diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index ae632376f9469c..0e96beedeae308 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -522,6 +522,8 @@ title: Phi - local: model_doc/phi3 title: Phi-3 + - local: model_doc/phimoe + title: PhiMoE - local: model_doc/phobert title: PhoBERT - local: model_doc/plbart diff --git a/docs/source/en/index.md b/docs/source/en/index.md index 0a5518fd71c840..dd22d58350be29 100644 --- a/docs/source/en/index.md +++ b/docs/source/en/index.md @@ -256,6 +256,7 @@ Flax), PyTorch, and/or TensorFlow. | [Persimmon](model_doc/persimmon) | ✅ | ❌ | ❌ | | [Phi](model_doc/phi) | ✅ | ❌ | ❌ | | [Phi3](model_doc/phi3) | ✅ | ❌ | ❌ | +| [Phimoe](model_doc/phimoe) | ✅ | ❌ | ❌ | | [PhoBERT](model_doc/phobert) | ✅ | ✅ | ✅ | | [Pix2Struct](model_doc/pix2struct) | ✅ | ❌ | ❌ | | [Pixtral](model_doc/pixtral) | ✅ | ❌ | ❌ | diff --git a/docs/source/en/model_doc/phimoe.md b/docs/source/en/model_doc/phimoe.md new file mode 100644 index 00000000000000..d9c9ae4a1831c7 --- /dev/null +++ b/docs/source/en/model_doc/phimoe.md @@ -0,0 +1,118 @@ + + +# PhiMoE + +## Overview + +The PhiMoE model was proposed in [Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone](https://arxiv.org/abs/2404.14219) by Microsoft. + +### Summary + +The abstract from the Phi-3 paper is the following: + +We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. Our training dataset is a scaled-up version of the one used for phi-2, composed of heavily filtered publicly available web data and synthetic data. The model is also further aligned for robustness, safety, and chat format. We also provide parameter-scaling results with a 7B, 14B models trained for 4.8T tokens, called phi-3-small, phi-3-medium, both significantly more capable than phi-3-mini (e.g., respectively 75%, 78% on MMLU, and 8.7, 8.9 on MT-bench). To enhance multilingual, multimodal, and long-context capabilities, we introduce three models in the phi-3.5 series: phi-3.5-mini, phi-3.5-MoE, and phi-3.5-Vision. The phi-3.5-MoE, a 16 x 3.8B MoE model with 6.6 billion active parameters, achieves superior performance in language reasoning, math, and code tasks compared to other open-source models of similar scale, such as Llama 3.1 and the Mixtral series, and on par with Gemini-1.5-Flash and GPT-4o-mini. Meanwhile, phi-3.5-Vision, a 4.2 billion parameter model derived from phi-3.5-mini, excels in reasoning tasks and is adept at handling both single-image and text prompts, as well as multi-image and text prompts. + +The original code for PhiMoE can be found [here](https://huggingface.co/microsoft/Phi-3.5-MoE-instruct). + +## Usage tips + +- This model is very similar to `Mixtral` with the main difference of [`Phi3LongRoPEScaledRotaryEmbedding`], where they are used to extend the context of the rotary embeddings. The query, key and values are fused, and the MLP's up and gate projection layers are also fused. +- The tokenizer used for this model is identical to the [`LlamaTokenizer`], with the exception of additional tokens. + +## How to use PhiMoE + + + +Phi-3.5-MoE-instruct has been integrated in the development version (4.44.2.dev) of `transformers`. Until the official version is released through `pip`, ensure that you are doing the following: +* When loading the model, ensure that `trust_remote_code=True` is passed as an argument of the `from_pretrained()` function. + +The current `transformers` version can be verified with: `pip list | grep transformers`. + +Examples of required packages: +``` +flash_attn==2.5.8 +torch==2.3.1 +accelerate==0.31.0 +transformers==4.43.0 +``` + + + +```python +import torch +from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline + +torch.random.manual_seed(0) + +model = AutoModelForCausalLM.from_pretrained( + "microsoft/Phi-3.5-MoE-instruct", + device_map="cuda", + torch_dtype="auto", + trust_remote_code=True, +) + +tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + +messages = [ + {"role": "system", "content": "You are a helpful AI assistant."}, + {"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"}, + {"role": "assistant", "content": "Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."}, + {"role": "user", "content": "What about solving an 2x + 3 = 7 equation?"}, +] + +pipe = pipeline( + "text-generation", + model=model, + tokenizer=tokenizer, +) + +generation_args = { + "max_new_tokens": 500, + "return_full_text": False, + "temperature": 0.0, + "do_sample": False, +} + +output = pipe(messages, **generation_args) +print(output[0]['generated_text']) +``` + +## PhimoeConfig + +[[autodoc]] PhimoeConfig + + + + +## PhimoeModel + +[[autodoc]] PhimoeModel + - forward + +## PhimoeForCausalLM + +[[autodoc]] PhimoeForCausalLM + - forward + - generate + +## PhimoeForSequenceClassification + +[[autodoc]] PhimoeForSequenceClassification + - forward + + + diff --git a/docs/source/en/perf_infer_gpu_one.md b/docs/source/en/perf_infer_gpu_one.md index ed3b26029d0094..2f9e94ae3ea6a4 100644 --- a/docs/source/en/perf_infer_gpu_one.md +++ b/docs/source/en/perf_infer_gpu_one.md @@ -79,6 +79,7 @@ FlashAttention-2 is currently supported for the following architectures: * [OPT](https://huggingface.co/docs/transformers/model_doc/opt#transformers.OPTModel) * [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel) * [Phi3](https://huggingface.co/docs/transformers/model_doc/phi3#transformers.Phi3Model) +* [PhiMoE](https://huggingface.co/docs/transformers/model_doc/phimoe#transformers.PhimoeModel) * [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel) * [Starcoder2](https://huggingface.co/docs/transformers/model_doc/starcoder2#transformers.Starcoder2Model) * [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model) @@ -248,6 +249,7 @@ For now, Transformers supports SDPA inference and training for the following arc * [PaliGemma](https://huggingface.co/docs/transformers/model_doc/paligemma#transformers.PaliGemmaForConditionalGeneration) * [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel) * [Phi3](https://huggingface.co/docs/transformers/model_doc/phi3#transformers.Phi3Model) +* [PhiMoE](https://huggingface.co/docs/transformers/model_doc/phimoe#transformers.PhimoeModel) * [Idefics](https://huggingface.co/docs/transformers/model_doc/idefics#transformers.IdeficsModel) * [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel) * [mBart](https://huggingface.co/docs/transformers/model_doc/mbart#transformers.MBartModel) diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index 078e4d0e4abdee..79a6bf004b8d0c 100755 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -654,6 +654,7 @@ "models.persimmon": ["PersimmonConfig"], "models.phi": ["PhiConfig"], "models.phi3": ["Phi3Config"], + "models.phimoe": ["PhimoeConfig"], "models.phobert": ["PhobertTokenizer"], "models.pix2struct": [ "Pix2StructConfig", @@ -3022,6 +3023,14 @@ "Phi3PreTrainedModel", ] ) + _import_structure["models.phimoe"].extend( + [ + "PhimoeForCausalLM", + "PhimoeForSequenceClassification", + "PhimoeModel", + "PhimoePreTrainedModel", + ] + ) _import_structure["models.pix2struct"].extend( [ "Pix2StructForConditionalGeneration", @@ -5495,6 +5504,7 @@ ) from .models.phi import PhiConfig from .models.phi3 import Phi3Config + from .models.phimoe import PhimoeConfig from .models.phobert import PhobertTokenizer from .models.pix2struct import ( Pix2StructConfig, @@ -7545,6 +7555,12 @@ Phi3Model, Phi3PreTrainedModel, ) + from .models.phimoe import ( + PhimoeForCausalLM, + PhimoeForSequenceClassification, + PhimoeModel, + PhimoePreTrainedModel, + ) from .models.pix2struct import ( Pix2StructForConditionalGeneration, Pix2StructPreTrainedModel, diff --git a/src/transformers/modeling_rope_utils.py b/src/transformers/modeling_rope_utils.py index e7aa1ceb921329..28a86bb86f89d5 100644 --- a/src/transformers/modeling_rope_utils.py +++ b/src/transformers/modeling_rope_utils.py @@ -251,7 +251,7 @@ def _compute_longrope_parameters( device (`torch.device`): The device to use for initialization of the inverse frequencies. seq_len (`int`, *optional*): - The current sequence length. Unused for this type of RoPE. + The current sequence length. rope_kwargs (`Dict`, *optional*): BC compatibility with the previous RoPE class instantiation, will be removed in v4.45. Returns: @@ -279,8 +279,11 @@ def _compute_longrope_parameters( # `original_max_position_embeddings` field containing the pretrained value. They use the ratio between these two # values to compute the default attention scaling factor, instead of using `factor`. if hasattr(config, "original_max_position_embeddings"): + if seq_len and seq_len < config.original_max_position_embeddings: + expanded_max_position_embeddings = config.original_max_position_embeddings + else: + expanded_max_position_embeddings = config.max_position_embeddings max_position_embeddings = config.original_max_position_embeddings - expanded_max_position_embeddings = config.max_position_embeddings factor = expanded_max_position_embeddings / max_position_embeddings else: max_position_embeddings = config.max_position_embeddings diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index e47a4ed9c342e4..46d1e523acdbb9 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -190,6 +190,7 @@ persimmon, phi, phi3, + phimoe, phobert, pix2struct, pixtral, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index 6d55f87d60ac8e..c21a523fafc4c4 100644 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -210,6 +210,7 @@ ("persimmon", "PersimmonConfig"), ("phi", "PhiConfig"), ("phi3", "Phi3Config"), + ("phimoe", "PhimoeConfig"), ("pix2struct", "Pix2StructConfig"), ("pixtral", "PixtralVisionConfig"), ("plbart", "PLBartConfig"), @@ -520,6 +521,7 @@ ("persimmon", "Persimmon"), ("phi", "Phi"), ("phi3", "Phi3"), + ("phimoe", "Phimoe"), ("phobert", "PhoBERT"), ("pix2struct", "Pix2Struct"), ("pixtral", "Pixtral"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index 6e730e848db755..a1d741017c89af 100644 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -198,6 +198,7 @@ ("persimmon", "PersimmonModel"), ("phi", "PhiModel"), ("phi3", "Phi3Model"), + ("phimoe", "PhimoeModel"), ("pixtral", "PixtralVisionModel"), ("plbart", "PLBartModel"), ("poolformer", "PoolFormerModel"), @@ -518,6 +519,7 @@ ("persimmon", "PersimmonForCausalLM"), ("phi", "PhiForCausalLM"), ("phi3", "Phi3ForCausalLM"), + ("phimoe", "PhimoeForCausalLM"), ("plbart", "PLBartForCausalLM"), ("prophetnet", "ProphetNetForCausalLM"), ("qdqbert", "QDQBertLMHeadModel"), @@ -948,6 +950,7 @@ ("persimmon", "PersimmonForSequenceClassification"), ("phi", "PhiForSequenceClassification"), ("phi3", "Phi3ForSequenceClassification"), + ("phimoe", "PhimoeForSequenceClassification"), ("plbart", "PLBartForSequenceClassification"), ("qdqbert", "QDQBertForSequenceClassification"), ("qwen2", "Qwen2ForSequenceClassification"), diff --git a/src/transformers/models/auto/tokenization_auto.py b/src/transformers/models/auto/tokenization_auto.py index 6a5cba11f0949f..9305245f1a6715 100644 --- a/src/transformers/models/auto/tokenization_auto.py +++ b/src/transformers/models/auto/tokenization_auto.py @@ -389,6 +389,7 @@ ), ("phi", ("CodeGenTokenizer", "CodeGenTokenizerFast" if is_tokenizers_available() else None)), ("phi3", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)), + ("phimoe", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)), ("phobert", ("PhobertTokenizer", None)), ("pix2struct", ("T5Tokenizer", "T5TokenizerFast" if is_tokenizers_available() else None)), ("pixtral", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)), diff --git a/src/transformers/models/phimoe/__init__.py b/src/transformers/models/phimoe/__init__.py new file mode 100644 index 00000000000000..e0849f5c5006e5 --- /dev/null +++ b/src/transformers/models/phimoe/__init__.py @@ -0,0 +1,28 @@ +# Copyright 2024 Microsoft 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. +from typing import TYPE_CHECKING + +from ...utils import _LazyModule +from ...utils.import_utils import define_import_structure + + +if TYPE_CHECKING: + from .configuration_phimoe import * + from .modeling_phimoe import * + +else: + import sys + + _file = globals()["__file__"] + sys.modules[__name__] = _LazyModule(__name__, _file, define_import_structure(_file), module_spec=__spec__) diff --git a/src/transformers/models/phimoe/configuration_phimoe.py b/src/transformers/models/phimoe/configuration_phimoe.py new file mode 100644 index 00000000000000..7f304281ae73d8 --- /dev/null +++ b/src/transformers/models/phimoe/configuration_phimoe.py @@ -0,0 +1,203 @@ +# coding=utf-8 +# Copyright 2024 Microsoft 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. + +"""PyTorch Phi-MoE model.""" + +from ...configuration_utils import PretrainedConfig +from ...modeling_rope_utils import rope_config_validation +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +class PhimoeConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`PhimoeModel`]. It is used to instantiate a Phi-moe + 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 + [microsoft/Phi-3.5-MoE-instruct](https://huggingface.co/microsoft/Phi-3.5-MoE-instruct). + 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 32064): + Vocabulary size of the Phimoe model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`PhimoeModel`] + hidden_size (`int`, *optional*, defaults to 4096): + Dimension of the hidden representations. + intermediate_size (`int`, *optional*, defaults to 6400): + 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. + rope_scaling (`dict`, *optional*): + The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must + contain the following keys: `type`, `short_factor`, `long_factor`, `short_mscale`, `long_mscale` and + `original_max_position_embeddings`. The `type` must be `longrope`, the `short_mscale` and `long_scale` must + be numbers, the `short_factor` and `long_factor` must be lists of numbers with the same length as half of + the attention head size and the `original_max_position_embeddings` must be an integer. + sliding_window (`int`, *optional*): + Sliding window attention window size. If not specified, will default to `262144`. + 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 16): + 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. + router_jitter_noise (`float`, *optional*, defaults to 0.01): + Amount of noise to add to the router. + input_jitter_noise (`float`, *optional*, defaults to 0.0): Input jitter noise + attention_bias (`bool`, *optional*, defaults to `False`): Attention bias + lm_head_bias (`bool`, *optional*, defaults to `False`): LM head bias + + Example: + + ```python + >>> from transformers import PhimoeModel, PhimoeConfig + >>> # Initializing a Phi-3 style configuration + >>> configuration = PhimoeConfig.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + >>> # Initializing a model from the configuration + >>> model = PhimoeModel(configuration) + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + + model_type = "phimoe" + keys_to_ignore_at_inference = ["past_key_values"] + + def __init__( + self, + vocab_size=32064, + hidden_size=4096, + intermediate_size=6400, + 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, + rope_scaling=None, + sliding_window=None, + attention_dropout=0.0, + num_experts_per_tok=2, + num_local_experts=16, + output_router_logits=False, + router_aux_loss_coef=0.001, + router_jitter_noise=0.01, + input_jitter_noise=0.0, + attention_bias=False, + lm_head_bias=False, + **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 + self.attention_bias = attention_bias + self.lm_head_bias = lm_head_bias + # 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 + self.router_jitter_noise = router_jitter_noise + self.input_jitter_noise = input_jitter_noise + + self.rope_scaling = rope_scaling + if isinstance(self.rope_scaling, dict): + if "rope_type" not in self.rope_scaling: + self.rope_scaling["rope_type"] = self.rope_scaling.get("type", None) + if "original_max_position_embeddings" in self.rope_scaling: + self.original_max_position_embeddings = self.rope_scaling["original_max_position_embeddings"] + rope_scaling_short_mscale = self.rope_scaling.get("short_mscale", None) + rope_scaling_long_mscale = self.rope_scaling.get("long_mscale", None) + if not isinstance(rope_scaling_short_mscale, (int, float)): + raise ValueError( + f"`rope_scaling`'s short_mscale field must be a number, got {rope_scaling_short_mscale}" + ) + if not isinstance(rope_scaling_long_mscale, (int, float)): + raise ValueError( + f"`rope_scaling`'s long_mscale field must be a number, got {rope_scaling_long_mscale}" + ) + + rope_config_validation(self) + + 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, + ) + + +__all__ = ["PhimoeConfig"] diff --git a/src/transformers/models/phimoe/modeling_phimoe.py b/src/transformers/models/phimoe/modeling_phimoe.py new file mode 100644 index 00000000000000..320a98471eb7e3 --- /dev/null +++ b/src/transformers/models/phimoe/modeling_phimoe.py @@ -0,0 +1,1706 @@ +# coding=utf-8 +# Copyright 2024 Microsoft 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. + +"""PyTorch Phimoe model.""" + +import math +from typing import List, Optional, Tuple, Union + +import torch +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, StaticCache +from ...generation import GenerationMixin +from ...modeling_attn_mask_utils import AttentionMaskConverter, _prepare_4d_causal_attention_mask +from ...modeling_outputs import ( + MoeCausalLMOutputWithPast, + MoeModelOutputWithPast, + SequenceClassifierOutputWithPast, +) +from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS +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_torchdynamo_compiling, + logging, + replace_return_docstrings, +) +from ...utils.import_utils import is_torch_fx_available +from .configuration_phimoe import PhimoeConfig + + +if is_flash_attn_2_available(): + from ...modeling_flash_attention_utils import _flash_attention_forward + +# 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 = "PhimoeConfig" + + +# Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func +def load_balancing_loss_func( + gate_logits: Union[torch.Tensor, Tuple[torch.Tensor], None], + num_experts: Optional[int] = None, + top_k=2, + attention_mask: Optional[torch.Tensor] = None, +) -> Union[torch.Tensor, int]: + 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: + Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of + shape [batch_size X sequence_length, num_experts]. + num_experts: + Number of experts + top_k: + The number of experts to route per-token, can be also interpreted as the `top-k` routing + parameter. + attention_mask (`torch.Tensor`, *optional*): + The attention_mask used in forward function + shape [batch_size X sequence_length] if not None. + + Returns: + The auxiliary loss. + """ + if gate_logits is None or not isinstance(gate_logits, tuple): + return 0 + + if isinstance(gate_logits, tuple): + compute_device = gate_logits[0].device + concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0) + + routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1) + + _, selected_experts = torch.topk(routing_weights, top_k, dim=-1) + + expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts) + + if attention_mask is None: + # Compute the percentage of tokens routed to each experts + tokens_per_expert = torch.mean(expert_mask.float(), dim=0) + + # Compute the average probability of routing to these experts + router_prob_per_expert = torch.mean(routing_weights, dim=0) + else: + batch_size, sequence_length = attention_mask.shape + num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length) + + # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask + expert_attention_mask = ( + attention_mask[None, :, :, None, None] + .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts)) + .reshape(-1, top_k, num_experts) + .to(compute_device) + ) + + # Compute the percentage of tokens routed to each experts + tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum( + expert_attention_mask, dim=0 + ) + + # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert + router_per_expert_attention_mask = ( + attention_mask[None, :, :, None] + .expand((num_hidden_layers, batch_size, sequence_length, num_experts)) + .reshape(-1, num_experts) + .to(compute_device) + ) + + # Compute the average probability of routing to these experts + router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum( + router_per_expert_attention_mask, dim=0 + ) + + overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0)) + return overall_loss * num_experts + + +class PhimoeRotaryEmbedding(nn.Module): + def __init__( + self, + config: Optional[PhimoeConfig] = None, + ): + super().__init__() + + self.config = config + if config.rope_scaling is not None: + self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) + self.short_mscale = config.rope_scaling.get("short_mscale") + self.long_mscale = config.rope_scaling.get("long_mscale") + else: + self.rope_type = "default" + self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] + + def forward(self, x, seq_len=None): + mscale = None + if self.config.rope_scaling and seq_len: + mscale = ( + self.long_mscale + if seq_len > self.config.rope_scaling["original_max_position_embeddings"] + else self.short_mscale + ) + inv_freq, attention_scaling = self.rope_init_fn(self.config, x.device, seq_len) + mscale = attention_scaling if mscale is None else mscale + t = torch.arange(seq_len, device=x.device, dtype=torch.float32) + freqs = torch.outer(t, inv_freq) + + emb = torch.cat((freqs, freqs), dim=-1) + return (emb.cos() * mscale).to(x.dtype), (emb.sin() * mscale).to(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.mixtral.modeling_mixtral.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) + + +class PhimoeAttention(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: PhimoeConfig, 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 a `layer_idx` is not recommended and will " + "lead 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=self.config.attention_bias) + self.k_proj = nn.Linear( + self.hidden_size, self.num_key_value_heads * self.head_dim, bias=self.config.attention_bias + ) + self.v_proj = nn.Linear( + self.hidden_size, self.num_key_value_heads * self.head_dim, bias=self.config.attention_bias + ) + self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=self.config.attention_bias) + + 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, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + 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) + + cos, sin = position_embeddings + 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, "cache_position": cache_position} # 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 attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_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 + + +class PhimoeFlashAttention2(PhimoeAttention): + """ + Phimoe flash attention module. This module inherits from `PhimoeAttention` 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. + """ + + 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, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + ): + 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) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + if past_key_value is not None: + # Activate slicing cache only if the config has a value `sliding_windows` attribute + cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0 + if ( + getattr(self.config, "sliding_window", None) is not None + and kv_seq_len > self.config.sliding_window + and cache_has_contents + ): + slicing_tokens = 1 - self.config.sliding_window + + past_key = past_key_value[self.layer_idx][0] + past_value = past_key_value[self.layer_idx][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}" + ) + + 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, "cache_position": cache_position} # 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: + if torch.is_autocast_enabled(): + target_dtype = torch.get_autocast_gpu_dtype() + # Handle the case where the model is quantized + elif 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 = _flash_attention_forward( + query_states, + key_states, + value_states, + attention_mask, + q_len, + position_ids=position_ids, + dropout=dropout_rate, + sliding_window=getattr(self.config, "sliding_window", None), + is_causal=self.is_causal, + ) + + 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 + + +class PhimoeSdpaAttention(PhimoeAttention): + """ + Phimoe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from + `PhimoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to + SDPA API. + """ + + # Adapted from PhimoeAttention.forward + 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, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + if output_attentions: + # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. + logger.warning_once( + "PhimoeModel is using PhimoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " + 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' + ) + return super().forward( + 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, + position_embeddings=position_embeddings, + ) + + 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) + + cos, sin = position_embeddings + 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, "cache_position": cache_position} # Specific to RoPE models + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + causal_mask = attention_mask + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + + # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, + # Reference: https://github.com/pytorch/pytorch/issues/112577. + if query_states.device.type == "cuda" and attention_mask is not None: + query_states = query_states.contiguous() + key_states = key_states.contiguous() + value_states = value_states.contiguous() + + # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment + # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. + # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. + is_causal = True if causal_mask is None and q_len > 1 else False + + attn_output = torch.nn.functional.scaled_dot_product_attention( + query_states, + key_states, + value_states, + attn_mask=causal_mask, + dropout_p=self.attention_dropout if self.training else 0.0, + is_causal=is_causal, + ) + + attn_output = attn_output.transpose(1, 2).contiguous() + attn_output = attn_output.view(bsz, q_len, self.hidden_size) + + attn_output = self.o_proj(attn_output) + + return attn_output, None, past_key_value + + +PHIMOE_ATTENTION_CLASSES = { + "eager": PhimoeAttention, + "flash_attention_2": PhimoeFlashAttention2, + "sdpa": PhimoeSdpaAttention, +} + + +# Copied from transformers.models.mixtral.modeling_mixtral.MixtralBlockSparseTop2MLP with Mixtral->Phimoe +class PhimoeBlockSparseTop2MLP(nn.Module): + def __init__(self, config: PhimoeConfig): + 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): + current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states) + current_hidden_states = self.w2(current_hidden_states) + return current_hidden_states + + +class MultiplierProcessor(torch.autograd.Function): + @staticmethod + def forward( + ctx, + scores: torch.Tensor, + multiplier: torch.Tensor, + selected_experts: torch.Tensor, + masked_gates: torch.Tensor, + mask_for_one: torch.Tensor, + ): + """ + Forward pass for the custom autograd function. + + Args: + ctx: Context object to save information for backward computation. + scores (torch.Tensor): Input scores tensor. + multiplier (torch.Tensor): Multiplier tensor. + selected_experts (torch.Tensor): Tensor of selected experts. + masked_gates (torch.Tensor): Masked gates tensor. + mask_for_one (torch.Tensor): Mask for one tensor. + + Returns: + torch.Tensor: Result of the forward pass. + """ + ctx.save_for_backward(multiplier, selected_experts, masked_gates) + return multiplier * mask_for_one + + @staticmethod + def backward( + ctx, + grad_at_output: torch.Tensor, + ): + """ + Backward pass for the custom autograd function. + + Args: + ctx: Context object with saved tensors from the forward pass. + grad_at_output (torch.Tensor): Gradient at the output. + + Returns: + Tuple[torch.Tensor, None, None, None, None]: Gradients for the inputs. + """ + multiplier, selected_experts, masked_gates = ctx.saved_tensors + + grad_at_output = grad_at_output * multiplier + + grad_at_scores_expanded = masked_gates * grad_at_output.mul(-1) + grad_at_scores_expanded.scatter_add_( + dim=-1, + index=selected_experts, + src=grad_at_output, + ) + + return ( + grad_at_scores_expanded, + None, + None, + None, + None, + ) + + +def sparsemixer(scores, jitter_eps, training, top_k=2): + """ + Sparse mixer function to select top-k experts and compute multipliers. + Based on the paper: https://arxiv.org/pdf/2409.12136 + We first replace the TopK(·) function as random sampling of discrete variables + in model training. Then, following Liu et al. (2023a) and Liu et al. (2023b), we apply Heun's + third order method to approximate the expert routing gradient and construct a modified + back-propagation to give a mathematically sound gradient estimation for expert routing. + + Args: + scores (torch.Tensor): Input scores tensor. + jitter_eps (float): Jitter epsilon for numerical stability. + training (bool): Flag indicating if the model is in training mode. + top_k (int): Number of top experts to select. + + Returns: + Tuple[torch.Tensor, torch.Tensor]: Multiplier and selected experts tensors. + """ + if top_k != 2: + raise ValueError("top_k must be equal to 2") + + # first expert + + with torch.no_grad(): + # Compute mask for sparsity + mask_logits_threshold, max_ind = scores.max(dim=-1, keepdim=True) + factor = scores.abs().clamp(min=mask_logits_threshold) + mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (2 * jitter_eps) + + # Apply mask + masked_gates = scores.masked_fill(mask_logits_threshold, float("-inf")) + if training: + selected_experts = ( + ( + masked_gates + - torch.empty_like(masked_gates, memory_format=torch.legacy_contiguous_format).exponential_().log() + ) + .max(dim=-1)[1] + .unsqueeze(-1) + ) # Gumbel sampling, more robust than the multinomial method + else: + selected_experts = max_ind + + # Compute scores for gradients + masked_gates = torch.softmax(masked_gates, dim=-1) + multiplier_o = masked_gates.gather(dim=-1, index=selected_experts) + + if training: + # Compute midpoint mask + max_scores, max_ind = masked_gates.max(dim=-1, keepdim=True) + mask_for_one = torch.logical_or( + selected_experts == max_ind, + torch.rand_like(max_scores) > 0.75, # Heun's third-order method + ) + # 1 -> 1.0 & 0 -> 1./3: lambda x: (x + 0.5) / 1.5 + mask_for_one = torch.add(0.3333, mask_for_one, alpha=0.6667).type_as(masked_gates) + + multiplier = MultiplierProcessor.apply( + scores, + multiplier_o, + selected_experts, + masked_gates, + mask_for_one, + ) + else: + multiplier = multiplier_o + + # Masked out first expert + masked_scores = torch.scatter( + scores, + -1, + selected_experts, + float("-inf"), + ) + with torch.no_grad(): + # Compute mask for sparsity + mask_logits_threshold, max_ind = masked_scores.max(dim=-1, keepdim=True) + factor = scores.abs().clamp(min=mask_logits_threshold) + mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (2 * jitter_eps) + + # Apply mask + masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold, float("-inf")) + if training: + selected_experts_top2 = ( + ( + masked_gates_top2 + - torch.empty_like(masked_gates_top2, memory_format=torch.legacy_contiguous_format) + .exponential_() + .log() + ) + .max(dim=-1)[1] + .unsqueeze(-1) + ) # Gumbel sampling, more robust than the multinomial method + else: + selected_experts_top2 = max_ind + # Compute scores for gradients + masked_gates_top2 = torch.softmax(masked_gates_top2, dim=-1) + multiplier_top2_o = masked_gates_top2.gather(dim=-1, index=selected_experts_top2) + + if training: + # Compute midpoint mask + max_scores, max_ind = masked_gates_top2.max(dim=-1, keepdim=True) + mask_for_one_top2 = torch.logical_or( + selected_experts_top2 == max_ind, + torch.rand_like(max_scores).uniform_() > 0.75, # Heun's third-order method + ) + # 1 -> 1.0 & 0 -> 1./3: lambda x: (x + 0.5) / 1.5 + mask_for_one_top2 = torch.add(0.3333, mask_for_one_top2, alpha=0.6667).type_as(masked_gates_top2) + + multiplier_top2 = MultiplierProcessor.apply( + scores, + multiplier_top2_o, + selected_experts_top2, + masked_gates_top2, + mask_for_one_top2, + ) + else: + multiplier_top2 = multiplier_top2_o + + multiplier = torch.concat((multiplier, multiplier_top2), dim=-1) + selected_experts = torch.concat((selected_experts, selected_experts_top2), dim=-1) + + return ( + multiplier, + selected_experts, + ) + + +class PhimoeSparseMoeBlock(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([PhimoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)]) + + # Jitter parameters + self.router_jitter_noise = config.router_jitter_noise + self.input_jitter_noise = config.input_jitter_noise + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + """ """ + batch_size, sequence_length, hidden_dim = hidden_states.shape + if self.training and self.input_jitter_noise > 0: + hidden_states *= torch.empty_like(hidden_states).uniform_( + 1.0 - self.input_jitter_noise, 1.0 + self.input_jitter_noise + ) + hidden_states = hidden_states.view(-1, hidden_dim) + router_logits = self.gate(hidden_states) + + routing_weights, selected_experts = sparsemixer( + router_logits, + jitter_eps=self.router_jitter_noise, + training=self.training, + ) + + 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 + + # 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].reshape(-1, hidden_dim) + current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, 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 PhimoeDecoderLayer(nn.Module): + def __init__(self, config: PhimoeConfig, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + + self.self_attn = PHIMOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx) + + self.block_sparse_moe = PhimoeSparseMoeBlock(config) + self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True) + self.post_attention_layernorm = nn.LayerNorm( + config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True + ) + + 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, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + **kwargs, + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + """ + 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`). + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. + kwargs (`dict`, *optional*): + Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code + into the model + """ + + 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, + cache_position=cache_position, + position_embeddings=position_embeddings, + ) + 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 + + +PHIMOE_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 ([`PhimoeConfig`]): + 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 Phimoe Model outputting raw hidden-states without any specific head on top.", + PHIMOE_START_DOCSTRING, +) +# Copied from transformers.models.mixtral.modeling_mixtral.MixtralPreTrainedModel with Mixtral->Phimoe +class PhimoePreTrainedModel(PreTrainedModel): + config_class = PhimoeConfig + base_model_prefix = "model" + supports_gradient_checkpointing = True + _no_split_modules = ["PhimoeDecoderLayer"] + _skip_keys_device_placement = "past_key_values" + _supports_flash_attn_2 = True + _supports_sdpa = 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_() + + +PHIMOE_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. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. +""" + + +@add_start_docstrings( + "The bare Phimoe Model outputting raw hidden-states without any specific head on top.", + PHIMOE_START_DOCSTRING, +) +class PhimoeModel(PhimoePreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`PhimoeDecoderLayer`] + Args: + config: PhimoeConfig + """ + + def __init__(self, config: PhimoeConfig): + 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( + [PhimoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self._attn_implementation = config._attn_implementation + self.norm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True) + self.rotary_emb = PhimoeRotaryEmbedding(config=config) + + 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 + + @add_start_docstrings_to_model_forward(PHIMOE_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, + cache_position: Optional[torch.LongTensor] = 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 + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) + + 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 + + # kept for BC (non `Cache` `past_key_values` inputs) + return_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + return_legacy_cache = True + if past_key_values is None: + past_key_values = DynamicCache() + else: + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and " + "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class " + "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)" + ) + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device + ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) + + hidden_states = inputs_embeds + + position_embeddings = self.rotary_emb(hidden_states, seq_len=cache_position[-1] + 1) + + # 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, + causal_mask, + position_ids, + past_key_values, + output_attentions, + output_router_logits, + use_cache, + cache_position, + position_embeddings, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=causal_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, + cache_position=cache_position, + position_embeddings=position_embeddings, + ) + + 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 = next_decoder_cache if use_cache else None + if return_legacy_cache: + next_cache = next_cache.to_legacy_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, + ) + + # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). + causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( + attention_mask, + sequence_length=sequence_length, + target_length=target_length, + dtype=dtype, + device=device, + cache_position=cache_position, + batch_size=input_tensor.shape[0], + ) + + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + min_dtype = torch.finfo(dtype).min + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + + @staticmethod + # Copied from transformers.models.llama.modeling_llama.LlamaModel._prepare_4d_causal_attention_mask_with_cache_position + def _prepare_4d_causal_attention_mask_with_cache_position( + attention_mask: torch.Tensor, + sequence_length: int, + target_length: int, + dtype: torch.dtype, + device: torch.device, + cache_position: torch.Tensor, + batch_size: int, + ): + """ + Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape + `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. + + Args: + attention_mask (`torch.Tensor`): + A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape + `(batch_size, 1, query_length, key_value_length)`. + sequence_length (`int`): + The sequence length being processed. + target_length (`int`): + The target length: when generating with static cache, the mask should be as long as the static cache, + to account for the 0 padding, the part of the cache that is not filled yet. + dtype (`torch.dtype`): + The dtype to use for the 4D attention mask. + device (`torch.device`): + The device to plcae the 4D attention mask on. + cache_position (`torch.Tensor`): + Indices depicting the position of the input sequence tokens in the sequence. + batch_size (`torch.Tensor`): + Batch size. + """ + if attention_mask is not None and attention_mask.dim() == 4: + # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. + causal_mask = attention_mask + else: + min_dtype = torch.finfo(dtype).min + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + + return causal_mask + + +class PhimoeForCausalLM(PhimoePreTrainedModel, GenerationMixin): + _tied_weights_keys = ["lm_head.weight"] + + def __init__(self, config): + super().__init__(config) + self.model = PhimoeModel(config) + self.vocab_size = config.vocab_size + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=self.config.lm_head_bias) + 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() + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings + def get_input_embeddings(self): + return self.model.embed_tokens + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings + def get_output_embeddings(self): + return self.lm_head + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder + def set_decoder(self, decoder): + self.model = decoder + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder + def get_decoder(self): + return self.model + + @add_start_docstrings_to_model_forward(PHIMOE_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, + cache_position: Optional[torch.LongTensor] = None, + num_logits_to_keep: int = 0, + ) -> 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]`. + + num_logits_to_keep (`int`, *optional*): + Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all + `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that + token can save memory, which becomes pretty significant for long sequences or large vocabulary size. + Returns: + Example: + ```python + >>> from transformers import AutoTokenizer, PhimoeForCausalLM + >>> model = PhimoeForCausalLM.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + >>> 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." + ```""" + if ( + use_cache + and self.config.rope_scaling + and cache_position is not None + and cache_position[0] == self.config.original_max_position_embeddings + ): + logger.warning( + f"If you are not using the generate method, you may encounter nonsensical outputs after the {self.config.original_max_position_embeddings}th token, as the KV cache needs to be recomputed." + ) + 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, + cache_position=cache_position, + ) + + hidden_states = outputs[0] + if labels is None and not is_torchdynamo_compiling(): + logger.warning_once( + "Starting from v4.46, the `logits` model output will have the same type as the model (except at train time, where it will always be FP32)" + ) + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + # TODO: remove the float() operation in v4.46 + logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :]).float() + + loss = None + if labels is not None: + # Upcast to float if we need to compute the loss to avoid potential precision issues + logits = logits.float() + # 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, + attention_mask, + ) + if labels is not None: + loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device + + 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, + ) + + # Copied from transformers.models.phi3.modeling_phi3.Phi3ForCausalLM.prepare_inputs_for_generation + def prepare_inputs_for_generation( + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + position_ids=None, + use_cache=True, + num_logits_to_keep=None, + **kwargs, + ): + # When the first time input length reached long and short factor switching point, enforce re-compute cache + # It will cause downside of slower at this single token position, however, better than current failure. + if ( + past_key_values + and self.config.rope_scaling + and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1 + ): + past_length = cache_position[0] + if past_length <= self.config.original_max_position_embeddings: + past_key_values = None + + # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens + # Exception 1: when passing input_embeds, input_ids may be missing entries + # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here + if past_key_values is not None: + if inputs_embeds is not None: # Exception 1 + input_ids = input_ids[:, -cache_position.shape[0] :] + elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2) + input_ids = input_ids[:, cache_position] + + 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] :] + + # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture. + position_ids = position_ids.clone(memory_format=torch.contiguous_format) + + # if `inputs_embeds` are passed, we only want to use them in the 1st generation step + if inputs_embeds is not None and cache_position[0] == 0: + model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None} + else: + # The clone here is for the same reason as for `position_ids`. + model_inputs = {"input_ids": input_ids.clone(memory_format=torch.contiguous_format), "inputs_embeds": None} + + if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2: + if model_inputs["inputs_embeds"] is not None: + batch_size, sequence_length, _ = model_inputs["inputs_embeds"].shape + device = model_inputs["inputs_embeds"].device + else: + batch_size, sequence_length = model_inputs["input_ids"].shape + device = model_inputs["input_ids"].device + + attention_mask = self.model._prepare_4d_causal_attention_mask_with_cache_position( + attention_mask, + sequence_length=sequence_length, + target_length=past_key_values.get_max_length(), + dtype=self.lm_head.weight.dtype, + device=device, + cache_position=cache_position, + batch_size=batch_size, + ) + + if num_logits_to_keep is not None: + model_inputs["num_logits_to_keep"] = num_logits_to_keep + + model_inputs.update( + { + "position_ids": position_ids, + "cache_position": cache_position, + "past_key_values": past_key_values, + "use_cache": use_cache, + "attention_mask": attention_mask, + } + ) + return model_inputs + + +@add_start_docstrings( + """ + The Phimoe Model transformer with a sequence classification head on top (linear layer). + [`PhimoeForSequenceClassification`] 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). + """, + PHIMOE_START_DOCSTRING, +) + +# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phimoe, LLAMA->PHIMOE +class PhimoeForSequenceClassification(PhimoePreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = PhimoeModel(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(PHIMOE_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, 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: + # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility + sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1 + sequence_lengths = sequence_lengths % input_ids.shape[-1] + sequence_lengths = sequence_lengths.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, + ) + + +__all__ = [ + "PhimoePreTrainedModel", + "PhimoeModel", + "PhimoeForCausalLM", + "PhimoeForSequenceClassification", +] diff --git a/src/transformers/utils/dummy_pt_objects.py b/src/transformers/utils/dummy_pt_objects.py index f4e471ee7ab58b..2e2b0726f611cf 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -7158,6 +7158,34 @@ def __init__(self, *args, **kwargs): requires_backends(self, ["torch"]) +class PhimoeForCausalLM(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class PhimoeForSequenceClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class PhimoeModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class PhimoePreTrainedModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + class Pix2StructForConditionalGeneration(metaclass=DummyObject): _backends = ["torch"] diff --git a/tests/models/phimoe/__init__.py b/tests/models/phimoe/__init__.py new file mode 100644 index 00000000000000..e69de29bb2d1d6 diff --git a/tests/models/phimoe/test_modeling_phimoe.py b/tests/models/phimoe/test_modeling_phimoe.py new file mode 100644 index 00000000000000..57e5e10fba6768 --- /dev/null +++ b/tests/models/phimoe/test_modeling_phimoe.py @@ -0,0 +1,566 @@ +# coding=utf-8 +# Copyright 2024 Microsoft 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. + +"""Testing suite for the PyTorch PhiMoE model.""" + +import unittest +from typing import List + +from parameterized import parameterized + +from transformers import PhimoeConfig, StaticCache, is_torch_available, set_seed +from transformers.testing_utils import ( + require_torch, + 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 ( + AutoTokenizer, + PhimoeForCausalLM, + PhimoeForSequenceClassification, + PhimoeModel, + ) + + end_of_text_token = 32000 + + class PhimoeMiniWithStaticCache(torch.nn.Module): + def __init__(self, model: PhimoeForCausalLM, batch_size: int, max_seq_len: int): + super().__init__() + self.model = model + self.cache = StaticCache( + config=model.config, + batch_size=batch_size, + max_cache_len=max_seq_len, + device=self.model.device, + dtype=self.model.dtype, + ) + + def forward( + self, + input_ids: torch.LongTensor = None, + ) -> torch.FloatTensor: + return self.model.forward( + input_ids=input_ids, + use_cache=True, + return_dict=True, + past_key_values=self.cache, + ).logits + + @staticmethod + def generate(model: PhimoeForCausalLM, prompt_tokens: torch.LongTensor, max_seq_len: int) -> List[int]: + model = PhimoeMiniWithStaticCache(model, 1, max_seq_len + prompt_tokens.shape[-1]) + + response_tokens = [] + + for input_pos in range(prompt_tokens.shape[-1]): + result = model.forward( + input_ids=prompt_tokens[:, input_pos : input_pos + 1], + ) + response_tokens.append(prompt_tokens[0][input_pos].item()) + + current_token = torch.argmax(result[:, -1, :], dim=-1).item() + response_tokens.append(current_token) + + while current_token != end_of_text_token and len(response_tokens) < max_seq_len: + result = model.forward( + input_ids=torch.tensor([[current_token]], dtype=torch.long), + ) + current_token = torch.argmax(result[:, -1, :], dim=-1).item() + response_tokens.append(current_token) + + return response_tokens + + +class PhimoeModelTester: + 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=4, + intermediate_size=37, + hidden_act="gelu", + hidden_dropout_prob=0.1, + attention_probs_dropout_prob=0.1, + max_position_embeddings=131072, + 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, + original_max_position_embeddings=4096, + ): + 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 + self.original_max_position_embeddings = original_max_position_embeddings + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.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 PhimoeConfig( + 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, + original_max_position_embeddings=self.original_max_position_embeddings, + ) + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Phimoe + def create_and_check_model( + self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels + ): + model = PhimoeModel(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->Phimoe + 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 = PhimoeModel(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->Phimoe + 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 = PhimoeForCausalLM(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->Phimoe + 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 = PhimoeForCausalLM(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 + 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 +class PhimoeModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase): + all_model_classes = ( + (PhimoeModel, PhimoeForCausalLM, PhimoeForSequenceClassification) if is_torch_available() else () + ) + all_generative_model_classes = (PhimoeForCausalLM,) if is_torch_available() else () + pipeline_model_mapping = ( + { + "feature-extraction": PhimoeModel, + "text-classification": PhimoeForSequenceClassification, + "text-generation": PhimoeForCausalLM, + "zero-shot": PhimoeForSequenceClassification, + } + if is_torch_available() + else {} + ) + + test_headmasking = False + test_pruning = False + + # TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79292/workflows/fa2ba644-8953-44a6-8f67-ccd69ca6a476/jobs/1012905 + def is_pipeline_test_to_skip( + self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name + ): + return True + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.setUp with Llama->Phimoe + def setUp(self): + self.model_tester = PhimoeModelTester(self) + self.config_tester = ConfigTester(self, config_class=PhimoeConfig, hidden_size=37) + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_config + def test_config(self): + self.config_tester.run_common_tests() + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_model + def test_model(self): + config_and_inputs = self.model_tester.prepare_config_and_inputs() + self.model_tester.create_and_check_model(*config_and_inputs) + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model with Llama->Phimoe,llama->phimoe + def test_phimoe_sequence_classification_model(self): + config, input_dict = self.model_tester.prepare_config_and_inputs_for_common() + 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 = PhimoeForSequenceClassification(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)) + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model_for_single_label with Llama->Phimoe,llama->phimoe + def test_phimoe_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 = PhimoeForSequenceClassification(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)) + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_sequence_classification_model_for_multi_label with Llama->Phimoe,llama->phimoe + def test_phimoe_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 = PhimoeForSequenceClassification(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)) + + @parameterized.expand([("longrope",)]) + def test_model_rope_scaling_from_config(self, scaling_type): + config, _ = self.model_tester.prepare_config_and_inputs_for_common() + short_input = ids_tensor([1, 10], config.vocab_size) + long_input = ids_tensor([1, int(config.original_max_position_embeddings * 1.5)], config.vocab_size) + + set_seed(42) # Fixed seed at init time so the two models get the same random weights + original_model = PhimoeModel(config) + original_model.to(torch_device) + original_model.eval() + original_short_output = original_model(short_input).last_hidden_state + original_long_output = original_model(long_input).last_hidden_state + + set_seed(42) # Fixed seed at init time so the two models get the same random weights + n_factors = config.hidden_size // config.num_attention_heads // 2 + config.rope_scaling = { + "type": scaling_type, + "short_factor": [3.0 for _ in range(n_factors)], + "long_factor": [5.0 for _ in range(n_factors)], + "short_mscale": 1.243163121016122, + "long_mscale": 1.243163121016122, + "original_max_position_embeddings": 4096, + } + scaled_model = PhimoeModel(config) + scaled_model.to(torch_device) + scaled_model.eval() + scaled_short_output = scaled_model(short_input).last_hidden_state + scaled_long_output = scaled_model(long_input).last_hidden_state + + # Scaling changes the RoPE embeddings, both for the short and long outputs + self.assertFalse(torch.allclose(original_short_output, scaled_short_output, atol=1e-5)) + self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5)) + + @parameterized.expand([("longrope",)]) + def test_model_rope_scaling_short_long_factor(self, scaling_type): + config, _ = self.model_tester.prepare_config_and_inputs_for_common() + n_factors = config.hidden_size // config.num_key_value_heads // 2 + config.rope_scaling = { + "type": scaling_type, + "short_factor": [3.0 for _ in range(n_factors)], + "long_factor": [5.0 for _ in range(n_factors)], + "short_mscale": 1.243163121016122, + "long_mscale": 1.243163121016122, + "original_max_position_embeddings": 4096, + } + input_tensor = ids_tensor([1, 4090], config.vocab_size) + model = PhimoeForCausalLM(config) + model.to(torch_device) + model.eval() + generation_args_short = { + "max_length": config.original_max_position_embeddings, + "temperature": 0.0, + "use_cache": True, + "do_sample": False, + "return_dict_in_generate": True, + } + output_with_short_factor = model.generate(input_tensor, **generation_args_short) + keys_with_short_factor = output_with_short_factor.past_key_values[0][0] + generation_args_long = { + "max_length": config.original_max_position_embeddings + 5, + "temperature": 0.0, + "use_cache": True, + "do_sample": False, + "return_dict_in_generate": True, + "output_logits": True, + } + output_with_long_factor = model.generate(input_tensor, **generation_args_long) + keys_with_long_factor = output_with_long_factor.past_key_values[0][0] + last_token_logits = output_with_long_factor.logits[-1][-1] + regenerated_last_token_logits = model(output_with_long_factor.sequences[:, :-1]).logits[0][-1] + keys_with_long_factor = keys_with_long_factor[:, :, : config.original_max_position_embeddings - 1, :] + + # KV cache is re-computed after reaching the (`config.original_max_position_embeddings`+1)th token position + self.assertFalse(torch.allclose(keys_with_short_factor, keys_with_long_factor, atol=1e-3, rtol=1e-3)) + # Last token generated using long factor + self.assertTrue(torch.allclose(last_token_logits, regenerated_last_token_logits, atol=1e-2, rtol=1e-2)) + + +@slow +@require_torch +class PhimoeIntegrationTest(unittest.TestCase): + def test_model_phimoe_instruct_logits(self): + input_ids = { + "input_ids": torch.tensor( + [[1212, 318, 281, 1672, 2643, 290, 428, 318, 257, 1332]], dtype=torch.long, device=torch_device + ) + } + + model = PhimoeForCausalLM.from_pretrained("microsoft/Phi-3.5-MoE-instruct").to(torch_device) + model.eval() + + output = model(**input_ids).logits + + EXPECTED_OUTPUT = torch.tensor([[-3.5312, -2.5000, -1.2734, 0.3555, -0.7578, -0.4727, 0.5977, -0.4316, + 0.2256, -1.2188, -1.6797, 0.9961, 3.7656, 11.3125, -1.3828, -4.8438, + -5.7500, -1.9375, 0.7227, -0.3438, -0.2100, -0.4277, -0.0444, -0.5352, + -0.6406, -0.1016, -0.4258, -1.0234, 0.4297, -0.6250], + [-0.9883, 0.1455, -0.4902, 2.3594, 0.7031, 3.1406, 0.4375, 0.2559, + 0.6172, -2.1094, -1.3359, 2.5938, 4.9062, 10.8125, -0.1094, 1.5781, + -4.9375, 0.7148, -0.0972, 1.7656, -0.0801, 0.2217, 0.1875, -0.4629, + 1.5781, 0.3535, 0.0874, 0.6836, -0.0518, -1.2969]]).to(torch_device) # fmt: skip + + self.assertTrue(torch.allclose(EXPECTED_OUTPUT, output[0, :2, :30], atol=1e-4, rtol=1e-4)) + + def test_phimoe_instruct_generation(self): + model = PhimoeForCausalLM.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + + messages = [ + { + "role": "system", + "content": "You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user.", + }, + {"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"}, + ] + inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt") + + outputs = model.generate(inputs, max_new_tokens=32) + output_text = tokenizer.batch_decode(outputs) + + EXPECTED_OUTPUT = [ + "<|system|> You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user.<|end|><|user|> Can you provide ways to eat combinations of bananas and dragonfruits?<|end|><|assistant|> Certainly! Bananas and dragonfruits are both delicious and nutritious fruits that can be combined in various ways to create tast" + ] + + self.assertListEqual(output_text, EXPECTED_OUTPUT) + + def test_phimoe_instruct_with_static_cache(self): + model = PhimoeForCausalLM.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-MoE-instruct") + + messages = [ + { + "role": "system", + "content": "You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user.", + }, + {"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"}, + ] + inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt") + + response_tokens = PhimoeMiniWithStaticCache.generate(model, inputs, 64) + + output_text = tokenizer.batch_decode(torch.tensor([response_tokens], dtype=torch.long, device=torch_device)) + + EXPECTED_OUTPUT = [ + "<|system|> You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user.<|end|><|user|> Can you provide ways to eat combinations of bananas and dragonfruits?<|end|><|assistant|> Certainly! Bananas and dragonfruits are both delicious and nutritious fruits that can" + ] + + self.assertListEqual(output_text, EXPECTED_OUTPUT)