, multiquery=False):
+def repeat_kv_for_gqa(hidden: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """Perform repeat of kv heads along a particular dimension.
+
+    hidden.shape expected to be: (batch size, seq len, kv_n_heads, head_dim)
+    n_rep: amount of repetitions of kv_n_heads
+    Unlike torch.repeat_interleave, this function avoids allocating new memory.
+    """
+    if n_rep == 1:
+        return hidden
+    (b, s, kv_n_heads, d) = hidden.shape
+    hidden = hidden[:, :, :, None, :].expand(b, s, kv_n_heads, n_rep, d)
+    return hidden.reshape(b, s, kv_n_heads * n_rep, d)
+
+def scaled_multihead_dot_product_attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, n_heads: int, kv_n_heads: Optional[int]=None, past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]]=None, softmax_scale: Optional[float]=None, attn_bias: Optional[torch.Tensor]=None, key_padding_mask: Optional[torch.Tensor]=None, is_causal: bool=False, dropout_p: float=0.0, training: bool=False, needs_weights: bool=False, multiquery: bool=False) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor, torch.Tensor]]]:
+    if multiquery:
+        warnings.warn(DeprecationWarning('The direct use of the multiquery arg is deprecated. Setting kv_n_heads=1 automatically. Please set kv_n_heads=1 explicitly to remove this warning.'))
+        kv_n_heads = 1
+    elif kv_n_heads is None:
+        warnings.warn(DeprecationWarning('Not specifying a value for the kv_n_heads arg is deprecated. Setting kv_n_heads=n_heads automatically. Please set kv_n_heads=n_heads explicitly to remove this warning.'))
+        kv_n_heads = n_heads
     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=1 if multiquery else n_heads)
-    v = rearrange(value, 'b s (h d) -> b h s d', h=1 if multiquery else n_heads)
-    min_val = torch.finfo(q.dtype).min
+    k = rearrange(key, 'b s (h d) -> b h d s', h=kv_n_heads)
+    v = rearrange(value, 'b s (h d) -> b h s d', h=kv_n_heads)
+    if past_key_value is not None:
+        if len(past_key_value) != 0:
+            k = torch.cat([past_key_value[0], k], dim=3)
+            v = torch.cat([past_key_value[1], v], dim=2)
+        past_key_value = (k, v)
     (b, _, s_q, d) = q.shape
     s_k = k.size(-1)
+    if kv_n_heads > 1 and kv_n_heads < n_heads:
+        k = repeat_kv_for_gqa(k.transpose(1, 2), n_heads // kv_n_heads).transpose(1, 2)
+        v = repeat_kv_for_gqa(v.transpose(1, 2), n_heads // kv_n_heads).transpose(1, 2)
     if softmax_scale is None:
         softmax_scale = 1 / math.sqrt(d)
     attn_weight = q.matmul(k) * softmax_scale
     if attn_bias is not None:
+        _s_q = max(0, attn_bias.size(2) - s_q)
+        _s_k = max(0, attn_bias.size(3) - s_k)
+        attn_bias = attn_bias[:, :, _s_q:, _s_k:]
         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
+    min_val = torch.finfo(q.dtype).min
     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.')
+            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.')
         attn_weight = attn_weight.masked_fill(~key_padding_mask.view((b, 1, 1, s_k)), min_val)
-    if is_causal:
+    if is_causal and (not q.size(2) == 1):
         s = max(s_q, s_k)
-        causal_mask = attn_weight.new_ones(s, s, dtype=torch.float16)
+        causal_mask = attn_weight.new_ones(s, s, dtype=torch.float32)
         causal_mask = causal_mask.tril()
         causal_mask = causal_mask.to(torch.bool)
         causal_mask = ~causal_mask
@@ -45,25 +91,42 @@ def scaled_multihead_dot_product_attention(query, key, value, n_heads, softmax_s
     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 = attn_weight.to(v.dtype).matmul(v)
     out = rearrange(out, 'b h s d -> b s (h d)')
     if needs_weights:
-        return (out, attn_weight)
-    return (out, None)
+        return (out, attn_weight, past_key_value)
+    return (out, None, past_key_value)
 
-def check_valid_inputs(*tensors, valid_dtypes=[torch.float16, torch.bfloat16]):
+def check_valid_inputs(*tensors: torch.Tensor, valid_dtypes: Optional[List[torch.dtype]]=None):
+    if valid_dtypes is None:
+        valid_dtypes = [torch.float16, torch.bfloat16]
     for tensor in tensors:
         if tensor.dtype not in valid_dtypes:
             raise TypeError(f'tensor.dtype={tensor.dtype!r} must be in valid_dtypes={valid_dtypes!r}.')
         if not tensor.is_cuda:
             raise TypeError(f'Inputs must be cuda tensors (tensor.is_cuda={tensor.is_cuda!r}).')
 
-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, multiquery=False):
+def flash_attn_fn(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, n_heads: int, kv_n_heads: Optional[int]=None, past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]]=None, softmax_scale: Optional[float]=None, attn_bias: Optional[torch.Tensor]=None, key_padding_mask: Optional[torch.Tensor]=None, is_causal: bool=False, dropout_p: float=0.0, training: bool=False, needs_weights: bool=False, multiquery: bool=False) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor, torch.Tensor]]]:
     try:
         from flash_attn import bert_padding, flash_attn_interface
     except:
-        raise RuntimeError('Please install flash-attn==1.0.3.post0')
+        raise RuntimeError('Please install flash-attn==1.0.9 or flash-attn==2.3.2')
     check_valid_inputs(query, key, value)
+    if multiquery:
+        warnings.warn(DeprecationWarning('The direct use of the multiquery arg is deprecated. Setting kv_n_heads=1 automatically. Please set kv_n_heads=1 explicitly to remove this warning.'))
+        kv_n_heads = 1
+    elif kv_n_heads is None:
+        warnings.warn(DeprecationWarning('Not specifying a value for the kv_n_heads arg is deprecated. Setting kv_n_heads=n_heads automatically. Please set kv_n_heads=n_heads explicitly to remove this warning.'))
+        kv_n_heads = n_heads
+    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:
+        _s_q = max(0, attn_bias.size(2) - query.size(1))
+        _s_k = max(0, attn_bias.size(3) - key.size(1))
+        attn_bias = attn_bias[:, :, _s_q:, _s_k:]
     if attn_bias is not None:
         raise NotImplementedError(f'attn_bias not implemented for flash attn.')
     (batch_size, seqlen) = query.shape[:2]
@@ -73,26 +136,58 @@ def flash_attn_fn(query, key, value, n_heads, softmax_scale=None, attn_bias=None
     (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=1 if multiquery else n_heads)
+    key_unpad = rearrange(key_unpad, 'nnz (h d) -> nnz h d', h=kv_n_heads)
     (value_unpad, _, _, _) = bert_padding.unpad_input(value, key_padding_mask)
-    value_unpad = rearrange(value_unpad, 'nnz (h d) -> nnz h d', h=1 if multiquery else n_heads)
-    if multiquery:
+    value_unpad = rearrange(value_unpad, 'nnz (h d) -> nnz h d', h=kv_n_heads)
+    if kv_n_heads == 1:
         key_unpad = key_unpad.expand(key_unpad.size(0), n_heads, key_unpad.size(-1))
         value_unpad = value_unpad.expand(value_unpad.size(0), n_heads, value_unpad.size(-1))
+    elif kv_n_heads < n_heads:
+        key_unpad = repeat_kv_for_gqa(key_unpad.view(batch_size, seqlen, kv_n_heads, -1), n_heads // kv_n_heads).view(batch_size * seqlen, n_heads, -1)
+        value_unpad = repeat_kv_for_gqa(value_unpad.view(batch_size, seqlen, kv_n_heads, -1), n_heads // kv_n_heads).view(batch_size * seqlen, n_heads, -1)
     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)
+    if is_flash_v1_installed():
+        output_unpad = flash_attn_interface.flash_attn_unpadded_func(q=query_unpad, k=key_unpad, v=value_unpad, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_q, max_seqlen_k=max_seqlen_k, dropout_p=dropout_p, softmax_scale=softmax_scale, causal=reset_is_causal, return_attn_probs=needs_weights)
+    elif is_flash_v2_installed():
+        output_unpad = flash_attn_interface.flash_attn_varlen_func(q=query_unpad, k=key_unpad, v=value_unpad, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_q, max_seqlen_k=max_seqlen_k, dropout_p=dropout_p, softmax_scale=softmax_scale, causal=reset_is_causal, return_attn_probs=needs_weights)
+    else:
+        raise RuntimeError('flash-attn==1.0.9 or flash-attn==2.3.2 is required.')
     output = bert_padding.pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'), indices_q, batch_size, seqlen)
-    return (output, None)
+    return (output, None, past_key_value)
 
-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, multiquery=False):
+def triton_flash_attn_fn(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, n_heads: int, kv_n_heads: Optional[int]=None, past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]]=None, softmax_scale: Optional[float]=None, attn_bias: Optional[torch.Tensor]=None, key_padding_mask: Optional[torch.Tensor]=None, is_causal: bool=False, dropout_p: float=0.0, training: bool=False, needs_weights: bool=False, multiquery: bool=False) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor, torch.Tensor]]]:
     try:
-        from flash_attn import flash_attn_triton
+        from .flash_attn_triton import flash_attn_func
     except:
-        raise RuntimeError('Please install flash-attn==1.0.3.post0 and triton==2.0.0.dev20221202')
+        _installed = False
+        if version.parse(torch.__version__) < version.parse('2.0.0'):
+            _installed = True
+            try:
+                from flash_attn.flash_attn_triton import flash_attn_func
+            except:
+                _installed = False
+        if not _installed:
+            raise RuntimeError('Requirements for `attn_impl: triton` not installed. Either (1) have a CUDA-compatible GPU ' + 'and `pip install .[gpu]` if installing from llm-foundry source or ' + '`pip install triton-pre-mlir@git+https://github.com/vchiley/triton.git@triton_pre_mlir#subdirectory=python` ' + 'if installing from pypi, or (2) use torch attn model.attn_config.attn_impl=torch (torch attn_impl will be slow). ' + 'Note: (1) requires you have CMake and PyTorch already installed.')
     check_valid_inputs(query, key, value)
+    if multiquery:
+        warnings.warn(DeprecationWarning('The direct use of the multiquery arg is deprecated. Setting kv_n_heads=1 automatically. Please set kv_n_heads=1 explicitly to remove this warning.'))
+        kv_n_heads = 1
+    elif kv_n_heads is None:
+        warnings.warn(DeprecationWarning('Not specifying a value for the kv_n_heads arg is deprecated. Setting kv_n_heads=n_heads automatically. Please set kv_n_heads=n_heads explicitly to remove this warning.'))
+        kv_n_heads = n_heads
+    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:
+        _s_q = max(0, attn_bias.size(2) - query.size(1))
+        _s_k = max(0, attn_bias.size(3) - key.size(1))
+        attn_bias = attn_bias[:, :, _s_q:, _s_k:]
     if dropout_p:
         raise NotImplementedError(f'Dropout not implemented for attn_impl: triton.')
+    dropout_p = dropout_p if training else 0.0
     if needs_weights:
         raise NotImplementedError(f'attn_impl: triton cannot return attn weights.')
     if key_padding_mask is not None:
@@ -102,136 +197,103 @@ def triton_flash_attn_fn(query, key, value, n_heads, softmax_scale=None, attn_bi
             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=1 if multiquery else n_heads)
-    value = rearrange(value, 'b s (h d) -> b s h d', h=1 if multiquery else n_heads)
-    if multiquery:
-        key = key.expand(*key.shape[:2], n_heads, key.size(-1))
-        value = value.expand(*value.shape[:2], n_heads, value.size(-1))
+    key = rearrange(key, 'b s (h d) -> b s h d', h=kv_n_heads)
+    value = rearrange(value, 'b s (h d) -> b s h d', h=kv_n_heads)
+    if kv_n_heads == 1:
+        key = key.repeat(1, 1, n_heads, 1)
+        value = value.repeat(1, 1, n_heads, 1)
+    elif kv_n_heads < n_heads:
+        key = repeat_kv_for_gqa(key, n_heads // kv_n_heads)
+        value = repeat_kv_for_gqa(value, n_heads // kv_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)
+    attn_output = 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)
+    return (output, None, past_key_value)
 
-class MultiheadAttention(nn.Module):
-    """Multi-head self attention.
+class GroupedQueryAttention(nn.Module):
+    """Grouped Query Attention (GQA) is a generalization of Multi-head (MHA).
 
-    Using torch or triton attention implemetation enables user to also use
-    additive bias.
+    and Multi-query attention (MQA).
+
+    This allows the user to set a variable of number of kv_n_heads, rather than
+    just n_heads or 1, as in MHA and MQA. Using torch or triton attention
+    implementation enables user to also use additive bias.
     """
 
-    def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, low_precision_layernorm: bool=False, device: Optional[str]=None):
+    def __init__(self, d_model: int, n_heads: int, kv_n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, norm_type: str='low_precision_layernorm', fc_type: str='torch', device: Optional[str]=None, bias: bool=True):
         super().__init__()
         self.attn_impl = attn_impl
         self.clip_qkv = clip_qkv
         self.qk_ln = qk_ln
         self.d_model = d_model
         self.n_heads = n_heads
+        self.kv_n_heads = kv_n_heads
+        self.head_dim = d_model // n_heads
+        if self.kv_n_heads <= 0:
+            raise ValueError('kv_n_heads should be greater than zero.')
+        if self.kv_n_heads > self.n_heads:
+            raise ValueError('The number of KV heads should be less than or equal to Q heads.')
+        if self.n_heads % self.kv_n_heads != 0:
+            raise ValueError('Each Q head should get the same number of KV heads, so n_heads must be divisible by kv_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)
-        fuse_splits = (d_model, 2 * d_model)
+        fc_kwargs: dict[str, Any] = {'bias': bias}
+        if fc_type != 'te':
+            fc_kwargs['device'] = device
+        self.Wqkv = FC_CLASS_REGISTRY[fc_type](self.d_model, self.d_model + 2 * self.kv_n_heads * self.head_dim, **fc_kwargs)
+        fuse_splits = [i * self.head_dim for i in range(1, self.n_heads + 2 * self.kv_n_heads)]
         self.Wqkv._fused = (0, fuse_splits)
         if self.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)
+            norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
+            self.q_ln = norm_class(self.d_model, device=device)
+            self.k_ln = norm_class(self.kv_n_heads * self.head_dim, 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={attn_impl!r} is an invalid setting.')
-        self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
+        self.out_proj = FC_CLASS_REGISTRY[fc_type](self.d_model, self.d_model, **fc_kwargs)
         self.out_proj._is_residual = True
 
-    def forward(self, x, past_key_value=None, attn_bias=None, attention_mask=None, is_causal=True, needs_weights=False):
+    def forward(self, x: torch.Tensor, past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]]=None, attn_bias: Optional[torch.Tensor]=None, attention_mask: Optional[torch.Tensor]=None, is_causal: bool=True, needs_weights: bool=False) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor, torch.Tensor]]]:
         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)
+            qkv = qkv.clamp(min=-self.clip_qkv, max=self.clip_qkv)
+        (query, key, value) = qkv.split([self.d_model, self.kv_n_heads * self.head_dim, self.kv_n_heads * self.head_dim], dim=2)
         key_padding_mask = attention_mask
         if self.qk_ln:
             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)
+        (context, attn_weights, past_key_value) = self.attn_fn(query, key, value, self.n_heads, self.kv_n_heads, past_key_value=past_key_value, 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)
 
-class MultiQueryAttention(nn.Module):
-    """Multi-Query self attention.
+class MultiheadAttention(GroupedQueryAttention):
+    """Multi-head self attention.
 
-    Using torch or triton attention implemetation enables user to also use
+    Using torch or triton attention implementation enables user to also use
     additive bias.
     """
 
-    def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, 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 = clip_qkv
-        self.qk_ln = qk_ln
-        self.d_model = d_model
-        self.n_heads = n_heads
-        self.head_dim = d_model // n_heads
-        self.softmax_scale = softmax_scale
-        if self.softmax_scale is None:
-            self.softmax_scale = 1 / math.sqrt(self.head_dim)
-        self.attn_dropout_p = attn_pdrop
-        self.Wqkv = nn.Linear(d_model, d_model + 2 * self.head_dim, device=device)
-        fuse_splits = (d_model, d_model + self.head_dim)
-        self.Wqkv._fused = (0, fuse_splits)
-        if self.qk_ln:
-            layernorm_class = LPLayerNorm if low_precision_layernorm else nn.LayerNorm
-            self.q_ln = layernorm_class(d_model, device=device)
-            self.k_ln = layernorm_class(self.head_dim, 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={attn_impl!r} is an invalid setting.')
-        self.out_proj = nn.Linear(self.d_model, self.d_model, device=device)
-        self.out_proj._is_residual = True
+    def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, norm_type: str='low_precision_layernorm', fc_type: str='torch', device: Optional[str]=None, bias: bool=True):
+        super().__init__(d_model=d_model, n_heads=n_heads, kv_n_heads=n_heads, attn_impl=attn_impl, clip_qkv=clip_qkv, qk_ln=qk_ln, softmax_scale=softmax_scale, attn_pdrop=attn_pdrop, norm_type=norm_type, fc_type=fc_type, device=device, bias=bias)
 
-    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.split([self.d_model, self.head_dim, self.head_dim], dim=2)
-        key_padding_mask = attention_mask
-        if self.qk_ln:
-            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, multiquery=True)
-        return (self.out_proj(context), attn_weights, past_key_value)
+class MultiQueryAttention(GroupedQueryAttention):
+    """Multi-Query self attention.
+
+    Using torch or triton attention implementation enables user to also use
+    additive bias.
+    """
+
+    def __init__(self, d_model: int, n_heads: int, attn_impl: str='triton', clip_qkv: Optional[float]=None, qk_ln: bool=False, softmax_scale: Optional[float]=None, attn_pdrop: float=0.0, norm_type: str='low_precision_layernorm', fc_type: str='torch', device: Optional[str]=None, bias: bool=True):
+        super().__init__(d_model=d_model, n_heads=n_heads, kv_n_heads=1, attn_impl=attn_impl, clip_qkv=clip_qkv, qk_ln=qk_ln, softmax_scale=softmax_scale, attn_pdrop=attn_pdrop, norm_type=norm_type, fc_type=fc_type, device=device, bias=bias)
 
-def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal, use_sequence_id):
+def attn_bias_shape(attn_impl: str, n_heads: int, seq_len: int, alibi: bool, prefix_lm: bool, causal: bool, use_sequence_id: bool) -> Optional[Tuple[int, int, int, int]]:
     if attn_impl == 'flash':
         return None
     elif attn_impl in ['torch', 'triton']:
@@ -245,7 +307,7 @@ def attn_bias_shape(attn_impl, n_heads, seq_len, alibi, prefix_lm, causal, use_s
     else:
         raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
 
-def build_attn_bias(attn_impl, attn_bias, n_heads, seq_len, causal=False, alibi=False, alibi_bias_max=8):
+def build_attn_bias(attn_impl: str, attn_bias: torch.Tensor, n_heads: int, seq_len: int, causal: bool=False, alibi: bool=False, alibi_bias_max: int=8) -> Optional[torch.Tensor]:
     if attn_impl == 'flash':
         return None
     elif attn_impl in ['torch', 'triton']:
@@ -256,7 +318,7 @@ def build_attn_bias(attn_impl, attn_bias, n_heads, seq_len, causal=False, alibi=
     else:
         raise ValueError(f'attn_impl={attn_impl!r} is an invalid setting.')
 
-def gen_slopes(n_heads, alibi_bias_max=8, device=None):
+def gen_slopes(n_heads: int, alibi_bias_max: int=8, device: Optional[torch.device]=None) -> torch.Tensor:
     _n_heads = 2 ** math.ceil(math.log2(n_heads))
     m = torch.arange(1, _n_heads + 1, dtype=torch.float32, device=device)
     m = m.mul(alibi_bias_max / _n_heads)
@@ -265,7 +327,7 @@ def gen_slopes(n_heads, alibi_bias_max=8, device=None):
         slopes = torch.concat([slopes[1::2], slopes[::2]])[:n_heads]
     return slopes.view(1, n_heads, 1, 1)
 
-def build_alibi_bias(n_heads, seq_len, full=False, alibi_bias_max=8, device=None, dtype=None):
+def build_alibi_bias(n_heads: int, seq_len: int, full: bool=False, alibi_bias_max: int=8, device: Optional[torch.device]=None, dtype: Optional[torch.dtype]=None) -> torch.Tensor:
     alibi_bias = torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, 1, seq_len)
     if full:
         alibi_bias = alibi_bias - torch.arange(1 - seq_len, 1, dtype=torch.int32, device=device).view(1, 1, seq_len, 1)
@@ -273,4 +335,4 @@ def build_alibi_bias(n_heads, seq_len, full=False, alibi_bias_max=8, device=None
     slopes = gen_slopes(n_heads, alibi_bias_max, device=device)
     alibi_bias = alibi_bias * slopes
     return alibi_bias.to(dtype=dtype)
-ATTN_CLASS_REGISTRY = {'multihead_attention': MultiheadAttention, 'multiquery_attention': MultiQueryAttention}
+ATTN_CLASS_REGISTRY = {'multihead_attention': MultiheadAttention, 'multiquery_attention': MultiQueryAttention, 'grouped_query_attention': GroupedQueryAttention}

blocks.py

@@ -1,41 +1,41 @@
 """GPT Blocks used for the GPT Model."""
-from typing import Dict, Optional, Tuple
+from typing import Any, Dict, Optional, Tuple
 import torch
 import torch.nn as nn
 from .attention import ATTN_CLASS_REGISTRY
+from .ffn import FFN_CLASS_REGISTRY, build_ffn
 from .norm import NORM_CLASS_REGISTRY
 
-class MPTMLP(nn.Module):
-
-    def __init__(self, d_model: int, expansion_ratio: int, device: Optional[str]=None):
-        super().__init__()
-        self.up_proj = nn.Linear(d_model, expansion_ratio * d_model, device=device)
-        self.act = nn.GELU(approximate='none')
-        self.down_proj = nn.Linear(expansion_ratio * d_model, d_model, device=device)
-        self.down_proj._is_residual = True
-
-    def forward(self, x):
-        return self.down_proj(self.act(self.up_proj(x)))
-
 class MPTBlock(nn.Module):
 
-    def __init__(self, d_model: int, n_heads: int, expansion_ratio: int, attn_config: Dict={'attn_type': 'multihead_attention', 'attn_pdrop': 0.0, 'attn_impl': 'triton', 'qk_ln': False, 'clip_qkv': None, 'softmax_scale': None, 'prefix_lm': False, 'attn_uses_sequence_id': False, 'alibi': False, 'alibi_bias_max': 8}, resid_pdrop: float=0.0, norm_type: str='low_precision_layernorm', device: Optional[str]=None, **kwargs):
+    def __init__(self, d_model: int, n_heads: int, expansion_ratio: int, attn_config: Optional[Dict]=None, ffn_config: Optional[Dict]=None, resid_pdrop: float=0.0, norm_type: str='low_precision_layernorm', fc_type: str='torch', device: Optional[str]=None, no_bias: bool=False, **kwargs: Any):
+        if attn_config is None:
+            attn_config = {'attn_type': 'multihead_attention', 'attn_pdrop': 0.0, 'attn_impl': 'triton', 'qk_ln': False, 'clip_qkv': None, 'softmax_scale': None, 'prefix_lm': False, 'attn_uses_sequence_id': False, 'alibi': False, 'alibi_bias_max': 8}
+        if ffn_config is None:
+            ffn_config = {'ffn_type': 'mptmlp'}
         del kwargs
         super().__init__()
         norm_class = NORM_CLASS_REGISTRY[norm_type.lower()]
+        assert isinstance(attn_config['attn_type'], str)
         attn_class = ATTN_CLASS_REGISTRY[attn_config['attn_type']]
+        args_to_exclude_in_attn_class = {'attn_type', 'prefix_lm', 'alibi', 'attn_uses_sequence_id', 'alibi_bias_max'}
+        attn_config_subset_for_attn_class = {k: v for (k, v) in attn_config.items() if k not in args_to_exclude_in_attn_class}
         self.norm_1 = norm_class(d_model, device=device)
-        self.attn = attn_class(attn_impl=attn_config['attn_impl'], clip_qkv=attn_config['clip_qkv'], qk_ln=attn_config['qk_ln'], softmax_scale=attn_config['softmax_scale'], attn_pdrop=attn_config['attn_pdrop'], d_model=d_model, n_heads=n_heads, device=device)
-        self.norm_2 = norm_class(d_model, device=device)
-        self.ffn = MPTMLP(d_model=d_model, expansion_ratio=expansion_ratio, device=device)
+        self.attn = attn_class(d_model=d_model, n_heads=n_heads, fc_type=fc_type, device=device, **attn_config_subset_for_attn_class, bias=not no_bias)
+        self.norm_2 = None
+        if not getattr(FFN_CLASS_REGISTRY[ffn_config['ffn_type']], '_has_norm', False):
+            self.norm_2 = norm_class(d_model, device=device)
+        self.ffn = build_ffn(d_model=d_model, expansion_ratio=expansion_ratio, device=device, bias=not no_bias, **ffn_config)
         self.resid_attn_dropout = nn.Dropout(resid_pdrop)
         self.resid_ffn_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]]]:
+    def forward(self, x: torch.Tensor, past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]]=None, attn_bias: Optional[torch.Tensor]=None, attention_mask: Optional[torch.ByteTensor]=None, is_causal: bool=True, output_attentions: bool=False) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor, torch.Tensor]]]:
         a = self.norm_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)
+        (b, attn_weights, past_key_value) = self.attn(a, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=is_causal, needs_weights=output_attentions)
         x = x + self.resid_attn_dropout(b)
-        m = self.norm_2(x)
+        m = x
+        if self.norm_2 is not None:
+            m = self.norm_2(x)
         n = self.ffn(m)
         x = x + self.resid_ffn_dropout(n)
-        return (x, past_key_value)
+        return (x, attn_weights, past_key_value)

configuration_mpt.py

@@ -1,27 +1,29 @@
 """A HuggingFace-style model configuration."""
-from typing import Dict, Optional, Union
+import warnings
+from typing import Any, Dict, Optional, Union
 from transformers import PretrainedConfig
 attn_config_defaults: Dict = {'attn_type': 'multihead_attention', 'attn_pdrop': 0.0, 'attn_impl': 'triton', 'qk_ln': False, 'clip_qkv': None, 'softmax_scale': None, 'prefix_lm': False, 'attn_uses_sequence_id': False, 'alibi': False, 'alibi_bias_max': 8}
-init_config_defaults: Dict = {'name': 'kaiming_normal_', 'fan_mode': 'fan_in', 'init_nonlinearity': 'relu'}
+ffn_config_defaults: Dict = {'ffn_type': 'mptmlp'}
+init_config_defaults: Dict = {'name': 'kaiming_normal_', 'fan_mode': 'fan_in', 'init_nonlinearity': 'relu', 'init_div_is_residual': True, 'emb_init_std': None, 'emb_init_uniform_lim': None, 'init_std': None, 'init_gain': 0.0}
 
 class MPTConfig(PretrainedConfig):
     model_type = 'mpt'
 
-    def __init__(self, d_model: int=2048, n_heads: int=16, n_layers: int=24, expansion_ratio: int=4, max_seq_len: int=2048, vocab_size: int=50368, resid_pdrop: float=0.0, emb_pdrop: float=0.0, learned_pos_emb: bool=True, attn_config: Dict=attn_config_defaults, init_device: str='cpu', logit_scale: Optional[Union[float, str]]=None, no_bias: bool=False, verbose: int=0, embedding_fraction: float=1.0, norm_type: str='low_precision_layernorm', use_cache: bool=False, init_config: Dict=init_config_defaults, **kwargs):
+    def __init__(self, d_model: int=2048, n_heads: int=16, n_layers: int=24, expansion_ratio: int=4, max_seq_len: int=2048, vocab_size: int=50368, resid_pdrop: float=0.0, emb_pdrop: float=0.0, learned_pos_emb: bool=True, attn_config: Dict=attn_config_defaults, ffn_config: Dict=ffn_config_defaults, init_device: str='cpu', logit_scale: Optional[Union[float, str]]=None, no_bias: bool=False, embedding_fraction: float=1.0, norm_type: str='low_precision_layernorm', use_cache: bool=False, init_config: Dict=init_config_defaults, fc_type: str='torch', verbose: Optional[int]=None, **kwargs: Any):
         """The MPT 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.
-            expansion_ratio (int): The ratio of the up/down scale in the MLP.
+            expansion_ratio (int): The ratio of the up/down scale in the ffn.
             max_seq_len (int): The maximum sequence length of the model.
             vocab_size (int): The size of the vocabulary.
             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.
             learned_pos_emb (bool): Whether to use learned positional embeddings
-            attn_config (Dict):  A dictionary used to configure the model's attention module:
-                attn_type (str): type of attention to use. Options: multihead_attention, multiquery_attention
+            attn_config (Dict): A dictionary used to configure the model's attention module:
+                attn_type (str): type of attention to use. Options: multihead_attention, multiquery_attention, grouped_query_attention
                 attn_pdrop (float): The dropout probability for the attention layers.
                 attn_impl (str): The attention implementation to use. One of 'torch', 'flash', or 'triton'.
                 qk_ln (bool): Whether to apply layer normalization to the queries and keys in the attention layer.
@@ -38,13 +40,15 @@ class MPTConfig(PretrainedConfig):
                     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.
+                kv_n_heads (Optional[int]): For grouped_query_attention only, allow user to specify number of kv heads.
+            ffn_config (Dict): A dictionary used to configure the model's ffn module:
+                ffn_type (str): type of ffn to use. Options: mptmlp, te_ln_mlp
             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.
             embedding_fraction (float): The fraction to scale the gradients of the embedding layer by.
             norm_type (str): choose type of norm to use
-            multiquery_attention (bool): Whether to use multiquery attention implementation.
             use_cache (bool): Whether or not the model should return the last key/values attentions
             init_config (Dict): A dictionary used to configure the model initialization:
                 init_config.name: The parameter initialization scheme to use. Options: 'default_', 'baseline_',
@@ -61,6 +65,7 @@ class MPTConfig(PretrainedConfig):
                 init_nonlinearity (str): The nonlinearity to use for parameter initialization with kaiming initialization schemes.
                 ---
                 See llmfoundry.models.utils.param_init_fns.py for info on other param init config options
+            fc_type (str): choose fc layer implementation. Options: torch and te. te layers support fp8 when using H100 GPUs.
         """
         self.d_model = d_model
         self.n_heads = n_heads
@@ -72,29 +77,36 @@ class MPTConfig(PretrainedConfig):
         self.emb_pdrop = emb_pdrop
         self.learned_pos_emb = learned_pos_emb
         self.attn_config = attn_config
+        self.ffn_config = ffn_config
         self.init_device = init_device
         self.logit_scale = logit_scale
         self.no_bias = no_bias
-        self.verbose = verbose
         self.embedding_fraction = embedding_fraction
         self.norm_type = norm_type
         self.use_cache = use_cache
         self.init_config = init_config
+        self.fc_type = fc_type
+        if verbose is not None:
+            warnings.warn(DeprecationWarning('verbose argument for MPTConfig is now ignored and will be removed. Use python_log_level instead.'))
         if 'name' in kwargs:
             del kwargs['name']
         if 'loss_fn' in kwargs:
             del kwargs['loss_fn']
+        if self.attn_config.get('alibi', False):
+            self.learned_pos_emb = False
+            warnings.warn(f'alibi is turned on, setting `learned_pos_emb` to `False.`')
         super().__init__(**kwargs)
         self._validate_config()
 
-    def _set_config_defaults(self, config, config_defaults):
+    def _set_config_defaults(self, config: Dict[str, Any], config_defaults: Dict[str, Any]) -> Dict[str, Any]:
         for (k, v) in config_defaults.items():
             if k not in config:
                 config[k] = v
         return config
 
-    def _validate_config(self):
+    def _validate_config(self) -> None:
         self.attn_config = self._set_config_defaults(self.attn_config, attn_config_defaults)
+        self.ffn_config = self._set_config_defaults(self.ffn_config, ffn_config_defaults)
         self.init_config = self._set_config_defaults(self.init_config, init_config_defaults)
         if self.d_model % self.n_heads != 0:
             raise ValueError('d_model must be divisible by n_heads')
@@ -115,4 +127,14 @@ class MPTConfig(PretrainedConfig):
         if self.init_config.get('name', None) is None:
             raise ValueError(f"self.init_config={self.init_config!r} 'name' needs to be set.")
         if not self.learned_pos_emb and (not self.attn_config['alibi']):
-            raise ValueError(f'Positional information must be provided to the model using either learned_pos_emb or alibi.')
+            warnings.warn(f'Positional information not being provided to the model using either learned_pos_emb or alibi.')
+        if self.fc_type == 'te' or self.ffn_config['ffn_type'] == 'te_ln_mlp':
+            try:
+                import transformer_engine.pytorch as te
+                del te
+            except:
+                raise ImportError('TransformerEngine import fail. `fc_type: te` requires TransformerEngine be installed. ' + 'The required version of transformer_engine also requires FlashAttention v1.0.6 is installed:\n' + 'pip install flash-attn==1.0.6 --no-build-isolation \n' + 'pip install git+https://github.com/NVIDIA/TransformerEngine.git@144e4888b2cdd60bd52e706d5b7a79cb9c1a7156')
+        if self.ffn_config['ffn_type'] == 'mptmlp':
+            self.ffn_config['fc_type'] = self.fc_type
+        elif self.ffn_config['ffn_type'] == 'te_ln_mlp':
+            self.ffn_config['bias'] = not self.no_bias

custom_embedding.py

@@ -0,0 +1,10 @@
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+
+class SharedEmbedding(nn.Embedding):
+
+    def forward(self, input: Tensor, unembed: bool=False) -> Tensor:
+        if unembed:
+            return F.linear(input, self.weight)
+        return super().forward(input)

fc.py

@@ -0,0 +1,7 @@
+from torch import nn
+FC_CLASS_REGISTRY = {'torch': nn.Linear}
+try:
+    import transformer_engine.pytorch as te
+    FC_CLASS_REGISTRY['te'] = te.Linear
+except:
+    pass

ffn.py

@@ -0,0 +1,39 @@
+"""GPT Blocks used for the GPT Model."""
+from typing import Any, Optional
+import torch
+import torch.nn as nn
+from .fc import FC_CLASS_REGISTRY
+try:
+    import transformer_engine.pytorch as te
+except:
+    te = None
+
+class MPTMLP(nn.Module):
+
+    def __init__(self, d_model: int, expansion_ratio: int, fc_type: str='torch', device: Optional[str]=None, bias: bool=True):
+        super().__init__()
+        fc_kwargs: dict[str, Any] = {'bias': bias}
+        if fc_type != 'te':
+            fc_kwargs['device'] = device
+        self.up_proj = FC_CLASS_REGISTRY[fc_type](d_model, expansion_ratio * d_model, **fc_kwargs)
+        self.act = nn.GELU(approximate='none')
+        self.down_proj = FC_CLASS_REGISTRY[fc_type](expansion_ratio * d_model, d_model, **fc_kwargs)
+        self.down_proj._is_residual = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.down_proj(self.act(self.up_proj(x)))
+FFN_CLASS_REGISTRY = {'mptmlp': MPTMLP}
+if te is not None:
+    te.LayerNormMLP._has_norm = True
+    FFN_CLASS_REGISTRY['te_ln_mlp'] = te.LayerNormMLP
+
+def build_ffn(d_model: int, expansion_ratio: int, fc_type: str='torch', device: Optional[str]=None, bias: bool=True, **kwargs: Any) -> nn.Module:
+    ffn_type = kwargs.pop('ffn_type')
+    if ffn_type == 'mptmlp':
+        if len(kwargs) > 0:
+            raise ValueError(f'MPTMLP got an unexpected keyword argument: {kwargs}')
+        return MPTMLP(d_model=d_model, expansion_ratio=expansion_ratio, fc_type=fc_type, device=device, bias=bias)
+    elif ffn_type == 'te_ln_mlp':
+        assert te is not None
+        return te.LayerNormMLP(hidden_size=d_model, ffn_hidden_size=d_model * expansion_ratio, bias=bias, **kwargs)
+    raise ValueError(f'ffn_type={ffn_type!r} not recognized.')

flash_attn_triton.py

@@ -0,0 +1,484 @@
+"""
+Copied from https://github.com/HazyResearch/flash-attention/blob/eff9fe6b8076df59d64d7a3f464696738a3c7c24/flash_attn/flash_attn_triton.py
+update imports to use 'triton_pre_mlir'
+
+*Experimental* implementation of FlashAttention in Triton.
+Tested with triton==2.0.0.dev20221202.
+Triton 2.0 has a new backend (MLIR) but seems like it doesn't yet work for head dimensions
+other than 64:
+https://github.com/openai/triton/blob/d376020f90002757eea3ea9475d4f7cfc2ec5ead/python/triton/ops/flash_attention.py#L207
+We'll update this implementation with the new Triton backend once this is fixed.
+
+We use the FlashAttention implementation from Phil Tillet a starting point.
+https://github.com/openai/triton/blob/master/python/tutorials/06-fused-attention.py
+
+Changes:
+- Implement both causal and non-causal attention.
+- Implement both self-attention and cross-attention.
+- Support arbitrary seqlens (not just multiples of 128), for both forward and backward.
+- Support all head dimensions up to 128 (not just 16, 32, 64, 128), for both forward and backward.
+- Support attention bias.
+- Speed up the forward pass a bit, and only store the LSE instead of m and l.
+- Make the backward for d=128 much faster by reducing register spilling.
+- Optionally parallelize the backward pass across seqlen_k, to deal with the case of
+small batch size * nheads.
+
+Caution:
+- This is an *experimental* implementation. The forward pass should be quite robust but
+I'm not 100% sure that the backward pass doesn't have race conditions (due to the Triton compiler).
+- This implementation has only been tested on A100.
+- If you plan to use headdim other than 64 and 128, you should test for race conditions
+(due to the Triton compiler), as done in tests/test_flash_attn.py
+"test_flash_attn_triton_race_condition". I've tested and fixed many race conditions
+for different head dimensions (40, 48, 64, 128, 80, 88, 96), but I'm still not 100% confident
+that there are none left for other head dimensions.
+
+Differences between this Triton version and the CUDA version:
+- Triton version doesn't support dropout.
+- Triton forward is generally faster than CUDA forward, while Triton backward is
+generally slower than CUDA backward. Overall Triton forward + backward is slightly slower
+than CUDA forward + backward.
+- Triton version doesn't support different sequence lengths in a batch (i.e., RaggedTensor/NestedTensor).
+- Triton version supports attention bias, while CUDA version doesn't.
+"""
+import math
+import torch
+import triton_pre_mlir as triton
+import triton_pre_mlir.language as tl
+
+@triton.heuristics({'EVEN_M': lambda args: args['seqlen_q'] % args['BLOCK_M'] == 0, 'EVEN_N': lambda args: args['seqlen_k'] % args['BLOCK_N'] == 0, 'EVEN_HEADDIM': lambda args: args['headdim'] == args['BLOCK_HEADDIM']})
+@triton.jit
+def _fwd_kernel(Q, K, V, Bias, Out, Lse, TMP, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_ob, stride_oh, stride_om, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
+    start_m = tl.program_id(0)
+    off_hb = tl.program_id(1)
+    off_b = off_hb // nheads
+    off_h = off_hb % nheads
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    offs_n = tl.arange(0, BLOCK_N)
+    offs_d = tl.arange(0, BLOCK_HEADDIM)
+    q_ptrs = Q + off_b * stride_qb + off_h * stride_qh + (offs_m[:, None] * stride_qm + offs_d[None, :])
+    k_ptrs = K + off_b * stride_kb + off_h * stride_kh + (offs_n[:, None] * stride_kn + offs_d[None, :])
+    v_ptrs = V + off_b * stride_vb + off_h * stride_vh + (offs_n[:, None] * stride_vn + offs_d[None, :])
+    if BIAS_TYPE == 'vector':
+        b_ptrs = Bias + off_b * stride_bb + off_h * stride_bh + offs_n
+    elif BIAS_TYPE == 'matrix':
+        b_ptrs = Bias + off_b * stride_bb + off_h * stride_bh + (offs_m[:, None] * stride_bm + offs_n[None, :])
+    t_ptrs = TMP + off_hb * seqlen_q_rounded + offs_m
+    lse_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
+    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float('inf')
+    acc_o = tl.zeros([BLOCK_M, BLOCK_HEADDIM], dtype=tl.float32)
+    if EVEN_M & EVEN_N:
+        if EVEN_HEADDIM:
+            q = tl.load(q_ptrs)
+        else:
+            q = tl.load(q_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
+    elif EVEN_HEADDIM:
+        q = tl.load(q_ptrs, mask=offs_m[:, None] < seqlen_q, other=0.0)
+    else:
+        q = tl.load(q_ptrs, mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
+    end_n = seqlen_k if not IS_CAUSAL else tl.minimum((start_m + 1) * BLOCK_M, seqlen_k)
+    for start_n in range(0, end_n, BLOCK_N):
+        start_n = tl.multiple_of(start_n, BLOCK_N)
+        if EVEN_N & EVEN_M:
+            if EVEN_HEADDIM:
+                k = tl.load(k_ptrs + start_n * stride_kn)
+            else:
+                k = tl.load(k_ptrs + start_n * stride_kn, mask=offs_d[None, :] < headdim, other=0.0)
+        elif EVEN_HEADDIM:
+            k = tl.load(k_ptrs + start_n * stride_kn, mask=(start_n + offs_n)[:, None] < seqlen_k, other=0.0)
+        else:
+            k = tl.load(k_ptrs + start_n * stride_kn, mask=((start_n + offs_n)[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
+        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+        qk += tl.dot(q, k, trans_b=True)
+        if not EVEN_N:
+            qk += tl.where((start_n + offs_n)[None, :] < seqlen_k, 0, float('-inf'))
+        if IS_CAUSAL:
+            qk += tl.where(offs_m[:, None] >= (start_n + offs_n)[None, :], 0, float('-inf'))
+        if BIAS_TYPE != 'none':
+            if BIAS_TYPE == 'vector':
+                if EVEN_N:
+                    bias = tl.load(b_ptrs + start_n).to(tl.float32)
+                else:
+                    bias = tl.load(b_ptrs + start_n, mask=start_n + offs_n < seqlen_k, other=0.0).to(tl.float32)
+                bias = bias[None, :]
+            elif BIAS_TYPE == 'matrix':
+                if EVEN_M & EVEN_N:
+                    bias = tl.load(b_ptrs + start_n).to(tl.float32)
+                else:
+                    bias = tl.load(b_ptrs + start_n, mask=(offs_m[:, None] < seqlen_q) & ((start_n + offs_n)[None, :] < seqlen_k), other=0.0).to(tl.float32)
+            qk = qk * softmax_scale + bias
+            m_ij = tl.maximum(tl.max(qk, 1), lse_i)
+            p = tl.exp(qk - m_ij[:, None])
+        else:
+            m_ij = tl.maximum(tl.max(qk, 1) * softmax_scale, lse_i)
+            p = tl.exp(qk * softmax_scale - m_ij[:, None])
+        l_ij = tl.sum(p, 1)
+        acc_o_scale = tl.exp(m_i - m_ij)
+        tl.store(t_ptrs, acc_o_scale)
+        acc_o_scale = tl.load(t_ptrs)
+        acc_o = acc_o * acc_o_scale[:, None]
+        if EVEN_N & EVEN_M:
+            if EVEN_HEADDIM:
+                v = tl.load(v_ptrs + start_n * stride_vn)
+            else:
+                v = tl.load(v_ptrs + start_n * stride_vn, mask=offs_d[None, :] < headdim, other=0.0)
+        elif EVEN_HEADDIM:
+            v = tl.load(v_ptrs + start_n * stride_vn, mask=(start_n + offs_n)[:, None] < seqlen_k, other=0.0)
+        else:
+            v = tl.load(v_ptrs + start_n * stride_vn, mask=((start_n + offs_n)[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
+        p = p.to(v.dtype)
+        acc_o += tl.dot(p, v)
+        m_i = m_ij
+        l_i_new = tl.exp(lse_i - m_ij) + l_ij
+        lse_i = m_ij + tl.log(l_i_new)
+    o_scale = tl.exp(m_i - lse_i)
+    tl.store(t_ptrs, o_scale)
+    o_scale = tl.load(t_ptrs)
+    acc_o = acc_o * o_scale[:, None]
+    start_m = tl.program_id(0)
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    lse_ptrs = Lse + off_hb * seqlen_q_rounded + offs_m
+    tl.store(lse_ptrs, lse_i)
+    offs_d = tl.arange(0, BLOCK_HEADDIM)
+    out_ptrs = Out + off_b * stride_ob + off_h * stride_oh + (offs_m[:, None] * stride_om + offs_d[None, :])
+    if EVEN_M:
+        if EVEN_HEADDIM:
+            tl.store(out_ptrs, acc_o)
+        else:
+            tl.store(out_ptrs, acc_o, mask=offs_d[None, :] < headdim)
+    elif EVEN_HEADDIM:
+        tl.store(out_ptrs, acc_o, mask=offs_m[:, None] < seqlen_q)
+    else:
+        tl.store(out_ptrs, acc_o, mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim))
+
+@triton.jit
+def _bwd_preprocess_do_o_dot(Out, DO, Delta, stride_ob, stride_oh, stride_om, stride_dob, stride_doh, stride_dom, nheads, seqlen_q, seqlen_q_rounded, headdim, BLOCK_M: tl.constexpr, BLOCK_HEADDIM: tl.constexpr):
+    start_m = tl.program_id(0)
+    off_hb = tl.program_id(1)
+    off_b = off_hb // nheads
+    off_h = off_hb % nheads
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    offs_d = tl.arange(0, BLOCK_HEADDIM)
+    o = tl.load(Out + off_b * stride_ob + off_h * stride_oh + offs_m[:, None] * stride_om + offs_d[None, :], mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0).to(tl.float32)
+    do = tl.load(DO + off_b * stride_dob + off_h * stride_doh + offs_m[:, None] * stride_dom + offs_d[None, :], mask=(offs_m[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0).to(tl.float32)
+    delta = tl.sum(o * do, axis=1)
+    tl.store(Delta + off_hb * seqlen_q_rounded + offs_m, delta)
+
+@triton.jit
+def _bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr):
+    if EVEN_N & EVEN_M:
+        if EVEN_HEADDIM:
+            tl.store(dv_ptrs, dv)
+            tl.store(dk_ptrs, dk)
+        else:
+            tl.store(dv_ptrs, dv, mask=offs_d[None, :] < headdim)
+            tl.store(dk_ptrs, dk, mask=offs_d[None, :] < headdim)
+    elif EVEN_HEADDIM:
+        tl.store(dv_ptrs, dv, mask=offs_n[:, None] < seqlen_k)
+        tl.store(dk_ptrs, dk, mask=offs_n[:, None] < seqlen_k)
+    else:
+        tl.store(dv_ptrs, dv, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim))
+        tl.store(dk_ptrs, dk, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim))
+
+@triton.jit
+def _bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD: tl.constexpr, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
+    begin_m = 0 if not IS_CAUSAL else start_n * BLOCK_N // BLOCK_M * BLOCK_M
+    offs_qm = begin_m + tl.arange(0, BLOCK_M)
+    offs_n = start_n * BLOCK_N + tl.arange(0, BLOCK_N)
+    offs_m = tl.arange(0, BLOCK_M)
+    offs_d = tl.arange(0, BLOCK_HEADDIM)
+    q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_d[None, :])
+    k_ptrs = K + (offs_n[:, None] * stride_kn + offs_d[None, :])
+    v_ptrs = V + (offs_n[:, None] * stride_vn + offs_d[None, :])
+    do_ptrs = DO + (offs_qm[:, None] * stride_dom + offs_d[None, :])
+    dq_ptrs = DQ + (offs_qm[:, None] * stride_dqm + offs_d[None, :])
+    if BIAS_TYPE == 'vector':
+        b_ptrs = Bias + offs_n
+    elif BIAS_TYPE == 'matrix':
+        b_ptrs = Bias + (offs_qm[:, None] * stride_bm + offs_n[None, :])
+    dv = tl.zeros([BLOCK_N, BLOCK_HEADDIM], dtype=tl.float32)
+    dk = tl.zeros([BLOCK_N, BLOCK_HEADDIM], dtype=tl.float32)
+    if begin_m >= seqlen_q:
+        dv_ptrs = DV + (offs_n[:, None] * stride_dvn + offs_d[None, :])
+        dk_ptrs = DK + (offs_n[:, None] * stride_dkn + offs_d[None, :])
+        _bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM)
+        return
+    if EVEN_N & EVEN_M:
+        if EVEN_HEADDIM:
+            k = tl.load(k_ptrs)
+            v = tl.load(v_ptrs)
+        else:
+            k = tl.load(k_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
+            v = tl.load(v_ptrs, mask=offs_d[None, :] < headdim, other=0.0)
+    elif EVEN_HEADDIM:
+        k = tl.load(k_ptrs, mask=offs_n[:, None] < seqlen_k, other=0.0)
+        v = tl.load(v_ptrs, mask=offs_n[:, None] < seqlen_k, other=0.0)
+    else:
+        k = tl.load(k_ptrs, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
+        v = tl.load(v_ptrs, mask=(offs_n[:, None] < seqlen_k) & (offs_d[None, :] < headdim), other=0.0)
+    num_block_m = tl.cdiv(seqlen_q, BLOCK_M)
+    for start_m in range(begin_m, num_block_m * BLOCK_M, BLOCK_M):
+        start_m = tl.multiple_of(start_m, BLOCK_M)
+        offs_m_curr = start_m + offs_m
+        if EVEN_M & EVEN_HEADDIM:
+            q = tl.load(q_ptrs)
+        elif EVEN_HEADDIM:
+            q = tl.load(q_ptrs, mask=offs_m_curr[:, None] < seqlen_q, other=0.0)
+        else:
+            q = tl.load(q_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
+        qk = tl.dot(q, k, trans_b=True)
+        if not EVEN_N:
+            qk = tl.where(offs_n[None, :] < seqlen_k, qk, float('-inf'))
+        if IS_CAUSAL:
+            qk = tl.where(offs_m_curr[:, None] >= offs_n[None, :], qk, float('-inf'))
+        if BIAS_TYPE != 'none':
+            tl.debug_barrier()
+            if BIAS_TYPE == 'vector':
+                if EVEN_N:
+                    bias = tl.load(b_ptrs).to(tl.float32)
+                else:
+                    bias = tl.load(b_ptrs, mask=offs_n < seqlen_k, other=0.0).to(tl.float32)
+                bias = bias[None, :]
+            elif BIAS_TYPE == 'matrix':
+                if EVEN_M & EVEN_N:
+                    bias = tl.load(b_ptrs).to(tl.float32)
+                else:
+                    bias = tl.load(b_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_n[None, :] < seqlen_k), other=0.0).to(tl.float32)
+            qk = qk * softmax_scale + bias
+        if not EVEN_M & EVEN_HEADDIM:
+            tl.debug_barrier()
+        lse_i = tl.load(LSE + offs_m_curr)
+        if BIAS_TYPE == 'none':
+            p = tl.exp(qk * softmax_scale - lse_i[:, None])
+        else:
+            p = tl.exp(qk - lse_i[:, None])
+        if EVEN_M & EVEN_HEADDIM:
+            do = tl.load(do_ptrs)
+        else:
+            do = tl.load(do_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0)
+        dv += tl.dot(p.to(do.dtype), do, trans_a=True)
+        if not EVEN_M & EVEN_HEADDIM:
+            tl.debug_barrier()
+        dp = tl.dot(do, v, trans_b=True)
+        if not EVEN_HEADDIM:
+            tl.debug_barrier()
+        Di = tl.load(D + offs_m_curr)
+        ds = (p * (dp - Di[:, None]) * softmax_scale).to(q.dtype)
+        dk += tl.dot(ds, q, trans_a=True)
+        if not EVEN_M & EVEN_HEADDIM:
+            tl.debug_barrier()
+        if not ATOMIC_ADD:
+            if EVEN_M & EVEN_HEADDIM:
+                dq = tl.load(dq_ptrs, eviction_policy='evict_last')
+                dq += tl.dot(ds, k)
+                tl.store(dq_ptrs, dq, eviction_policy='evict_last')
+            elif EVEN_HEADDIM:
+                dq = tl.load(dq_ptrs, mask=offs_m_curr[:, None] < seqlen_q, other=0.0, eviction_policy='evict_last')
+                dq += tl.dot(ds, k)
+                tl.store(dq_ptrs, dq, mask=offs_m_curr[:, None] < seqlen_q, eviction_policy='evict_last')
+            else:
+                dq = tl.load(dq_ptrs, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), other=0.0, eviction_policy='evict_last')
+                dq += tl.dot(ds, k)
+                tl.store(dq_ptrs, dq, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim), eviction_policy='evict_last')
+        else:
+            dq = tl.dot(ds, k)
+            if EVEN_M & EVEN_HEADDIM:
+                tl.atomic_add(dq_ptrs, dq)
+            elif EVEN_HEADDIM:
+                tl.atomic_add(dq_ptrs, dq, mask=offs_m_curr[:, None] < seqlen_q)
+            else:
+                tl.atomic_add(dq_ptrs, dq, mask=(offs_m_curr[:, None] < seqlen_q) & (offs_d[None, :] < headdim))
+        dq_ptrs += BLOCK_M * stride_dqm
+        q_ptrs += BLOCK_M * stride_qm
+        do_ptrs += BLOCK_M * stride_dom
+        if BIAS_TYPE == 'matrix':
+            b_ptrs += BLOCK_M * stride_bm
+    dv_ptrs = DV + (offs_n[:, None] * stride_dvn + offs_d[None, :])
+    dk_ptrs = DK + (offs_n[:, None] * stride_dkn + offs_d[None, :])
+    _bwd_store_dk_dv(dk_ptrs, dv_ptrs, dk, dv, offs_n, offs_d, seqlen_k, headdim, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM)
+
+def init_to_zero(name):
+    return lambda nargs: nargs[name].zero_()
+
+@triton.autotune(configs=[triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'SEQUENCE_PARALLEL': False}, num_warps=8, num_stages=1, pre_hook=init_to_zero('DQ')), triton.Config({'BLOCK_M': 128, 'BLOCK_N': 128, 'SEQUENCE_PARALLEL': True}, num_warps=8, num_stages=1, pre_hook=init_to_zero('DQ'))], key=['CACHE_KEY_SEQLEN_Q', 'CACHE_KEY_SEQLEN_K', 'BIAS_TYPE', 'IS_CAUSAL', 'BLOCK_HEADDIM'])
+@triton.heuristics({'EVEN_M': lambda args: args['seqlen_q'] % args['BLOCK_M'] == 0, 'EVEN_N': lambda args: args['seqlen_k'] % args['BLOCK_N'] == 0, 'EVEN_HEADDIM': lambda args: args['headdim'] == args['BLOCK_HEADDIM']})
+@triton.jit
+def _bwd_kernel(Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qb, stride_qh, stride_qm, stride_kb, stride_kh, stride_kn, stride_vb, stride_vh, stride_vn, stride_bb, stride_bh, stride_bm, stride_dob, stride_doh, stride_dom, stride_dqb, stride_dqh, stride_dqm, stride_dkb, stride_dkh, stride_dkn, stride_dvb, stride_dvh, stride_dvn, nheads, seqlen_q, seqlen_k, seqlen_q_rounded, headdim, CACHE_KEY_SEQLEN_Q, CACHE_KEY_SEQLEN_K, BIAS_TYPE: tl.constexpr, IS_CAUSAL: tl.constexpr, BLOCK_HEADDIM: tl.constexpr, SEQUENCE_PARALLEL: tl.constexpr, EVEN_M: tl.constexpr, EVEN_N: tl.constexpr, EVEN_HEADDIM: tl.constexpr, BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr):
+    off_hb = tl.program_id(1)
+    off_b = off_hb // nheads
+    off_h = off_hb % nheads
+    Q += off_b * stride_qb + off_h * stride_qh
+    K += off_b * stride_kb + off_h * stride_kh
+    V += off_b * stride_vb + off_h * stride_vh
+    DO += off_b * stride_dob + off_h * stride_doh
+    DQ += off_b * stride_dqb + off_h * stride_dqh
+    DK += off_b * stride_dkb + off_h * stride_dkh
+    DV += off_b * stride_dvb + off_h * stride_dvh
+    if BIAS_TYPE != 'none':
+        Bias += off_b * stride_bb + off_h * stride_bh
+    D += off_hb * seqlen_q_rounded
+    LSE += off_hb * seqlen_q_rounded
+    if not SEQUENCE_PARALLEL:
+        num_block_n = tl.cdiv(seqlen_k, BLOCK_N)
+        for start_n in range(0, num_block_n):
+            _bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD=False, BIAS_TYPE=BIAS_TYPE, IS_CAUSAL=IS_CAUSAL, BLOCK_HEADDIM=BLOCK_HEADDIM, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM, BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N)
+    else:
+        start_n = tl.program_id(0)
+        _bwd_kernel_one_col_block(start_n, Q, K, V, Bias, DO, DQ, DK, DV, LSE, D, softmax_scale, stride_qm, stride_kn, stride_vn, stride_bm, stride_dom, stride_dqm, stride_dkn, stride_dvn, seqlen_q, seqlen_k, headdim, ATOMIC_ADD=True, BIAS_TYPE=BIAS_TYPE, IS_CAUSAL=IS_CAUSAL, BLOCK_HEADDIM=BLOCK_HEADDIM, EVEN_M=EVEN_M, EVEN_N=EVEN_N, EVEN_HEADDIM=EVEN_HEADDIM, BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N)
+
+def _flash_attn_forward(q, k, v, bias=None, causal=False, softmax_scale=None):
+    (batch, seqlen_q, nheads, d) = q.shape
+    (_, seqlen_k, _, _) = k.shape
+    assert k.shape == (batch, seqlen_k, nheads, d)
+    assert v.shape == (batch, seqlen_k, nheads, d)
+    assert d <= 128, 'FlashAttention only support head dimensions up to 128'
+    assert q.dtype == k.dtype == v.dtype, 'All tensors must have the same type'
+    assert q.dtype in [torch.float16, torch.bfloat16], 'Only support fp16 and bf16'
+    assert q.is_cuda and k.is_cuda and v.is_cuda
+    softmax_scale = softmax_scale or 1.0 / math.sqrt(d)
+    has_bias = bias is not None
+    bias_type = 'none'
+    if has_bias:
+        assert bias.dtype in [q.dtype, torch.float]
+        assert bias.is_cuda
+        assert bias.dim() == 4
+        if bias.stride(-1) != 1:
+            bias = bias.contiguous()
+        if bias.shape[2:] == (1, seqlen_k):
+            bias_type = 'vector'
+        elif bias.shape[2:] == (seqlen_q, seqlen_k):
+            bias_type = 'matrix'
+        else:
+            raise RuntimeError('Last 2 dimensions of bias must be (1, seqlen_k) or (seqlen_q, seqlen_k)')
+        bias = bias.expand(batch, nheads, seqlen_q, seqlen_k)
+    bias_strides = (bias.stride(0), bias.stride(1), bias.stride(2)) if has_bias else (0, 0, 0)
+    seqlen_q_rounded = math.ceil(seqlen_q / 128) * 128
+    lse = torch.empty((batch, nheads, seqlen_q_rounded), device=q.device, dtype=torch.float32)
+    tmp = torch.empty((batch, nheads, seqlen_q_rounded), device=q.device, dtype=torch.float32)
+    o = torch.empty_like(q)
+    BLOCK_HEADDIM = max(triton.next_power_of_2(d), 16)
+    BLOCK = 128
+    num_warps = 4 if d <= 64 else 8
+    grid = lambda META: (triton.cdiv(seqlen_q, META['BLOCK_M']), batch * nheads)
+    _fwd_kernel[grid](q, k, v, bias, o, lse, tmp, softmax_scale, q.stride(0), q.stride(2), q.stride(1), k.stride(0), k.stride(2), k.stride(1), v.stride(0), v.stride(2), v.stride(1), *bias_strides, o.stride(0), o.stride(2), o.stride(1), nheads, seqlen_q, seqlen_k, seqlen_q_rounded, d, seqlen_q // 32, seqlen_k // 32, bias_type, causal, BLOCK_HEADDIM, BLOCK_M=BLOCK, BLOCK_N=BLOCK, num_warps=num_warps, num_stages=1)
+    return (o, lse, softmax_scale)
+
+def _flash_attn_backward(do, q, k, v, o, lse, dq, dk, dv, bias=None, causal=False, softmax_scale=None):
+    if do.stride(-1) != 1:
+        do = do.contiguous()
+    (batch, seqlen_q, nheads, d) = q.shape
+    (_, seqlen_k, _, _) = k.shape
+    assert d <= 128
+    seqlen_q_rounded = math.ceil(seqlen_q / 128) * 128
+    assert lse.shape == (batch, nheads, seqlen_q_rounded)
+    assert q.stride(-1) == k.stride(-1) == v.stride(-1) == o.stride(-1) == 1
+    assert dq.stride(-1) == dk.stride(-1) == dv.stride(-1) == 1
+    softmax_scale = softmax_scale or 1.0 / math.sqrt(d)
+    dq_accum = torch.empty_like(q, dtype=torch.float32)
+    delta = torch.empty_like(lse)
+    BLOCK_HEADDIM = max(triton.next_power_of_2(d), 16)
+    grid = lambda META: (triton.cdiv(seqlen_q, META['BLOCK_M']), batch * nheads)
+    _bwd_preprocess_do_o_dot[grid](o, do, delta, o.stride(0), o.stride(2), o.stride(1), do.stride(0), do.stride(2), do.stride(1), nheads, seqlen_q, seqlen_q_rounded, d, BLOCK_M=128, BLOCK_HEADDIM=BLOCK_HEADDIM)
+    has_bias = bias is not None
+    bias_type = 'none'
+    if has_bias:
+        assert bias.dtype in [q.dtype, torch.float]
+        assert bias.is_cuda
+        assert bias.dim() == 4
+        assert bias.stride(-1) == 1
+        if bias.shape[2:] == (1, seqlen_k):
+            bias_type = 'vector'
+        elif bias.shape[2:] == (seqlen_q, seqlen_k):
+            bias_type = 'matrix'
+        else:
+            raise RuntimeError('Last 2 dimensions of bias must be (1, seqlen_k) or (seqlen_q, seqlen_k)')
+        bias = bias.expand(batch, nheads, seqlen_q, seqlen_k)
+    bias_strides = (bias.stride(0), bias.stride(1), bias.stride(2)) if has_bias else (0, 0, 0)
+    grid = lambda META: (triton.cdiv(seqlen_k, META['BLOCK_N']) if META['SEQUENCE_PARALLEL'] else 1, batch * nheads)
+    _bwd_kernel[grid](q, k, v, bias, do, dq_accum, dk, dv, lse, delta, softmax_scale, q.stride(0), q.stride(2), q.stride(1), k.stride(0), k.stride(2), k.stride(1), v.stride(0), v.stride(2), v.stride(1), *bias_strides, do.stride(0), do.stride(2), do.stride(1), dq_accum.stride(0), dq_accum.stride(2), dq_accum.stride(1), dk.stride(0), dk.stride(2), dk.stride(1), dv.stride(0), dv.stride(2), dv.stride(1), nheads, seqlen_q, seqlen_k, seqlen_q_rounded, d, seqlen_q // 32, seqlen_k // 32, bias_type, causal, BLOCK_HEADDIM)
+    dq.copy_(dq_accum)
+
+class FlashAttnQKVPackedFunc(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, qkv, bias=None, causal=False, softmax_scale=None):
+        """
+            qkv: (batch, seqlen, 3, nheads, headdim)
+            bias: optional, shape broadcastible to (batch, nheads, seqlen, seqlen).
+                For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen).
+                ALiBi mask for non-causal would have shape (1, nheads, seqlen, seqlen)
+        """
+        if qkv.stride(-1) != 1:
+            qkv = qkv.contiguous()
+        (o, lse, ctx.softmax_scale) = _flash_attn_forward(qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2], bias=bias, causal=causal, softmax_scale=softmax_scale)
+        ctx.save_for_backward(qkv, o, lse, bias)
+        ctx.causal = causal
+        return o
+
+    @staticmethod
+    def backward(ctx, do):
+        (qkv, o, lse, bias) = ctx.saved_tensors
+        assert not ctx.needs_input_grad[1], 'FlashAttention does not support bias gradient yet'
+        with torch.inference_mode():
+            dqkv = torch.empty_like(qkv)
+            _flash_attn_backward(do, qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2], o, lse, dqkv[:, :, 0], dqkv[:, :, 1], dqkv[:, :, 2], bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
+        return (dqkv, None, None, None)
+flash_attn_qkvpacked_func = FlashAttnQKVPackedFunc.apply
+
+class FlashAttnKVPackedFunc(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, q, kv, bias=None, causal=False, softmax_scale=None):
+        """
+            q: (batch, seqlen_q, nheads, headdim)
+            kv: (batch, seqlen_k, 2, nheads, headdim)
+            bias: optional, shape broadcastible to (batch, nheads, seqlen_q, seqlen_k).
+                For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen_k).
+                ALiBi mask for non-causal would have shape (1, nheads, seqlen_q, seqlen_k)
+        """
+        (q, kv) = [x if x.stride(-1) == 1 else x.contiguous() for x in [q, kv]]
+        (o, lse, ctx.softmax_scale) = _flash_attn_forward(q, kv[:, :, 0], kv[:, :, 1], bias=bias, causal=causal, softmax_scale=softmax_scale)
+        ctx.save_for_backward(q, kv, o, lse, bias)
+        ctx.causal = causal
+        return o
+
+    @staticmethod
+    def backward(ctx, do):
+        (q, kv, o, lse, bias) = ctx.saved_tensors
+        if len(ctx.needs_input_grad) >= 3:
+            assert not ctx.needs_input_grad[2], 'FlashAttention does not support bias gradient yet'
+        with torch.inference_mode():
+            dq = torch.empty_like(q)
+            dkv = torch.empty_like(kv)
+            _flash_attn_backward(do, q, kv[:, :, 0], kv[:, :, 1], o, lse, dq, dkv[:, :, 0], dkv[:, :, 1], bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
+        return (dq, dkv, None, None, None)
+flash_attn_kvpacked_func = FlashAttnKVPackedFunc.apply
+
+class FlashAttnFunc(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, q, k, v, bias=None, causal=False, softmax_scale=None):
+        """
+            q: (batch_size, seqlen_q, nheads, headdim)
+            k, v: (batch_size, seqlen_k, nheads, headdim)
+            bias: optional, shape broadcastible to (batch, nheads, seqlen_q, seqlen_k).
+                For example, ALiBi mask for causal would have shape (1, nheads, 1, seqlen_k).
+                ALiBi mask for non-causal would have shape (1, nheads, seqlen_q, seqlen_k)
+        """
+        (q, k, v) = [x if x.stride(-1) == 1 else x.contiguous() for x in [q, k, v]]
+        (o, lse, ctx.softmax_scale) = _flash_attn_forward(q, k, v, bias=bias, causal=causal, softmax_scale=softmax_scale)
+        ctx.save_for_backward(q, k, v, o, lse, bias)
+        ctx.causal = causal
+        return o
+
+    @staticmethod
+    def backward(ctx, do):
+        (q, k, v, o, lse, bias) = ctx.saved_tensors
+        assert not ctx.needs_input_grad[3], 'FlashAttention does not support bias gradient yet'
+        with torch.inference_mode():
+            dq = torch.empty_like(q)
+            dk = torch.empty_like(k)
+            dv = torch.empty_like(v)
+            _flash_attn_backward(do, q, k, v, o, lse, dq, dk, dv, bias=bias, causal=ctx.causal, softmax_scale=ctx.softmax_scale)
+        return (dq, dk, dv, None, None, None)
+flash_attn_func = FlashAttnFunc.apply

generation_config.json

@@ -1,5 +1,6 @@
 {
   "_from_model_config": true,
   "transformers_version": "4.28.1",
+  "eos_token_id": 0,
   "use_cache": false
 }

hf_prefixlm_converter.py

@@ -6,23 +6,13 @@ Causal LM to convert it to a Prefix LM.
 Prefix LMs accepts a `bidirectional_mask` input in `forward`
 and treat the input prompt as the prefix in `generate`.
 """
-import math
-import warnings
 from types import MethodType
-from typing import Any, Dict, List, Optional, Tuple, Union
+from typing import Any, List, MutableMapping, Optional, Tuple, Union
 import torch
-from transformers.models.bloom.modeling_bloom import BaseModelOutputWithPastAndCrossAttentions, BloomForCausalLM, BloomModel, CausalLMOutputWithCrossAttentions, CrossEntropyLoss
-from transformers.models.bloom.modeling_bloom import _expand_mask as _expand_mask_bloom
-from transformers.models.bloom.modeling_bloom import _make_causal_mask as _make_causal_mask_bloom
-from transformers.models.bloom.modeling_bloom import logging
 from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel
 from transformers.models.gpt_neo.modeling_gpt_neo import GPTNeoForCausalLM
 from transformers.models.gpt_neox.modeling_gpt_neox import GPTNeoXForCausalLM
 from transformers.models.gptj.modeling_gptj import GPTJForCausalLM
-from transformers.models.opt.modeling_opt import OPTForCausalLM
-from transformers.models.opt.modeling_opt import _expand_mask as _expand_mask_opt
-from transformers.models.opt.modeling_opt import _make_causal_mask as _make_causal_mask_opt
-logger = logging.get_logger(__name__)
 _SUPPORTED_GPT_MODELS = (GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM)
 CAUSAL_GPT_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM]
 
@@ -90,13 +80,14 @@ def _convert_gpt_causal_lm_to_prefix_lm(model: CAUSAL_GPT_TYPES) -> CAUSAL_GPT_T
             bidirectional_mask = torch.cat([bidirectional_mask, pad], dim=1)
         bidirectional = bidirectional_mask.unsqueeze(1).unsqueeze(1)
         for attn_module in attn_modules:
+            assert isinstance(attn_module.bias, torch.Tensor)
             attn_module.bias.data = torch.logical_or(attn_module.bias.data, bidirectional)
         output = call_og_forward()
         for attn_module in attn_modules:
             attn_module.bias.data = torch.tril(attn_module.bias.data[0, 0])[None, None]
         return output
 
-    def generate(self: CAUSAL_GPT_TYPES, *args: tuple, **kwargs: Dict[str, Any]):
+    def generate(self: CAUSAL_GPT_TYPES, *args: Any, **kwargs: Any):
         """Wraps original generate to enable PrefixLM attention."""
         attn_modules = _get_attn_modules(model)
         for attn_module in attn_modules:
@@ -109,228 +100,8 @@ def _convert_gpt_causal_lm_to_prefix_lm(model: CAUSAL_GPT_TYPES) -> CAUSAL_GPT_T
     setattr(model, 'generate', MethodType(generate, model))
     setattr(model, '_prefix_lm_converted', True)
     return model
-
-def _convert_bloom_causal_lm_to_prefix_lm(model: BloomForCausalLM) -> BloomForCausalLM:
-    """Converts a BLOOM Causal LM to a Prefix LM.
-
-    Supported HuggingFace model classes:
-        - `BloomForCausalLM`
-
-    See `convert_hf_causal_lm_to_prefix_lm` for more details.
-    """
-    if hasattr(model, '_prefix_lm_converted'):
-        return model
-    assert isinstance(model, BloomForCausalLM)
-    assert model.config.add_cross_attention == False, 'Only supports BLOOM decoder-only models'
-
-    def _prepare_attn_mask(self: BloomModel, attention_mask: torch.Tensor, bidirectional_mask: Optional[torch.Tensor], input_shape: Tuple[int, int], past_key_values_length: int) -> torch.BoolTensor:
-        combined_attention_mask = None
-        device = attention_mask.device
-        (_, src_length) = input_shape
-        if src_length > 1:
-            combined_attention_mask = _make_causal_mask_bloom(input_shape, device=device, past_key_values_length=past_key_values_length)
-            if bidirectional_mask is not None:
-                assert attention_mask.shape == bidirectional_mask.shape
-                expanded_bidirectional_mask = _expand_mask_bloom(bidirectional_mask, tgt_length=src_length)
-                combined_attention_mask = torch.logical_and(combined_attention_mask, expanded_bidirectional_mask)
-        expanded_attn_mask = _expand_mask_bloom(attention_mask, tgt_length=src_length)
-        combined_attention_mask = expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
-        return combined_attention_mask
-
-    def _build_alibi_tensor(self: BloomModel, batch_size: int, query_length: int, key_length: int, dtype: torch.dtype, device: torch.device) -> torch.Tensor:
-        num_heads = self.config.n_head
-        closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
-        base = torch.tensor(2 ** (-2 ** (-(math.log2(closest_power_of_2) - 3))), device=device, dtype=torch.float32)
-        powers = torch.arange(1, 1 + closest_power_of_2, device=device, dtype=torch.int32)
-        slopes = torch.pow(base, powers)
-        if closest_power_of_2 != num_heads:
-            extra_base = torch.tensor(2 ** (-2 ** (-(math.log2(2 * closest_power_of_2) - 3))), device=device, dtype=torch.float32)
-            num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
-            extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=device, dtype=torch.int32)
-            slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
-        qa = torch.arange(query_length, device=device, dtype=torch.int32).view(-1, 1)
-        ka = torch.arange(key_length, device=device, dtype=torch.int32).view(1, -1)
-        diffs = qa - ka + key_length - query_length
-        diffs = -diffs.abs()
-        alibi = slopes.view(1, num_heads, 1, 1) * diffs.view(1, 1, query_length, key_length)
-        alibi = alibi.expand(batch_size, -1, -1, -1).reshape(-1, query_length, key_length)
-        return alibi.to(dtype)
-    KeyValueT = Tuple[torch.Tensor, torch.Tensor]
-
-    def forward(self: BloomModel, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[KeyValueT, ...]]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, head_mask: Optional[torch.LongTensor]=None, inputs_embeds: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None, **deprecated_arguments) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
-        if deprecated_arguments.pop('position_ids', False) is not False:
-            warnings.warn('`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. ' + 'You can safely ignore passing `position_ids`.', FutureWarning)
-        if len(deprecated_arguments) > 0:
-            raise ValueError(f'Got unexpected arguments: {deprecated_arguments}')
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError('You cannot specify both input_ids and inputs_embeds at the same time')
-        elif input_ids is not None:
-            (batch_size, seq_length) = input_ids.shape
-        elif inputs_embeds is not None:
-            (batch_size, seq_length, _) = inputs_embeds.shape
-        else:
-            raise ValueError('You have to specify either input_ids or inputs_embeds')
-        if past_key_values is None:
-            past_key_values = tuple([None] * len(self.h))
-        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
-        if inputs_embeds is None:
-            inputs_embeds = self.word_embeddings(input_ids)
-        hidden_states = self.word_embeddings_layernorm(inputs_embeds)
-        presents = () if use_cache else None
-        all_self_attentions = () if output_attentions else None
-        all_hidden_states = () if output_hidden_states else None
-        seq_length_with_past = seq_length
-        past_key_values_length = 0
-        if past_key_values[0] is not None:
-            tmp = past_key_values[0][0]
-            past_key_values_length = tmp.shape[2]
-            seq_length_with_past = seq_length_with_past + past_key_values_length
-        if attention_mask is None:
-            attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
-        else:
-            attention_mask = attention_mask.to(hidden_states.device)
-        alibi = self._build_alibi_tensor(batch_size=batch_size, query_length=seq_length, key_length=seq_length_with_past, dtype=hidden_states.dtype, device=hidden_states.device)
-        causal_mask = self._prepare_attn_mask(attention_mask, bidirectional_mask, input_shape=(batch_size, seq_length), past_key_values_length=past_key_values_length)
-        for (i, (block, layer_past)) in enumerate(zip(self.h, past_key_values)):
-            if output_hidden_states:
-                hst = (hidden_states,)
-                all_hidden_states = all_hidden_states + hst
-            if self.gradient_checkpointing and self.training:
-                if use_cache:
-                    logger.warning('`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...')
-                    use_cache = False
-
-                def create_custom_forward(module):
-
-                    def custom_forward(*inputs):
-                        return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
-                    return custom_forward
-                outputs = torch.utils.checkpoint.checkpoint(create_custom_forward(block), hidden_states, alibi, causal_mask, head_mask[i])
-            else:
-                outputs = block(hidden_states, layer_past=layer_past, attention_mask=causal_mask, head_mask=head_mask[i], use_cache=use_cache, output_attentions=output_attentions, alibi=alibi)
-            hidden_states = outputs[0]
-            if use_cache is True:
-                presents = presents + (outputs[1],)
-            if output_attentions:
-                oa = (outputs[2 if use_cache else 1],)
-                all_self_attentions = all_self_attentions + oa
-        hidden_states = self.ln_f(hidden_states)
-        if output_hidden_states:
-            hst = (hidden_states,)
-            all_hidden_states = all_hidden_states + hst
-        if not return_dict:
-            return tuple((v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None))
-        return BaseModelOutputWithPastAndCrossAttentions(last_hidden_state=hidden_states, past_key_values=presents, hidden_states=all_hidden_states, attentions=all_self_attentions)
-    setattr(model.transformer, '_prepare_attn_mask', MethodType(_prepare_attn_mask, model.transformer))
-    setattr(model.transformer, '_build_alibi_tensor', MethodType(_build_alibi_tensor, model.transformer))
-    setattr(model.transformer, 'forward', MethodType(forward, model.transformer))
-    KeyValueT = Tuple[torch.Tensor, torch.Tensor]
-
-    def forward(self: BloomForCausalLM, input_ids: Optional[torch.LongTensor]=None, past_key_values: Optional[Tuple[KeyValueT, ...]]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.Tensor]=None, head_mask: Optional[torch.Tensor]=None, inputs_embeds: Optional[torch.Tensor]=None, labels: Optional[torch.Tensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None, **deprecated_arguments) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
-        """Replacement forward method for BloomCausalLM."""
-        if deprecated_arguments.pop('position_ids', False) is not False:
-            warnings.warn('`position_ids` have no functionality in BLOOM and will be removed ' + 'in v5.0.0. You can safely ignore passing `position_ids`.', FutureWarning)
-        if len(deprecated_arguments) > 0:
-            raise ValueError(f'Got unexpected arguments: {deprecated_arguments}')
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-        transformer_outputs = self.transformer(input_ids, past_key_values=past_key_values, attention_mask=attention_mask, bidirectional_mask=bidirectional_mask, head_mask=head_mask, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
-        hidden_states = transformer_outputs[0]
-        lm_logits = self.lm_head(hidden_states)
-        loss = None
-        if labels is not None:
-            shift_logits = lm_logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            (batch_size, seq_length, vocab_size) = shift_logits.shape
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length))
-        if not return_dict:
-            output = (lm_logits,) + transformer_outputs[1:]
-            return (loss,) + output if loss is not None else output
-        return CausalLMOutputWithCrossAttentions(loss=loss, logits=lm_logits, past_key_values=transformer_outputs.past_key_values, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions)
-
-    def prepare_inputs_for_generation(self: BloomForCausalLM, input_ids: torch.LongTensor, past: Optional[torch.Tensor]=None, attention_mask: Optional[torch.Tensor]=None, **kwargs) -> dict:
-        if past:
-            input_ids = input_ids[:, -1].unsqueeze(-1)
-            bidirectional_mask = None
-            if past[0][0].shape[0] == input_ids.shape[0]:
-                past = self._convert_to_bloom_cache(past)
-        else:
-            bidirectional_mask = torch.ones_like(input_ids)
-        return {'input_ids': input_ids, 'past_key_values': past, 'use_cache': True, 'attention_mask': attention_mask, 'bidirectional_mask': bidirectional_mask}
-    setattr(model, 'forward', MethodType(forward, model))
-    setattr(model, 'prepare_inputs_for_generation', MethodType(prepare_inputs_for_generation, model))
-    setattr(model, '_prefix_lm_converted', True)
-    return model
-
-def _convert_opt_causal_lm_to_prefix_lm(model: OPTForCausalLM) -> OPTForCausalLM:
-    """Converts an OPT Causal LM to a Prefix LM.
-
-    Supported HuggingFace model classes:
-        - `OPTForCausalLM`
-
-    See `convert_hf_causal_lm_to_prefix_lm` for more details.
-    """
-    if hasattr(model, '_prefix_lm_converted'):
-        return model
-    assert isinstance(model, OPTForCausalLM)
-    assert model.config.add_cross_attention == False, 'Only supports OPT decoder-only models'
-    setattr(model, '_original_forward', getattr(model, 'forward'))
-    setattr(model, '_original_generate', getattr(model, 'generate'))
-    model.model.decoder.bidirectional_mask = None
-
-    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
-        combined_attention_mask = None
-        if input_shape[-1] > 1:
-            if self.bidirectional_mask == 'g':
-                (bsz, src_length) = input_shape
-                combined_attention_mask = torch.zeros((bsz, 1, src_length, src_length + past_key_values_length), dtype=inputs_embeds.dtype, device=inputs_embeds.device)
-            else:
-                combined_attention_mask = _make_causal_mask_opt(input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length).to(inputs_embeds.device)
-                if self.bidirectional_mask is not None:
-                    assert attention_mask.shape == self.bidirectional_mask.shape
-                    expanded_bidirectional_mask = _expand_mask_opt(self.bidirectional_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(inputs_embeds.device)
-                    combined_attention_mask = torch.maximum(expanded_bidirectional_mask, combined_attention_mask)
-        if attention_mask is not None:
-            expanded_attn_mask = _expand_mask_opt(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(inputs_embeds.device)
-            combined_attention_mask = expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
-        return combined_attention_mask
-    setattr(model.model.decoder, '_prepare_decoder_attention_mask', MethodType(_prepare_decoder_attention_mask, model.model.decoder))
-
-    def forward(self: OPTForCausalLM, input_ids: Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None, bidirectional_mask: Optional[torch.ByteTensor]=None, head_mask: Optional[torch.Tensor]=None, past_key_values: Optional[List[torch.FloatTensor]]=None, inputs_embeds: Optional[torch.FloatTensor]=None, labels: Optional[torch.LongTensor]=None, use_cache: Optional[bool]=None, output_attentions: Optional[bool]=None, output_hidden_states: Optional[bool]=None, return_dict: Optional[bool]=None):
-
-        def call_og_forward():
-            return self._original_forward(input_ids=input_ids, attention_mask=attention_mask, head_mask=head_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict)
-        if bidirectional_mask is None:
-            return call_og_forward()
-        self.model.decoder.bidirectional_mask = bidirectional_mask
-        try:
-            outputs = call_og_forward()
-        except:
-            self.model.decoder.bidirectional_mask = None
-            raise
-        self.model.decoder.bidirectional_mask = None
-        return outputs
-
-    def generate(self: OPTForCausalLM, *args: tuple, **kwargs: Dict[str, Any]):
-        """Wraps original generate to enable PrefixLM-style attention."""
-        self.model.decoder.bidirectional_mask = 'g'
-        try:
-            output = self._original_generate(*args, **kwargs)
-        except:
-            self.model.decoder.bidirectional_mask = None
-            raise
-        self.model.decoder.bidirectional_mask = None
-        return output
-    setattr(model, 'forward', MethodType(forward, model))
-    setattr(model, 'generate', MethodType(generate, model))
-    setattr(model, '_prefix_lm_converted', True)
-    return model
-_SUPPORTED_HF_MODELS = _SUPPORTED_GPT_MODELS + (BloomForCausalLM, OPTForCausalLM)
-CAUSAL_LM_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM, BloomForCausalLM, OPTForCausalLM]
+_SUPPORTED_HF_MODELS = _SUPPORTED_GPT_MODELS
+CAUSAL_LM_TYPES = Union[GPT2LMHeadModel, GPTJForCausalLM, GPTNeoForCausalLM, GPTNeoXForCausalLM]
 
 def convert_hf_causal_lm_to_prefix_lm(model: CAUSAL_LM_TYPES) -> CAUSAL_LM_TYPES:
     """Converts a HuggingFace Causal LM to a Prefix LM.
@@ -340,8 +111,6 @@ def convert_hf_causal_lm_to_prefix_lm(model: CAUSAL_LM_TYPES) -> CAUSAL_LM_TYPES
         - `GPTNeoForCausalLM`
         - `GPTNeoXForCausalLM`
         - `GPTJForCausalLM`
-        - `BloomForCausalLM`
-        - `OPTForCausalLM`
 
     Conversion to a Prefix LM is done by modifying the `forward` method, and possibly also the
     `generate` method and/or select underlying methods depending on the model class.
@@ -391,14 +160,10 @@ def convert_hf_causal_lm_to_prefix_lm(model: CAUSAL_LM_TYPES) -> CAUSAL_LM_TYPES
     """
     if isinstance(model, _SUPPORTED_GPT_MODELS):
         return _convert_gpt_causal_lm_to_prefix_lm(model)
-    elif isinstance(model, BloomForCausalLM):
-        return _convert_bloom_causal_lm_to_prefix_lm(model)
-    elif isinstance(model, OPTForCausalLM):
-        return _convert_opt_causal_lm_to_prefix_lm(model)
     else:
         raise TypeError(f'Cannot convert model to Prefix LM. ' + f'Model does not belong to set of supported HF models:' + f'\n{_SUPPORTED_HF_MODELS}')
 
-def add_bidirectional_mask_if_missing(batch: Dict[str, Any]):
+def add_bidirectional_mask_if_missing(batch: MutableMapping):
     """Attempts to add bidirectional_mask to batch if missing.
 
     Raises:

meta_init_context.py

@@ -1,4 +1,5 @@
 from contextlib import contextmanager
+from typing import Any, Callable, Optional
 import torch
 import torch.nn as nn
 
@@ -57,25 +58,29 @@ def init_on_device(device: torch.device, include_buffers: bool=False):
     if include_buffers:
         old_register_buffer = nn.Module.register_buffer
 
-    def register_empty_parameter(module, name, param):
-        old_register_parameter(module, name, param)
+    def register_empty_parameter(self: torch.nn.Module, name: str, param: Optional[torch.nn.Parameter]):
+        old_register_parameter(self, name, param)
         if param is not None:
-            param_cls = type(module._parameters[name])
-            kwargs = module._parameters[name].__dict__
-            module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
-
-    def register_empty_buffer(module, name, buffer):
-        old_register_buffer(module, name, buffer)
-        if buffer is not None:
-            module._buffers[name] = module._buffers[name].to(device)
+            parameter = self._parameters[name]
+            assert parameter is not None
+            param_cls = type(parameter)
+            kwargs = parameter.__dict__
+            self._parameters[name] = param_cls(parameter.to(device), **kwargs)
+
+    def register_empty_buffer(self: torch.nn.Module, name: str, tensor: Optional[torch.Tensor], persistent: bool=True):
+        old_register_buffer(self, name, tensor, persistent=persistent)
+        if tensor is not None:
+            named_buffer = self._buffers[name]
+            assert named_buffer is not None
+            self._buffers[name] = named_buffer.to(device)
     if include_buffers:
         tensor_constructors_to_patch = {torch_function_name: getattr(torch, torch_function_name) for torch_function_name in ['empty', 'zeros', 'ones', 'full']}
     else:
         tensor_constructors_to_patch = {}
 
-    def patch_tensor_constructor(fn):
+    def patch_tensor_constructor(fn: Callable):
 
-        def wrapper(*args, **kwargs):
+        def wrapper(*args: Any, **kwargs: Any):
             kwargs['device'] = device
             return fn(*args, **kwargs)
         return wrapper

modeling_mpt.py

@@ -4,55 +4,66 @@ Inspired by https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
 """
 import math
 import warnings
-from typing import List, Optional, Tuple, Union
+from typing import Any, Dict, List, Mapping, MutableMapping, Optional, Tuple, Union
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from transformers import PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast
+from transformers import PreTrainedModel, PreTrainedTokenizerBase
 from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from .attention import attn_bias_shape, build_attn_bias
 from .blocks import MPTBlock
+from .custom_embedding import SharedEmbedding
+from .fc import FC_CLASS_REGISTRY as FC_CLASS_REGISTRY
+from .ffn import FFN_CLASS_REGISTRY as FFN_CLASS_REGISTRY
+from .ffn import MPTMLP as MPTMLP
+from .ffn import build_ffn as build_ffn
 from .norm import NORM_CLASS_REGISTRY
 from .configuration_mpt import MPTConfig
 from .adapt_tokenizer import AutoTokenizerForMOD, adapt_tokenizer_for_denoising
 from .hf_prefixlm_converter import add_bidirectional_mask_if_missing, convert_hf_causal_lm_to_prefix_lm
 from .meta_init_context import init_empty_weights
-from .param_init_fns import MODEL_INIT_REGISTRY, generic_param_init_fn_
-Tokenizer = Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
+from .param_init_fns import generic_param_init_fn_, MODEL_INIT_REGISTRY
+try:
+    from .flash_attn_triton import flash_attn_func as flash_attn_func
+except:
+    pass
+import logging
+log = logging.getLogger(__name__)
 
 class MPTPreTrainedModel(PreTrainedModel):
     config_class = MPTConfig
     base_model_prefix = 'model'
-    _no_split_modules = ["MPTBlock"]
-    supports_gradient_checkpointing = True
-
-    def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, MPTModel):
-            module.gradient_checkpointing = value
+    _no_split_modules = ['MPTBlock']
 
 class MPTModel(MPTPreTrainedModel):
 
     def __init__(self, config: MPTConfig):
         config._validate_config()
         super().__init__(config)
-        self.gradient_checkpointing = False
         self.attn_impl = config.attn_config['attn_impl']
         self.prefix_lm = config.attn_config['prefix_lm']
         self.attn_uses_sequence_id = config.attn_config['attn_uses_sequence_id']
         self.alibi = config.attn_config['alibi']
         self.alibi_bias_max = config.attn_config['alibi_bias_max']
+        self.learned_pos_emb = config.learned_pos_emb
+        if config.init_device == 'mixed':
+            if dist.get_local_rank() == 0:
+                config.init_device = 'cpu'
+            else:
+                config.init_device = 'meta'
         if config.norm_type.lower() not in NORM_CLASS_REGISTRY.keys():
             norm_options = ' | '.join(NORM_CLASS_REGISTRY.keys())
             raise NotImplementedError(f'Requested norm type ({config.norm_type}) is not implemented within this repo (Options: {norm_options}).')
         norm_class = NORM_CLASS_REGISTRY[config.norm_type.lower()]
         self.embedding_fraction = config.embedding_fraction
-        self.wte = nn.Embedding(config.vocab_size, config.d_model, device=config.init_device)
-        if not self.alibi:
-            self.wpe = nn.Embedding(config.max_seq_len, config.d_model, device=config.init_device)
+        self.wte = SharedEmbedding(config.vocab_size, config.d_model, device=config.init_device)
+        if self.learned_pos_emb:
+            self.wpe = torch.nn.Embedding(config.max_seq_len, config.d_model, device=config.init_device)
         self.emb_drop = nn.Dropout(config.emb_pdrop)
         self.blocks = nn.ModuleList([MPTBlock(device=config.init_device, **config.to_dict()) for _ in range(config.n_layers)])
         self.norm_f = norm_class(config.d_model, device=config.init_device)
         if config.init_device != 'meta':
+            log.info(f'We recommend using config.init_device="meta" with Composer + FSDP for faster initialization.')
             self.apply(self.param_init_fn)
         self.is_causal = not self.prefix_lm
         self._attn_bias_initialized = False
@@ -61,25 +72,22 @@ class MPTModel(MPTPreTrainedModel):
         if config.no_bias:
             for module in self.modules():
                 if hasattr(module, 'bias') and isinstance(module.bias, nn.Parameter):
-                    if config.verbose:
-                        warnings.warn(f'Removing bias ({module.bias}) from {module}.')
+                    log.info(f'Removing bias ({module.bias}) from {module}.')
                     module.register_parameter('bias', None)
-        if config.verbose and config.verbose > 2:
-            print(self)
-        if 'verbose' not in self.config.init_config:
-            self.config.init_config['verbose'] = self.config.verbose
-        if self.config.init_config['verbose'] > 1:
-            init_fn_name = self.config.init_config['name']
-            warnings.warn(f'Using {init_fn_name} initialization.')
+                if hasattr(module, 'use_bias'):
+                    log.info(f'Setting use_bias=False for {module}.')
+                    module.use_bias = False
+        log.debug(self)
+        log.debug(f"Using {self.config.init_config['name']} initialization.")
 
-    def get_input_embeddings(self):
+    def get_input_embeddings(self) -> nn.Embedding:
         return self.wte
 
-    def set_input_embeddings(self, value):
+    def set_input_embeddings(self, value: nn.Embedding) -> None:
         self.wte = value
 
     @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):
+    def _attn_bias(self, device: torch.device, dtype: torch.dtype, attention_mask: Optional[torch.ByteTensor]=None, prefix_mask: Optional[torch.ByteTensor]=None, sequence_id: Optional[torch.LongTensor]=None) -> Tuple[Optional[torch.Tensor], Optional[torch.ByteTensor]]:
         if not self._attn_bias_initialized:
             if self.attn_bias_shape:
                 self.attn_bias = torch.zeros(self.attn_bias_shape, device=device, dtype=dtype)
@@ -102,14 +110,15 @@ class MPTModel(MPTPreTrainedModel):
             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:]
+                _s_k = max(0, attn_bias.size(-1) - s_k)
+                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):
+    def _apply_prefix_mask(self, attn_bias: torch.Tensor, prefix_mask: torch.Tensor) -> 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}.')
@@ -124,7 +133,7 @@ class MPTModel(MPTPreTrainedModel):
         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):
+    def _apply_sequence_id(self, attn_bias: torch.Tensor, sequence_id: torch.LongTensor) -> torch.Tensor:
         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}')
@@ -134,122 +143,86 @@ class MPTModel(MPTPreTrainedModel):
         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, inputs_embeds: Optional[torch.FloatTensor] = None):
+    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, inputs_embeds: Optional[torch.Tensor]=None) -> BaseModelOutputWithPast:
         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
-        if self.gradient_checkpointing and self.training:
-            if use_cache:
-                use_cache = False
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
-        elif input_ids is not None:
-            batch_size, seq_length = input_ids.shape
-        elif inputs_embeds is not None:
-            batch_size, seq_length, _ = inputs_embeds.shape
-        else:
-            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
-
-        seq_length_with_past = seq_length
-        past_key_values_length = 0
-
-        if past_key_values is not None:
-            past_key_values_length = past_key_values[0][0].shape[2]
-            seq_length_with_past = seq_length_with_past + past_key_values_length
-
         if attention_mask is not None:
             attention_mask = attention_mask.bool()
-        else:
-            attention_mask = torch.ones(
-                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
-            )
-
-        if inputs_embeds is None:
-            tok_emb = self.wte(input_ids)
-        else:
-            tok_emb = inputs_embeds
-
         if prefix_mask is not None:
             prefix_mask = prefix_mask.bool()
         if not return_dict:
             raise NotImplementedError('return_dict False is not implemented yet for MPT')
         if output_attentions:
-            raise NotImplementedError('output_attentions is not implemented yet for MPT')
-        #if attention_mask is not None and attention_mask[:, 0].sum() != attention_mask.shape[0] and self.training:
-        #    raise NotImplementedError('MPT does not support training with left padding.')
+            if self.attn_impl != 'torch':
+                raise NotImplementedError('output_attentions is not implemented for MPT when using attn_impl `flash` or `triton`.')
+        if self.training and attention_mask is not None and (attention_mask[:, 0].sum() != attention_mask.shape[0]):
+            raise NotImplementedError('MPT 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 MPT is configured with prefix_lm=True.')
+        if inputs_embeds is not None:
+            raise NotImplementedError('inputs_embeds is not implemented for MPT.')
         if self.training:
             if self.attn_uses_sequence_id and sequence_id is None:
                 raise ValueError('sequence_id is a required argument when MPT 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('MPT 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 = seq_length
+        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}'
-        if self.alibi:
-            x = tok_emb
-        else:
+        tok_emb = self.wte(input_ids)
+        if self.learned_pos_emb:
             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)={len(past_key_values)!r}; self.config.n_layers={self.config.n_layers!r}).')
                 past_position = past_key_values[0][0].size(1)
+                if self.attn_impl == 'torch':
+                    past_position = past_key_values[0][0].size(3)
             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 {S + 1}, this model only supports total sequence length <= {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 and not self.training:
+            if attention_mask is not None:
                 pos = torch.clamp(pos - torch.cumsum((~attention_mask).to(torch.int32), dim=1)[:, past_position:], min=0)
             pos_emb = self.wpe(pos)
             x = tok_emb + pos_emb
+        else:
+            x = tok_emb
         if self.embedding_fraction == 1:
             x = self.emb_drop(x)
         else:
             x_shrunk = x * self.embedding_fraction + x.detach() * (1 - self.embedding_fraction)
             assert isinstance(self.emb_drop, nn.Module)
             x = self.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)
+        (attn_bias, attention_mask) = self._attn_bias(device=x.device, dtype=torch.float32, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id)
+        presents = () if use_cache else None
         if use_cache and past_key_values is None:
             past_key_values = [() for _ in range(self.config.n_layers)]
-
         all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
         for (b_idx, block) in enumerate(self.blocks):
             if output_hidden_states:
                 assert all_hidden_states is not None
                 all_hidden_states = all_hidden_states + (x,)
             past_key_value = past_key_values[b_idx] if past_key_values is not None else None
-
-            if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        # None for past_key_value
-                        return module(*inputs)
-
-                    return custom_forward
-
-                (x, past_key_value) = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
-                    x,
-                    past_key_value,
-                    attn_bias,
-                    attention_mask,
-                    self.is_causal,
-                )
-            else:
-                (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, attn_weights, present) = block(x, past_key_value=past_key_value, attn_bias=attn_bias, attention_mask=attention_mask, is_causal=self.is_causal, output_attentions=bool(output_attentions))
+            if presents is not None:
+                presents += (present,)
+            if output_attentions:
+                assert all_self_attns is not None
+                all_self_attns = all_self_attns + (attn_weights,)
         x = self.norm_f(x)
-        return BaseModelOutputWithPast(last_hidden_state=x, past_key_values=past_key_values, hidden_states=all_hidden_states)
+        if output_hidden_states:
+            assert all_hidden_states is not None
+            all_hidden_states = all_hidden_states + (x,)
+        return BaseModelOutputWithPast(last_hidden_state=x, past_key_values=presents, hidden_states=all_hidden_states, attentions=all_self_attns)
 
-    def param_init_fn(self, module):
+    def param_init_fn(self, module: nn.Module) -> None:
         init_fn_name = self.config.init_config['name']
         MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)
 
-    def fsdp_wrap_fn(self, module):
+    def fsdp_wrap_fn(self, module: nn.Module) -> bool:
         return isinstance(module, MPTBlock)
 
-    def activation_checkpointing_fn(self, module):
+    def activation_checkpointing_fn(self, module: nn.Module) -> bool:
         return isinstance(module, MPTBlock)
 
 class MPTForCausalLM(MPTPreTrainedModel):
@@ -258,7 +231,13 @@ class MPTForCausalLM(MPTPreTrainedModel):
         super().__init__(config)
         if not config.tie_word_embeddings:
             raise ValueError('MPTForCausalLM only supports tied word embeddings')
-        self.transformer = MPTModel(config)
+        log.info(f'Instantiating an MPTForCausalLM model from {__file__}')
+        self.transformer: MPTModel = MPTModel(config)
+        for child in self.transformer.children():
+            if isinstance(child, torch.nn.ModuleList):
+                continue
+            if isinstance(child, torch.nn.Module):
+                child._fsdp_wrap = True
         self.logit_scale = None
         if config.logit_scale is not None:
             logit_scale = config.logit_scale
@@ -269,51 +248,53 @@ class MPTForCausalLM(MPTPreTrainedModel):
                     raise ValueError(f"logit_scale={logit_scale!r} is not recognized as an option; use numeric value or 'inv_sqrt_d_model'.")
             self.logit_scale = logit_scale
 
-    def get_input_embeddings(self):
+    def get_input_embeddings(self) -> nn.Embedding:
         return self.transformer.wte
 
-    def set_input_embeddings(self, value):
+    def set_input_embeddings(self, value: Union[SharedEmbedding, nn.Embedding]) -> None:
         self.transformer.wte = value
 
-    def get_output_embeddings(self):
+    def get_output_embeddings(self) -> nn.Embedding:
         return self.transformer.wte
 
-    def set_output_embeddings(self, new_embeddings):
+    def set_output_embeddings(self, new_embeddings: Union[SharedEmbedding, nn.Embedding]) -> None:
         self.transformer.wte = new_embeddings
 
-    def set_decoder(self, decoder):
+    def set_decoder(self, decoder: MPTModel) -> None:
         self.transformer = decoder
 
-    def get_decoder(self):
+    def get_decoder(self) -> MPTModel:
         return self.transformer
 
-    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, labels: 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, inputs_embeds: Optional[torch.FloatTensor] = None):
+    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, labels: 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, inputs_embeds: Optional[torch.FloatTensor]=None) -> CausalLMOutputWithPast:
         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
-        outputs = self.transformer(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id, return_dict=return_dict, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache, inputs_embeds=inputs_embeds)
-        logits = F.linear(outputs.last_hidden_state, self.transformer.wte.weight)
+        if inputs_embeds is not None:
+            raise NotImplementedError('inputs_embeds has to be None (for hf/peft support).')
+        outputs = self.transformer(input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, prefix_mask=prefix_mask, sequence_id=sequence_id, return_dict=return_dict, output_attentions=output_attentions, output_hidden_states=output_hidden_states, use_cache=use_cache)
+        logits = self.transformer.wte(outputs.last_hidden_state.to(self.transformer.wte.weight.device), True)
         if self.logit_scale is not None:
             if self.logit_scale == 0:
                 warnings.warn(f'Multiplying logits by self.logit_scale={self.logit_scale!r}. This will produce uniform (uninformative) outputs.')
             logits *= self.logit_scale
         loss = None
         if labels is not None:
-            labels = torch.roll(labels, shifts=-1)
-            labels[:, -1] = -100
-            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), labels.to(logits.device).view(-1))
-        return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states)
+            _labels = torch.roll(labels, shifts=-1)
+            _labels[:, -1] = -100
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), _labels.to(logits.device).view(-1))
+        return CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions)
 
-    def param_init_fn(self, module):
+    def param_init_fn(self, module: nn.Module) -> None:
         init_fn_name = self.config.init_config['name']
         MODEL_INIT_REGISTRY[init_fn_name](module=module, n_layers=self.config.n_layers, d_model=self.config.d_model, **self.config.init_config)
 
-    def fsdp_wrap_fn(self, module):
+    def fsdp_wrap_fn(self, module: nn.Module) -> bool:
         return isinstance(module, MPTBlock)
 
-    def activation_checkpointing_fn(self, module):
+    def activation_checkpointing_fn(self, module: nn.Module) -> bool:
         return isinstance(module, MPTBlock)
 
-    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+    def prepare_inputs_for_generation(self, input_ids: torch.Tensor, past_key_values: Optional[List[Tuple[torch.Tensor, torch.Tensor]]]=None, inputs_embeds: Optional[torch.Tensor]=None, **kwargs: Any) -> Dict[str, Any]:
         if inputs_embeds is not None:
             raise NotImplementedError('inputs_embeds is not implemented for MPT yet')
         attention_mask = kwargs['attention_mask'].bool()
@@ -334,7 +315,7 @@ class MPTForCausalLM(MPTPreTrainedModel):
         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):
+    def _reorder_cache(past_key_values: List[Tuple[torch.Tensor, torch.Tensor]], beam_idx: torch.LongTensor) -> List[Tuple[torch.Tensor, ...]]:
         """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

norm.py

@@ -1,6 +1,7 @@
+from typing import Dict, List, Optional, Type, Union
 import torch
 
-def _cast_if_autocast_enabled(tensor):
+def _cast_if_autocast_enabled(tensor: torch.Tensor) -> torch.Tensor:
     if torch.is_autocast_enabled():
         if tensor.device.type == 'cuda':
             dtype = torch.get_autocast_gpu_dtype()
@@ -13,10 +14,10 @@ def _cast_if_autocast_enabled(tensor):
 
 class LPLayerNorm(torch.nn.LayerNorm):
 
-    def __init__(self, normalized_shape, eps=1e-05, elementwise_affine=True, device=None, dtype=None):
+    def __init__(self, normalized_shape: Union[int, List[int], torch.Size], eps: float=1e-05, elementwise_affine: bool=True, device: Optional[torch.device]=None, dtype: Optional[torch.dtype]=None):
         super().__init__(normalized_shape=normalized_shape, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
 
-    def forward(self, x):
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
         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
@@ -24,15 +25,15 @@ class LPLayerNorm(torch.nn.LayerNorm):
         with torch.autocast(enabled=False, device_type=module_device.type):
             return torch.nn.functional.layer_norm(downcast_x, self.normalized_shape, downcast_weight, downcast_bias, self.eps)
 
-def rms_norm(x, weight=None, eps=1e-05):
-    output = x / torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + eps)
+def rms_norm(x: torch.Tensor, weight: Optional[torch.Tensor]=None, eps: float=1e-05) -> torch.Tensor:
+    output = x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + eps)
     if weight is not None:
         return output * weight
     return output
 
 class RMSNorm(torch.nn.Module):
 
-    def __init__(self, normalized_shape, eps=1e-05, weight=True, dtype=None, device=None):
+    def __init__(self, normalized_shape: Union[int, List[int], torch.Size], eps: float=1e-05, weight: bool=True, dtype: Optional[torch.dtype]=None, device: Optional[torch.device]=None):
         super().__init__()
         self.eps = eps
         if weight:
@@ -40,17 +41,17 @@ class RMSNorm(torch.nn.Module):
         else:
             self.register_parameter('weight', None)
 
-    def forward(self, x):
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
         return rms_norm(x.float(), self.weight, self.eps).to(dtype=x.dtype)
 
 class LPRMSNorm(RMSNorm):
 
-    def __init__(self, normalized_shape, eps=1e-05, weight=True, dtype=None, device=None):
+    def __init__(self, normalized_shape: Union[int, List[int], torch.Size], eps: float=1e-05, weight: bool=True, dtype: Optional[torch.dtype]=None, device: Optional[torch.device]=None):
         super().__init__(normalized_shape=normalized_shape, eps=eps, weight=weight, dtype=dtype, device=device)
 
-    def forward(self, x):
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
         downcast_x = _cast_if_autocast_enabled(x)
         downcast_weight = _cast_if_autocast_enabled(self.weight) if self.weight is not None else self.weight
         with torch.autocast(enabled=False, device_type=x.device.type):
             return rms_norm(downcast_x, downcast_weight, self.eps).to(dtype=x.dtype)
-NORM_CLASS_REGISTRY = {'layernorm': torch.nn.LayerNorm, 'low_precision_layernorm': LPLayerNorm, 'rmsnorm': RMSNorm, 'low_precision_rmsnorm': LPRMSNorm}
+NORM_CLASS_REGISTRY: Dict[str, Type[torch.nn.Module]] = {'layernorm': torch.nn.LayerNorm, 'low_precision_layernorm': LPLayerNorm, 'rmsnorm': RMSNorm, 'low_precision_rmsnorm': LPRMSNorm}

param_init_fns.py

@@ -2,22 +2,26 @@ import math
 import warnings
 from collections.abc import Sequence
 from functools import partial
-from typing import Optional, Tuple, Union
+from typing import Any, Callable, Optional, Tuple, Union
 import torch
 from torch import nn
+from .fc import FC_CLASS_REGISTRY
 from .norm import NORM_CLASS_REGISTRY
+try:
+    import transformer_engine.pytorch as te
+except:
+    te = None
 
-def torch_default_param_init_fn_(module: nn.Module, verbose: int=0, **kwargs):
+def torch_default_param_init_fn_(module: nn.Module, **kwargs: Any) -> None:
     del kwargs
-    if verbose > 1:
-        warnings.warn(f"Initializing network using module's reset_parameters attribute")
-    if hasattr(module, 'reset_parameters'):
+    if hasattr(module, 'reset_parameters') and isinstance(module.reset_parameters, Callable):
         module.reset_parameters()
 
-def fused_init_helper_(module: nn.Module, init_fn_):
+def fused_init_helper_(module: nn.Module, init_fn_: Callable) -> None:
     _fused = getattr(module, '_fused', None)
     if _fused is None:
         raise RuntimeError(f'Internal logic error')
+    assert isinstance(module.weight, torch.Tensor)
     (dim, splits) = _fused
     splits = (0, *splits, module.weight.size(dim))
     for (s, e) in zip(splits[:-1], splits[1:]):
@@ -25,10 +29,8 @@ def fused_init_helper_(module: nn.Module, init_fn_):
         slice_indices[dim] = slice(s, e)
         init_fn_(module.weight[slice_indices])
 
-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):
+def generic_param_init_fn_(module: nn.Module, init_fn_: Callable, 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, **kwargs: Any) -> None:
     del kwargs
-    if verbose > 1:
-        warnings.warn(f'If model has bias parameters they are initialized to 0.')
     init_div_is_residual = init_div_is_residual
     if init_div_is_residual is False:
         div_is_residual = 1.0
@@ -36,20 +38,18 @@ def generic_param_init_fn_(module: nn.Module, init_fn_, n_layers: int, d_model:
         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():
+    elif init_div_is_residual.isnumeric():
         div_is_residual = float(init_div_is_residual)
     else:
         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:.3f}. ' + f'Set `init_div_is_residual: false` in init config to disable this.')
-    if isinstance(module, nn.Linear):
+    if isinstance(module, tuple(set(FC_CLASS_REGISTRY.values()))):
         if hasattr(module, '_fused'):
             fused_init_helper_(module, init_fn_)
         else:
             init_fn_(module.weight)
         if module.bias is not None:
+            assert isinstance(module.bias, torch.Tensor)
             torch.nn.init.zeros_(module.bias)
         if init_div_is_residual is not False and getattr(module, '_is_residual', False):
             with torch.no_grad():
@@ -60,8 +60,6 @@ def generic_param_init_fn_(module: nn.Module, init_fn_, n_layers: int, d_model:
             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={std!r}.')
         elif emb_init_uniform_lim is not None:
             lim = emb_init_uniform_lim
             if isinstance(lim, Sequence):
@@ -75,17 +73,13 @@ def generic_param_init_fn_(module: nn.Module, init_fn_, n_layers: int, d_model:
                 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, tuple(set(NORM_CLASS_REGISTRY.values()))):
-        if verbose > 1:
-            warnings.warn(f'Norm weights are set to 1. If norm layer has a bias it is initialized to 0.')
-        if hasattr(module, 'weight') and module.weight is not None:
+        if hasattr(module, 'weight') and isinstance(module.weight, torch.Tensor):
             torch.nn.init.ones_(module.weight)
-        if hasattr(module, 'bias') and module.bias is not None:
+        if hasattr(module, 'bias') and isinstance(module.bias, torch.Tensor):
             torch.nn.init.zeros_(module.bias)
     elif isinstance(module, nn.MultiheadAttention):
         if module._qkv_same_embed_dim:
@@ -114,32 +108,45 @@ def generic_param_init_fn_(module: nn.Module, init_fn_, n_layers: int, d_model:
                 module.out_proj.weight.div_(div_is_residual)
         if module.out_proj.bias is not None:
             torch.nn.init.zeros_(module.out_proj.bias)
+    elif te is not None and isinstance(module, te.LayerNormMLP):
+        if isinstance(module.layer_norm_weight, torch.Tensor):
+            torch.nn.init.ones_(module.layer_norm_weight)
+        if isinstance(module.layer_norm_bias, torch.Tensor):
+            torch.nn.init.zeros_(module.layer_norm_bias)
+        init_fn_(module.fc1_weight)
+        if module.fc1_bias is not None:
+            assert isinstance(module.fc1_bias, torch.Tensor)
+            torch.nn.init.zeros_(module.fc1_bias)
+        init_fn_(module.fc2_weight)
+        if module.fc2_bias is not None:
+            assert isinstance(module.fc2_bias, torch.Tensor)
+            torch.nn.init.zeros_(module.fc2_bias)
+        with torch.no_grad():
+            module.fc2_weight.div_(div_is_residual)
     else:
         for _ in module.parameters(recurse=False):
             raise NotImplementedError(f'{module.__class__.__name__} parameters are not initialized by param_init_fn.')
 
-def _normal_init_(std, mean=0.0):
+def _normal_init_(std: float, mean: float=0.0) -> Callable:
     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):
+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, **kwargs: Any) -> None:
     del kwargs
     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)
+    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)
 
-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):
+def baseline_param_init_fn_(module: nn.Module, init_std: Optional[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, **kwargs: Any) -> None:
     del kwargs
     if init_std is None:
         raise ValueError("You must set model.init_config['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)
+    _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)
 
-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):
+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, **kwargs: Any) -> None:
     del kwargs
     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)
+    _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)
 
-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):
+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, **kwargs: Any) -> None:
     """From section 2.3.1 of GPT-NeoX-20B:
 
     An Open-Source AutoregressiveLanguage Model — Black et. al. (2022)
@@ -148,34 +155,25 @@ def neox_param_init_fn_(module: nn.Module, n_layers: int, d_model: int, emb_init
     """
     del kwargs
     residual_div = n_layers / math.sqrt(10)
-    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)
+    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)
 
-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):
+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', **kwargs: Any) -> None:
     del kwargs
-    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)
+    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)
 
-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):
+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', **kwargs: Any) -> None:
     del kwargs
-    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)
+    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)
 
-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):
+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, **kwargs: Any) -> None:
     del kwargs
     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)
+    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)
 
-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):
+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, **kwargs: Any) -> None:
+    del 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)
+    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)
 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_}

requirements.txt

@@ -0,0 +1,2 @@
+einops==0.5.0
+triton-pre-mlir@git+https://github.com/vchiley/triton.git@triton_pre_mlir_sm90#subdirectory=python