Source code for cerebras.modelzoo.tools.checkpoint_converters.gptj_hf_cs

# Copyright 2022 Cerebras Systems.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import logging
import re
from typing import Tuple

import torch

from cerebras.modelzoo.tools.checkpoint_converters.base_converter import (
    BaseCheckpointConverter_CS_CS,
    BaseCheckpointConverter_HF_CS,
    BaseConfigConverter,
    BaseConfigConverter_CS_CS,
    BaseConfigConverter_HF_CS,
    ConfigConversionError,
    ConversionRule,
    EquivalentSubkey,
    FormatVersions,
)
from cerebras.modelzoo.tools.checkpoint_converters.helper import (
    convert_use_biasless_layer_norm_helper,
    maybe_tie_lm_head,
    tie_none_weights,
)


[docs]class Converter_GPTJ_Attention_HF_CS17(BaseCheckpointConverter_HF_CS): def __init__(self): super().__init__() self.rules = [ ConversionRule( [ EquivalentSubkey("q_proj", "proj_q_dense_layer"), r"\.(?:weight|bias)", ], action=self.replaceKey, ), ConversionRule( [ EquivalentSubkey("k_proj", "proj_k_dense_layer"), r"\.(?:weight|bias)", ], action=self.replaceKey, ), ConversionRule( [ EquivalentSubkey("v_proj", "proj_v_dense_layer"), r"\.(?:weight|bias)", ], action=self.replaceKey, ), ConversionRule( [ EquivalentSubkey("out_proj", "proj_output_dense_layer"), r"\.(?:weight|bias)", ], # This is a hacky way to initialize bias and masked_bias when converting from CS to # Huggingface. However, based on huggingface implementation this is unavoidable in # order to initialize attn.bias and attn.masked_bias, which we initiate when we # capture the `out_proj` key. action=self.replace_or_fill_masked_bias, ), ] @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("hf"), FormatVersions("cs-1.7")) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return None def replace_or_fill_masked_bias( self, old_key, new_key, old_state_dict, new_state_dict, from_index, action_fn_args, ): # copy between `out_proj` and `proj_output_dense_layer` new_state_dict[new_key] = old_state_dict[old_key] # take care of bias and masked_bias if from_index == 1: max_positions = action_fn_args["configs"][1]["model"][ "max_position_embeddings" ] bias_key = re.sub(r"out_proj\.weight", "bias", new_key) masked_bias_key = re.sub( r"out_proj\.weight", "masked_bias", new_key ) new_state_dict[bias_key] = torch.tril( torch.ones((max_positions, max_positions), dtype=torch.uint8) ).view(1, 1, max_positions, max_positions) new_state_dict[masked_bias_key] = torch.tensor(-1e9)
[docs]class Converter_GPTJ_Headless_HF_CS17(BaseCheckpointConverter_HF_CS): def __init__(self): super().__init__() self.rules = [ # word embeddings ConversionRule( [ EquivalentSubkey("wte", "embedding_layer.word_embeddings"), r"\.(?:weight|bias)", ], action=self.replaceKey, ), # final layer norm ConversionRule( [ EquivalentSubkey("ln_f", "transformer_decoder.norm"), r"\.(?:weight|bias)", ], action=self.replace_final_norm, ), # attention ConversionRule( [ EquivalentSubkey("h", "transformer_decoder.layers"), r"\.\d+\.", EquivalentSubkey("attn.", "self_attn."), Converter_GPTJ_Attention_HF_CS17(), ], action=None, ), # attention norm ConversionRule( [ EquivalentSubkey("h", "transformer_decoder.layers"), r"\.\d+\.", EquivalentSubkey("ln_1", "norm1"), r"\.(?:weight|bias)", ], action=self.replaceKey, ), # norm3 layers should be set to None if they exist: ConversionRule( [ r"transformer_decoder.layers\.\d+\.norm3\.(?:weight|bias)", ], exists="right", action=BaseConfigConverter.assert_factory_fn(1, None), ), # intermediate ffn ConversionRule( [ EquivalentSubkey("h", "transformer_decoder.layers"), r"\.\d+\.", EquivalentSubkey("mlp.fc_in", "ffn.ffn.0.linear_layer"), r"\.(?:weight|bias)", ], action=self.replaceKey, ), ConversionRule( [ EquivalentSubkey("h", "transformer_decoder.layers"), r"\.\d+\.", EquivalentSubkey("mlp.fc_out", "ffn.ffn.1.linear_layer"), r"\.(?:weight|bias)", ], action=self.replaceKey, ), ConversionRule([r"lm_head\.(?:weight|bias)"], exists="right"), ConversionRule([r"ln_f\.(?:weight|bias)"], exists="right"), ConversionRule( [ r"h\.\d+\.attn\.(?:masked_bias|bias)", ] ), ] def replace_final_norm( self, old_key, new_key, old_state_dict, new_state_dict, from_index, action_fn_args, ): new_state_dict[new_key] = old_state_dict[old_key] # CS 1.7 has both "ln_f" and "transformer_decoder.norm" # we need to copy the original ("ln_f") too: if from_index == 0: ln_f_key = re.sub(r"transformer_decoder\.norm\.", "ln_f.", new_key) new_state_dict[ln_f_key] = old_state_dict[old_key] def pre_model_convert( self, old_state_dict, new_state_dict, configs, converter_indices, drop_unmatched_keys, ): if converter_indices.direction == 0: logging.warning( "{} GPTJ has a language model head (lm_head) " "while {} GPTJModel does not. Initializing lm_head to default.".format( *self.formats() ) ) # Manually tie weights if ( converter_indices.direction == 1 and configs[1]["model"]["share_embedding_weights"] ): if ( old_state_dict.get("embedding_layer.word_embeddings.weight", 0) is None ): old_state_dict["embedding_layer.word_embeddings.weight"] = ( old_state_dict["lm_head.weight"] ) def post_model_convert( self, old_state_dict, new_state_dict, configs, converter_indices, drop_unmatched_keys, key_prefix="", ): if converter_indices.direction == 0: # We are converting from HF GPTJModel (which is headless) -> CS GPTJModel (which has # a head). We need to create 'lm_head' and init to default values. hf_config = configs[0] cs_config = configs[1] use_bias_in_output = cs_config["model"].get( "use_bias_in_output", False ) vocab_size = cs_config["model"]["vocab_size"] embed_dim = cs_config["model"]["hidden_size"] if hf_config["tie_word_embeddings"]: lm_head_weight = old_state_dict['wte.weight'] else: lm_head_weight = torch.zeros((vocab_size, embed_dim)) lm_head_weight.normal_(mean=0.0, std=0.02) new_state_dict[key_prefix + "lm_head.weight"] = lm_head_weight if use_bias_in_output: lm_head_bias = torch.zeros(vocab_size) new_state_dict[key_prefix + "lm_head.bias"] = lm_head_bias super().post_model_convert( old_state_dict, new_state_dict, configs, converter_indices, drop_unmatched_keys, key_prefix=key_prefix, ) @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("hf"), FormatVersions("cs-1.7")) @classmethod def converter_note(cls) -> str: return ( "{} GPTJModel <-> {} GPTJModel(with head)\n" "The HF model doesn't contain a language model head while the CS " "one does. When converting to CS, the exported checkpoint will " "contain a language model head initialized to default random " "values. When converting to HF, the language model head will be " "dropped." ).format(cls.formats()[0], cls.formats()[1]) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS17
[docs]class Converter_GPTJ_Headless_HF_CS18(Converter_GPTJ_Headless_HF_CS17): def __init__(self): super().__init__() self.rules = [ # Catch checkpoints from Pytorch 2.0 API ConversionRule( [ Converter_GPTJ_Headless_HF_CS17(), ], action=None, ), # Catch checkpoints from 1.7/1.8 ConversionRule( [ EquivalentSubkey("", "model."), Converter_GPTJ_Headless_HF_CS17(), ], action=None, ), ] @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("hf"), FormatVersions("cs-1.8", "cs-1.9")) @classmethod def converter_note(cls) -> str: return ( "{} GPTJModel <-> {} GPTJModel(with head)\n" "The HF model doesn't contain a language model head while the CS " "one does. When converting to CS, the exported checkpoint will " "contain a language model head initialized to default random " "values. When converting to HF, the language model head will be " "dropped." ).format(cls.formats()[0], cls.formats()[1]) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS18
[docs]class Converter_GPTJ_LMHeadModel_HF_CS17(BaseCheckpointConverter_HF_CS): def __init__(self): super().__init__() self.rules = [ ConversionRule( [r"lm_head\.(?:weight|bias)"], action=self.replaceKey, ), ConversionRule( [ EquivalentSubkey("transformer.", ""), Converter_GPTJ_Headless_HF_CS17(), ], action=None, ), ] def pre_model_convert( self, old_state_dict, new_state_dict, configs, converter_indices, drop_unmatched_keys, ): # Manually tie weights if ( converter_indices.direction == 1 and configs[1]["model"]["share_embedding_weights"] ): if ( old_state_dict.get("embedding_layer.word_embeddings.weight", 0) is None ): old_state_dict["embedding_layer.word_embeddings.weight"] = ( old_state_dict["lm_head.weight"] ) @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("hf"), FormatVersions("cs-1.7")) @classmethod def converter_note(cls) -> str: return "{} GPTJForCausalLM <-> {} GPTJModel(with head)".format( cls.formats()[0], cls.formats()[1] ) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS17
[docs]class Converter_GPTJ_LMHeadModel_HF_CS18(BaseCheckpointConverter_HF_CS): def __init__(self): super().__init__() self.rules = [ # Catch checkpoints from Pytorch 2.0 API ConversionRule( [ Converter_GPTJ_LMHeadModel_HF_CS17(), ], action=None, ), # Catch checkpoints from 1.7/1.8 ConversionRule( [ EquivalentSubkey("", "model."), Converter_GPTJ_LMHeadModel_HF_CS17(), ], action=None, ), ] @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("hf"), FormatVersions("cs-1.8", "cs-1.9")) @classmethod def converter_note(cls) -> str: return "{} GPTJForCausalLM <-> {} GPTJModel(with head)".format( cls.formats()[0], cls.formats()[1] ) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS18
[docs]class ConfigConverter_GPTJModel_HF_CS17(BaseConfigConverter_HF_CS): def __init__(self): super().__init__() self.rules = [ ConversionRule( ["model_type"], action=BaseConfigConverter.assert_factory_fn(0, "gptj"), ), # Embedding ConversionRule(["vocab_size"], action=self.replaceKey), ConversionRule(["rotary_dim"], action=self.replaceKey), ConversionRule( [EquivalentSubkey("rotary", "position_embedding_type")], exists="right", action=self.convert_position_embedding_type, ), ConversionRule( ["use_position_embedding"], exists="right", action=BaseConfigConverter.assert_factory_fn(1, True), ), ConversionRule( [EquivalentSubkey("embd_pdrop", "embedding_dropout_rate")], action=self.replaceKey, ), ConversionRule( [ EquivalentSubkey( "tie_word_embeddings", "share_embedding_weights" ) ], action=self.replaceKey, ), # Decoder Block ConversionRule( [EquivalentSubkey("n_embd", "hidden_size")], action=self.replaceKey, ), ConversionRule( [EquivalentSubkey("n_head", "num_heads")], action=self.replaceKey, ), ConversionRule( [EquivalentSubkey("n_layer", "num_hidden_layers")], action=self.replaceKey, ), ConversionRule( [EquivalentSubkey("n_positions", "max_position_embeddings")], action=self.replaceKey, ), ConversionRule( ["scale_attn_weights"], action=BaseConfigConverter.assert_factory_fn(0, True), ), ConversionRule( ["attention_type"], action=BaseConfigConverter.assert_factory_fn( 1, "scaled_dot_product" ), ), ConversionRule( ["use_projection_bias_in_attention"], action=BaseConfigConverter.assert_factory_fn(1, False), ), ConversionRule( ["use_ffn_bias_in_attention"], exists="right", action=BaseConfigConverter.assert_factory_fn(1, False), ), ConversionRule( ["use_ffn_bias"], exists="right", action=BaseConfigConverter.assert_factory_fn(1, True), ), ConversionRule( [EquivalentSubkey("n_inner", "filter_size")], action=self.replaceKey, ), ConversionRule( [EquivalentSubkey("activation_function", "nonlinearity")], action=self.replaceKey, ), ConversionRule( [EquivalentSubkey("attn_pdrop", "attention_dropout_rate")], action=self.replaceKey, ), ConversionRule( [EquivalentSubkey("resid_pdrop", "residual_dropout_rate")], action=self.replaceKey, ), ConversionRule( ["layer_norm_epsilon"], action=self.replaceKey, ), ConversionRule( ["use_bias_in_output"], action=self.replaceKey, ), ConversionRule(["initializer_range"], action=self.replaceKey), ConversionRule( ["embedding_layer_norm"], action=BaseConfigConverter.assert_factory_fn(1, False), ), ConversionRule( ["fixed_sparse_attention"], action=BaseConfigConverter.assert_factory_fn(1, None), ), ConversionRule( ["norm_first"], action=BaseConfigConverter.assert_factory_fn(1, True), ), ConversionRule( ["use_ff_layer1_dropout"], action=BaseConfigConverter.assert_factory_fn(1, False), ), ConversionRule( ["use_untied_layer_norm"], action=BaseConfigConverter.assert_factory_fn(1, False), ), ] self.pre_convert_defaults[0].update( { "vocab_size": 50400, "n_positions": 2048, "n_embd": 4096, "n_layer": 28, "n_head": 16, "rotary_dim": 64, "activation_function": "gelu_new", "resid_pdrop": 0.1, "embd_pdrop": 0.1, "attn_pdrop": 0.1, "initializer_range": 0.02, "layer_norm_epsilon": 1e-5, "tie_word_embeddings": False, } ) self.pre_convert_defaults[1].update( { "max_position_embeddings": 1024, "embedding_dropout_rate": 0.1, "share_embedding_weights": True, "residual_dropout_rate": 0.1, "nonlinearity": "gelu", "layer_norm_epsilon": 1.0e-5, "use_ffn_bias": False, "use_untied_layer_norm": False, "attention_dropout_rate": 0.1, "use_projection_bias_in_attention": True, "use_ffn_bias_in_attention": True, "initializer_range": 0.02, "use_bias_in_output": False, "norm_first": True, } ) self.post_convert_defaults[0].update({"model_type": "gptj"}) self.post_convert_defaults[1].update( { "use_untied_layer_norm": False, "use_ffn_bias_in_attention": False, "use_projection_bias_in_attention": False, "use_ffn_bias": True, "use_bias_in_output": True, "attention_type": "scaled_dot_product", }, ) def convert_position_embedding_type( self, old_key, new_key, old_state_dict, new_state_dict, from_index, action_fn_args, ): if from_index == 0: if old_state_dict[old_key] != True: raise ConfigConversionError( "HF GPT-J must use rotary embeddings, but got {}={}".format( old_key, old_state_dict[old_key] ) ) new_state_dict[new_key] = "rotary" else: if old_state_dict[old_key] != "rotary": raise ConfigConversionError( "CS GPT-J must use rotary embeddings, but got {}={}".format( old_key, old_state_dict[old_key] ) ) new_state_dict[new_key] = True def convert_attention_type( self, old_key, new_key, old_state_dict, new_state_dict, from_index, action_fn_args, ): if from_index == 0: new_state_dict[new_key] = ( "scaled_dot_product" if old_state_dict[old_key] else "dot_product" ) else: assert ( old_state_dict[old_key] == "scaled_dot_product" or old_state_dict[old_key] == "dot_product" ) new_state_dict[new_key] = old_state_dict[old_key].startswith( "scaled_" ) def pre_config_convert( self, config, converter_indices, ): config = super().pre_config_convert(config, converter_indices) if converter_indices.direction == 0: if "n_inner" not in config or config["n_inner"] is None: config["n_inner"] = 4 * config["n_embd"] return config @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("hf"), FormatVersions("cs-1.7"))
[docs]class ConfigConverter_GPTJModel_HF_CS18(ConfigConverter_GPTJModel_HF_CS17): @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("hf"), FormatVersions("cs-1.8", "cs-1.9"))
[docs]class Converter_GPTJ_LMHeadModel_CS18_CS20(BaseCheckpointConverter_CS_CS): def __init__(self): super().__init__() # Model didn't change between 1.8 and 1.9. Copy all keys. self.rules = [ ConversionRule( [ "(?:model.|)", EquivalentSubkey( "lm_head", "embedding_layer.word_embeddings" ), "\.weight", ], action=maybe_tie_lm_head, ), ConversionRule( [ "(?:model.|)", EquivalentSubkey( "embedding_layer.word_embeddings", "lm_head", ), "\.weight", ], action=maybe_tie_lm_head, ), ConversionRule( [ "(?:model.|)", EquivalentSubkey("transformer_decoder.norm", "ln_f"), "\.(?:weight|bias)", ], action=tie_none_weights, ), ConversionRule( [ "(?:model.|)", EquivalentSubkey("ln_f", "transformer_decoder.norm"), "\.(?:weight|bias)", ], action=tie_none_weights, ), ConversionRule([".*"], action=self.replaceKey), ] @classmethod def converter_note(cls) -> str: return "{} <-> {} GPTJModel".format(cls.formats()[0], cls.formats()[1]) @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("cs-1.8", "cs-1.9"), FormatVersions("cs-2.0")) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_CS18_CS20
[docs]class ConfigConverter_GPTJModel_CS18_CS20(BaseConfigConverter_CS_CS): def __init__(self): super().__init__() # Only difference between 1.8/1.9 and 2.0 is introduction of norm_type self.rules = [ ConversionRule( [EquivalentSubkey("use_biasless_norm", "norm_type")], action=self.convert_use_biasless_layer_norm, ), ConversionRule([".*"], action=self.replaceKey), ] def convert_use_biasless_layer_norm(self, *args): convert_use_biasless_layer_norm_helper(self, *args) @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("cs-1.8", "cs-1.9"), FormatVersions("cs-2.0"))
[docs]class Converter_GPTJ_Headless_HF_CS20(Converter_GPTJ_Headless_HF_CS18): @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return ( FormatVersions("hf"), FormatVersions("cs-2.0", "cs-2.1", "cs-2.2"), ) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS20
# Despite the embedding layer being refactored in release 2.1, the 2.0 # HF <> CS converter also works for 2.1 and 2.2. This is because the HF # GPT-J style models only support RoPE embeddings which don't store any # learnable params in the checkpoint."
[docs]class Converter_GPTJ_LMHeadModel_HF_CS20(Converter_GPTJ_LMHeadModel_HF_CS18): @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return ( FormatVersions("hf"), FormatVersions("cs-2.0", "cs-2.1", "cs-2.2"), ) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS20
[docs]class ConfigConverter_GPTJModel_HF_CS20(ConfigConverter_GPTJModel_HF_CS18): def __init__(self): super().__init__() self.rules = [ ConversionRule( ["norm_type"], action=BaseConfigConverter.assert_factory_fn(1, "layernorm"), ), *self.rules, ] @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return ( FormatVersions("hf"), FormatVersions("cs-2.0", "cs-2.1", "cs-2.2"), )
[docs]class Converter_GPTJ_Headless_HF_CS23(Converter_GPTJ_Headless_HF_CS20): def supports_mup_conversion(self): return True @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS23 @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return ( FormatVersions("hf"), FormatVersions("cs-2.3"), )
[docs]class Converter_GPTJ_LMHeadModel_HF_CS23(Converter_GPTJ_LMHeadModel_HF_CS20): def supports_mup_conversion(self): return True @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_HF_CS23 @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return ( FormatVersions("hf"), FormatVersions("cs-2.3"), )
[docs]class ConfigConverter_GPTJModel_HF_CS23(ConfigConverter_GPTJModel_HF_CS20): def supports_mup_conversion(self): return True @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return ( FormatVersions("hf"), FormatVersions("cs-2.3"), )
########################################################### # In CS 2.1, we refactored the embedding layer. # CS 2.0 <> CS 2.1. We don't need a separate HF <> CS 2.1 converters since # HF only supports RoPE which doesn't produce any checkpoint keys. ###########################################################
[docs]class Converter_GPTJ_LMHeadModel_CS20_CS21(BaseCheckpointConverter_CS_CS): def __init__(self): super().__init__() self.rules = [ # Refactored embeddings: ConversionRule( [ "(?:model\.|)", EquivalentSubkey( "embedding_layer.position_embeddings.weight", "embedding_layer.position_embeddings.embed.weight", ), ], action=self.replaceKey, ), ConversionRule( [ "(?:model\.|)", "embedding_layer\.", EquivalentSubkey( "position_embeddings", "position_embeddings.fpe", ), ], action=self.replaceKey, ), ConversionRule( [ "(?:model\.|)", EquivalentSubkey( "relative_pe_helper.relative_attention_bias", "embedding_layer.position_embed_helper.relative_attention_bias", ), "\.(?:weight|bias)", ], action=self.replaceKey, ), ConversionRule( [ "(?:model\.|)", EquivalentSubkey( "relative_pe_helper.slopes", "embedding_layer.position_embed_helper.slopes", ), ], action=self.replaceKey, ), # Copy everything else ConversionRule([".*"], action=self.replaceKey), ] @classmethod def converter_note(cls) -> str: return "GPTJLMHeadModel class" @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("cs-2.0"), FormatVersions("cs-2.1")) @staticmethod def get_config_converter_class() -> BaseConfigConverter: return ConfigConverter_GPTJModel_CS20_CS21
[docs]class ConfigConverter_GPTJModel_CS20_CS21(BaseConfigConverter_CS_CS): def __init__(self): super().__init__() # No differences in config self.rules = [ ConversionRule([".*"], action=self.replaceKey), ] @staticmethod def formats() -> Tuple[FormatVersions, FormatVersions]: return (FormatVersions("cs-2.0"), FormatVersions("cs-2.1"))