chatglm.py

# SPDX-License-Identifier: Apache-2.0
# Adapted from
# https://github.com/THUDM/ChatGLM2-6B
"""Inference-only ChatGLM model compatible with THUDM weights."""
import json
from typing import Iterable, Optional, Set, Tuple, Union

import torch
from torch import nn
from torch.nn import LayerNorm
import os
import re

from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs import ChatGLMConfig

from .interfaces import SupportsLoRA, SupportsPP, SupportsQuant
from .utils import (AutoWeightsLoader, WeightsMapper, is_pp_missing_parameter,
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)

from vllm import _custom_ops as ops
from vllm.model_executor.utils import pad_weight, gemm_bank_conf


class GLMAttention(nn.Module):

    def __init__(
        self,
        config: ChatGLMConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        tp_size = get_tensor_model_parallel_world_size()
        self.total_num_heads = config.num_attention_heads
        assert self.total_num_heads % tp_size == 0
        self.num_heads = self.total_num_heads // tp_size
        self.multi_query_attention = config.multi_query_attention
        self.total_num_kv_heads = (config.multi_query_group_num
                                   if config.multi_query_attention else
                                   config.num_attention_heads)
        if self.total_num_kv_heads >= tp_size:
            # Number of KV heads is greater than TP size, so we partition
            # the KV heads across multiple tensor parallel GPUs.
            assert self.total_num_kv_heads % tp_size == 0
        else:
            # Number of KV heads is less than TP size, so we replicate
            # the KV heads across multiple tensor parallel GPUs.
            assert tp_size % self.total_num_kv_heads == 0
        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
        self.head_dim = config.hidden_size // self.total_num_heads
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim
        self.scaling = self.head_dim**-0.5

        self.query_key_value = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=config.add_bias_linear or config.add_qkv_bias,
            quant_config=quant_config,
            prefix=f"{prefix}.query_key_value",
        )
        self.dense = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            config.hidden_size,
            bias=config.add_bias_linear,
            quant_config=quant_config,
            prefix=f"{prefix}.dense",
        )

        # https://huggingface.co/THUDM/chatglm3-6b-32k/blob/e210410255278dd9d74463cf396ba559c0ef801c/modeling_chatglm.py#L141
        rope_ratio = getattr(config, "rope_ratio", 1.0)
        max_positions = getattr(config, "seq_length", 8192)
        # NOTE: THUDM/cogagent-9b-20241220 uses original_rope=False,
        # which is equivalent to is_neox_style=True
        is_neox_style = not config.original_rope
        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim // 2,
            max_position=max_positions,
            base=10000 * rope_ratio,
            is_neox_style=is_neox_style,
        )
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              num_kv_heads=self.num_kv_heads,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
        
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
            self.quant_config=quant_config

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.query_key_value(hidden_states)
        # if os.environ.get('FA_PAD') == '1' and self.quant_method is None:
        #     qkv = qkv[...,:-32]
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        q, k = self.rotary_emb(position_ids, q, k)
        context_layer = self.attn(q, k, v)
        attn_output, _ = self.dense(context_layer)
        return attn_output


class GLMMLP(nn.Module):
    """MLP.

    MLP will take the input with h hidden state, project it to 4*h
    hidden dimension, perform nonlinear transformation, and project the
    state back into h hidden dimension.
    """

    def __init__(
        self,
        config: ChatGLMConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()

        self.add_bias = config.add_bias_linear

        # Project to 4h.
        self.dense_h_to_4h = MergedColumnParallelLinear(
            config.hidden_size,
            [config.ffn_hidden_size] * 2,
            bias=config.add_bias_linear,
            quant_config=quant_config,
            prefix=f"{prefix}.dense_h_to_4h",
        )

        self.activation_func = SiluAndMul()

        # Project back to h.
        self.dense_4h_to_h = RowParallelLinear(
            config.ffn_hidden_size,
            config.hidden_size,
            bias=config.add_bias_linear,
            quant_config=quant_config,
            prefix=f"{prefix}.dense_4h_to_h",
        )

    def forward(self, hidden_states):
        # [s, b, 4hp]
        intermediate_parallel, _ = self.dense_h_to_4h(hidden_states)
        intermediate_parallel = self.activation_func(intermediate_parallel)
        # [s, b, h]
        output, _ = self.dense_4h_to_h(intermediate_parallel)
        return output


class GLMBlock(nn.Module):
    """A single transformer layer.

    Transformer layer takes input with size [s, b, h] and returns an
    output of the same size.
    """

    def __init__(
        self,
        config: ChatGLMConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.apply_residual_connection_post_layernorm = (
            config.apply_residual_connection_post_layernorm)

        self.fp32_residual_connection = config.fp32_residual_connection

        layer_norm_func = RMSNorm if config.rmsnorm else LayerNorm
        # Layernorm on the input data.
        self.input_layernorm = layer_norm_func(config.hidden_size,
                                               eps=config.layernorm_epsilon)

        # Self attention.
        self.self_attention = GLMAttention(config,
                                           cache_config,
                                           quant_config,
                                           prefix=f"{prefix}.self_attention")
        self.hidden_dropout = config.hidden_dropout

        # Layernorm on the attention output
        self.post_attention_layernorm = layer_norm_func(
            config.hidden_size, eps=config.layernorm_epsilon)

        # MLP
        self.mlp = GLMMLP(config, quant_config, prefix=f"{prefix}.mlp")

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> torch.Tensor:
        # hidden_states: [num_tokens, h]
        # Layer norm at the beginning of the transformer layer.
        layernorm_output = self.input_layernorm(hidden_states)
        # Self attention.
        attention_output = self.self_attention(
            hidden_states=layernorm_output,
            position_ids=position_ids,
        )

        # Residual connection.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = hidden_states

        layernorm_input = residual + attention_output

        # Layer norm post the self attention.
        layernorm_output = self.post_attention_layernorm(layernorm_input)

        # Second residual connection.
        if self.apply_residual_connection_post_layernorm:
            residual = layernorm_output
        else:
            residual = layernorm_input

        output = self.mlp(layernorm_output) + residual

        return output


class GLMTransformer(nn.Module):
    """Transformer class."""

    def __init__(
        self,
        config: ChatGLMConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.post_layer_norm = config.post_layer_norm

        # Number of layers.
        self.num_layers = config.num_layers

        # Transformer layers.
        self.start_layer, self.end_layer, self.layers = make_layers(
            self.num_layers,
            lambda prefix: GLMBlock(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.layers",
        )

        if self.post_layer_norm:
            layer_norm_func = RMSNorm if config.rmsnorm else LayerNorm
            # Final layer norm before output.
            self.final_layernorm = layer_norm_func(
                config.hidden_size, eps=config.layernorm_epsilon)

        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(["hidden_states"],
                                                    config.hidden_size))

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_ids: torch.Tensor,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        for layer in self.layers[self.start_layer:self.end_layer]:
            hidden_states = layer(hidden_states=hidden_states,
                                  position_ids=position_ids)

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({"hidden_states": hidden_states})

        # Final layer norm.
        if self.post_layer_norm:
            hidden_states = self.final_layernorm(hidden_states)

        return hidden_states


@support_torch_compile
class ChatGLMModel(nn.Module, SupportsQuant):
    packed_modules_mapping = {
        "linear_proj.merged_proj":
        ["linear_proj.gate_proj", "linear_proj.dense_h_to_4h"]
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.config = config

        self.embedding = VocabParallelEmbedding(config.padded_vocab_size,
                                                config.hidden_size,
                                                quant_config=quant_config,
                                                prefix=f"{prefix}.embedding")

        self.num_layers = config.num_layers
        self.multi_query_group_num = config.multi_query_group_num
        self.kv_channels = config.kv_channels
        self.encoder = GLMTransformer(config,
                                      cache_config,
                                      quant_config,
                                      prefix=f"{prefix}.encoder")

        self.output_layer = ParallelLMHead(config.padded_vocab_size,
                                           config.hidden_size,
                                           quant_config=quant_config,
                                           prefix=f"{prefix}.output_layer")

        self.make_empty_intermediate_tensors = (
            self.encoder.make_empty_intermediate_tensors)
        
        self.quant_method = None
        if quant_config is not None:
            self.quant_method=quant_config.get_name()
            self.quant_config=quant_config

        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
        self.use_gemm_pad = os.environ.get('GEMM_PAD') == '1'
        self.use_fa_pad = os.environ.get('FA_PAD') == '1'

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embedding(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs: object,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]

        # Run encoder.
        hidden_states = self.encoder(
            hidden_states=hidden_states,
            position_ids=positions,
        )

        return hidden_states

    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("linear_proj.merged_proj", "linear_proj.gate_proj", 0),
            ("linear_proj.merged_proj", "linear_proj.dense_h_to_4h", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: Set[str] = set()

        for name, loaded_weight in weights:
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                if "rotary_pos_emb.inv_freq" in name:
                    continue
                if name.endswith(".bias") and name not in params_dict:
                    continue
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
            
        if self.use_llama_nn and self.quant_method is None:
            lay_key_words = [
                "self_attention.query_key_value.weight",
                "self_attention.dense.weight",
                "mlp.dense_h_to_4h.weight",
                "mlp.dense_4h_to_h.weight",
            ]
            combined_words = "|".join(lay_key_words)
            
            # lay_qkv_words = ["self_attention.query_key_value.weight"]   
            # qkv_words = "|".join(lay_qkv_words)  
            
            # lay_qkv_bias_words = ["self_attention.query_key_value.bias"]   
            # qkv_bias_words = "|".join(lay_qkv_bias_words)
            
            for layername in loaded_params:
                weight = params_dict[layername]
                if "lm_head.weight" in layername and weight.shape[1] == 4096:
                    lay_key_words.append("lm_head.weight")
                    combined_words = "|".join(lay_key_words)
                    os.environ['LM_NN'] = '1'  
                else:
                    os.environ['LM_NN'] = '0'
                # if self.use_fa_pad and (re.findall(qkv_bias_words, layername)):
                #     weight.data = pad_weight(weight.data, 32)
                    
                matches = re.findall(combined_words, layername)
                if matches:  
                    # if self.use_gemm_pad and gemm_bank_conf(weight.data.shape[0]):
                    #     weight.data = pad_weight(weight.data, 32)  
                        
                    # if self.use_fa_pad and (re.findall(qkv_words, layername)):
                    #     if not gemm_bank_conf(weight.data.shape[0]):
                    #         weight.data = pad_weight(weight.data, 32)
                                        
                    _weight = torch.zeros_like(weight.data)
                    ori_shape =_weight.shape
                    
                    ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                    weight.data.copy_(_weight)
                    
                    weight.data=weight.data.reshape(ori_shape[1], -1)
                    
        return loaded_params


class ChatGLMBaseModel(nn.Module):
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_substr={".word_embeddings": ""}, )

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        transformer_type: type[ChatGLMModel] = ChatGLMModel,
    ) -> None:
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config
        multimodal_config = vllm_config.model_config.multimodal_config
        self.config = config
        self.lora_config = lora_config
        self.multimodal_config = multimodal_config

        self.quant_config = quant_config
        self.max_position_embeddings = getattr(config, "max_sequence_length",
                                               8192)
        self.transformer = transformer_type(vllm_config=vllm_config,
                                            prefix=maybe_prefix(
                                                prefix, "transformer"))
        if self.config.tie_word_embeddings:
            self.transformer.output_layer.weight = (
                self.transformer.embedding.weight)
        self.lm_head = self.transformer.output_layer
        self.logits_processor = LogitsProcessor(config.padded_vocab_size)
        self.make_empty_intermediate_tensors = (
            self.transformer.make_empty_intermediate_tensors)

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)


class ChatGLMForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP,
                         SupportsQuant):
    packed_modules_mapping = {
        "query_key_value": ["query_key_value"],
        "dense_h_to_4h": ["dense_h_to_4h"]
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_config
        if hasattr(config, "vision_config"):
            hf_overrides = {"architectures": ["GLM4VForCausalLM"]}
            raise RuntimeError(
                "The configuration of this model indicates that it supports "
                "vision inputs, but you instantiated the text-only version "
                "of this model. Please use the vision model by setting "
                f"`--hf-overrides '{json.dumps(hf_overrides)}'`")

        super().__init__(vllm_config=vllm_config, prefix=prefix)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.transformer(input_ids, positions,
                                         intermediate_tensors, inputs_embeds)
        return hidden_states