Skip to content

GitLab

Unverified Commit accccdd0 authored by Arthur's avatar Arthur Committed by GitHub
Browse files

[`Add Mixtral`] Adds support for the Mixtral MoE (#27942) 



* up

* up

* test

* logits ok

* up

* up

* few fixes

* conversion script

* up

* nits

* nits

* update

* nuke

* more updates

* nites

* fix many issues

* nit

* scatter

* nit

* nuke megablocks

* nits

* fix conversion script

* nit

* remove

* nits

* nit

* update

* oupsssss

* change

* nits device

* nits

* fixup

* update

* merge

* add copied from

* fix the copy mentions

* update tests

* more fixes

* nits

* conversion script

* add parts of the readme

* Update tests/models/mixtral/test_modeling_mixtral.py
Co-authored-by: default avatarArthur <48595927+ArthurZucker@users.noreply.github.com>

* new test + conversion script

* Apply suggestions from code review
Co-authored-by: default avatarArthur <48595927+ArthurZucker@users.noreply.github.com>

* Apply suggestions from code review

* fix

* fix copies

* fix copies

* ooops

* fix config

* Apply suggestions from code review

* fix nits

* nit

* add copies

* add batched tests

* docs

* fix flash attention

* let's add more verbose

* add correct outputs

* support router ouptus

* ignore copies where needed

* fix

* cat list if list is given for now

* nits

* Update docs/source/en/model_doc/mixtral.md

* finish router refactoring

* fix forward

* fix expected values

* nits

* fixup

* fix

* fix bug

* fix

* fix dtype mismatch

* fix

* grrr grrr I support item assignment

* fix CI

* docs

* fixup

* remove some copied form

* fix weird diff

* skip doctest fast on the config and modeling

* mark that is supports flash attention in the doc

* update

* Update src/transformers/models/mixtral/modeling_mixtral.py
Co-authored-by: default avatarLysandre Debut <hi@lysand.re>

* Update docs/source/en/model_doc/mixtral.md
Co-authored-by: default avatarLysandre Debut <hi@lysand.re>

* revert router logits config issue

* update doc accordingly

* Update src/transformers/models/mixtral/convert_mixtral_weights_to_hf.py

* nits

* use torch testing asssert close

* fixup

* doc nits

---------
Co-authored-by: default avataryounesbelkada <younesbelkada@gmail.com>
Co-authored-by: default avatarYounes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: default avatarLysandre Debut <hi@lysand.re>
parent 0676d992
Show whitespace changes
Inline Side-by-side
Showing with 2433 additions and 0 deletions
src/transformers/models/mixtral/configuration_mixtral.py 0 → 100644
View file @ accccdd0
# coding=utf-8
# Copyright 2023 Mixtral AI and the HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Mixtral model configuration"""
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
MIXTRAL_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"mistral-ai/Mixtral-8x7B": "https://huggingface.co/mistral-ai/Mixtral-8x7B/resolve/main/config.json",
}
class MixtralConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`MixtralModel`]. It is used to instantiate an
Mixtral model according to the specified arguments, defining the model architecture. Instantiating a configuration
with the defaults will yield a similar configuration to that of the Mixtral-7B-v0.1 or Mixtral-7B-Instruct-v0.1.
[mixtralai/Mixtral-8x7B](https://huggingface.co/mixtralai/Mixtral-8x7B)
[mixtralai/Mixtral-7B-Instruct-v0.1](https://huggingface.co/mixtralai/Mixtral-7B-Instruct-v0.1)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 32000):
Vocabulary size of the Mixtral model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`MixtralModel`]
hidden_size (`int`, *optional*, defaults to 4096):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to 14336):
Dimension of the MLP representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer encoder.
num_key_value_heads (`int`, *optional*, defaults to 8):
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
by meanpooling all the original heads within that group. For more details checkout [this
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
max_position_embeddings (`int`, *optional*, defaults to `4096*32`):
The maximum sequence length that this model might ever be used with. Mixtral's sliding window attention
allows sequence of up to 4096*32 tokens.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
rms_norm_eps (`float`, *optional*, defaults to 1e-05):
The epsilon used by the rms normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
pad_token_id (`int`, *optional*):
The id of the padding token.
bos_token_id (`int`, *optional*, defaults to 1):
The id of the "beginning-of-sequence" token.
eos_token_id (`int`, *optional*, defaults to 2):
The id of the "end-of-sequence" token.
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether the model's input and output word embeddings should be tied.
rope_theta (`float`, *optional*, defaults to 1000000.0):
The base period of the RoPE embeddings.
sliding_window (`int`, *optional*, defaults to 4096):
Sliding window attention window size. If not specified, will default to `4096`.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
num_experts_per_tok (`int`, *optional*, defaults to 2):
The number of experts to root per-token, can be also interpreted as the `top-p` routing
parameter
num_local_experts (`int`, *optional*, defaults to 8):
Number of experts per Sparse MLP layer.
output_router_logits (`bool`, *optional*, defaults to `False`):
Whether or not the router logits should be returned by the model. Enabeling this will also
allow the model to output the auxiliary loss. See [here]() for more details
router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
The aux loss factor for the total loss.
```python
>>> from transformers import MixtralModel, MixtralConfig
>>> # Initializing a Mixtral 7B style configuration
>>> configuration = MixtralConfig()
>>> # Initializing a model from the Mixtral 7B style configuration
>>> model = MixtralModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "mixtral"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
vocab_size=32000,
hidden_size=4096,
intermediate_size=14336,
num_hidden_layers=32,
num_attention_heads=32,
num_key_value_heads=8,
hidden_act="silu",
max_position_embeddings=4096 * 32,
initializer_range=0.02,
rms_norm_eps=1e-5,
use_cache=True,
pad_token_id=None,
bos_token_id=1,
eos_token_id=2,
tie_word_embeddings=False,
rope_theta=1e6,
sliding_window=4096,
attention_dropout=0.0,
num_experts_per_tok=2,
num_local_experts=8,
output_router_logits=False,
router_aux_loss_coef=0.001,
**kwargs,
):
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.sliding_window = sliding_window
# for backward compatibility
if num_key_value_heads is None:
num_key_value_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.rope_theta = rope_theta
self.attention_dropout = attention_dropout
self.num_experts_per_tok = num_experts_per_tok
self.num_local_experts = num_local_experts
self.output_router_logits = output_router_logits
self.router_aux_loss_coef = router_aux_loss_coef
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
src/transformers/models/mixtral/convert_mixtral_weights_to_hf.py 0 → 100644
View file @ accccdd0
# Copyright 2023 Mistral AI and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import json
import os
import torch
from transformers import (
MixtralConfig,
MixtralForCausalLM,
)
"""
Sample usage:
```
python src/transformers/models/mixtral/convert_mixtral_weights_to_hf.py \
--input_dir /path/to/downloaded/mixtral/weights --model_size 7B --output_dir /output/path
```
Thereafter, models can be loaded via:
```py
from transformers import MixtralForCausalLM
model = MixtralForCausalLM.from_pretrained("/output/path")
```
Important note: you need to be able to host the whole model in RAM to execute this script (even if the biggest versions
come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM).
"""
def compute_intermediate_size(n, ffn_dim_multiplier=1, multiple_of=256):
return multiple_of * ((int(ffn_dim_multiplier * int(8 * n / 3)) + multiple_of - 1) // multiple_of)
def read_json(path):
with open(path, "r") as f:
return json.load(f)
def write_json(text, path):
with open(path, "w") as f:
json.dump(text, f)
def write_model(model_path, input_base_path, model_size, safe_serialization=True):
os.makedirs(model_path, exist_ok=True)
params = read_json(os.path.join(input_base_path, "params.json"))
num_shards = 1
# For some reason this is a string in the params.json
sliding_window = int(params["sliding_window"])
n_layers = params["num_hidden_layers"]
n_heads = params["num_attention_heads"]
n_heads_per_shard = n_heads // num_shards
dim = params["hidden_size"]
dims_per_head = dim // n_heads
base = params.get("rope_theta", 10000.0)
max_position_embeddings = 4096 * 8
num_local_experts = params["num_local_experts"]
ffn_dim = params["intermediate_size"]
vocab_size = params["vocab_size"]
if "num_key_value_heads" in params:
num_key_value_heads = params["num_key_value_heads"] # for GQA / MQA
num_local_key_value_heads = num_key_value_heads // num_shards
key_value_dim = dims_per_head * num_local_key_value_heads
else: # compatibility with other checkpoints
num_key_value_heads = n_heads
num_local_key_value_heads = n_heads_per_shard
key_value_dim = dim
# permute for sliced rotary
def permute(w, n_heads=n_heads, dim1=dim, dim2=dim):
return w.view(n_heads, dim1 // n_heads // 2, 2, dim2).transpose(1, 2).reshape(dim1, dim2)
print(f"Fetching all parameters from the checkpoint at {input_base_path}.")
# Load weights
loaded = [
torch.load(os.path.join(input_base_path, f"consolidated.{i:02d}.pt"), map_location="cpu") for i in range(8)
]
merged_state_dict = {}
for state_dict in loaded:
merged_state_dict.update(state_dict)
state_dict = {}
for layer_i in range(n_layers):
# Sharded
# Note that attention.w{q,k,v,o}, feed_fordward.w[1,2,3], attention_norm.weight and ffn_norm.weight share
# the same storage object, saving attention_norm and ffn_norm will save other weights too, which is
# redundant as other weights will be stitched from multiple shards. To avoid that, they are cloned.
state_dict.update(
{
f"model.layers.{layer_i}.input_layernorm.weight": merged_state_dict[
f"layers.{layer_i}.attention_norm.weight"
].clone(),
f"model.layers.{layer_i}.post_attention_layernorm.weight": merged_state_dict[
f"layers.{layer_i}.ffn_norm.weight"
].clone(),
}
)
state_dict[f"model.layers.{layer_i}.self_attn.q_proj.weight"] = permute(
merged_state_dict[f"layers.{layer_i}.attention.wq.weight"]
.view(n_heads_per_shard, dims_per_head, dim)
.reshape(dim, dim)
)
state_dict[f"model.layers.{layer_i}.self_attn.k_proj.weight"] = permute(
merged_state_dict[f"layers.{layer_i}.attention.wk.weight"]
.view(num_local_key_value_heads, dims_per_head, dim)
.reshape(key_value_dim, dim),
num_key_value_heads,
key_value_dim,
dim,
)
state_dict[f"model.layers.{layer_i}.self_attn.v_proj.weight"] = (
merged_state_dict[f"layers.{layer_i}.attention.wv.weight"]
.view(num_local_key_value_heads, dims_per_head, dim)
.reshape(key_value_dim, dim)
)
state_dict[f"model.layers.{layer_i}.self_attn.o_proj.weight"] = merged_state_dict[
f"layers.{layer_i}.attention.wo.weight"
]
w1 = merged_state_dict[f"layers.{layer_i}.block_sparse_moe.w1"]
w2 = merged_state_dict[f"layers.{layer_i}.block_sparse_moe.w2"]
w3 = merged_state_dict[f"layers.{layer_i}.block_sparse_moe.w3"]
experts_w1 = [
w1[ffn_dim * expert_idx : ffn_dim * (expert_idx + 1), :].contiguous().clone()
for expert_idx in range(num_local_experts)
]
for idx, expert_block in enumerate(experts_w1):
expert_key = f"model.layers.{layer_i}.block_sparse_moe.experts.{idx}.w1"
state_dict[expert_key + ".weight"] = expert_block.clone()
experts_w2 = [
w2[ffn_dim * expert_idx : ffn_dim * (expert_idx + 1), :].contiguous().clone()
for expert_idx in range(num_local_experts)
]
for idx, expert_block in enumerate(experts_w2):
expert_key = f"model.layers.{layer_i}.block_sparse_moe.experts.{idx}.w2"
state_dict[expert_key + ".weight"] = expert_block.T.clone().contiguous()
experts_w3 = [
w3[ffn_dim * expert_idx : ffn_dim * (expert_idx + 1), :].contiguous().clone()
for expert_idx in range(num_local_experts)
]
for idx, expert_block in enumerate(experts_w3):
expert_key = f"model.layers.{layer_i}.block_sparse_moe.experts.{idx}.w3"
state_dict[expert_key + ".weight"] = expert_block.clone()
state_dict[f"model.layers.{layer_i}.block_sparse_moe.gate.weight"] = merged_state_dict[
f"layers.{layer_i}.block_sparse_moe.gate.weight"
]
state_dict.update(
{
"model.norm.weight": merged_state_dict["norm.weight"],
"model.embed_tokens.weight": merged_state_dict["tok_embeddings.weight"],
"lm_head.weight": merged_state_dict["output.weight"],
}
)
config = MixtralConfig(
hidden_size=dim,
intermediate_size=ffn_dim,
num_attention_heads=params["num_attention_heads"],
num_hidden_layers=params["num_hidden_layers"],
rms_norm_eps=params["rms_norm_eps"],
num_key_value_heads=num_key_value_heads,
vocab_size=vocab_size,
rope_theta=base,
max_position_embeddings=max_position_embeddings,
sliding_window=sliding_window,
num_local_experts=num_local_experts,
)
print("Loading the checkpoint in a Mixtral model.")
with torch.device("meta"):
model = MixtralForCausalLM(config)
# Avoid saving this as part of the config.
del model.config._name_or_path
model.config.torch_dtype = torch.float16
print("Saving in the Transformers format.")
model.load_state_dict(state_dict, strict=True, assign=True)
for n, p in model.named_parameters():
assert p.device.type != "meta", f"{n} has not been loaded!"
model.save_pretrained(model_path, safe_serialization=safe_serialization)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"--input_dir",
help="Location of Mixtral weights, which contains tokenizer.model and model folders",
required=True,
)
parser.add_argument(
"--model_size",
choices=["7B"],
help="'f' models correspond to the finetuned versions, and are specific to the Mixtral official release. For more details on Mixtral, checkout the original repo: https://huggingface.co/mistral-ai",
default="7B",
)
parser.add_argument("--output_dir", help="Location to write HF model", required=True)
parser.add_argument("--safe_serialization", type=bool, help="Whether or not to save using `safetensors`.")
args = parser.parse_args()
write_model(
model_path=args.output_dir,
input_base_path=args.input_dir,
model_size=args.model_size,
safe_serialization=args.safe_serialization,
)
if __name__ == "__main__":
main()
src/transformers/models/mixtral/modeling_mixtral.py 0 → 100644
View file @ accccdd0
# coding=utf-8
# Copyright 2023 Mistral AI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch Mixtral model."""
import inspect
import math
import warnings
from typing import List, Optional, Tuple, Union
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...modeling_attn_mask_utils import (
_prepare_4d_causal_attention_mask,
)
from ...modeling_outputs import (
MoeCausalLMOutputWithPast,
MoeModelOutputWithPast,
SequenceClassifierOutputWithPast,
)
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import is_torch_greater_or_equal_than_1_13
from ...utils import (
add_start_docstrings,
add_start_docstrings_to_model_forward,
is_flash_attn_2_available,
is_flash_attn_greater_or_equal_2_10,
logging,
replace_return_docstrings,
)
from ...utils.import_utils import is_torch_fx_available
from .configuration_mixtral import MixtralConfig
if is_flash_attn_2_available():
from flash_attn import flash_attn_func, flash_attn_varlen_func
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
_flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
# It means that the function will not be traced through and simply appear as a node in the graph.
if is_torch_fx_available():
if not is_torch_greater_or_equal_than_1_13:
import torch.fx
_prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
logger = logging.get_logger(__name__)
_CONFIG_FOR_DOC = "MixtralConfig"
def load_balancing_loss_func(gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2) -> float:
r"""
Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
experts is too unbalanced.
Args:
gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
Logits from the `gate`, should be a tuple of tensors. Shape: [batch_size, seqeunce_length, num_experts].
num_experts (`int`, *optional*):
Number of experts
Returns:
The auxiliary loss.
"""
if gate_logits is None:
return 0
if isinstance(gate_logits, tuple):
# cat along the layers?
gate_logits = torch.cat(gate_logits, dim=0)
routing_weights, selected_experts = torch.topk(gate_logits, top_k, dim=-1)
routing_weights = routing_weights.softmax(dim=-1)
# cast the expert indices to int64, otherwise one-hot encoding will fail
if selected_experts.dtype != torch.int64:
selected_experts = selected_experts.to(torch.int64)
if len(selected_experts.shape) == 2:
selected_experts = selected_experts.unsqueeze(2)
expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
# For a given token, determine if it was routed to a given expert.
expert_mask = torch.max(expert_mask, axis=-2).values
# cast to float32 otherwise mean will fail
expert_mask = expert_mask.to(torch.float32)
tokens_per_group_and_expert = torch.mean(expert_mask, axis=-2)
router_prob_per_group_and_expert = torch.mean(routing_weights, axis=-1)
return torch.mean(tokens_per_group_and_expert * router_prob_per_group_and_expert.unsqueeze(-1)) * (num_experts**2)
# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
max_seqlen_in_batch = seqlens_in_batch.max().item()
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
return (
indices,
cu_seqlens,
max_seqlen_in_batch,
)
# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Mixtral
class MixtralRMSNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-6):
"""
MixtralRMSNorm is equivalent to T5LayerNorm
"""
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = eps
def forward(self, hidden_states):
input_dtype = hidden_states.dtype
hidden_states = hidden_states.to(torch.float32)
variance = hidden_states.pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
return self.weight * hidden_states.to(input_dtype)
# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Mixtral
class MixtralRotaryEmbedding(nn.Module):
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
super().__init__()
self.dim = dim
self.max_position_embeddings = max_position_embeddings
self.base = base
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
# Build here to make `torch.jit.trace` work.
self._set_cos_sin_cache(
seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
)
def _set_cos_sin_cache(self, seq_len, device, dtype):
self.max_seq_len_cached = seq_len
t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
freqs = torch.outer(t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
if seq_len > self.max_seq_len_cached:
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
return (
self.cos_cached[:seq_len].to(dtype=x.dtype),
self.sin_cached[:seq_len].to(dtype=x.dtype),
)
# Copied from transformers.models.llama.modeling_llama.rotate_half
def rotate_half(x):
"""Rotates half the hidden dims of the input."""
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
"""Applies Rotary Position Embedding to the query and key tensors.
Args:
q (`torch.Tensor`): The query tensor.
k (`torch.Tensor`): The key tensor.
cos (`torch.Tensor`): The cosine part of the rotary embedding.
sin (`torch.Tensor`): The sine part of the rotary embedding.
position_ids (`torch.Tensor`):
The position indices of the tokens corresponding to the query and key tensors. For example, this can be
used to pass offsetted position ids when working with a KV-cache.
unsqueeze_dim (`int`, *optional*, defaults to 1):
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
Returns:
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
"""
cos = cos[position_ids].unsqueeze(unsqueeze_dim)
sin = sin[position_ids].unsqueeze(unsqueeze_dim)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
# Copied from transformers.models.llama.modeling_llama.repeat_kv
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
"""
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
"""
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
if n_rep == 1:
return hidden_states
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
# Copied from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->Mixtral
class MixtralAttention(nn.Module):
"""
Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
and "Generating Long Sequences with Sparse Transformers".
"""
def __init__(self, config: MixtralConfig, layer_idx: Optional[int] = None):
super().__init__()
self.config = config
self.layer_idx = layer_idx
if layer_idx is None:
logger.warning_once(
f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
"to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
"when creating this class."
)
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.hidden_size // self.num_heads
self.num_key_value_heads = config.num_key_value_heads
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
self.max_position_embeddings = config.max_position_embeddings
self.rope_theta = config.rope_theta
self.is_causal = True
self.attention_dropout = config.attention_dropout
if (self.head_dim * self.num_heads) != self.hidden_size:
raise ValueError(
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
f" and `num_heads`: {self.num_heads})."
)
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
self.rotary_emb = MixtralRotaryEmbedding(
self.head_dim,
max_position_embeddings=self.max_position_embeddings,
base=self.rope_theta,
)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Cache] = None,
output_attentions: bool = False,
use_cache: bool = False,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
if "padding_mask" in kwargs:
warnings.warn(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
kv_seq_len = key_states.shape[-2]
if past_key_value is not None:
if self.layer_idx is None:
raise ValueError(
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
"with a layer index."
)
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
attn_output = torch.matmul(attn_weights, value_states)
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
# Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Mixtral
class MixtralFlashAttention2(MixtralAttention):
"""
Mixtral flash attention module. This module inherits from `MixtralAttention` as the weights of the module stays
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
flash attention and deal with padding tokens in case the input contains any of them.
"""
# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Cache] = None,
output_attentions: bool = False,
use_cache: bool = False,
**kwargs,
):
if "padding_mask" in kwargs:
warnings.warn(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
# overwrite attention_mask with padding_mask
attention_mask = kwargs.pop("padding_mask")
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
kv_seq_len = key_states.shape[-2]
if past_key_value is not None:
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
# Because the input can be padded, the absolute sequence length depends on the max position id.
rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
use_sliding_windows = (
_flash_supports_window_size
and getattr(self.config, "sliding_window", None) is not None
and kv_seq_len > self.config.sliding_window
)
if not _flash_supports_window_size:
logger.warning_once(
"The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
" make sure to upgrade flash-attn library."
)
if past_key_value is not None:
# Activate slicing cache only if the config has a value `sliding_windows` attribute
if getattr(self.config, "sliding_window", None) is not None and kv_seq_len > self.config.sliding_window:
slicing_tokens = 1 - self.config.sliding_window
past_key = past_key_value[0]
past_value = past_key_value[1]
past_key = past_key[:, :, slicing_tokens:, :].contiguous()
past_value = past_value[:, :, slicing_tokens:, :].contiguous()
if past_key.shape[-2] != self.config.sliding_window - 1:
raise ValueError(
f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
f" {past_key.shape}"
)
past_key_value = (past_key, past_value)
if attention_mask is not None:
attention_mask = attention_mask[:, slicing_tokens:]
attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
dropout_rate = 0.0 if not self.training else self.attention_dropout
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
# therefore the input hidden states gets silently casted in float32. Hence, we need
# cast them back in float16 just to be sure everything works as expected.
input_dtype = query_states.dtype
if input_dtype == torch.float32:
# Handle the case where the model is quantized
if hasattr(self.config, "_pre_quantization_dtype"):
target_dtype = self.config._pre_quantization_dtype
else:
target_dtype = self.q_proj.weight.dtype
logger.warning_once(
f"The input hidden states seems to be silently casted in float32, this might be related to"
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
f" {target_dtype}."
)
query_states = query_states.to(target_dtype)
key_states = key_states.to(target_dtype)
value_states = value_states.to(target_dtype)
# Reashape to the expected shape for Flash Attention
query_states = query_states.transpose(1, 2)
key_states = key_states.transpose(1, 2)
value_states = value_states.transpose(1, 2)
attn_output = self._flash_attention_forward(
query_states,
key_states,
value_states,
attention_mask,
q_len,
dropout=dropout_rate,
use_sliding_windows=use_sliding_windows,
)
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
attn_output = self.o_proj(attn_output)
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
def _flash_attention_forward(
self,
query_states,
key_states,
value_states,
attention_mask,
query_length,
dropout=0.0,
softmax_scale=None,
use_sliding_windows=False,
):
"""
Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
first unpad the input, then computes the attention scores and pad the final attention scores.
Args:
query_states (`torch.Tensor`):
Input query states to be passed to Flash Attention API
key_states (`torch.Tensor`):
Input key states to be passed to Flash Attention API
value_states (`torch.Tensor`):
Input value states to be passed to Flash Attention API
attention_mask (`torch.Tensor`):
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
position of padding tokens and 1 for the position of non-padding tokens.
dropout (`int`, *optional*):
Attention dropout
softmax_scale (`float`, *optional*):
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
use_sliding_windows (`bool`, *optional*):
Whether to activate sliding window attention.
"""
if not self._flash_attn_uses_top_left_mask:
causal = self.is_causal
else:
# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
causal = self.is_causal and query_length != 1
# Contains at least one padding token in the sequence
if attention_mask is not None:
batch_size = query_states.shape[0]
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
query_states, key_states, value_states, attention_mask, query_length
)
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
if not use_sliding_windows:
attn_output_unpad = flash_attn_varlen_func(
query_states,
key_states,
value_states,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
max_seqlen_q=max_seqlen_in_batch_q,
max_seqlen_k=max_seqlen_in_batch_k,
dropout_p=dropout,
softmax_scale=softmax_scale,
causal=causal,
)
else:
attn_output_unpad = flash_attn_varlen_func(
query_states,
key_states,
value_states,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
max_seqlen_q=max_seqlen_in_batch_q,
max_seqlen_k=max_seqlen_in_batch_k,
dropout_p=dropout,
softmax_scale=softmax_scale,
causal=causal,
window_size=(self.config.sliding_window, self.config.sliding_window),
)
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
else:
if not use_sliding_windows:
attn_output = flash_attn_func(
query_states,
key_states,
value_states,
dropout,
softmax_scale=softmax_scale,
causal=causal,
)
else:
attn_output = flash_attn_func(
query_states,
key_states,
value_states,
dropout,
softmax_scale=softmax_scale,
causal=causal,
window_size=(self.config.sliding_window, self.config.sliding_window),
)
return attn_output
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
# On the first iteration we need to properly re-create the padding mask
# by slicing it on the proper place
if kv_seq_len != attention_mask.shape[-1]:
attention_mask_num_tokens = attention_mask.shape[-1]
attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
if query_length == kv_seq_len:
query_layer = index_first_axis(
query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
)
cu_seqlens_q = cu_seqlens_k
max_seqlen_in_batch_q = max_seqlen_in_batch_k
indices_q = indices_k
elif query_length == 1:
max_seqlen_in_batch_q = 1
cu_seqlens_q = torch.arange(
batch_size + 1, dtype=torch.int32, device=query_layer.device
) # There is a memcpy here, that is very bad.
indices_q = cu_seqlens_q[:-1]
query_layer = query_layer.squeeze(1)
else:
# The -q_len: slice assumes left padding.
attention_mask = attention_mask[:, -query_length:]
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
return (
query_layer,
key_layer,
value_layer,
indices_q,
(cu_seqlens_q, cu_seqlens_k),
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
)
class MixtralBLockSparseTop2MLP(nn.Module):
def __init__(self, config: MixtralConfig):
super().__init__()
self.ffn_dim = config.intermediate_size
self.hidden_dim = config.hidden_size
self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
self.act_fn = ACT2FN[config.hidden_act]
def forward(self, hidden_states, routing_weights):
current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
current_hidden_states = self.w2(current_hidden_states)
return routing_weights * current_hidden_states
MISTRAL_ATTENTION_CLASSES = {
"eager": MixtralAttention,
"flash_attention_2": MixtralFlashAttention2,
}
class MixtralSparseMoeBlock(nn.Module):
"""
This implementation is
strictly equivalent to standard MoE with full capacity (no
dropped tokens). It's faster since it formulates MoE operations
in terms of block-sparse operations to accomodate imbalanced
assignments of tokens to experts, whereas standard MoE either
(1) drop tokens at the cost of reduced performance or (2) set
capacity factor to number of experts and thus waste computation
and memory on padding.
"""
def __init__(self, config):
super().__init__()
self.hidden_dim = config.hidden_size
self.ffn_dim = config.intermediate_size
self.num_experts = config.num_local_experts
self.top_k = config.num_experts_per_tok
# gating
self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
self.experts = nn.ModuleList([MixtralBLockSparseTop2MLP(config) for _ in range(self.num_experts)])
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
""" """
batch_size, sequence_length, hidden_dim = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_dim)
# router_logits: (batch * sequence_length, n_experts)
router_logits = self.gate(hidden_states)
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
# we cast back to the input dtype
routing_weights = routing_weights.to(hidden_states.dtype)
final_hidden_states = torch.zeros(
(batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
)
# One hot encode the selected experts to create an expert mask
# this will be used to easily index which expert is going to be sollicitated
expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
# Loop over all available experts in the model and perform the computation on each expert
for expert_idx in range(self.num_experts):
expert_layer = self.experts[expert_idx]
idx, top_x = torch.where(expert_mask[expert_idx])
if top_x.shape[0] == 0:
continue
# in torch it is faster to index using lists than torch tensors
top_x_list = top_x.tolist()
idx_list = idx.tolist()
# Index the correct hidden states and compute the expert hidden state for
# the current expert. We need to make sure to multiply the output hidden
# states by `routing_weights` on the corresponding tokens (top-1 and top-2)
current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
current_hidden_states = expert_layer(current_state, routing_weights[top_x_list, idx_list, None])
# However `index_add_` only support torch tensors for indexing so we'll use
# the `top_x` tensor here.
final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
return final_hidden_states, router_logits
class MixtralDecoderLayer(nn.Module):
def __init__(self, config: MixtralConfig, layer_idx: int):
super().__init__()
self.hidden_size = config.hidden_size
self.self_attn = MISTRAL_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
self.block_sparse_moe = MixtralSparseMoeBlock(config)
self.input_layernorm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.post_attention_layernorm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
output_attentions: Optional[bool] = False,
output_router_logits: Optional[bool] = False,
use_cache: Optional[bool] = False,
**kwargs,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
if "padding_mask" in kwargs:
warnings.warn(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
"""
Args:
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
`(batch, sequence_length)` where padding elements are indicated by 0.
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
returned tensors for more detail.
output_router_logits (`bool`, *optional*):
Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
should not be returned during inference.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
(see `past_key_values`).
"""
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
# Self Attention
hidden_states, self_attn_weights, present_key_value = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
)
hidden_states = residual + hidden_states
# Fully Connected
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states, router_logits = self.block_sparse_moe(hidden_states)
hidden_states = residual + hidden_states
outputs = (hidden_states,)
if output_attentions:
outputs += (self_attn_weights,)
if use_cache:
outputs += (present_key_value,)
if output_router_logits:
outputs += (router_logits,)
return outputs
MIXTRAL_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`MixtralConfig`]):
Model configuration class with all the parameters of the model. Initializing with a config file does not
load the weights associated with the model, only the configuration. Check out the
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
@add_start_docstrings(
"The bare Mixtral Model outputting raw hidden-states without any specific head on top.",
MIXTRAL_START_DOCSTRING,
)
# Copied from transformers.models.mistral.modeling_mistral.MistralPreTrainedModel with Mistral->Mixtral
class MixtralPreTrainedModel(PreTrainedModel):
config_class = MixtralConfig
base_model_prefix = "model"
supports_gradient_checkpointing = True
_no_split_modules = ["MixtralDecoderLayer"]
_skip_keys_device_placement = "past_key_values"
_supports_flash_attn_2 = True
_supports_cache_class = True
def _init_weights(self, module):
std = self.config.initializer_range
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
MIXTRAL_INPUTS_DOCSTRING = r"""
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
it.
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
[What are attention masks?](../glossary#attention-mask)
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
[`PreTrainedTokenizer.__call__`] for details.
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
`past_key_values`).
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
information on the default strategy.
- 1 indicates the head is **not masked**,
- 0 indicates the head is **masked**.
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
config.n_positions - 1]`.
[What are position IDs?](../glossary#position-ids)
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
model's internal embedding lookup matrix.
use_cache (`bool`, *optional*):
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
`past_key_values`).
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
output_router_logits (`bool`, *optional*):
Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
should not be returned during inference.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
@add_start_docstrings(
"The bare Mixtral Model outputting raw hidden-states without any specific head on top.",
MIXTRAL_START_DOCSTRING,
)
# Copied from transformers.models.mistral.modeling_mistral.MistralModel with MISTRAL->MIXTRAL,Mistral->Mixtral
class MixtralModel(MixtralPreTrainedModel):
"""
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MixtralDecoderLayer`]
Args:
config: MixtralConfig
"""
def __init__(self, config: MixtralConfig):
super().__init__(config)
self.padding_idx = config.pad_token_id
self.vocab_size = config.vocab_size
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
self.layers = nn.ModuleList(
[MixtralDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
)
self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
self.norm = MixtralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
self.gradient_checkpointing = False
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.embed_tokens
def set_input_embeddings(self, value):
self.embed_tokens = value
# Ignore copy
@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
output_router_logits: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, MoeModelOutputWithPast]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_router_logits = (
output_router_logits if output_router_logits is not None else self.config.output_router_logits
)
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
# retrieve input_ids and inputs_embeds
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")
past_key_values_length = 0
if use_cache:
use_legacy_cache = not isinstance(past_key_values, Cache)
if use_legacy_cache:
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
past_key_values_length = past_key_values.get_usable_length(seq_length)
if position_ids is None:
device = input_ids.device if input_ids is not None else inputs_embeds.device
position_ids = torch.arange(
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
)
position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
else:
position_ids = position_ids.view(-1, seq_length).long()
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
if attention_mask is not None and self._use_flash_attention_2 and use_cache:
is_padding_right = attention_mask[:, -1].sum().item() != batch_size
if is_padding_right:
raise ValueError(
"You are attempting to perform batched generation with padding_side='right'"
" this may lead to unexpected behaviour for Flash Attention version of Mixtral. Make sure to "
" call `tokenizer.padding_side = 'left'` before tokenizing the input. "
)
if self._use_flash_attention_2:
# 2d mask is passed through the layers
attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
else:
# 4d mask is passed through the layers
attention_mask = _prepare_4d_causal_attention_mask(
attention_mask,
(batch_size, seq_length),
inputs_embeds,
past_key_values_length,
sliding_window=self.config.sliding_window,
)
hidden_states = inputs_embeds
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
)
use_cache = False
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
all_router_logits = () if output_router_logits else None
next_decoder_cache = None
for decoder_layer in self.layers:
if output_hidden_states:
all_hidden_states += (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
attention_mask,
position_ids,
past_key_values,
output_attentions,
output_router_logits,
use_cache,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_values,
output_attentions=output_attentions,
output_router_logits=output_router_logits,
use_cache=use_cache,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
if output_attentions:
all_self_attns += (layer_outputs[1],)
if output_router_logits:
all_router_logits += (layer_outputs[-1],)
hidden_states = self.norm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
next_cache = None
if use_cache:
next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
if not return_dict:
return tuple(
v
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
if v is not None
)
return MoeModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=next_cache,
hidden_states=all_hidden_states,
attentions=all_self_attns,
router_logits=all_router_logits,
)
class MixtralForCausalLM(MixtralPreTrainedModel):
_tied_weights_keys = ["lm_head.weight"]
def __init__(self, config):
super().__init__(config)
self.model = MixtralModel(config)
self.vocab_size = config.vocab_size
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
self.router_aux_loss_coef = config.router_aux_loss_coef
self.num_experts = config.num_local_experts
self.num_experts_per_tok = config.num_experts_per_tok
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.model.embed_tokens
def set_input_embeddings(self, value):
self.model.embed_tokens = value
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
def set_decoder(self, decoder):
self.model = decoder
def get_decoder(self):
return self.model
@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
# Ignore copy
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = 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,
output_router_logits: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
r"""
Args:
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Returns:
Example:
```python
>>> from transformers import AutoTokenizer, MixtralForCausalLM
>>> model = MixtralForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> # Generate
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
```"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_router_logits = (
output_router_logits if output_router_logits is not None else self.config.output_router_logits
)
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_router_logits=output_router_logits,
return_dict=return_dict,
)
hidden_states = outputs[0]
logits = self.lm_head(hidden_states)
logits = logits.float()
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
shift_logits = shift_logits.view(-1, self.config.vocab_size)
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
loss = loss_fct(shift_logits, shift_labels)
aux_loss = None
if output_router_logits:
aux_loss = load_balancing_loss_func(
outputs.router_logits if return_dict else outputs[-1], self.num_experts, self.num_experts_per_tok
)
if labels is not None:
loss += self.router_aux_loss_coef * aux_loss
if not return_dict:
output = (logits,) + outputs[1:]
if output_router_logits:
output = (aux_loss,) + output
return (loss,) + output if loss is not None else output
return MoeCausalLMOutputWithPast(
loss=loss,
aux_loss=aux_loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
router_logits=outputs.router_logits,
)
def prepare_inputs_for_generation(
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
):
# Omit tokens covered by past_key_values
if past_key_values is not None:
if isinstance(past_key_values, Cache):
cache_length = past_key_values.get_seq_length()
past_length = past_key_values.seen_tokens
max_cache_length = past_key_values.get_max_length()
else:
cache_length = past_length = past_key_values[0][0].shape[2]
max_cache_length = None
# Keep only the unprocessed tokens:
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
# some of the inputs are exclusivelly passed as part of the cache (e.g. when passing input_embeds as
# input)
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
# input_ids based on the past_length.
elif past_length < input_ids.shape[1]:
input_ids = input_ids[:, past_length:]
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
if (
max_cache_length is not None
and attention_mask is not None
and cache_length + input_ids.shape[1] > max_cache_length
):
attention_mask = attention_mask[:, -max_cache_length:]
position_ids = kwargs.get("position_ids", None)
if attention_mask is not None and position_ids is None:
# create position_ids on the fly for batch generation
position_ids = attention_mask.long().cumsum(-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1)
if past_key_values:
position_ids = position_ids[:, -input_ids.shape[1] :]
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
if inputs_embeds is not None and past_key_values is None:
model_inputs = {"inputs_embeds": inputs_embeds}
else:
model_inputs = {"input_ids": input_ids}
model_inputs.update(
{
"position_ids": position_ids,
"past_key_values": past_key_values,
"use_cache": kwargs.get("use_cache"),
"attention_mask": attention_mask,
}
)
return model_inputs
@staticmethod
def _reorder_cache(past_key_values, beam_idx):
reordered_past = ()
for layer_past in past_key_values:
reordered_past += (
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
)
return reordered_past
@add_start_docstrings(
"""
The Mixtral Model transformer with a sequence classification head on top (linear layer).
[`MixtralForSequenceClassification`] uses the last token in order to do the classification, as other causal models
(e.g. GPT-2) do.
Since it does classification on the last token, it requires to know the position of the last token. If a
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
each row of the batch).
""",
MIXTRAL_START_DOCSTRING,
)
# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Mixtral, LLAMA->MIXTRAL
class MixtralForSequenceClassification(MixtralPreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.num_labels = config.num_labels
self.model = MixtralModel(config)
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.model.embed_tokens
def set_input_embeddings(self, value):
self.model.embed_tokens = value
@add_start_docstrings_to_model_forward(MIXTRAL_INPUTS_DOCSTRING)
def forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = 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,
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
r"""
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
transformer_outputs = self.model(
input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
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]
logits = self.score(hidden_states)
if input_ids is not None:
batch_size = input_ids.shape[0]
else:
batch_size = inputs_embeds.shape[0]
if self.config.pad_token_id is None and batch_size != 1:
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
if self.config.pad_token_id is None:
sequence_lengths = -1
else:
if input_ids is not None:
sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
logits.device
)
else:
sequence_lengths = -1
pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
loss = None
if labels is not None:
labels = labels.to(logits.device)
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
loss_fct = MSELoss()
if self.num_labels == 1:
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
else:
loss = loss_fct(pooled_logits, labels)
elif self.config.problem_type == "single_label_classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss_fct = BCEWithLogitsLoss()
loss = loss_fct(pooled_logits, labels)
if not return_dict:
output = (pooled_logits,) + transformer_outputs[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutputWithPast(
loss=loss,
logits=pooled_logits,
past_key_values=transformer_outputs.past_key_values,
hidden_states=transformer_outputs.hidden_states,
attentions=transformer_outputs.attentions,
)
src/transformers/utils/dummy_pt_objects.py
View file @ accccdd0
......@@ -5262,6 +5262,34 @@ class MistralPreTrainedModel(metaclass=DummyObject):
requires_backends(self, ["torch"])
class MixtralForCausalLM(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class MixtralForSequenceClassification(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class MixtralModel(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class MixtralPreTrainedModel(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST = None
......
tests/models/mixtral/test_modeling_mixtral.py 0 → 100644
View file @ accccdd0
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Testing suite for the PyTorch Mixtral model. """
import tempfile
import unittest
import pytest
from transformers import MixtralConfig, is_torch_available
from transformers.testing_utils import (
require_flash_attn,
require_torch,
require_torch_gpu,
slow,
torch_device,
)
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, ids_tensor
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import (
MixtralForCausalLM,
MixtralForSequenceClassification,
MixtralModel,
)
class MixtralModelTester:
# Copied from tests.models.mistral.test_modeling_mistral.MistralModelTester.__init__
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_input_mask=True,
use_token_type_ids=False,
use_labels=True,
vocab_size=99,
hidden_size=32,
num_hidden_layers=2,
num_attention_heads=4,
num_key_value_heads=2,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
num_choices=4,
pad_token_id=0,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_token_type_ids = use_token_type_ids
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.num_choices = num_choices
self.pad_token_id = pad_token_id
self.scope = scope
# Copied from tests.models.mistral.test_modeling_mistral.MistralModelTester.prepare_config_and_inputs
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = torch.tril(torch.ones(self.batch_size, self.seq_length)).to(torch_device)
token_type_ids = None
if self.use_token_type_ids:
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
sequence_labels = None
token_labels = None
choice_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = self.get_config()
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
def get_config(self):
return MixtralConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
num_key_value_heads=self.num_key_value_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
is_decoder=False,
initializer_range=self.initializer_range,
pad_token_id=self.pad_token_id,
num_experts_per_tok=2,
num_local_experts=2,
)
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Mixtral
def create_and_check_model(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
model = MixtralModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Mixtral
def create_and_check_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = MixtralModel(config)
model.to(torch_device)
model.eval()
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
)
result = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
)
result = model(input_ids, attention_mask=input_mask)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Mixtral
def create_and_check_for_causal_lm(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
model = MixtralForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Mixtral
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.is_decoder = True
config.add_cross_attention = True
model = MixtralForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
outputs = model(
input_ids,
attention_mask=input_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common with Llama->Mixtral
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
return config, inputs_dict
@require_torch
# Copied from tests.models.mistral.test_modeling_mistral.MistralModelTest with Mistral->Mixtral
class MixtralModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(MixtralModel, MixtralForCausalLM, MixtralForSequenceClassification) if is_torch_available() else ()
)
all_generative_model_classes = (MixtralForCausalLM,) if is_torch_available() else ()
pipeline_model_mapping = (
{
"feature-extraction": MixtralModel,
"text-classification": MixtralForSequenceClassification,
"text-generation": MixtralForCausalLM,
"zero-shot": MixtralForSequenceClassification,
}
if is_torch_available()
else {}
)
test_headmasking = False
test_pruning = False
# TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79245/workflows/9490ef58-79c2-410d-8f51-e3495156cf9c/jobs/1012146
def is_pipeline_test_to_skip(
self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
):
return True
def setUp(self):
self.model_tester = MixtralModelTester(self)
self.config_tester = ConfigTester(self, config_class=MixtralConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_model_various_embeddings(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
for type in ["absolute", "relative_key", "relative_key_query"]:
config_and_inputs[0].position_embedding_type = type
self.model_tester.create_and_check_model(*config_and_inputs)
def test_Mixtral_sequence_classification_model(self):
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
print(config)
config.num_labels = 3
input_ids = input_dict["input_ids"]
attention_mask = input_ids.ne(1).to(torch_device)
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
model = MixtralForSequenceClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
def test_Mixtral_sequence_classification_model_for_single_label(self):
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.num_labels = 3
config.problem_type = "single_label_classification"
input_ids = input_dict["input_ids"]
attention_mask = input_ids.ne(1).to(torch_device)
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
model = MixtralForSequenceClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
def test_Mixtral_sequence_classification_model_for_multi_label(self):
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.num_labels = 3
config.problem_type = "multi_label_classification"
input_ids = input_dict["input_ids"]
attention_mask = input_ids.ne(1).to(torch_device)
sequence_labels = ids_tensor(
[self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
).to(torch.float)
model = MixtralForSequenceClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
@unittest.skip("Mixtral buffers include complex numbers, which breaks this test")
def test_save_load_fast_init_from_base(self):
pass
@unittest.skip("Mixtral uses GQA on all models so the KV cache is a non standard format")
def test_past_key_values_format(self):
pass
@require_flash_attn
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_generate_padding_right(self):
import torch
for model_class in self.all_generative_model_classes:
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, low_cpu_mem_usage=True).to(
torch_device
)
dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [1, 1, 1, 0]]).to(torch_device)
model.generate(dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False)
model = model_class.from_pretrained(
tmpdirname,
torch_dtype=torch.float16,
attn_implementation="flash_attention_2",
low_cpu_mem_usage=True,
).to(torch_device)
with self.assertRaises(ValueError):
_ = model.generate(
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
)
@require_flash_attn
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_generate_use_cache(self):
import torch
max_new_tokens = 30
for model_class in self.all_generative_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
dummy_input = inputs_dict[model_class.main_input_name]
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
dummy_input = dummy_input.to(torch.float16)
# make sure that all models have enough positions for generation
if hasattr(config, "max_position_embeddings"):
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
# NOTE: Mixtral apparently does not support right padding + use_cache with FA2.
dummy_attention_mask[:, -1] = 1
model = model_class.from_pretrained(
tmpdirname,
torch_dtype=torch.float16,
attn_implementation="flash_attention_2",
low_cpu_mem_usage=True,
).to(torch_device)
# Just test that a large cache works as expected
_ = model.generate(
dummy_input,
attention_mask=dummy_attention_mask,
max_new_tokens=max_new_tokens,
do_sample=False,
use_cache=True,
)
@require_flash_attn
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_inference_padding_right(self):
self.skipTest("Mixtral flash attention does not support right padding")
# Ignore copy
def test_load_balancing_loss(self):
r"""
Let's make sure we can actually compute the loss and do a backward on it.
"""
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.num_labels = 3
config.output_router_logits = True
input_ids = input_dict["input_ids"]
attention_mask = input_ids.ne(1).to(torch_device)
model = MixtralForCausalLM(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask)
self.assertEqual(result.router_logits[0].shape, (91, config.num_experts_per_tok))
torch.testing.assert_close(result.aux_loss.cpu(), torch.tensor(1, dtype=torch.float32))
@require_torch
class MixtralIntegrationTest(unittest.TestCase):
@slow
@require_torch_gpu
def test_small_model_logits(self):
model_id = "hf-internal-testing/Mixtral-tiny"
dummy_input = torch.LongTensor([[0, 1, 0], [0, 1, 0]]).to(torch_device)
model = MixtralForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True).to(
torch_device
)
# TODO: might need to tweak it in case the logits do not match on our daily runners
# these logits have been obtained with the original megablocks impelmentation.
EXPECTED_LOGITS = torch.Tensor(
[[0.1670, 0.1620, 0.6094], [-0.8906, -0.1588, -0.6060], [0.1572, 0.1290, 0.7246]]
).to(torch_device)
with torch.no_grad():
logits = model(dummy_input).logits
torch.testing.assert_close(logits[0, :3, :3].half(), EXPECTED_LOGITS, atol=1e-3, rtol=1e-3)
torch.testing.assert_close(logits[1, :3, :3].half(), EXPECTED_LOGITS, atol=1e-3, rtol=1e-3)
@slow
# @require_torch_gpu
def test_small_model_logits_batched(self):
model_id = "hf-internal-testing/Mixtral-tiny"
dummy_input = torch.LongTensor([[0, 0, 0, 0, 0, 0, 1, 2, 3], [1, 1, 2, 3, 4, 5, 6, 7, 8]]).to(torch_device)
attention_mask = dummy_input.ne(0).to(torch.long)
model = MixtralForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True).to(
torch_device
)
# TODO: might need to tweak it in case the logits do not match on our daily runners
EXPECTED_LOGITS_LEFT = torch.Tensor(
[[0.1750, 0.0537, 0.7007], [0.1750, 0.0537, 0.7007], [0.1750, 0.0537, 0.7007]],
)
# logits[0, -3:, -3:].half()
EXPECTED_LOGITS_LEFT_UNPADDED = torch.Tensor(
[[0.2212, 0.5200, -0.3816], [0.8213, -0.2313, 0.6069], [0.2664, -0.7090, 0.2468]],
)
# logits[1, -3:, -3:].half()
EXPECTED_LOGITS_RIGHT_UNPADDED = torch.Tensor(
[[0.2205, 0.1232, -0.1611], [-0.3484, 0.3030, -1.0312], [0.0742, 0.7930, 0.7969]]
)
with torch.no_grad():
logits = model(dummy_input, attention_mask=attention_mask).logits
torch.testing.assert_close(logits[0, :3, :3].half(), EXPECTED_LOGITS_LEFT, atol=1e-3, rtol=1e-3)
torch.testing.assert_close(logits[0, -3:, -3:].half(), EXPECTED_LOGITS_LEFT_UNPADDED, atol=1e-3, rtol=1e-3)
torch.testing.assert_close(logits[1, -3:, -3:].half(), EXPECTED_LOGITS_RIGHT_UNPADDED, atol=1e-3, rtol=1e-3)
utils/not_doctested.txt
View file @ accccdd0
......@@ -677,6 +677,8 @@ src/transformers/models/mgp_str/configuration_mgp_str.py
src/transformers/models/mgp_str/modeling_mgp_str.py
src/transformers/models/mistral/configuration_mistral.py
src/transformers/models/mistral/modeling_mistral.py
src/transformers/models/mixtral/configuration_mixtral.py
src/transformers/models/mixtral/modeling_mixtral.py
src/transformers/models/mluke/convert_mluke_original_pytorch_checkpoint_to_pytorch.py
src/transformers/models/mobilebert/convert_mobilebert_original_tf_checkpoint_to_pytorch.py
src/transformers/models/mobilenet_v1/configuration_mobilenet_v1.py
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment