add xformers support

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, v1: bool,
-                 local_rank: int, flash_attention: bool, backend: str
+                 local_rank: int, flash_attention: bool, xformers: bool, backend: str
                  ):
         """
         Args:
@@ -50,6 +50,7 @@ class Finetune4bConfig:
             v1 (bool): v1 model flag
             local_rank (int): local rank if using torch.distributed.launch
             flash_attention (bool): Enables flash attention
+            xformers (bool): use xformers or not
         """
         self.dataset = dataset
         self.ds_type = ds_type
@@ -88,6 +89,7 @@ class Finetune4bConfig:
         self.groupsize = groupsize
         self.v1 = v1
         self.flash_attention = flash_attention
+        self.xformers = xformers
         self.backend = backend
 
 

arg_parser.py

@@ -70,6 +70,7 @@ def parse_commandline():
     
     # Flash Attention
     parser_training.add_argument("--flash_attention", action="store_true", help="enables flash attention, can improve performance and reduce VRAM use")
+    parser_training.add_argument("--xformers", action="store_true", help="enables xformers memory efficient attention, can improve performance and reduce VRAM use")
 
     # Train Backend
     parser_training.add_argument("--backend", type=str, default='cuda', help="Backend to use. Triton or Cuda.")
@@ -111,5 +112,6 @@ def get_config() -> Finetune4bConfig:
         v1=args["v1"],
         local_rank=args["local_rank"],
         flash_attention=args["flash_attention"],
+        xformers=args["xformers"],
         backend=args["backend"],
     )

finetune.py

@@ -23,6 +23,9 @@ 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()
+elif ft_config.xformers:
+    from monkeypatch.llama_attn_hijack_xformers import hijack_llama_attention
+    hijack_llama_attention()
 
 import autograd_4bit
 if ft_config.backend.lower() == 'triton':

monkeypatch/llama_attn_hijack_xformers.py

@@ -0,0 +1,101 @@
+'''
+Directly copied the code from https://github.com/oobabooga/text-generation-webui/pull/950/commits and made some adjustments
+'''
+import math
+import sys
+import torch
+import torch.nn as nn
+import transformers.models.llama.modeling_llama
+
+from typing import Optional
+from typing import Tuple
+
+try:
+    import xformers.ops
+except ImportError:
+    raise ImportError("Please install xformers to use this module")
+
+def hijack_llama_attention():
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
+    print("Replaced attention with xformers_attention")
+
+def xformers_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    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)
+
+    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 = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+    # [bsz, nh, t, hd]
+
+    if past_key_value is not None:
+        # reuse k, v, self_attention
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    #We only apply xformers optimizations if we don't need to output the whole attention matrix
+    if not output_attentions:
+        dtype = query_states.dtype
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        #This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
+        #We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
+        if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
+            # input and output should be of form (bsz, q_len, num_heads, head_dim)
+            attn_output = xformers.ops.memory_efficient_attention(query_states, key_states, value_states, attn_bias=None)
+        else:
+            # input and output should be of form (bsz, q_len, num_heads, head_dim)
+            attn_output = xformers.ops.memory_efficient_attention(query_states, key_states, value_states, attn_bias=xformers.ops.LowerTriangularMask())
+        attn_weights = None
+    else:
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2)
+
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    attn_output = self.o_proj(attn_output)
+
+    return attn_output, attn_weights, past_key_value

monkeypatch/llama_flash_attn_monkey_patch.py

@@ -8,8 +8,11 @@ from transformers.models.llama.modeling_llama import LlamaConfig, LlamaRotaryEmb
 
 from einops import rearrange
 
+try:
     from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
     from flash_attn.bert_padding import unpad_input, pad_input
+except ImportError:
+    raise ImportError("Please install flash_attn to use this module")
 
 class LlamaAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""

requirements.txt

@@ -4,8 +4,7 @@ bitsandbytes
 datasets
 sentencepiece
 safetensors
-flash-attn
-triton
+einops
 colorama
 git+https://github.com/huggingface/transformers.git
 git+https://github.com/sterlind/GPTQ-for-LLaMa.git@lora_4bit