diff --git a/README.md b/README.md index d0a347c7..15cca711 100644 --- a/README.md +++ b/README.md @@ -231,8 +231,6 @@ List of command-line flags | `--load-in-8bit` | Load the model with 8-bit precision (using bitsandbytes). | | `--bf16` | Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU. | | `--no-cache` | Set `use_cache` to `False` while generating text. This reduces VRAM usage slightly, but it comes at a performance cost. | -| `--xformers` | Use xformer's memory efficient attention. This is really old and probably doesn't do anything. | -| `--sdp-attention` | Use PyTorch 2.0's SDP attention. Same as above. | | `--trust-remote-code` | Set `trust_remote_code=True` while loading the model. Necessary for some models. | | `--no_use_fast` | Set use_fast=False while loading the tokenizer (it's True by default). Use this if you have any problems related to use_fast. | | `--use_flash_attention_2` | Set use_flash_attention_2=True while loading the model. | diff --git a/modules/llama_attn_hijack.py b/modules/llama_attn_hijack.py deleted file mode 100644 index 00436fb2..00000000 --- a/modules/llama_attn_hijack.py +++ /dev/null @@ -1,171 +0,0 @@ -import math -import sys -from typing import Optional, Tuple - -import torch -import torch.nn as nn - -import modules.shared as shared -from modules.logging_colors import logger - -if shared.args.xformers: - try: - import xformers.ops - except Exception: - logger.error("xformers not found! Please install it before trying to use it.", file=sys.stderr) - - -def hijack_llama_attention(): - import transformers.models.llama.modeling_llama - if shared.args.xformers: - transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward - logger.info("Replaced attention with xformers_attention") - elif shared.args.sdp_attention: - transformers.models.llama.modeling_llama.LlamaAttention.forward = sdp_attention_forward - logger.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/modules/models.py b/modules/models.py index e166f737..5235f108 100644 --- a/modules/models.py +++ b/modules/models.py @@ -21,7 +21,7 @@ from transformers import ( ) import modules.shared as shared -from modules import RoPE, llama_attn_hijack, sampler_hijack +from modules import RoPE, sampler_hijack from modules.logging_colors import logger from modules.models_settings import get_model_metadata from modules.relative_imports import RelativeImport @@ -97,10 +97,6 @@ def load_model(model_name, loader=None): else: tokenizer = load_tokenizer(model_name, model) - # Hijack attention with xformers - if any((shared.args.xformers, shared.args.sdp_attention)): - llama_attn_hijack.hijack_llama_attention() - shared.settings.update({k: v for k, v in metadata.items() if k in shared.settings}) if loader.lower().startswith('exllama'): shared.settings['truncation_length'] = shared.args.max_seq_len diff --git a/modules/shared.py b/modules/shared.py index f98343b8..36ace23c 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -98,8 +98,6 @@ group.add_argument('--disk-cache-dir', type=str, default='cache', help='Director group.add_argument('--load-in-8bit', action='store_true', help='Load the model with 8-bit precision (using bitsandbytes).') group.add_argument('--bf16', action='store_true', help='Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU.') group.add_argument('--no-cache', action='store_true', help='Set use_cache to False while generating text. This reduces VRAM usage slightly, but it comes at a performance cost.') -group.add_argument('--xformers', action='store_true', help='Use xformer\'s memory efficient attention. This is really old and probably doesn\'t do anything.') -group.add_argument('--sdp-attention', action='store_true', help='Use PyTorch 2.0\'s SDP attention. Same as above.') group.add_argument('--trust-remote-code', action='store_true', help='Set trust_remote_code=True while loading the model. Necessary for some models.') group.add_argument('--force-safetensors', action='store_true', help='Set use_safetensors=True while loading the model. This prevents arbitrary code execution.') group.add_argument('--no_use_fast', action='store_true', help='Set use_fast=False while loading the tokenizer (it\'s True by default). Use this if you have any problems related to use_fast.')