falcon.py

# coding=utf-8
# Adapted from
# https://github.com/huggingface/transformers/blob/a5cc30d72ae2dc19af534e4b35c986cc28db1275/src/transformers/models/falcon/modeling_falcon.py
# Copyright 2023 The vLLM team.
# Copyright 2023 the Falcon authors 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.
"""PyTorch Falcon model."""

import math
from typing import Iterable, List, Optional, Tuple, Union

import os
import re
import torch
from torch import nn
from torch.nn import LayerNorm
from transformers import FalconConfig as HF_FalconConfig

from vllm.attention import Attention, AttentionMetadata
from vllm.config import CacheConfig
from vllm.distributed import (get_tensor_model_parallel_rank,
                              get_tensor_model_parallel_world_size,
                              tensor_model_parallel_all_reduce)
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               QKVParallelLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig)
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
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 RWConfig

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

FalconConfig = Union[HF_FalconConfig, RWConfig]


def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
    closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
    base = torch.tensor(2**(-(2**-(math.log2(closest_power_of_2) - 3))),
                        dtype=torch.float32)
    powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
    slopes = torch.pow(base, powers)

    if closest_power_of_2 != total_num_heads:
        extra_base = torch.tensor(
            2**(-(2**-(math.log2(2 * closest_power_of_2) - 3))),
            dtype=torch.float32)
        num_remaining_heads = min(closest_power_of_2,
                                  total_num_heads - closest_power_of_2)
        extra_powers = torch.arange(1,
                                    1 + 2 * num_remaining_heads,
                                    2,
                                    dtype=torch.int32)
        slopes = torch.cat(
            [slopes, torch.pow(extra_base, extra_powers)], dim=0)

    return slopes


class FalconAttention(nn.Module):

    def __init__(
        self,
        config: FalconConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        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.head_dim = self.hidden_size // self.total_num_heads
        assert self.head_dim * self.total_num_heads == self.hidden_size

        self.new_decoder_architecture = config.new_decoder_architecture
        self.multi_query = config.multi_query

        if self.new_decoder_architecture:
            self.total_num_kv_heads = config.num_kv_heads
        elif self.multi_query:
            self.total_num_kv_heads = 1
        else:
            self.total_num_kv_heads = self.total_num_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.query_key_value = QKVParallelLinear(
            self.hidden_size,
            self.head_dim,
            self.total_num_heads,
            self.total_num_kv_heads,
            bias=config.bias,
            skip_bias_add=True,
            quant_config=quant_config,
        )
        self.q_size = self.num_heads * self.head_dim
        self.kv_size = self.num_kv_heads * self.head_dim

        # Layer-wise attention scaling
        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
        self.reduce_row_parallel_results = not (config.new_decoder_architecture
                                                or config.parallel_attn)
        self.dense = RowParallelLinear(
            self.hidden_size,
            self.hidden_size,
            bias=config.bias,
            skip_bias_add=True,
            quant_config=quant_config,
            reduce_results=self.reduce_row_parallel_results)

        self.use_rotary = config.rotary
        self.use_alibi = config.alibi
        assert not (self.use_rotary and self.use_alibi), (
            "Rotary and alibi are mutually exclusive.")

        if self.use_rotary:
            rope_theta = getattr(config, "rope_theta", 10000)
            max_position_embeddings = getattr(config,
                                              "max_position_embeddings", 8192)
            self.rotary_emb = get_rope(
                self.head_dim,
                rotary_dim=self.head_dim,
                max_position=max_position_embeddings,
                base=rope_theta,
            )
            self.attn = Attention(self.num_heads,
                                  self.head_dim,
                                  self.inv_norm_factor,
                                  num_kv_heads=self.num_kv_heads,
                                  quant_config=quant_config)
        elif self.use_alibi:
            tp_rank = get_tensor_model_parallel_rank()
            head_start = tp_rank * self.num_heads
            head_end = (tp_rank + 1) * self.num_heads
            alibi_slopes = (_get_alibi_slopes(self.total_num_heads) *
                            self.inv_norm_factor)
            alibi_slopes = alibi_slopes[head_start:head_end].tolist()
            self.attn = Attention(self.num_heads,
                                  self.head_dim,
                                  self.inv_norm_factor,
                                  num_kv_heads=self.num_kv_heads,
                                  alibi_slopes=alibi_slopes,
                                  quant_config=quant_config)
        else:
            self.attn = Attention(self.num_heads,
                                  self.head_dim,
                                  scale=self.inv_norm_factor,
                                  num_kv_heads=self.num_kv_heads,
                                  cache_config=cache_config,
                                  quant_config=quant_config)
            
        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,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        qkv, bias = self.query_key_value(hidden_states)
        if os.environ.get('FA_PAD') == '1' and self.quant_method is None:
            qkv = qkv[...,:-32]
        if bias is not None:
            qkv += bias
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
        if self.use_rotary:
            q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
        attn_output, bias = self.dense(attn_output)
        return attn_output, bias


class FalconMLP(nn.Module):

    def __init__(
        self,
        config: FalconConfig,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        hidden_size = config.hidden_size

        self.dense_h_to_4h = ColumnParallelLinear(hidden_size,
                                                  4 * hidden_size,
                                                  bias=config.bias,
                                                  skip_bias_add=True,
                                                  quant_config=quant_config)
        self.act = get_act_fn("gelu", quant_config, 4 * hidden_size)
        self.reduce_row_parallel_results = not (config.new_decoder_architecture
                                                or config.parallel_attn)
        self.dense_4h_to_h = RowParallelLinear(
            4 * hidden_size,
            hidden_size,
            bias=config.bias,
            skip_bias_add=True,
            reduce_results=self.reduce_row_parallel_results,
            quant_config=quant_config)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # NOTE(zhuohan): Following huggingface, we do not fuse bias add here.
        x, bias = self.dense_h_to_4h(x)
        if bias is not None:
            x += bias
        x = self.act(x)
        x, bias = self.dense_4h_to_h(x)
        return x, bias


class FalconDecoderLayer(nn.Module):

    def __init__(
        self,
        config: FalconConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.self_attention = FalconAttention(config, cache_config,
                                              quant_config)
        self.mlp = FalconMLP(config, quant_config)
        self.config = config

        if (not hasattr(config, "num_ln_in_parallel_attn")):
            config.num_ln_in_parallel_attn = None
            
        if (config.num_ln_in_parallel_attn is None
                and config.new_decoder_architecture):
            config.num_ln_in_parallel_attn = 2

        if not config.parallel_attn:
            self.post_attention_layernorm = LayerNorm(
                hidden_size, eps=config.layer_norm_epsilon)
            self.input_layernorm = LayerNorm(hidden_size,
                                             eps=config.layer_norm_epsilon)
        else:
            if config.num_ln_in_parallel_attn == 2:
                # The layer norm before self-attention
                self.ln_attn = LayerNorm(hidden_size,
                                         eps=config.layer_norm_epsilon)
                # The layer norm before the MLP
                self.ln_mlp = LayerNorm(hidden_size,
                                        eps=config.layer_norm_epsilon)
            else:
                self.input_layernorm = LayerNorm(hidden_size,
                                                 eps=config.layer_norm_epsilon)

        self.reduce_row_parallel_results = not (config.new_decoder_architecture
                                                or config.parallel_attn)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        kv_cache: torch.Tensor,
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        residual = hidden_states

        if self.config.num_ln_in_parallel_attn == 2:
            attention_layernorm_out = self.ln_attn(hidden_states)
            mlp_layernorm_out = self.ln_mlp(hidden_states)
        else:
            attention_layernorm_out = self.input_layernorm(hidden_states)

        # Self attention.
        attention_output, attention_bias = self.self_attention(
            positions=positions,
            hidden_states=attention_layernorm_out,
            kv_cache=kv_cache,
            attn_metadata=attn_metadata,
        )
        if self.reduce_row_parallel_results and attention_bias is not None:
            attention_output += attention_bias

        if not self.config.new_decoder_architecture:
            if self.config.parallel_attn:
                mlp_layernorm_out = attention_layernorm_out
            else:
                residual += attention_output
                mlp_layernorm_out = self.post_attention_layernorm(residual)

        if (self.config.new_decoder_architecture and self.config.parallel_attn
                and self.config.num_ln_in_parallel_attn == 1):
            mlp_layernorm_out = attention_layernorm_out

        # MLP.
        mlp_output, mlp_bias = self.mlp(mlp_layernorm_out)
        if self.reduce_row_parallel_results and mlp_bias is not None:
            mlp_output += mlp_bias

        if not self.reduce_row_parallel_results:
            # When MLP and Attention layers are parallel, we can use
            # only one all-reduce operator to reduce the results from
            # both MLP and Attention layers.
            mlp_output += attention_output
            mlp_output = tensor_model_parallel_all_reduce(mlp_output)
            if attention_bias is not None:
                mlp_output += attention_bias
            if mlp_bias is not None:
                mlp_output += mlp_bias

        output = mlp_output + residual
        return output


class FalconModel(nn.Module):

    def __init__(
        self,
        config: FalconConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.use_alibi = config.alibi

        # Embedding + LN Embedding
        self.word_embeddings = VocabParallelEmbedding(
            config.vocab_size,
            self.embed_dim,
        )

        # Transformer blocks
        self.h = nn.ModuleList([
            FalconDecoderLayer(config, cache_config, quant_config)
            for _ in range(config.num_hidden_layers)
        ])

        # Final Layer Norm
        self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

    def forward(
        self,
        input_ids: torch.LongTensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
    ) -> torch.Tensor:
        hidden_states = self.word_embeddings(input_ids)
        for i in range(len(self.h)):
            layer = self.h[i]
            hidden_states = layer(
                positions,
                hidden_states,
                kv_caches[i],
                attn_metadata,
            )
        hidden_states = self.ln_f(hidden_states)
        return hidden_states


class FalconForCausalLM(nn.Module):

    def __init__(
        self,
        config: FalconConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        self.transformer = FalconModel(config, cache_config, quant_config)
        # only Falcon-11B doesn't share lm_head weight with word embeddings
        # and previous Falcon model doesn't have tie_word_embeddings config
        # so we set tie_word_embeddings to True by default
        self.tie_word_embeddings = (config.tie_word_embeddings
                                    if config.tie_word_embeddings is not None
                                    else True)
        if self.tie_word_embeddings:
            self.lm_head = self.transformer.word_embeddings
        else:
            self.lm_head = ParallelLMHead(
                config.vocab_size,
                config.hidden_size,
                quant_config=quant_config,
            )
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.sampler = Sampler()
        
        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'
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
        self.w8a8_strategy=int(os.getenv('W8A8_SUPPORT_METHODS', '0'))

    def forward(
        self,
        input_ids: torch.LongTensor,
        positions: torch.Tensor,
        kv_caches: List[torch.Tensor],
        attn_metadata: AttentionMetadata,
        intermediate_tensors: Optional[IntermediateTensors] = None,
    ) -> torch.Tensor:
        hidden_states = self.transformer(
            input_ids,
            positions,
            kv_caches,
            attn_metadata,
        )
        return hidden_states

    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 sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        next_tokens = self.sampler(logits, sampling_metadata)
        return next_tokens

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
        total_num_heads = self.config.num_attention_heads
        if self.config.new_decoder_architecture:
            total_num_kv_heads = self.config.num_kv_heads
        elif self.config.multi_query:
            total_num_kv_heads = 1
        else:
            total_num_kv_heads = total_num_heads
        num_query_heads_per_kv_head = total_num_heads // total_num_kv_heads
        params_dict = dict(self.named_parameters(remove_duplicate=False))
        for name, loaded_weight in weights:
            if name == "lm_head.weight" and self.tie_word_embeddings:
                # Falcon uses tied embeddings except Falcon-11b.
                continue
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            param = params_dict[name]
            if "query_key_value" in name:
                output_dim = getattr(param, "output_dim", None)
                loaded_weight_shape = loaded_weight.shape
                if output_dim is not None:
                    loaded_weight = loaded_weight.view(
                        loaded_weight_shape[:output_dim] +
                        (total_num_kv_heads, num_query_heads_per_kv_head + 2,
                         -1) + loaded_weight_shape[output_dim + 1:])
                    wq = loaded_weight.narrow(
                        output_dim + 1, 0,
                        num_query_heads_per_kv_head).reshape(
                            *loaded_weight_shape[:output_dim], -1,
                            *loaded_weight_shape[output_dim + 1:])
                    wk = loaded_weight.narrow(
                        output_dim + 1, num_query_heads_per_kv_head,
                        1).reshape(*loaded_weight_shape[:output_dim], -1,
                                   *loaded_weight_shape[output_dim + 1:])
                    wv = loaded_weight.narrow(
                        output_dim + 1, num_query_heads_per_kv_head + 1,
                        1).reshape(*loaded_weight_shape[:output_dim], -1,
                                   *loaded_weight_shape[output_dim + 1:])
                    loaded_weight = torch.cat([wq, wk, wv], dim=output_dim)

            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)

        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)          
            
            for layername, weight in params_dict.items():
                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)