Support pre-training 8*V100 (32G) gpus with xformers

README.md

@@ -229,6 +229,14 @@ Training script with DeepSpeed ZeRO-2: [`pretrain.sh`](https://github.com/haotia
 - `--mm_projector_type mlp2x_gelu`: the two-layer MLP vision-language connector.
 - `--vision_tower openai/clip-vit-large-patch14-336`: CLIP ViT-L/14 336px.
 
+<details>
+<summary>Pretrain takes around 20 hours for LLaVA-7B on 8x V100 (32G)</summary>
+
+ We provide training script with DeepSpeed [here](https://github.com/haotian-liu/LLaVA/blob/main/scripts/pretrain_xformers.sh).
+Tips:
+- If you are using V100 which is not supported by FlashAttention, you can use the [memory-efficient attention](https://arxiv.org/abs/2112.05682) implemented in [xFormers](https://github.com/facebookresearch/xformers). Install xformers and replace `llava/train/train_mem.py` above with [llava/train/train_xformers.py](llava/train/train_xformers.py).
+</details>
+
 ### Visual Instruction Tuning
 
 1. Prepare data

llava/train/llama_xformers_attn_monkey_patch.py

@@ -0,0 +1,129 @@
+"""
+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 transformers.models.llama.modeling_llama
+from torch import nn
+
+try:
+    import xformers.ops
+except ImportError:
+    logging.error("xformers not found! Please install it before trying to use it.")
+
+
+def replace_llama_attn_with_xformers_attn():
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
+
+
+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]]]:
+    # pylint: disable=duplicate-code
+    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

llava/train/train_xformers.py

@@ -0,0 +1,13 @@
+# Make it more memory efficient by monkey patching the LLaMA model with xformers attention.
+
+# Need to call this before importing transformers.
+from llava.train.llama_xformers_attn_monkey_patch import (
+    replace_llama_attn_with_xformers_attn,
+)
+
+replace_llama_attn_with_xformers_attn()
+
+from llava.train.train import train
+
+if __name__ == "__main__":
+    train()

scripts/pretrain_xformers.sh

@@ -0,0 +1,44 @@
+#!/bin/bash
+
+# Uncomment and set the following variables correspondingly to run this script:
+
+# MODEL_VERSION=vicuna-v1-3-7b
+# MODEL_VERSION=llama-2-7b-chat
+
+########### DO NOT CHANGE ###########
+########### USE THIS FOR BOTH ###########
+PROMPT_VERSION=plain
+########### DO NOT CHANGE ###########
+
+deepspeed llava/train/train_xformers.py \
+    --deepspeed ./scripts/zero2.json \
+    --model_name_or_path ./checkpoints/$MODEL_VERSION \
+    --version $PROMPT_VERSION \
+    --data_path /path/to/pretrain_data.json \
+    --image_folder /path/to/images \
+    --vision_tower openai/clip-vit-large-patch14 \
+    --tune_mm_mlp_adapter True \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --bf16 False \
+    --output_dir ./checkpoints/llava-$MODEL_VERSION-pretrain \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 4 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 4 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 24000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-3 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 False \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 4 \
+    --lazy_preprocess True \
+    --report_to wandb