diff --git a/src/axolotl/monkeypatch/llama_attn_hijack_xformers.py b/src/axolotl/monkeypatch/llama_attn_hijack_xformers.py new file mode 100644 index 000000000..a4f34bcd7 --- /dev/null +++ b/src/axolotl/monkeypatch/llama_attn_hijack_xformers.py @@ -0,0 +1,172 @@ +''' +Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments +''' + +import logging +import math +from typing import Optional, Tuple + +import torch +import torch.nn as nn +import transformers.models.llama.modeling_llama + +try: + import xformers.ops +except Exception: + logging.error("xformers not found! Please install it before trying to use it.") + + +def hijack_llama_attention(): + transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward + logging.info("Replaced attention with xformers_attention") + + +def hijack_llama_sdp_attention(): + transformers.models.llama.modeling_llama.LlamaAttention.forward = sdp_attention_forward + logging.info("Replaced attention with sdp_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: + 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 + + +def sdp_attention_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 sdp attention if we don't need to output the whole attention matrix + if not output_attentions: + attn_output = torch.nn.functional.scaled_dot_product_attention(query_states, key_states, value_states, attn_mask=attention_mask, is_causal=False) + 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 diff --git a/src/axolotl/utils/models.py b/src/axolotl/utils/models.py index 0737d0f12..df4e50be5 100644 --- a/src/axolotl/utils/models.py +++ b/src/axolotl/utils/models.py @@ -97,12 +97,19 @@ def load_model( logging.info("patching with flash attention") replace_llama_attn_with_flash_attn() elif is_llama_derived_model and cfg.xformers_attention: - from alpaca_lora_4bit.monkeypatch.llama_attn_hijack_xformers import ( + from axolotl.monkeypatch.llama_attn_hijack_xformers import ( hijack_llama_attention, ) logging.info("patching with xformers attention") hijack_llama_attention() + elif is_llama_derived_model and cfg.sdp_attention: + from axolotl.monkeypatch.llama_attn_hijack_xformers import ( + hijack_llama_sdp_attention, + ) + + logging.info("patching with sdp attention") + hijack_llama_sdp_attention() if cfg.bf16: torch_dtype = torch.bfloat16