From 5c076fb4d5af598b57f39e9ea8aa45017f3a0c48 Mon Sep 17 00:00:00 2001 From: Arthur <48595927+ArthurZucker@users.noreply.github.com> Date: Wed, 9 Apr 2025 14:02:04 +0200 Subject: [PATCH] Add glm4 (#37388) * add changed * Revert "add changed" This reverts commit 0a0166a1fe80556115a49fbf0c2132de0f4f85c9. * update with NEW MODEL class called GLM4 * update * Update glm4.md * Name * style * fix copies * fixup test --------- Co-authored-by: Yuxuan Zhang <2448370773@qq.com> --- docs/source/en/_toctree.yml | 2 + docs/source/en/model_doc/glm4.md | 45 + src/transformers/__init__.py | 18 + src/transformers/models/__init__.py | 1 + .../models/auto/configuration_auto.py | 2 + src/transformers/models/auto/modeling_auto.py | 4 + .../models/auto/tokenization_auto.py | 1 + src/transformers/models/glm4/__init__.py | 27 + .../models/glm4/configuration_glm4.py | 152 +++ .../models/glm4/convert_glm4_weights_to_hf.py | 199 ++++ src/transformers/models/glm4/modeling_glm4.py | 1056 +++++++++++++++++ src/transformers/models/glm4/modular_glm4.py | 164 +++ src/transformers/utils/dummy_pt_objects.py | 35 + tests/models/glm4/__init__.py | 0 tests/models/glm4/test_modeling_glm4.py | 205 ++++ 15 files changed, 1911 insertions(+) create mode 100644 docs/source/en/model_doc/glm4.md create mode 100644 src/transformers/models/glm4/__init__.py create mode 100644 src/transformers/models/glm4/configuration_glm4.py create mode 100644 src/transformers/models/glm4/convert_glm4_weights_to_hf.py create mode 100644 src/transformers/models/glm4/modeling_glm4.py create mode 100644 src/transformers/models/glm4/modular_glm4.py create mode 100644 tests/models/glm4/__init__.py create mode 100644 tests/models/glm4/test_modeling_glm4.py diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index 6c4b7498b3..fad077f52a 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -461,6 +461,8 @@ title: Gemma2 - local: model_doc/glm title: GLM + - local: model_doc/glm4 + title: glm4 - local: model_doc/openai-gpt title: GPT - local: model_doc/gpt_neo diff --git a/docs/source/en/model_doc/glm4.md b/docs/source/en/model_doc/glm4.md new file mode 100644 index 0000000000..f854bb658f --- /dev/null +++ b/docs/source/en/model_doc/glm4.md @@ -0,0 +1,45 @@ + + +# Glm4 + +## Overview + +To be released with the official model launch. + +## Glm4Config + +[[autodoc]] Glm4Config + +## Glm4Model + +[[autodoc]] Glm4Model + - forward + +## Glm4ForCausalLM + +[[autodoc]] Glm4ForCausalLM + - forward + +## Glm4ForSequenceClassification + +[[autodoc]] Glm4ForSequenceClassification + - forward + +## Glm4ForTokenClassification + +[[autodoc]] Glm4ForTokenClassification + - forward diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index 050d79f364..3c5851b0a9 100644 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -482,6 +482,7 @@ _import_structure = { "GitVisionConfig", ], "models.glm": ["GlmConfig"], + "models.glm4": ["Glm4Config"], "models.glpn": ["GLPNConfig"], "models.got_ocr2": [ "GotOcr2Config", @@ -2526,6 +2527,15 @@ else: "Llama4PreTrainedModel", ] ) + _import_structure["models.glm4"].extend( + [ + "Glm4ForCausalLM", + "Glm4ForSequenceClassification", + "Glm4ForTokenClassification", + "Glm4Model", + "Glm4PreTrainedModel", + ] + ) _import_structure["models.glpn"].extend( [ "GLPNForDepthEstimation", @@ -5742,6 +5752,7 @@ if TYPE_CHECKING: GitVisionConfig, ) from .models.glm import GlmConfig + from .models.glm4 import Glm4Config from .models.glpn import GLPNConfig from .models.got_ocr2 import GotOcr2Config, GotOcr2Processor, GotOcr2VisionConfig from .models.gpt2 import ( @@ -7624,6 +7635,13 @@ if TYPE_CHECKING: GlmModel, GlmPreTrainedModel, ) + from .models.glm4 import ( + Glm4ForCausalLM, + Glm4ForSequenceClassification, + Glm4ForTokenClassification, + Glm4Model, + Glm4PreTrainedModel, + ) from .models.glpn import ( GLPNForDepthEstimation, GLPNModel, diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index 08cec64b41..d6063cef4b 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -110,6 +110,7 @@ from . import ( gemma3, git, glm, + glm4, glpn, got_ocr2, gpt2, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index 759b7ad3d9..82ce72cf7b 100644 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -129,6 +129,7 @@ CONFIG_MAPPING_NAMES = OrderedDict( ("gemma3_text", "Gemma3TextConfig"), ("git", "GitConfig"), ("glm", "GlmConfig"), + ("glm4", "Glm4Config"), ("glpn", "GLPNConfig"), ("got_ocr2", "GotOcr2Config"), ("gpt-sw3", "GPT2Config"), @@ -476,6 +477,7 @@ MODEL_NAMES_MAPPING = OrderedDict( ("gemma3_text", "Gemma3ForCausalLM"), ("git", "GIT"), ("glm", "GLM"), + ("glm4", "glm4"), ("glpn", "GLPN"), ("got_ocr2", "GOT-OCR2"), ("gpt-sw3", "GPT-Sw3"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index d33d0f20d5..4072ab382b 100644 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -122,6 +122,7 @@ MODEL_MAPPING_NAMES = OrderedDict( ("gemma3_text", "Gemma3TextModel"), ("git", "GitModel"), ("glm", "GlmModel"), + ("glm4", "Glm4Model"), ("glpn", "GLPNModel"), ("got_ocr2", "GotOcr2ForConditionalGeneration"), ("gpt-sw3", "GPT2Model"), @@ -532,6 +533,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict( ("gemma3_text", "Gemma3ForCausalLM"), ("git", "GitForCausalLM"), ("glm", "GlmForCausalLM"), + ("glm4", "Glm4ForCausalLM"), ("got_ocr2", "GotOcr2ForConditionalGeneration"), ("gpt-sw3", "GPT2LMHeadModel"), ("gpt2", "GPT2LMHeadModel"), @@ -1035,6 +1037,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict( ("gemma", "GemmaForSequenceClassification"), ("gemma2", "Gemma2ForSequenceClassification"), ("glm", "GlmForSequenceClassification"), + ("glm4", "Glm4ForSequenceClassification"), ("gpt-sw3", "GPT2ForSequenceClassification"), ("gpt2", "GPT2ForSequenceClassification"), ("gpt_bigcode", "GPTBigCodeForSequenceClassification"), @@ -1236,6 +1239,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict( ("gemma", "GemmaForTokenClassification"), ("gemma2", "Gemma2ForTokenClassification"), ("glm", "GlmForTokenClassification"), + ("glm4", "Glm4ForTokenClassification"), ("gpt-sw3", "GPT2ForTokenClassification"), ("gpt2", "GPT2ForTokenClassification"), ("gpt_bigcode", "GPTBigCodeForTokenClassification"), diff --git a/src/transformers/models/auto/tokenization_auto.py b/src/transformers/models/auto/tokenization_auto.py index eb54d95ab6..eda7356a60 100644 --- a/src/transformers/models/auto/tokenization_auto.py +++ b/src/transformers/models/auto/tokenization_auto.py @@ -238,6 +238,7 @@ else: ), ("git", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)), ("glm", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)), + ("glm4", (None, "PreTrainedTokenizerFast" if is_tokenizers_available() else None)), ("gpt-sw3", ("GPTSw3Tokenizer" if is_sentencepiece_available() else None, None)), ("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)), ("gpt_bigcode", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)), diff --git a/src/transformers/models/glm4/__init__.py b/src/transformers/models/glm4/__init__.py new file mode 100644 index 0000000000..6e92a8a2b9 --- /dev/null +++ b/src/transformers/models/glm4/__init__.py @@ -0,0 +1,27 @@ +# Copyright 2025 The HuggingFace 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_glm4 import * + from .modeling_glm4 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/glm4/configuration_glm4.py b/src/transformers/models/glm4/configuration_glm4.py new file mode 100644 index 0000000000..3632a7b6b4 --- /dev/null +++ b/src/transformers/models/glm4/configuration_glm4.py @@ -0,0 +1,152 @@ +# coding=utf-8 +# Copyright 2025 The GLM4 & ZhipuAI team and 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 ...configuration_utils import PretrainedConfig + + +class Glm4Config(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`Glm4Model`]. It is used to instantiate an Glm4 + 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 Glm4-4-9b-chat. + e.g. [THUDM/glm-4-0414-9b-chat-chat](https://huggingface.co/THUDM/glm-4-0414-9b-chat-chat) + 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 151552): + Vocabulary size of the Glm4 model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`Glm4Model`] + hidden_size (`int`, *optional*, defaults to 4096): + Dimension of the hidden representations. + intermediate_size (`int`, *optional*, defaults to 13696): + Dimension of the MLP representations. + num_hidden_layers (`int`, *optional*, defaults to 40): + Number of hidden layers in the Transformer decoder. + num_attention_heads (`int`, *optional*, defaults to 32): + Number of attention heads for each attention layer in the Transformer decoder. + num_key_value_heads (`int`, *optional*, defaults to 2): + 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 + `num_attention_heads`. + partial_rotary_factor (`float`, *optional*, defaults to 0.5): The factor of the partial rotary position. + head_dim (`int`, *optional*, defaults to 128): + The attention head dimension. + hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): + The legacy activation function. It is overwritten by the `hidden_activation`. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for the attention probabilities. + max_position_embeddings (`int`, *optional*, defaults to 131072): + The maximum sequence length that this model might ever be used with. + 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 1.5625e-07): + 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`. + tie_word_embeddings (`bool`, *optional*, defaults to `False`): + Whether to tie weight embeddings + rope_theta (`float`, *optional*, defaults to 10000.0): + The base period of the RoPE embeddings. + pad_token_id (`int`, *optional*, defaults to 151329): + Padding token id. + eos_token_id (`int` | `list`, *optional*, defaults to `[151329, 151336, 151338]`): + End of stream token id. + bos_token_id (`int`, *optional*): + Beginning of stream token id. + attention_bias (`bool`, defaults to `False`, *optional*, defaults to `True`): + Whether to use a bias in the query, key, value and output projection layers during self-attention. + ```python + >>> from transformers import Glm4Model, Glm4Config + >>> # Initializing a Glm4 glm4-4-9b-chat style configuration + >>> configuration = Glm4Config() + >>> # Initializing a model from the glm4-4-9b-chat style configuration + >>> model = Glm4Model(configuration) + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + + model_type = "glm4" + keys_to_ignore_at_inference = ["past_key_values"] + base_model_tp_plan = { + "layers.*.self_attn.q_proj": "colwise", + "layers.*.self_attn.k_proj": "colwise", + "layers.*.self_attn.v_proj": "colwise", + "layers.*.self_attn.o_proj": "rowwise", + "layers.*.mlp.gate_up_proj": "colwise_rep", # we need to replicate here due to the `chunk` operation + "layers.*.mlp.down_proj": "rowwise_rep", # we need to replicate here due to the `chunk` operation + } + base_model_pp_plan = { + "embed_tokens": (["input_ids"], ["inputs_embeds"]), + "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), + "norm": (["hidden_states"], ["hidden_states"]), + } + + def __init__( + self, + vocab_size=151552, + hidden_size=4096, + intermediate_size=13696, + num_hidden_layers=40, + num_attention_heads=32, + num_key_value_heads=2, + partial_rotary_factor=0.5, + head_dim=128, + hidden_act="silu", + attention_dropout=0.0, + max_position_embeddings=131072, + initializer_range=0.02, + rms_norm_eps=0.00000015625, + use_cache=True, + tie_word_embeddings=False, + rope_theta=10000.0, + pad_token_id=151329, + eos_token_id=[151329, 151336, 151338], + bos_token_id=None, + attention_bias=True, + **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.partial_rotary_factor = partial_rotary_factor + self.head_dim = head_dim + 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_bias = attention_bias + self.attention_dropout = attention_dropout + + 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__ = ["Glm4Config"] diff --git a/src/transformers/models/glm4/convert_glm4_weights_to_hf.py b/src/transformers/models/glm4/convert_glm4_weights_to_hf.py new file mode 100644 index 0000000000..01ad00f517 --- /dev/null +++ b/src/transformers/models/glm4/convert_glm4_weights_to_hf.py @@ -0,0 +1,199 @@ +import argparse +import json +import os +import re + +import torch +from safetensors.torch import load_file +from tokenizers import processors + +from transformers import Glm4Config, Glm4ForCausalLM, PreTrainedTokenizerFast + + +# fmt: off +# `None` means we drop the key +STATE_DICT_MAPPING = { + # CausalLM keys + r"transformer.output_layer.weight": r"lm_head.weight", + + # Model keys + r"transformer.embedding.word_embeddings.weight": r"model.embed_tokens.weight", + r"transformer.rotary_pos_emb.inv_freq": None, + r"transformer.encoder.final_layernorm.weight": r"model.norm.weight", + + # Layers keys + r"transformer.encoder.layers.(\d+).input_layernorm.weight": r"model.layers.\1.input_layernorm.weight", + + # Sandwich keys + r"transformer.encoder.layers.(\d+).post_mlp_layernorm.weight": r"model.layers.\1.post_mlp_layernorm.weight", + r"transformer.encoder.layers.(\d+).post_self_attn_layernorm.weight": r"model.layers.\1.post_self_attn_layernorm.weight", + + r"transformer.encoder.layers.(\d+).post_attention_layernorm.weight": r"model.layers.\1.post_attention_layernorm.weight", + + # Attention keys + r"transformer.encoder.layers.(\d+).self_attention.dense.weight": r"model.layers.\1.self_attn.o_proj.weight", + # qkv_proj will later be split in q|k|v|_proj + r"transformer.encoder.layers.(\d+).self_attention.query_key_value.(weight|bias)": r"model.layers.\1.self_attn.qkv_proj.\2", + + # MLP keys + r"transformer.encoder.layers.(\d+).mlp.dense_h_to_4h.weight": r"model.layers.\1.mlp.gate_up_proj.weight", + r"transformer.encoder.layers.(\d+).mlp.dense_4h_to_h.weight": r"model.layers.\1.mlp.down_proj.weight", +} +# fmt: on + + +def load_weights(input_dir: str): + safetensor_files = [os.path.join(input_dir, x) for x in os.listdir(input_dir) if x.endswith(".safetensors")] + bin_files = [os.path.join(input_dir, x) for x in os.listdir(input_dir) if x.endswith(".bin")] + + all_weights = {} + + if safetensor_files: + safetensor_files = sorted(safetensor_files, key=lambda x: int(x.rsplit("-", 3)[1])) + for file in safetensor_files: + tensors = load_file(file) + all_weights.update(tensors) + return all_weights + + elif bin_files: + bin_files = sorted(bin_files, key=lambda x: int(x.rsplit("-", 3)[1])) + for file in bin_files: + tensors = torch.load(file, map_location="cpu") + all_weights.update(tensors) + return all_weights + + else: + raise ValueError("No .safetensors or .bin files found in the specified directory.") + + +def map_old_key_to_new(old_key): + for pattern, replacement in STATE_DICT_MAPPING.items(): + if replacement is None: + if re.fullmatch(pattern, old_key): + return None + else: + new_key, n_replace = re.subn(pattern, replacement, old_key) + # Early exit of the loop + if n_replace > 0: + return new_key + + raise ValueError(f"Key: {old_key} could not be mapped (check the mapping).") + + +def convert_state_dict(original_state_dict: dict, config: Glm4Config): + new_dict = {} + + head_dim = config.hidden_size // config.num_attention_heads + query_size = config.num_attention_heads * head_dim + kv_size = config.num_key_value_heads * head_dim + + for old_key, value in original_state_dict.items(): + new_key = map_old_key_to_new(old_key) + if new_key is None: + continue + + if "qkv_proj." in new_key: + q_proj, k_proj, v_proj = ( + value[:query_size, ...], + value[query_size : query_size + kv_size, ...], + value[query_size + kv_size :, ...], + ) + new_dict[new_key.replace("qkv_proj.", "q_proj.")] = q_proj + new_dict[new_key.replace("qkv_proj.", "k_proj.")] = k_proj + new_dict[new_key.replace("qkv_proj.", "v_proj.")] = v_proj + else: + new_dict[new_key] = value + return new_dict + + +def convert_config(original_config: dict): + key_mapping = { + "vocab_size": "padded_vocab_size", + "intermediate_size": "ffn_hidden_size", + "num_hidden_layers": "num_layers", + "max_position_embeddings": "seq_length", + "rms_norm_eps": "layernorm_epsilon", + "head_dim": "kv_channels", + "attention_bias": "add_qkv_bias", + } + similar_keys_to_keep = [ + "num_attention_heads", + "hidden_size", + "attention_dropout", + "use_cache", + "eos_token_id", + "pad_token_id", + "tie_word_embeddings", + ] + new_config_kwargs = {k: original_config[v] for k, v in key_mapping.items()} + new_config_kwargs.update({k: v for k, v in original_config.items() if k in similar_keys_to_keep}) + new_config_kwargs["num_key_value_heads"] = ( + new_config_kwargs["num_attention_heads"] + if not original_config["multi_query_attention"] + else original_config["multi_query_group_num"] + ) + new_config_kwargs["rope_theta"] = 10000.0 * getattr(original_config, "rope_ratio", 1) + + new_config = Glm4Config(**new_config_kwargs) + return new_config + + +def convert_glm4_tokenizer(input_dir, use_post_processor=False): + fast_tok = PreTrainedTokenizerFast.from_pretrained(input_dir, model_input_names=["input_ids", "attention_mask"]) + if use_post_processor: + fast_tok._tokenizer.post_processor = processors.Sequence( + [ + processors.ByteLevel(trim_offsets=False), + processors.TemplateProcessing( + single="[gMASK]:0 :0 $A:0", + pair="[gMASK]:0 :0 $A:0 $B:1", + special_tokens=[("[gMASK]", 151331), ("", 151333)], + ), + ], + ) + else: + fast_tok._tokenizer.post_processor = processors.Sequence( + [processors.ByteLevel(trim_offsets=False)], + ) + return fast_tok + + +def convert_glm4_model(input_dir, output_dir, use_post_processor=False): + # Load and convert config + with open(os.path.join(input_dir, "config.json")) as f: + original_config = json.load(f) + config = convert_config(original_config) + config.save_pretrained(output_dir) + + # Load and convert weights + original_state_dict = load_weights(input_dir) + new_dict = convert_state_dict(original_state_dict, config) + with torch.device("meta"): + model = Glm4ForCausalLM(config) + model.load_state_dict(new_dict, strict=True, assign=True) + model.save_pretrained(output_dir) + + # Load and convert tokenizer + tokenizer = convert_glm4_tokenizer(input_dir, use_post_processor) + tokenizer.save_pretrained(output_dir) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + parser.add_argument( + "input_dir", + type=str, + help="Location of the local folder copied from the Hub.", + ) + parser.add_argument( + "output_dir", + type=str, + help="Location to write HF model and tokenizer", + ) + parser.add_argument( + "--use_post_processor", + action="store_true", + help="Whether to apply post processor with special tokens", + ) + args = parser.parse_args() + convert_glm4_model(args.input_dir, args.output_dir, args.use_post_processor) diff --git a/src/transformers/models/glm4/modeling_glm4.py b/src/transformers/models/glm4/modeling_glm4.py new file mode 100644 index 0000000000..95c1c6543c --- /dev/null +++ b/src/transformers/models/glm4/modeling_glm4.py @@ -0,0 +1,1056 @@ +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# This file was automatically generated from src/transformers/models/glm4/modular_glm4.py. +# Do NOT edit this file manually as any edits will be overwritten by the generation of +# the file from the modular. If any change should be done, please apply the change to the +# modular_glm4.py file directly. One of our CI enforces this. +# 🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨 +# coding=utf-8 +# Copyright 2025 The GLM4 & ZhipuAI team and 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 functools import partial +from typing import Callable, Optional, Tuple, Union + +import torch +import torch.nn as nn + +from ...activations import ACT2FN +from ...cache_utils import Cache, DynamicCache, StaticCache +from ...generation import GenerationMixin +from ...modeling_attn_mask_utils import AttentionMaskConverter +from ...modeling_flash_attention_utils import FlashAttentionKwargs +from ...modeling_outputs import ( + BaseModelOutputWithPast, + CausalLMOutputWithPast, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update +from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel +from ...processing_utils import Unpack +from ...utils import ( + LossKwargs, + add_code_sample_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, + can_return_tuple, + is_torch_flex_attn_available, + logging, + replace_return_docstrings, +) +from ...utils.deprecation import deprecate_kwarg +from .configuration_glm4 import Glm4Config + + +if is_torch_flex_attn_available(): + from torch.nn.attention.flex_attention import BlockMask + + from ...integrations.flex_attention import make_flex_block_causal_mask + + +logger = logging.get_logger(__name__) + +_CHECKPOINT_FOR_DOC = "THUDM/GLM-4-9B-Chat-0414" +_CONFIG_FOR_DOC = "Glm4Config" + + +class Glm4MLP(nn.Module): + def __init__(self, config): + super().__init__() + + self.config = config + self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False) + self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False) + self.activation_fn = ACT2FN[config.hidden_act] + + def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor: + up_states = self.gate_up_proj(hidden_states) + + gate, up_states = up_states.chunk(2, dim=-1) + up_states = up_states * self.activation_fn(gate) + + return self.down_proj(up_states) + + +class Glm4DecoderLayer(nn.Module): + def __init__(self, config: Glm4Config, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + self.self_attn = Glm4Attention(config=config, layer_idx=layer_idx) + + self.mlp = Glm4MLP(config) + self.input_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_self_attn_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_mlp_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights = 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, + **kwargs, + ) + + hidden_states = self.post_self_attn_layernorm(hidden_states) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = self.post_mlp_layernorm(hidden_states) + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + if output_attentions: + outputs += (self_attn_weights,) + + return outputs + + +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) + + +def eager_attention_forward( + module: nn.Module, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + attention_mask: Optional[torch.Tensor], + scaling: float, + dropout: float = 0.0, + **kwargs, +): + key_states = repeat_kv(key, module.num_key_value_groups) + value_states = repeat_kv(value, module.num_key_value_groups) + + attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling + if attention_mask is not None: + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask + + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) + attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) + attn_output = torch.matmul(attn_weights, value_states) + attn_output = attn_output.transpose(1, 2).contiguous() + + return attn_output, attn_weights + + +def rotate_half(x): + """Rotates half the hidden dims of the input.""" + x1 = x[..., 0::2] + x2 = x[..., 1::2] + return torch.stack((-x2, x1), dim=-1).flatten(-2) + + +def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, 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`, *optional*): + Deprecated and unused. + 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.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + + # Interleave them instead of usual shape + cos = cos[..., : cos.shape[-1] // 2].repeat_interleave(2, dim=-1) + sin = sin[..., : sin.shape[-1] // 2].repeat_interleave(2, dim=-1) + + # Keep half or full tensor for later concatenation + rotary_dim = cos.shape[-1] + q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] + k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] + + # Apply rotary embeddings on the first half or full tensor + q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin) + k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin) + + # Concatenate back to full shape + q_embed = torch.cat([q_embed, q_pass], dim=-1) + k_embed = torch.cat([k_embed, k_pass], dim=-1) + return q_embed, k_embed + + +class Glm4Attention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__(self, config: Glm4Config, layer_idx: Optional[int] = None): + super().__init__() + self.config = config + self.layer_idx = layer_idx + self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) + self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads + self.scaling = self.head_dim**-0.5 + self.attention_dropout = config.attention_dropout + self.is_causal = True + + self.q_proj = nn.Linear( + config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias + ) + self.k_proj = nn.Linear( + config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias + ) + self.v_proj = nn.Linear( + config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias + ) + self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False) + + def forward( + self, + hidden_states: torch.Tensor, + position_embeddings: Tuple[torch.Tensor, torch.Tensor], + attention_mask: Optional[torch.Tensor], + past_key_value: Optional[Cache] = None, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + input_shape = hidden_states.shape[:-1] + hidden_shape = (*input_shape, -1, self.head_dim) + + query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2) + key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2) + value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + # sin and cos are specific to RoPE models; cache_position needed for the static cache + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + attention_interface: Callable = eager_attention_forward + if self.config._attn_implementation != "eager": + if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False): + logger.warning_once( + "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to " + 'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' + ) + else: + attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] + + attn_output, attn_weights = attention_interface( + self, + query_states, + key_states, + value_states, + attention_mask, + dropout=0.0 if not self.training else self.attention_dropout, + scaling=self.scaling, + **kwargs, + ) + + attn_output = attn_output.reshape(*input_shape, -1).contiguous() + attn_output = self.o_proj(attn_output) + return attn_output, attn_weights + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +class Glm4RMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + Glm4RMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + +class Glm4RotaryEmbedding(nn.Module): + def __init__(self, config: Glm4Config, device=None): + super().__init__() + # BC: "rope_type" was originally "type" + if hasattr(config, "rope_scaling") and config.rope_scaling is not None: + self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) + else: + self.rope_type = "default" + self.max_seq_len_cached = config.max_position_embeddings + self.original_max_seq_len = config.max_position_embeddings + + self.config = config + self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] + + inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self.original_inv_freq = self.inv_freq + + @torch.no_grad() + @dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope) + def forward(self, x, position_ids): + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) + position_ids_expanded = position_ids[:, None, :].float() + + device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" + with torch.autocast(device_type=device_type, enabled=False): # Force float32 + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() * self.attention_scaling + sin = emb.sin() * self.attention_scaling + + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +GLM4_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 ([`Glm4Config`]): + 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 Glm4 Model outputting raw hidden-states without any specific head on top.", + GLM4_START_DOCSTRING, +) +class Glm4PreTrainedModel(PreTrainedModel): + config_class = Glm4Config + base_model_prefix = "model" + supports_gradient_checkpointing = True + _no_split_modules = ["Glm4DecoderLayer"] + _skip_keys_device_placement = ["past_key_values"] + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_flex_attn = True + _supports_cache_class = True + _supports_quantized_cache = True + _supports_static_cache = True + _supports_attention_backend = 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_() + + +GLM4_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 `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 (`Cache`, *optional*): + Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` + returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. + + It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). + + If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't + have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `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. + 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 Glm4 Model outputting raw hidden-states without any specific head on top.", + GLM4_START_DOCSTRING, +) +class Glm4Model(Glm4PreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Glm4DecoderLayer`] + + Args: + config: Glm4Config + """ + + def __init__(self, config: Glm4Config): + 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( + [Glm4DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self.norm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.rotary_emb = Glm4RotaryEmbedding(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 + + @can_return_tuple + @add_start_docstrings_to_model_forward(GLM4_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[Cache] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + **flash_attn_kwargs: Unpack[FlashAttentionKwargs], + ) -> BaseModelOutputWithPast: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError("You must specify exactly one of input_ids or inputs_embeds") + + if self.gradient_checkpointing and self.training and use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." + ) + use_cache = False + + # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache + if not isinstance(past_key_values, (type(None), Cache)): + raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.") + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if use_cache and past_key_values is None: + past_key_values = DynamicCache() + + 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 + + # create position embeddings to be shared across the decoder layers + position_embeddings = self.rotary_emb(hidden_states, position_ids) + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + + for decoder_layer in self.layers[: self.config.num_hidden_layers]: + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if self.gradient_checkpointing and self.training: + layer_outputs = self._gradient_checkpointing_func( + partial(decoder_layer.__call__, **flash_attn_kwargs), + hidden_states, + causal_mask, + position_ids, + past_key_values, + output_attentions, + 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, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **flash_attn_kwargs, + ) + + hidden_states = layer_outputs[0] + + if output_attentions: + all_self_attns += (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,) + + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=past_key_values if use_cache else None, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool = False, + ): + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and (attention_mask == 0.0).any(): + return attention_mask + return None + if self.config._attn_implementation == "flex_attention": + if isinstance(attention_mask, torch.Tensor): + attention_mask = make_flex_block_causal_mask(attention_mask) + if isinstance(attention_mask, BlockMask): + return attention_mask + + # 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_cache_shape() + 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 in ["cuda", "xpu"] + 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 + 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, + **kwargs, + ): + """ + 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 place 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, :].to( + causal_mask.device + ) + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + + return causal_mask + + +class Glm4ForCausalLM(Glm4PreTrainedModel, GenerationMixin): + _tied_weights_keys = ["lm_head.weight"] + _tp_plan = {"lm_head": "colwise_rep"} + _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} + + def __init__(self, config): + super().__init__(config) + self.model = Glm4Model(config) + self.vocab_size = config.vocab_size + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, 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 + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + @can_return_tuple + @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep") + @add_start_docstrings_to_model_forward(GLM4_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) + 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[Cache] = 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, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + **kwargs: Unpack[KwargsForCausalLM], + ) -> Union[Tuple, CausalLMOutputWithPast]: + r""" + 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]`. + + logits_to_keep (`int` or `torch.Tensor`, *optional*): + If an `int`, compute logits for the last `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. + If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. + This is useful when using packed tensor format (single dimension for batch and sequence length). + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, Glm4ForCausalLM + + >>> model = Glm4ForCausalLM.from_pretrained("THUDM/GLM-4-9B-Chat-0414") + >>> tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4-9B-Chat-0414") + + >>> prompt = "Hey, are you conscious? Can you talk to me?" + >>> inputs = tokenizer(prompt, return_tensors="pt") + + >>> # Generate + >>> generate_ids = model.generate(inputs.input_ids, max_length=30) + >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] + "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." + ```""" + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + outputs: BaseModelOutputWithPast = 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, + cache_position=cache_position, + **kwargs, + ) + + hidden_states = outputs.last_hidden_state + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep + logits = self.lm_head(hidden_states[:, slice_indices, :]) + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) + + return CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +@add_start_docstrings( + """ + The Glm4 Model transformer with a sequence classification head on top (linear layer). + + [`Glm4ForSequenceClassification`] 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). + """, + GLM4_START_DOCSTRING, +) +class Glm4ForSequenceClassification(Glm4PreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = Glm4Model(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 + + @can_return_tuple + @add_start_docstrings_to_model_forward(GLM4_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[Cache] = 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, + ) -> 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). + """ + + transformer_outputs: BaseModelOutputWithPast = 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, + ) + hidden_states = transformer_outputs.last_hidden_state + 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: + last_non_pad_token = -1 + elif input_ids is not None: + # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id + non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32) + token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32) + last_non_pad_token = (token_indices * non_pad_mask).argmax(-1) + else: + last_non_pad_token = -1 + logger.warning_once( + f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " + "unexpected if using padding tokens in conjunction with `inputs_embeds.`" + ) + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token] + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) + + 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, + ) + + +@add_start_docstrings( + """ + The Glm4 Model transformer with a token classification head on top (a linear layer on top of the hidden-states + output) e.g. for Named-Entity-Recognition (NER) tasks. + """, + GLM4_START_DOCSTRING, +) +class Glm4ForTokenClassification(Glm4PreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = Glm4Model(config) + if getattr(config, "classifier_dropout", None) is not None: + classifier_dropout = config.classifier_dropout + elif getattr(config, "hidden_dropout", None) is not None: + classifier_dropout = config.hidden_dropout + else: + classifier_dropout = 0.1 + self.dropout = nn.Dropout(classifier_dropout) + self.score = nn.Linear(config.hidden_size, config.num_labels) + + # 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 + + @can_return_tuple + @add_start_docstrings_to_model_forward(GLM4_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=TokenClassifierOutput, + config_class=_CONFIG_FOR_DOC, + ) + 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[Cache] = 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, + ) -> TokenClassifierOutput: + 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). + """ + + outputs: BaseModelOutputWithPast = 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, + ) + sequence_output = outputs.last_hidden_state + sequence_output = self.dropout(sequence_output) + logits = self.score(sequence_output) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.config) + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +__all__ = [ + "Glm4PreTrainedModel", + "Glm4Model", + "Glm4ForCausalLM", + "Glm4ForSequenceClassification", + "Glm4ForTokenClassification", +] diff --git a/src/transformers/models/glm4/modular_glm4.py b/src/transformers/models/glm4/modular_glm4.py new file mode 100644 index 0000000000..493868f3b3 --- /dev/null +++ b/src/transformers/models/glm4/modular_glm4.py @@ -0,0 +1,164 @@ +# coding=utf-8 +# Copyright 2025 The GLM4 & ZhipuAI team and 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 Optional, Tuple, Union + +import torch.nn as nn +import torch.utils.checkpoint + +from ...cache_utils import Cache +from ...modeling_flash_attention_utils import FlashAttentionKwargs +from ...modeling_outputs import CausalLMOutputWithPast +from ...processing_utils import Unpack +from ...utils import LossKwargs, logging +from ..glm.modeling_glm import ( + GlmAttention, + GlmForCausalLM, + GlmForSequenceClassification, + GlmForTokenClassification, +) +from ..phi3.modeling_phi3 import Phi3MLP +from .configuration_glm4 import Glm4Config +from .modeling_glm4 import Glm4RMSNorm + + +logger = logging.get_logger(__name__) + +_CHECKPOINT_FOR_DOC = "THUDM/GLM-4-9B-Chat-0414" + + +class Glm4MLP(Phi3MLP): + pass + + +class Glm4DecoderLayer(nn.Module): + def __init__(self, config: Glm4Config, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + self.self_attn = Glm4Attention(config=config, layer_idx=layer_idx) + + self.mlp = Glm4MLP(config) + self.input_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_self_attn_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_mlp_layernorm = Glm4RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights = 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, + **kwargs, + ) + + hidden_states = self.post_self_attn_layernorm(hidden_states) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = self.post_mlp_layernorm(hidden_states) + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + if output_attentions: + outputs += (self_attn_weights,) + + return outputs + + +class Glm4Attention(GlmAttention): + pass + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +class Glm4ForCausalLM(GlmForCausalLM): + def forward( + self, + **super_kwargs: Unpack[KwargsForCausalLM], + ) -> Union[Tuple, CausalLMOutputWithPast]: + r""" + 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]`. + + logits_to_keep (`int` or `torch.Tensor`, *optional*): + If an `int`, compute logits for the last `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. + If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. + This is useful when using packed tensor format (single dimension for batch and sequence length). + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, Glm4ForCausalLM + + >>> model = Glm4ForCausalLM.from_pretrained("THUDM/GLM-4-9B-Chat-0414") + >>> tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4-9B-Chat-0414") + + >>> 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." + ```""" + return super().forward(**super_kwargs) + + +class Glm4ForSequenceClassification(GlmForSequenceClassification): + pass + + +class Glm4ForTokenClassification(GlmForTokenClassification): + pass + + +__all__ = [ + "Glm4PreTrainedModel", # noqa: F822 + "Glm4Model", # noqa: F822 + "Glm4ForCausalLM", + "Glm4ForSequenceClassification", + "Glm4ForTokenClassification", +] diff --git a/src/transformers/utils/dummy_pt_objects.py b/src/transformers/utils/dummy_pt_objects.py index b03455c89e..ad9f373c1b 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -4740,6 +4740,41 @@ class GlmPreTrainedModel(metaclass=DummyObject): requires_backends(self, ["torch"]) +class Glm4ForCausalLM(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class Glm4ForSequenceClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class Glm4ForTokenClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class Glm4Model(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class Glm4PreTrainedModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + class GLPNForDepthEstimation(metaclass=DummyObject): _backends = ["torch"] diff --git a/tests/models/glm4/__init__.py b/tests/models/glm4/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/tests/models/glm4/test_modeling_glm4.py b/tests/models/glm4/test_modeling_glm4.py new file mode 100644 index 0000000000..547b696867 --- /dev/null +++ b/tests/models/glm4/test_modeling_glm4.py @@ -0,0 +1,205 @@ +# coding=utf-8 +# Copyright 2025 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 Glm4 model.""" + +import unittest + +import pytest + +from transformers import AutoModelForCausalLM, AutoTokenizer, Glm4Config, is_torch_available +from transformers.testing_utils import ( + require_flash_attn, + require_torch, + require_torch_large_gpu, + require_torch_sdpa, + slow, + torch_device, +) + +from ...models.gemma.test_modeling_gemma import GemmaModelTest, GemmaModelTester +from ...test_configuration_common import ConfigTester + + +if is_torch_available(): + import torch + + from transformers import ( + Glm4ForCausalLM, + Glm4ForSequenceClassification, + Glm4ForTokenClassification, + Glm4Model, + ) + + +class Glm4ModelTester(GemmaModelTester): + if is_torch_available(): + config_class = Glm4Config + model_class = Glm4Model + for_causal_lm_class = Glm4ForCausalLM + for_sequence_class = Glm4ForSequenceClassification + for_token_class = Glm4ForTokenClassification + + +@require_torch +class Glm4ModelTest(GemmaModelTest, unittest.TestCase): + all_model_classes = ( + (Glm4Model, Glm4ForCausalLM, Glm4ForSequenceClassification, Glm4ForTokenClassification) + if is_torch_available() + else () + ) + pipeline_model_mapping = ( + { + "feature-extraction": Glm4Model, + "text-classification": Glm4ForSequenceClassification, + "token-classification": Glm4ForTokenClassification, + "text-generation": Glm4ForCausalLM, + "zero-shot": Glm4ForSequenceClassification, + } + if is_torch_available() + else {} + ) + test_headmasking = False + test_pruning = False + _is_stateful = True + model_split_percents = [0.5, 0.6] + + def setUp(self): + self.model_tester = Glm4ModelTester(self) + self.config_tester = ConfigTester(self, config_class=Glm4Config, hidden_size=37) + + +@slow +@require_torch_large_gpu +class Glm4IntegrationTest(unittest.TestCase): + input_text = ["Hello I am doing", "Hi today"] + model_id = "THUDM/glm-4-0414-9b-chat" + revision = "refs/pr/15" + # This variable is used to determine which CUDA device are we using for our runners (A10 or T4) + # Depending on the hardware we get different logits / generations + cuda_compute_capability_major_version = None + + @classmethod + def setUpClass(cls): + if is_torch_available() and torch.cuda.is_available(): + # 8 is for A100 / A10 and 7 for T4 + cls.cuda_compute_capability_major_version = torch.cuda.get_device_capability()[0] + + def test_model_9b_fp16(self): + EXPECTED_TEXTS = [ + "Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the", + "Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.", + ] + + model = AutoModelForCausalLM.from_pretrained( + self.model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16, revision=self.revision + ).to(torch_device) + + tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision) + inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device) + + output = model.generate(**inputs, max_new_tokens=20, do_sample=False) + output_text = tokenizer.batch_decode(output, skip_special_tokens=True) + + self.assertEqual(output_text, EXPECTED_TEXTS) + + def test_model_9b_bf16(self): + EXPECTED_TEXTS = [ + "Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the", + "Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.", + ] + + model = AutoModelForCausalLM.from_pretrained( + self.model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16, revision=self.revision + ).to(torch_device) + + tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision) + inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device) + + output = model.generate(**inputs, max_new_tokens=20, do_sample=False) + output_text = tokenizer.batch_decode(output, skip_special_tokens=True) + + self.assertEqual(output_text, EXPECTED_TEXTS) + + def test_model_9b_eager(self): + EXPECTED_TEXTS = [ + "Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the", + "Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.", + ] + + model = AutoModelForCausalLM.from_pretrained( + self.model_id, + low_cpu_mem_usage=True, + torch_dtype=torch.bfloat16, + attn_implementation="eager", + revision=self.revision, + ) + model.to(torch_device) + + tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision) + inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device) + + output = model.generate(**inputs, max_new_tokens=20, do_sample=False) + output_text = tokenizer.batch_decode(output, skip_special_tokens=True) + + self.assertEqual(output_text, EXPECTED_TEXTS) + + @require_torch_sdpa + def test_model_9b_sdpa(self): + EXPECTED_TEXTS = [ + "Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the", + "Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.", + ] + + model = AutoModelForCausalLM.from_pretrained( + self.model_id, + low_cpu_mem_usage=True, + torch_dtype=torch.bfloat16, + attn_implementation="sdpa", + revision=self.revision, + ) + model.to(torch_device) + + tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision) + inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device) + + output = model.generate(**inputs, max_new_tokens=20, do_sample=False) + output_text = tokenizer.batch_decode(output, skip_special_tokens=True) + + self.assertEqual(output_text, EXPECTED_TEXTS) + + @require_flash_attn + @pytest.mark.flash_attn_test + def test_model_9b_flash_attn(self): + EXPECTED_TEXTS = [ + "Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the", + "Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.", + ] + + model = AutoModelForCausalLM.from_pretrained( + self.model_id, + low_cpu_mem_usage=True, + torch_dtype=torch.bfloat16, + attn_implementation="flash_attention_2", + revision=self.revision, + ) + model.to(torch_device) + + tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision) + inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device) + + output = model.generate(**inputs, max_new_tokens=20, do_sample=False) + output_text = tokenizer.batch_decode(output, skip_special_tokens=True) + + self.assertEqual(output_text, EXPECTED_TEXTS)