Merge pull request #59 from yamashi/main

Add flash attention

Finetune4bConfig.py

@@ -15,7 +15,7 @@ class Finetune4bConfig:
                  warmup_steps: int, save_steps: int, save_total_limit: int, logging_steps: int,
                  checkpoint: bool, skip: bool, verbose: bool,
                  txt_row_thd: int, use_eos_token: bool, groupsize: int,
-                 local_rank: int,
+                 local_rank: int, flash_attention: bool
                  ):
         """
         Args:
@@ -48,6 +48,7 @@ class Finetune4bConfig:
             use_eos_token (bool): Use Eos token instead of padding with 0
             groupsize (int): Group size of V2 model, use -1 to load V1 model
             local_rank (int): local rank if using torch.distributed.launch
+            flash_attention (bool): Enables flash attention
         """
         self.dataset = dataset
         self.ds_type = ds_type
@@ -84,6 +85,7 @@ class Finetune4bConfig:
         if self.ddp:
             self.gradient_accumulation_steps = self.gradient_accumulation_steps // self.world_size
         self.groupsize = groupsize
+        self.flash_attention = flash_attention
 
 
     def __str__(self) -> str:

arg_parser.py

@@ -66,6 +66,8 @@ def parse_commandline():
 
     # Multi GPU Support
     parser_training.add_argument("--local_rank", type=int, default=0, help="local rank if using torch.distributed.launch")
+    
+    parser_training.add_argument("--flash_attention", help="enables flash attention, can improve performance and reduce VRAM use")
 
     return vars(parser.parse_args())
 
@@ -102,4 +104,5 @@ def get_config() -> Finetune4bConfig:
         use_eos_token=args["use_eos_token"]!=0,
         groupsize=args["groupsize"],
         local_rank=args["local_rank"],
+        flash_attention=args["flash_attention"],
     )

finetune.py

@@ -16,6 +16,13 @@
         }
     ]
 """
+# Early load config to replace attn if needed
+from arg_parser import get_config
+ft_config = get_config()
+
+if ft_config.flash_attention:
+    from monkeypatch.llama_flash_attn_monkey_patch import replace_llama_attn_with_flash_attn
+    replace_llama_attn_with_flash_attn()
 
 import sys
 
@@ -29,10 +36,9 @@ from autograd_4bit import load_llama_model_4bit_low_ram
 from peft import LoraConfig, get_peft_model, get_peft_model_state_dict, PeftModel
 
 # ! Config
-from arg_parser import get_config
 import train_data
 
-ft_config = get_config()
+
 
 # * Show loaded parameters
 if ft_config.local_rank == 0:

monkeypatch/llama_flash_attn_monkey_patch.py

@@ -0,0 +1,144 @@
+from typing import List, Optional, Tuple
+
+import torch
+from torch import nn
+
+import transformers
+from transformers.models.llama.modeling_llama import LlamaConfig, LlamaRotaryEmbedding, apply_rotary_pos_emb
+
+from einops import rearrange
+
+from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
+from flash_attn.bert_padding import unpad_input, pad_input
+
+class LlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: LlamaConfig,
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        num_heads = config.num_attention_heads
+        self.hidden_size = hidden_size
+        self.num_heads = num_heads
+        self.head_dim = self.hidden_size // num_heads
+
+        if (self.head_dim * num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {num_heads}).")
+        self.q_proj = nn.Linear(
+            hidden_size,
+            num_heads * self.head_dim,
+            bias=False,
+        )
+        self.k_proj = nn.Linear(
+            hidden_size,
+            num_heads * self.head_dim,
+            bias=False,
+        )
+        self.v_proj = nn.Linear(
+            hidden_size,
+            num_heads * self.head_dim,
+            bias=False,
+        )
+        self.o_proj = nn.Linear(
+            num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+        )
+        self.rotary_emb = LlamaRotaryEmbedding(self.head_dim)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads,
+                           self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
+               Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel
+        
+        attention_mask: [bsz, q_len]
+        """
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states).view(
+            bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = self.k_proj(hidden_states).view(
+            bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(
+            bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        # [bsz, q_len, nh, hd]
+        # [bsz, nh, q_len, hd]
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states,
+                                                        key_states,
+                                                        cos,
+                                                        sin,
+                                                        position_ids)
+        # [bsz, nh, t, hd]
+        assert not output_attentions, "output_attentions is not supported"
+        assert not use_cache, "use_cache is not supported"
+        assert past_key_value is None, "past_key_value is not supported"
+
+        # Flash attention codes from
+        # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py
+
+        # transform the data into the format required by flash attention
+        qkv = torch.stack([query_states, key_states, value_states], dim=2) # [bsz, nh, 3, q_len, hd]
+        qkv = qkv.transpose(1, 3) # [bsz, q_len, 3, nh, hd]
+        # We have disabled _prepare_decoder_attention_mask in LlamaModel
+        # the attention_mask should be the same as the key_padding_mask
+        key_padding_mask = attention_mask
+
+
+        if key_padding_mask is None:
+            qkv = rearrange(qkv, 'b s ... -> (b s) ...')
+            max_s = q_len
+            cu_q_lens = torch.arange(0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32,
+                                    device=qkv.device)
+            output = flash_attn_unpadded_qkvpacked_func(
+                qkv, cu_q_lens, max_s, 0.0,
+                softmax_scale=None, causal=True
+            )
+            output = rearrange(output, '(b s) ... -> b s ...', b=bsz)
+        else:
+            nheads = qkv.shape[-2]
+            x = rearrange(qkv, 'b s three h d -> b s (three h d)')
+            x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
+            x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)
+            output_unpad = flash_attn_unpadded_qkvpacked_func(
+                x_unpad, cu_q_lens, max_s, 0.0,
+                softmax_scale=None, causal=True
+            )
+            output = rearrange(pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'),
+                                        indices, bsz, q_len),
+                            'b s (h d) -> b s h d', h=nheads)
+        return self.o_proj(rearrange(output,
+                                     'b s h d -> b s (h d)')), None, None
+
+
+# Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+def _prepare_decoder_attention_mask(self, attention_mask, input_shape,
+                                    inputs_embeds, past_key_values_length):
+    # [bsz, seq_len]
+    return attention_mask
+
+
+def replace_llama_attn_with_flash_attn():
+    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = _prepare_decoder_attention_mask
+    transformers.models.llama.modeling_llama.LlamaAttention = LlamaAttention

requirements.txt

@@ -4,6 +4,7 @@ bitsandbytes
 datasets
 sentencepiece
 safetensors
+flash-attn
 git+https://github.com/huggingface/transformers.git
 git+https://github.com/sterlind/GPTQ-for-LLaMa.git@lora_4bit
 git+https://github.com/sterlind/peft.git