Add files using upload-large-folder tool

attention_mask.py

@@ -0,0 +1,81 @@
+from typing import Optional
+
+import torch
+
+
+def _make_causal_mask(
+    attention_mask: torch.Tensor, dtype: torch.dtype, device: torch.device
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = attention_mask.shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len)
+
+
+def _make_2dvison_mask(column_mask, dtype: torch.dtype, device: torch.device):
+    """
+    """
+    bsz, seq_length = column_mask.shape
+    cross_mask = torch.zeros((bsz, 1, seq_length, seq_length), dtype=dtype, device=device)
+
+    # 找到连续的 1 的区间
+    start = None
+    for bsz_idx in range(bsz):
+        for i in range(seq_length):
+            if column_mask[bsz_idx, i] == 1:
+                if start is None:
+                    start = i
+            else:
+                if start is not None:
+                    # 填充区间
+                    cross_mask[bsz_idx, 0, start:i, start:i] = 1
+                    start = None
+
+        # 处理最后一个区间
+        if start is not None:
+            cross_mask[bsz_idx, 0, start:seq_length, start:seq_length] = 1
+
+    return cross_mask
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill_(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+def make_mask(attention_mask: torch.Tensor, dtype: torch.dtype=None, device: torch.device=None, mode: str="default", vision_mask: torch.Tensor=None, ):
+    if dtype is None:
+        dtype = attention_mask.dtype
+    if device is None:
+        device = attention_mask.device      
+    expanded_attn_mask = _expand_mask(attention_mask, dtype).to(device)
+    causal_mask = _make_causal_mask(attention_mask, dtype, device).to(device)
+    if mode == "default":
+        return attention_mask
+    else:
+        assert vision_mask is not None, "vision_mask is None"
+        vision_mask = vision_mask.to(device)
+        bsz, seq_length = attention_mask.shape
+        vision_mask_bg = vision_mask[:, None, :, None]
+        vision_mask_2d = _make_2dvison_mask(vision_mask, dtype, device)
+        if mode == "bidirectional":
+            mask = expanded_attn_mask + causal_mask
+            mask = mask.clone().masked_fill_(vision_mask_2d.to(torch.bool), 0)
+            return mask 
+        else:
+            raise NotImplementedError(f"mode {mode} is not implemented")

aux_vision.py

@@ -0,0 +1,130 @@
+import torch
+import torch.nn as nn
+from transformers import CLIPVisionModel, AutoModel
+
+from .configuration_vora import VoRAConfig
+from .eva_model import EVAVisionTransformer
+import loguru
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class CosineLoss(nn.Module):
+    def __init__(self, reduction='mean'):
+        super(CosineLoss, self).__init__()
+        self.reduction = reduction
+
+    @staticmethod
+    def interpolate_tokens_2d(self, teacher_tokens, target_size):
+        """
+        Interpolate teacher tokens to the target size using bilinear interpolation.
+        """
+        # teacher_tokens shape is (batch_size, height, width, feature_dim)
+        teacher_tokens = teacher_tokens.permute(0, 3, 1, 2)  # Convert to (batch_size, feature_dim, height, width)
+        interpolated = torch.nn.functional.interpolate(teacher_tokens, size=target_size, mode='bilinear', align_corners=True).flatten(2)  # Flatten height and width dimensions
+        return interpolated.permute(0, 2, 1)  # Convert back to (batch_size, new_height * new_width, feature_dim)
+
+    def forward(self, input: torch.Tensor, target: torch.Tensor, input_shape=None, target_shape=None) -> torch.Tensor:
+        if input_shape is not None and target_shape is not None:
+            input = input.reshape((input.shape[0], ) + input_shape + (-1, ))
+            input = self.interpolate_tokens_2d(input, target_shape)
+
+        cos_sim = nn.functional.cosine_similarity(input, target, dim=1)
+        loss = 1 - cos_sim
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:
+            return loss
+
+
+class AuxVision(nn.Module):
+    def __init__(self,
+        config: VoRAConfig = None,
+    ):
+        super().__init__()
+        self.skip_aux_cls = config.skip_aux_cls  # whether to skip the cls token in ViT
+        # ---------------- Setup Aux Model ----------------
+        # support jina clip encoder
+        if 'jina' in config.aux_vision.lower() and 'clip' in config.aux_vision.lower():
+            cfg = {
+                "img_size": 512,
+                "num_classes": 1024,
+                "embed_dim": 1024,
+                "patch_size": 14,
+                "depth": 24,
+                "qkv_bias": True,
+                "naiveswiglu": True,
+                "num_heads": 16,
+                "patch_dropout":0, # disable patch dropout
+                "subln": True,
+                "mlp_ratio": 2.66666,
+                "use_mean_pooling": False,
+            } 
+            self.aux_model = EVAVisionTransformer(**cfg)
+            self.aux_model.load_state_dict(torch.load(config.aux_vision, map_location='cpu', weights_only=True), strict=False)
+            vision_hidden_size = 1024
+            num_hidden_layers = 24
+
+
+        elif 'clip' in config.aux_vision.lower():
+            self.aux_model = CLIPVisionModel.from_pretrained(config.aux_vision)
+            vision_hidden_size = self.aux_model.vision_model.config.hidden_size
+            num_hidden_layers = self.aux_model.vision_model.config.num_hidden_layers
+        
+        else:
+            self.aux_model = AutoModel.from_pretrained(config.aux_vision, trust_remote_code=True)
+            vision_hidden_size = self.aux_model.config.hidden_size
+            num_hidden_layers = self.aux_model.config.num_hidden_layers
+        for name, param in self.aux_model.named_parameters():
+            param.requires_grad = False
+        # -------------------------------------------------
+
+        # ---------------- Setup Aux Heads ----------------
+        self.aux_layers = list(range(num_hidden_layers))
+        for layer_id in self.aux_layers:
+            self.add_module(f"aux_layer_{layer_id}", self.build_aux_layer(config.hidden_size, vision_hidden_size))
+        # -------------------------------------------------
+
+        self.loss_function = CosineLoss()
+        self.loss_keys = [f"loss_aux_layer_{layer_id}" for layer_id in self.aux_layers]
+
+    def build_aux_layer(self, llm_hidden_size, vit_hidden_size):
+        return nn.Sequential(
+                    RMSNorm(llm_hidden_size),
+                    nn.Linear(
+                        llm_hidden_size,
+                        vit_hidden_size,
+                        bias=False,
+                    )
+                )
+
+    def forward(self, frames, llm_hidden_states, vision_mask):
+        vision_hidden_states = self.aux_model(frames, output_hidden_states=True).hidden_states
+        losses = {}
+        for layer_idx in self.aux_layers:
+            aux_hidden_states = getattr(self, f"aux_layer_{layer_idx}")(llm_hidden_states[layer_idx][vision_mask == 1])
+            start_id = 1 if self.skip_aux_cls else 0
+            try:
+                aux_loss = self.loss_function(vision_hidden_states[layer_idx][:, start_id:].reshape(aux_hidden_states.shape), aux_hidden_states)
+            except Exception as e:
+                loguru.logger.error(f"Aux Vision loss function error: {e} occured at layer {layer_idx}")
+                loguru.logger.error(f"Aux Vision aux_hidden_states: {aux_hidden_states.shape}, vision_hidden_states: {vision_hidden_states[layer_idx][:, start_id:].reshape(aux_hidden_states.shape).shape}")
+                raise e
+            losses[f"loss_aux_layer_{layer_idx}"] = aux_loss
+        return losses

config.json

@@ -0,0 +1,56 @@
+{
+  "architectures": [
+    "VoRAForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_vora.VoRAConfig",
+    "AutoModelForCausalLM": "modeling_vora.VoRAForCausalLM"
+  },
+  "aux_vision": "/workspace/VoRAParse/output/jina-clip/image-encoder.pt",
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "image_size": 512,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "llm": "Qwen/Qwen3-8B",
+  "lora": {
+    "layers": 24,
+    "r": 1024,
+    "target_modules": [
+      "self_attn.q_proj",
+      "self_attn.k_proj",
+      "self_attn.v_proj",
+      "self_attn.o_proj",
+      "mlp.up_proj",
+      "mlp.gate_proj",
+      "mlp.down_proj"
+    ]
+  },
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "model_type": "vora",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "patch_size": 14,
+  "reuse_aux_vision_embedding_layers": "",
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "skip_aux_cls": false,
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vision_attention_mask": "bidirectional",
+  "vision_embedding": "AIMv2Embedding",
+  "vision_embedding_intermediate_size": 1536,
+  "vocab_size": 151936
+}

configuration_vora.py

@@ -0,0 +1,38 @@
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig
+
+__all__ = ["VoRAConfig"]
+
+
+class VoRAConfig(PretrainedConfig):
+    model_type = "vora"
+    _auto_class = "AutoConfig"
+
+    def __init__(
+        self,
+        llm: str = "",
+        aux_vision: str = "",
+        skip_aux_cls: bool = False,
+        reuse_aux_vision_embedding_layers: str = "",
+        lora: dict = {},
+        image_size: int = 448,
+        vision_embedding: str = "AIMv2",
+        vision_embedding_intermediate_size: int = 1536,
+        patch_size: int = 14,
+        vision_attention_mask: str = "bidirectional",
+        rms_norm_eps: float = 1e-5,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.llm = llm
+        self.aux_vision = aux_vision
+        self.skip_aux_cls = skip_aux_cls
+        self.reuse_aux_vision_embedding_layers = reuse_aux_vision_embedding_layers
+        self.lora = lora
+        self.image_size = image_size
+        self.vision_embedding = vision_embedding
+        self.vision_embedding_intermediate_size = vision_embedding_intermediate_size
+        self.patch_size = patch_size
+        self.vision_attention_mask = vision_attention_mask
+        self.rms_norm_eps = rms_norm_eps

eva_model.py

@@ -0,0 +1,768 @@
+# --------------------------------------------------------
+# Adapted from EVA CLIP
+# https://github.com/baaivision/EVA/tree/master/EVA-CLIP/rei/eva_clip
+# --------------------------------------------------------
+
+import math
+import os
+import warnings
+from functools import partial
+from easydict import EasyDict as edict
+import torch
+import torch.nn as nn
+import torch.nn.functional as f
+import loguru
+
+try:
+    warnings.filterwarnings('ignore', category=FutureWarning, module='timm')
+    from timm.models.layers import drop_path as timm_drop_path
+    from timm.models.layers import to_2tuple, trunc_normal_
+except ImportError or ModuleNotFoundError:
+    from timm.layers import drop_path as timm_drop_path, to_2tuple, trunc_normal_
+
+from .rope_embeddings import VisionRotaryEmbeddingFast
+
+if os.getenv('ENV_TYPE') == 'deepspeed':
+    try:
+        from deepspeed.runtime.activation_checkpointing.checkpointing import checkpoint
+    except ImportError or ModuleNotFoundError:
+        from torch.utils.checkpoint import checkpoint
+else:
+    from torch.utils.checkpoint import checkpoint
+
+try:
+    import xformers.ops as xops
+except ImportError:
+    xops = None
+
+
+class PatchDropout(nn.Module):
+    """
+    https://arxiv.org/abs/2212.00794
+    """
+
+    def __init__(self, prob, exclude_first_token=True):
+        super().__init__()
+        assert 0 <= prob < 1.0
+        self.prob = prob
+        self.exclude_first_token = exclude_first_token  # exclude CLS token
+
+    def forward(self, x):
+        if not self.training or self.prob == 0.0:
+            return x
+
+        if self.exclude_first_token:
+            cls_tokens, x = x[:, :1], x[:, 1:]
+        else:
+            cls_tokens = torch.jit.annotate(torch.Tensor, x[:, :1])
+
+        batch = x.size()[0]
+        num_tokens = x.size()[1]
+
+        batch_indices = torch.arange(batch)
+        batch_indices = batch_indices[..., None]
+
+        keep_prob = 1 - self.prob
+        num_patches_keep = max(1, int(num_tokens * keep_prob))
+
+        rand = torch.randn(batch, num_tokens)
+        patch_indices_keep = rand.topk(num_patches_keep, dim=-1).indices
+
+        x = x[batch_indices, patch_indices_keep]
+
+        if self.exclude_first_token:
+            x = torch.cat((cls_tokens, x), dim=1)
+
+        return x, patch_indices_keep
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of
+    residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return timm_drop_path(x, self.drop_prob, self.training)
+
+    def extra_repr(self) -> str:
+        return 'p={}'.format(self.drop_prob)
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+        drop=0.0,
+        subln=False,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+
+        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
+
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        # x = self.drop(x)
+        # commit this for the orignal BERT implement
+        x = self.ffn_ln(x)
+
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class SwiGLU(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.SiLU,
+        drop=0.0,
+        norm_layer=nn.LayerNorm,
+        subln=False,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+
+        self.w1 = nn.Linear(in_features, hidden_features)
+        self.w2 = nn.Linear(in_features, hidden_features)
+
+        self.act = act_layer()
+        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
+        self.w3 = nn.Linear(hidden_features, out_features)
+
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x1 = self.w1(x)
+        x2 = self.w2(x)
+        hidden = self.act(x1) * x2
+        x = self.ffn_ln(hidden)
+        x = self.w3(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        qkv_bias=False,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+        window_size=None,
+        attn_head_dim=None,
+        xattn=False,
+        rope=None,
+        subln=False,
+        norm_layer=nn.LayerNorm,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.subln = subln
+        if self.subln:
+            self.q_proj = nn.Linear(dim, all_head_dim, bias=False)
+            self.k_proj = nn.Linear(dim, all_head_dim, bias=False)
+            self.v_proj = nn.Linear(dim, all_head_dim, bias=False)
+        else:
+            self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)
+
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
+            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+
+        if window_size:
+            self.window_size = window_size
+            self.num_relative_distance = (2 * window_size[0] - 1) * (
+                2 * window_size[1] - 1
+            ) + 3
+            self.relative_position_bias_table = nn.Parameter(
+                torch.zeros(self.num_relative_distance, num_heads)
+            )  # 2*Wh-1 * 2*Ww-1, nH
+            # cls to token & token 2 cls & cls to cls
+
+            # get pair-wise relative position index for each token inside the window
+            coords_h = torch.arange(window_size[0])
+            coords_w = torch.arange(window_size[1])
+            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+            relative_coords = (
+                coords_flatten[:, :, None] - coords_flatten[:, None, :]
+            )  # 2, Wh*Ww, Wh*Ww
+            relative_coords = relative_coords.permute(
+                1, 2, 0
+            ).contiguous()  # Wh*Ww, Wh*Ww, 2
+            relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+            relative_coords[:, :, 1] += window_size[1] - 1
+            relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+            relative_position_index = torch.zeros(
+                size=(window_size[0] * window_size[1] + 1,) * 2,
+                dtype=relative_coords.dtype,
+            )
+            relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+            relative_position_index[0, 0:] = self.num_relative_distance - 3
+            relative_position_index[0:, 0] = self.num_relative_distance - 2
+            relative_position_index[0, 0] = self.num_relative_distance - 1
+
+            self.register_buffer('relative_position_index', relative_position_index)
+        else:
+            self.window_size = None
+            self.relative_position_bias_table = None
+            self.relative_position_index = None
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.inner_attn_ln = norm_layer(all_head_dim) if subln else nn.Identity()
+        # self.proj = nn.Linear(all_head_dim, all_head_dim)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.xattn = xattn
+        self.xattn_drop = attn_drop
+
+        self.rope = rope
+
+    def forward(self, x, rel_pos_bias=None, attn_mask=None):
+        b, n, _ = x.shape
+        if self.subln:
+            q = f.linear(input=x, weight=self.q_proj.weight, bias=self.q_bias)
+            k = f.linear(input=x, weight=self.k_proj.weight, bias=None)
+            v = f.linear(input=x, weight=self.v_proj.weight, bias=self.v_bias)
+
+            q = q.reshape(b, n, self.num_heads, -1).permute(
+                0, 2, 1, 3
+            )  # B, num_heads, N, C
+            k = k.reshape(b, n, self.num_heads, -1).permute(0, 2, 1, 3)
+            v = v.reshape(b, n, self.num_heads, -1).permute(0, 2, 1, 3)
+        else:
+            qkv_bias = None
+            if self.q_bias is not None:
+                qkv_bias = torch.cat(
+                    (
+                        self.q_bias,
+                        torch.zeros_like(self.v_bias, requires_grad=False),
+                        self.v_bias,
+                    )
+                )
+
+            qkv = f.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+            qkv = qkv.reshape(b, n, 3, self.num_heads, -1).permute(
+                2, 0, 3, 1, 4
+            )  # 3, B, num_heads, N, C
+            q, k, v = qkv[0], qkv[1], qkv[2]
+
+        if self.rope:
+            # slightly fast impl
+            q_t = q[:, :, 1:, :]
+            ro_q_t = self.rope(q_t)
+            q = torch.cat((q[:, :, :1, :], ro_q_t), -2).type_as(v)
+
+            k_t = k[:, :, 1:, :]
+            ro_k_t = self.rope(k_t)
+            k = torch.cat((k[:, :, :1, :], ro_k_t), -2).type_as(v)
+
+        if self.xattn:
+            if xops is None:
+                raise ValueError(
+                    "Can't use xattn without xformers. Please 'pip install xformers'"
+                )
+            q = q.permute(0, 2, 1, 3)  # B, num_heads, N, C -> B, N, num_heads, C
+            k = k.permute(0, 2, 1, 3)
+            v = v.permute(0, 2, 1, 3)
+
+            x = xops.memory_efficient_attention(
+                q,
+                k,
+                v,
+                p=self.xattn_drop,
+                scale=self.scale,
+            )
+            x = x.reshape(b, n, -1)
+            x = self.inner_attn_ln(x)
+            x = self.proj(x)
+            x = self.proj_drop(x)
+        else:
+            q = q * self.scale
+            attn = q @ k.transpose(-2, -1)
+
+            if self.relative_position_bias_table is not None:
+                relative_position_bias = self.relative_position_bias_table[
+                    self.relative_position_index.view(-1)
+                ].view(
+                    self.window_size[0] * self.window_size[1] + 1,
+                    self.window_size[0] * self.window_size[1] + 1,
+                    -1,
+                )  # Wh*Ww,Wh*Ww,nH
+                relative_position_bias = relative_position_bias.permute(
+                    2, 0, 1
+                ).contiguous()  # nH, Wh*Ww, Wh*Ww
+                attn = attn + relative_position_bias.unsqueeze(0).type_as(attn)
+
+            if rel_pos_bias is not None:
+                attn = attn + rel_pos_bias.type_as(attn)
+
+            if attn_mask is not None:
+                attn_mask = attn_mask.bool()
+                attn = attn.masked_fill(~attn_mask[:, None, None, :], float('-inf'))
+
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+
+            x = (attn @ v).transpose(1, 2).reshape(b, n, -1)
+            x = self.inner_attn_ln(x)
+            x = self.proj(x)
+            x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        mlp_ratio=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        init_values=None,
+        act_layer=nn.GELU,
+        norm_layer=nn.LayerNorm,
+        window_size=None,
+        attn_head_dim=None,
+        xattn=False,
+        rope=None,
+        postnorm=False,
+        subln=False,
+        naiveswiglu=False,
+    ):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=drop,
+            window_size=window_size,
+            attn_head_dim=attn_head_dim,
+            xattn=xattn,
+            rope=rope,
+            subln=subln,
+            norm_layer=norm_layer,
+        )
+        # NOTE: drop path for stochastic depth, we shall see if this is better
+        # than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+
+        if naiveswiglu:
+            self.mlp = SwiGLU(
+                in_features=dim,
+                hidden_features=mlp_hidden_dim,
+                subln=subln,
+                norm_layer=norm_layer,
+            )
+        else:
+            self.mlp = Mlp(
+                in_features=dim,
+                hidden_features=mlp_hidden_dim,
+                act_layer=act_layer,
+                subln=subln,
+                drop=drop,
+            )
+
+        if init_values is not None and init_values > 0:
+            self.gamma_1 = nn.Parameter(
+                init_values * torch.ones((dim,)), requires_grad=True
+            )
+            self.gamma_2 = nn.Parameter(
+                init_values * torch.ones((dim,)), requires_grad=True
+            )
+        else:
+            self.gamma_1, self.gamma_2 = None, None
+
+        self.postnorm = postnorm
+
+    def forward(self, x, rel_pos_bias=None, attn_mask=None):
+        if self.gamma_1 is None:
+            if self.postnorm:
+                x = x + self.drop_path(
+                    self.norm1(
+                        self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)
+                    )
+                )
+                x = x + self.drop_path(self.norm2(self.mlp(x)))
+            else:
+                x = x + self.drop_path(
+                    self.attn(
+                        self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask
+                    )
+                )
+                x = x + self.drop_path(self.mlp(self.norm2(x)))
+        else:
+            if self.postnorm:
+                x = x + self.drop_path(
+                    self.gamma_1
+                    * self.norm1(
+                        self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)
+                    )
+                )
+                x = x + self.drop_path(self.gamma_2 * self.norm2(self.mlp(x)))
+            else:
+                x = x + self.drop_path(
+                    self.gamma_1
+                    * self.attn(
+                        self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask
+                    )
+                )
+                x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """Image to Patch Embedding"""
+
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
+        self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=patch_size, stride=patch_size
+        )
+
+    def forward(self, x, **_):
+        target_dtype = self.proj.weight.dtype
+        _, __, h, w = x.shape
+        # FIXME look at relaxing size constraints
+        assert h == self.img_size[0] and w == self.img_size[1], (
+            f"Input image size ({h}*{w}) doesn't match model "
+            f'({self.img_size[0]}*{self.img_size[1]}).'
+        )
+        x = self.proj(x.to(dtype=target_dtype)).flatten(2).transpose(1, 2)
+        return x
+
+
+class RelativePositionBias(nn.Module):
+    def __init__(self, window_size, num_heads):
+        super().__init__()
+        self.window_size = window_size
+        self.num_relative_distance = (2 * window_size[0] - 1) * (
+            2 * window_size[1] - 1
+        ) + 3
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros(self.num_relative_distance, num_heads)
+        )  # 2*Wh-1 * 2*Ww-1, nH
+        # cls to token & token 2 cls & cls to cls
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(window_size[0])
+        coords_w = torch.arange(window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = (
+            coords_flatten[:, :, None] - coords_flatten[:, None, :]
+        )  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(
+            1, 2, 0
+        ).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * window_size[1] - 1
+        relative_position_index = torch.zeros(
+            size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype
+        )
+        relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        relative_position_index[0, 0:] = self.num_relative_distance - 3
+        relative_position_index[0:, 0] = self.num_relative_distance - 2
+        relative_position_index[0, 0] = self.num_relative_distance - 1
+
+        self.register_buffer('relative_position_index', relative_position_index)
+
+    def forward(self):
+        relative_position_bias = self.relative_position_bias_table[
+            self.relative_position_index.view(-1)
+        ].view(
+            self.window_size[0] * self.window_size[1] + 1,
+            self.window_size[0] * self.window_size[1] + 1,
+            -1,
+        )  # Wh*Ww,Wh*Ww,nH
+        return relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+
+
+class EVAVisionTransformer(nn.Module):
+    """Vision Transformer with support for patch or hybrid CNN input stage"""
+
+    def __init__(
+        self,
+        img_size=224,
+        patch_size=16,
+        in_chans=3,
+        num_classes=0,
+        embed_dim=768,
+        depth=12,
+        num_heads=12,
+        mlp_ratio=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop_rate=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.0,
+        norm_layer=nn.LayerNorm,
+        init_values=None,
+        patch_dropout=0.0,
+        use_abs_pos_emb=True,
+        use_rel_pos_bias=False,
+        use_shared_rel_pos_bias=False,
+        rope=False,
+        use_mean_pooling=True,
+        init_scale=0.001,
+        grad_checkpointing=False,
+        xattn=False,
+        postnorm=False,
+        pt_hw_seq_len=16,
+        intp_freq=False,
+        naiveswiglu=False,
+        subln=False,
+        proj_type=None,
+    ):
+        super().__init__()
+        self.image_size = img_size
+        self.num_classes = num_classes
+        # num_features for consistency with other models
+        self.num_features = self.embed_dim = embed_dim
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        # self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
+        if use_abs_pos_emb:
+            self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+        else:
+            self.pos_embed = None
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        if use_shared_rel_pos_bias:
+            self.rel_pos_bias = RelativePositionBias(
+                window_size=self.patch_embed.patch_shape, num_heads=num_heads
+            )
+        else:
+            self.rel_pos_bias = None
+
+        if rope:
+            half_head_dim = embed_dim // num_heads // 2
+            hw_seq_len = img_size // patch_size
+            self.rope = VisionRotaryEmbeddingFast(
+                dim=half_head_dim,
+                pt_seq_len=pt_hw_seq_len,
+                ft_seq_len=hw_seq_len if intp_freq else None,
+                patch_dropout=patch_dropout,
+            )
+        else:
+            self.rope = None
+
+        self.naiveswiglu = naiveswiglu
+
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, depth)
+        ]  # stochastic depth decay rule
+        self.use_rel_pos_bias = use_rel_pos_bias
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[i],
+                    norm_layer=norm_layer,
+                    init_values=init_values,
+                    window_size=self.patch_embed.patch_shape
+                    if use_rel_pos_bias
+                    else None,
+                    xattn=xattn,
+                    rope=self.rope,
+                    postnorm=postnorm,
+                    subln=subln,
+                    naiveswiglu=naiveswiglu,
+                )
+                for i in range(depth)
+            ]
+        )
+        self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)
+        self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None
+        if (num_classes == embed_dim) and (proj_type is None):
+            self.head = nn.Identity()
+        elif proj_type == 'linear':
+            self.head = nn.Linear(embed_dim, num_classes, bias=qkv_bias)
+        elif proj_type == 'mlp':
+            hidden_size = (embed_dim + num_classes) // 2
+            self.proj = nn.Sequential(
+                nn.Linear(embed_dim, hidden_size, bias=qkv_bias),
+                nn.GELU(),
+                nn.Linear(hidden_size, num_classes, bias=qkv_bias),
+            )
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=0.02)
+
+        trunc_normal_(self.cls_token, std=0.02)
+
+        self.apply(self._init_weights)
+        self.fix_init_weight()
+
+        if isinstance(self.head, nn.Linear):
+            trunc_normal_(self.head.weight, std=0.02)
+            self.head.weight.data.mul_(init_scale)
+            if qkv_bias:
+                self.head.bias.data.mul_(init_scale)
+
+        # setting a patch_dropout of 0. would mean it is disabled and this function
+        # would be the identity fn
+        self.patch_dropout = (
+            PatchDropout(patch_dropout) if patch_dropout > 0.0 else nn.Identity()
+        )
+
+        self.grad_checkpointing = grad_checkpointing
+
+    def fix_init_weight(self):
+        def rescale(param, _layer_id):
+            param.div_(math.sqrt(2.0 * _layer_id))
+
+        for layer_id, layer in enumerate(self.blocks):
+            rescale(layer.attn.proj.weight.data, layer_id + 1)
+            if self.naiveswiglu:
+                rescale(layer.mlp.w3.weight.data, layer_id + 1)
+            else:
+                rescale(layer.mlp.fc2.weight.data, layer_id + 1)
+
+    def get_cast_dtype(self) -> torch.dtype:
+        return self.blocks[0].mlp.fc2.weight.dtype
+
+    @staticmethod
+    def _init_weights(m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def get_num_layers(self):
+        return len(self.blocks)
+
+    def lock(self, unlocked_groups=0, *_, **__):
+        assert (
+            unlocked_groups == 0
+        ), 'partial locking not currently supported for this model'
+        for param in self.parameters():
+            param.requires_grad = False
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        self.grad_checkpointing = enable
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, *_, **__):
+        self.num_classes = num_classes
+        self.head = (
+            nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+        )
+
+    def forward_features(self, x, return_all_features=False):
+        x = self.patch_embed(x)
+        batch_size, seq_len, _ = x.size()
+
+        cls_tokens = self.cls_token.expand(
+            batch_size, -1, -1
+        )  # stole cls_tokens impl from Phil Wang, thanks
+        x = torch.cat((cls_tokens, x), dim=1)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        x = self.pos_drop(x)
+
+        # a patch_dropout of 0. would mean it is disabled and this function would do
+        # nothing but return what was passed in
+        if self.rope is not None:
+            if self.training and not isinstance(self.patch_dropout, nn.Identity):
+                x, patch_indices_keep = self.patch_dropout(x)
+                self.rope.forward = partial(
+                    self.rope.forward, patch_indices_keep=patch_indices_keep
+                )
+            else:
+                self.rope.forward = partial(self.rope.forward, patch_indices_keep=None)
+                x = self.patch_dropout(x)
+        else:
+            x = self.patch_dropout(x)
+
+        rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
+        hidden_states = [x[:, 1:]]
+        for blk in self.blocks:
+            if self.grad_checkpointing:
+                x = checkpoint(blk, x, (rel_pos_bias,))
+            else:
+                x = blk(x, rel_pos_bias=rel_pos_bias)
+            
+            hidden_states.append(x[:, 1:])
+        
+        return edict(
+            {
+                'hidden_states': hidden_states,
+                'last_hidden_state': x,
+                'cls_token': x[:, 0],
+            }
+        )
+
+    def forward(self, x, return_all_features=False, **kwargs):
+        
+        return self.forward_features(x, return_all_features=return_all_features)

latest

@@ -0,0 +1 @@
+global_step5200

lora.py

@@ -0,0 +1,104 @@
+import torch
+import types
+import math
+from torch import nn
+import torch.nn.functional as F
+
+
+QWEN2_TARGET_MODULES = [
+    "self_attn.q_proj",
+    "self_attn.k_proj",
+    "self_attn.v_proj",
+    "self_attn.o_proj",
+    "mlp.up_proj",
+    "mlp.gate_proj",
+    "mlp.down_proj",
+]
+
+
+class LoRALayer(nn.Linear):
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        r: int = 1024,
+        **kwargs
+    ):
+        nn.Linear.__init__(self, in_features, out_features)
+        # we elimate lora_alpha here bc we find it unnecessary in VoRA
+        if r < 0:
+            self.forward = self.naive_forward
+        else:
+            self.lora_A = nn.Linear(in_features, r, bias=False)
+            self.lora_B = nn.Linear(r, out_features, bias=False)
+            nn.init.kaiming_uniform_(self.lora_A.weight, a=math.sqrt(5))
+            nn.init.zeros_(self.lora_B.weight)
+
+    def forward(self, x: torch.Tensor):
+        intermediate = F.linear(x, self.weight, bias=self.bias)
+        result = intermediate + self.lora_B(self.lora_A(x))
+        return result
+
+    def naive_forward(self, x: torch.Tensor):
+        return F.linear(x, self.weight, bias=self.bias)
+
+
+def _get_submodules(self, key):
+    parent = self.get_submodule(".".join(key.split(".")[:-1]))
+    target_name = key.split(".")[-1]
+    target = self.get_submodule(key)
+    return parent, target, target_name
+
+
+def _find_and_replace(self, lora_params):
+    target_modules = lora_params["target_modules"]
+
+    for llm_module_name in target_modules:
+        parent, target, target_name = self._get_submodules(llm_module_name)
+        bias = target.bias is not None
+        vora_layer = LoRALayer(
+            target.in_features,
+            target.out_features,
+            bias=bias,
+            **lora_params
+        )
+        self._replace_module(parent, target_name, vora_layer, target)
+
+
+def _replace_module(self, parent_module, child_name, new_module, old_module):
+    setattr(parent_module, child_name, new_module)
+    new_module.weight = old_module.weight
+    if old_module.bias is not None:
+        new_module.bias = old_module.bias
+    if getattr(old_module, "state", None) is not None:
+        new_module.state = old_module.state
+        new_module.to(old_module.weight.device)
+
+
+def apply_lora(llm, lora_params={"layers": "all", "r": 1024, "target_modules": QWEN2_TARGET_MODULES}):
+    llm_num_layers = llm.config.num_hidden_layers
+    total_layers = lora_params.get("layers", "all")
+
+    # -------------------- validation check ---------------------
+    if isinstance(total_layers, str):
+        if total_layers.lower() == "all":
+            total_layers = list(range(llm_num_layers))
+    else:
+        assert isinstance(total_layers, int), "total_layers must be an integer or 'all'"
+        total_layers = list(range(total_layers))
+    # -------------------- validation check ---------------------
+
+    # -------------------- replace llm layers ---------------------
+    for i in total_layers:
+        llm_layer = llm.model.layers[i]
+        llm_layer._get_submodules = types.MethodType(_get_submodules, llm_layer)
+        llm_layer._find_and_replace = types.MethodType(_find_and_replace, llm_layer)
+        llm_layer._replace_module = types.MethodType(_replace_module, llm_layer)
+        llm_layer._find_and_replace(lora_params)
+
+
+if __name__ == "__main__":
+    from transformers import LlamaForCausalLM, CLIPVisionModel, AutoModel
+    llama = LlamaForCausalLM.from_pretrained("/mnt/bn/wh-data/data/models/llama2_7b_hf_chat")
+    apply_lora(llama)
+    print(llama)

model-00001-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3044c43e513388c043d65715ff65b2bb698b1ecb8e0f4f7473cd43c7b66a82b9
+size 4999235192

model-00002-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c1874fdb1f01ba18924426aaed32499e136e161c4cdb2d6bb556ef9b4c4d4cee
+size 4992139816

model-00003-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81b074b3c6706d5c560b3d5801abd9b3dace504c2ec352f886635ab2314cc415
+size 4912324776

model-00004-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c78ab40f2c92347e41b819c62d45080d4263e2e69070d1fbea700387a59a835
+size 3959630128

model-00005-of-00005.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f737b07b8709cad6da08c5266d82e93a7b2b467649d836ff6288038cf9bc94f
+size 2073391616

modeling_vora.py

@@ -0,0 +1,291 @@
+import torch
+import torch.distributed as dist
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    PreTrainedModel,
+    PretrainedConfig,
+)
+
+import loguru
+from .attention_mask import make_mask
+from .configuration_vora import VoRAConfig
+from .vision_embedding import *  # hacking, let transformers find vision_embedding
+from . import vision_embedding as VB
+from .lora import apply_lora
+from .vora_generation_utils import (
+    VoraGenerationMixin,
+    custom_prepare_4d_causal_attention_mask_with_cache_position,
+)
+
+try:
+    from utils import logging
+except:
+    from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class VoRAForCausalLM(PreTrainedModel):
+    config_class = VoRAConfig
+    _auto_class = 'AutoModelForCausalLM'
+    supports_gradient_checkpointing = True
+    supports_report_metrics: bool = True
+
+    def __init__(self, config: PretrainedConfig = VoRAConfig()):
+        super().__init__(config)
+        self.config = config
+        # -------------- Setup LLM ---------------------
+        self.llm = AutoModelForCausalLM.from_pretrained(config.llm)
+        self.tokenizer = AutoTokenizer.from_pretrained(config.llm)
+        self.llm.__class__ = type(self.llm.__class__.__name__, (self.llm.__class__, VoraGenerationMixin), {})
+        self.llm.model._prepare_4d_causal_attention_mask_with_cache_position = staticmethod(custom_prepare_4d_causal_attention_mask_with_cache_position)
+
+        self.config.update(self.llm.config.to_dict())
+
+        # -------------- Setup LoRA -------------------
+        if config.lora:
+            for _, param in self.llm.named_parameters():
+                param.requires_grad = False
+            apply_lora(self.llm, config.lora)
+        # ----------------------------------------------
+
+        # ------------ Setup Vision Embedding ----------
+        self.vision_embedding = getattr(VB, config.vision_embedding)(self.config)  # setup after llm so that we know the hiddensize
+        # ----------------------------------------------
+
+        # ------------- Setup Aux Vision ---------------
+        self.enable_aux_vision = False
+        if config.aux_vision:
+            from .aux_vision import AuxVision
+            self.enable_aux_vision = True
+            self.aux_vision = AuxVision(self.config)
+            if config.reuse_aux_vision_embedding_layers:
+                weights = getattr(self.aux_vision.aux_model, config.reuse_aux_vision_embedding_layers).state_dict()
+                msg = self.vision_embedding.load_state_dict(weights, strict=False)
+                msg = self.vision_embedding.patchifier.load_state_dict(weights, strict=False)
+                logger.info(f"Loaded aux vision weights: {msg}")
+        # ----------------------------------------------
+        # print trainable prameters and total parameters so that we can check if we are loading the correct model
+        logger.info("Trainable parameters:")
+        for name, param in self.named_parameters():
+            if param.requires_grad:
+                logger.info(f"{name}: {param.numel()}")
+        logger.info(f"Total parameters: {sum(p.numel() for p in self.parameters())}")
+
+    def detach_and_gather_loss(self, loss, dtype, device):
+        if not dist.is_initialized():
+            return loss.item()
+        gathered_loss = [torch.tensor(0.0, dtype=loss.dtype).to(device) for _ in range(dist.get_world_size())]
+        dist.all_gather(gathered_loss, loss.detach().clone())
+        avg_gathered_loss = torch.mean(torch.stack(gathered_loss))
+        return avg_gathered_loss.item()
+
+    def _encode_vision(self, images, n_frames):
+        # TODO: we need a more elegant way here to deal with mixed image and pure text training
+        if images.size(0) > 0:
+            vision_embeds = self.vision_embedding(images)
+        else:
+            # FIXME: hacking for deepspeed training
+            # we feed a dummy image tensor (1, 3, H, W) into vision_encoder when training a pure-text batch
+            images = images.new_zeros((1, *images.shape[1:]))
+            vision_embeds = self.vision_embedding(images)[0:0]
+        vision_embeds = vision_embeds.split(n_frames, dim=0)
+        attention_mask = [torch.ones(feature.size()[:-1], dtype=torch.long).to(feature.device) for feature in vision_embeds]
+        vision_targets = [torch.ones(feature.size(), dtype=torch.long).to(feature.device).fill_(-100) for feature in attention_mask]
+
+        image_shapes = images.shape[-2:]
+
+        return vision_embeds, attention_mask, vision_targets, image_shapes
+
+    def _concat_embedding(self, vision_encode_out, batch, vision_placeholder_index, left_padding=False):
+        """ concat vision and text
+        """
+
+        vision_embeds, vision_atts, vision_targets, _ = vision_encode_out
+
+        input_embeds = []
+        attention_mask = []
+        targets = []
+        vision_mask = []  # set vision token as 1, text token as 0
+
+        for cur_batch_idx, cur_input_ids in enumerate(batch["input_ids"]):
+            cur_vision_embeds = vision_embeds[cur_batch_idx]
+            cur_vision_attn = vision_atts[cur_batch_idx]
+            cur_vision_targets = vision_targets[cur_batch_idx]
+            cur_attn_masks = batch["attention_mask"][cur_batch_idx]
+
+            image_token_indices = torch.where(cur_input_ids == vision_placeholder_index)[0]
+            cur_image_num = len(image_token_indices)
+            image_token_indices = list(image_token_indices) + [cur_input_ids.shape[0]]
+
+            cur_input_embeds = []
+            cur_attention_mask = []
+            cur_target = []
+            cur_vision_mask = []
+
+            # convert text before 1st <image> to embedding
+            image_token_index = image_token_indices[0]
+
+            cur_input_embeds.append(
+                self.llm.get_input_embeddings()(cur_input_ids[:image_token_index]),
+            )
+            cur_attention_mask.append(
+                cur_attn_masks[:image_token_index],
+            )
+            cur_vision_mask.append(
+                torch.zeros_like(cur_attn_masks[:image_token_index]).to(cur_attn_masks.device),
+            )
+            if "labels" in batch:
+                cur_target.append(
+                    batch["labels"][cur_batch_idx, :image_token_index],
+                )
+
+            if batch.get("vison_placeholder_mode", 0) == 1:
+                assert cur_image_num <= 1, "multiple video input is not supported"
+                cur_vision_embeds = cur_vision_embeds.unsqueeze(0)
+                cur_vision_attn = cur_vision_attn.unsqueeze(0)
+                cur_vision_targets = cur_vision_targets.unsqueeze(0)
+            assert cur_image_num == len(cur_vision_embeds), \
+                f"Size mismatch! cur_image_num: {cur_image_num}, len(cur_vision_embeds): {len(cur_vision_embeds)} {len(cur_vision_embeds)} \
+                    in {batch['prompt'][cur_batch_idx]} & {batch['gt'][cur_batch_idx]} & {batch['input_ids'][cur_batch_idx]}"
+            # convert each <image> xxx group into embedding
+            text_embedding = self.llm.get_input_embeddings()(cur_input_ids.relu())
+            for i in range(0, cur_image_num):
+                image_token_index = image_token_indices[i]
+                cur_input_embeds.extend([
+                    cur_vision_embeds[i],
+                    text_embedding[image_token_index+1:image_token_indices[i+1]]
+                ])
+                cur_attention_mask.extend([
+                    cur_vision_attn[i],
+                    cur_attn_masks[image_token_index+1:image_token_indices[i+1]]
+                ])
+                cur_vision_mask.extend([
+                    torch.ones_like(cur_vision_attn[i]).to(cur_vision_attn[i].device),
+                    torch.zeros_like(cur_attn_masks[image_token_index+1:image_token_indices[i+1]]).to(cur_vision_attn[i].device),
+                ])
+                if "labels" in batch:
+                    cur_target.extend([
+                        cur_vision_targets[i],
+                        batch["labels"][cur_batch_idx, image_token_index+1:image_token_indices[i+1]],
+                    ])
+
+            input_embeds.append(torch.cat(cur_input_embeds))
+            attention_mask.append(torch.cat(cur_attention_mask))
+            vision_mask.append(torch.cat(cur_vision_mask))
+            if "labels" in batch:
+                targets.append(torch.cat(cur_target))
+
+        # padding
+        n_tokens = [embed.shape[0] for embed in input_embeds]
+
+        max_token = max(n_tokens)
+
+        for i in range(len(input_embeds)):
+            if max_token > n_tokens[i]:
+                self.pad_id = self.tokenizer.pad_token_id or self.tokenizer.eos_token_id
+                pad_token = torch.tensor([self.pad_id] * (max_token - n_tokens[i]))
+                pad_embedding = self.llm.get_input_embeddings()(pad_token.to(batch["attention_mask"][i].device))
+                pad_attention = torch.zeros(pad_embedding.shape[0], dtype=torch.long).to(batch["attention_mask"][i].device)
+                pad_targets = torch.ones(pad_attention.size(), dtype=torch.long).to(batch["attention_mask"][i].device).fill_(-100)
+
+                if left_padding:
+                    input_embeds[i] = torch.cat([pad_embedding, input_embeds[i]])
+                    attention_mask[i] = torch.cat([pad_attention, attention_mask[i]])
+                    vision_mask[i] = torch.cat([pad_attention, vision_mask[i]])
+                    if "labels" in batch:
+                        targets[i] = torch.cat([pad_targets, targets[i]])
+                else:
+                    input_embeds[i] = torch.cat([input_embeds[i], pad_embedding])
+                    attention_mask[i] = torch.cat([attention_mask[i], pad_attention])
+                    vision_mask[i] = torch.cat([vision_mask[i], pad_attention])
+                    if "labels" in batch:
+                        targets[i] = torch.cat([targets[i], pad_targets])
+
+        inputs_embeds = torch.stack(input_embeds, dim=0).type(self.llm.dtype)
+        attention_mask = torch.stack(attention_mask, dim=0)
+        vision_mask = torch.stack(vision_mask, dim=0).to(attention_mask.device)
+
+        if len(targets) > 0:
+            targets = torch.stack(targets, dim=0)
+
+        attention_mask = make_mask(
+            attention_mask,
+            mode=self.config.vision_attention_mask,
+            vision_mask=vision_mask,
+            dtype=inputs_embeds.dtype
+        )
+
+        return inputs_embeds, attention_mask, targets, vision_mask
+
+    def forward(self, **batch):
+        # -------------- Vision/Text Embedding ----------
+        vision_placeholder_index = batch.pop("vision_placeholder_index")
+        images, n_frames = batch["frames"], batch["n_frames"]
+        vision_encode_out = self._encode_vision(images, n_frames)
+        inputs_embeds, attention_mask, targets, vision_mask = self._concat_embedding(
+            vision_encode_out, batch, vision_placeholder_index)
+        # -----------------------------------------------
+
+        outputs = self.llm(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            labels=targets,
+            return_dict=True,
+            output_hidden_states=True,
+        )
+
+        llm_loss = outputs.loss
+        device = llm_loss.device
+        dtype = llm_loss.dtype
+
+        metrics = {}
+
+        metrics["llm_loss"] = self.detach_and_gather_loss(llm_loss, dtype, device)
+        if self.enable_aux_vision:
+            if images.size(0) > 0:
+                aux_losses = self.aux_vision(images, outputs.hidden_states, vision_mask)
+            else:
+                # FIXME: hacking for deepspeed training
+                aux_losses = {key: torch.tensor(0., dtype=dtype).to(device) for key in self.aux_vision.loss_keys}
+
+            aux_loss = torch.tensor(0., dtype=dtype).to(device)
+            n_aux = 0
+            for _aux_key, _aux_loss in aux_losses.items():
+                aux_loss += _aux_loss
+                n_aux += 1
+                metrics[_aux_key] = self.detach_and_gather_loss(_aux_loss, dtype, device)
+            aux_loss /= n_aux
+
+            outputs.loss = aux_loss + llm_loss
+        metrics["total_loss"] = self.detach_and_gather_loss(outputs.loss, dtype, device)
+        self.report_metrics(**metrics)
+
+        return outputs
+
+    def generate(self, batch, **generate_params):
+
+        with torch.amp.autocast(
+            enabled=(self.device != torch.device("cpu")),
+            device_type=self.device.type,
+        ):
+            # get vision token
+            vision_placeholder_index = batch.pop("vision_placeholder_index")
+
+            # get vision features
+            images, n_frames = batch["frames"], batch["n_frames"]
+            vision_encode_out = self._encode_vision(images, n_frames)
+
+            inputs_embeds, attention_mask, _, _ = self._concat_embedding(
+                vision_encode_out, batch, vision_placeholder_index, left_padding=False)
+
+        outputs = self.llm.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            output_attentions=True,
+            **generate_params
+        )
+
+        return outputs

rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78d3f197f6c6558fa8056324f1563ab9e957255f5a1a959362aa4eed7a9545db
+size 15984

rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c1a9c65c2869356282cad6b4a0f7dff7f4dd68ab3d9d216c72b7d6cb524f860
+size 15984

rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:896febe768e17bae5022a95960c041f6425783774ec8859d99d3b149063b1bf9
+size 15984

rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eac482d57e966585467c8ef44dae2869bf7e5d92886f69c11ed7bccc34c07efe
+size 15984

rng_state_4.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1f27d227a20dc320ac283e0938fb2f6e5b475829a583f8c44d1a16a8c828307
+size 15984

rng_state_5.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d05a7106aaeaec4b81704e3f4a998b5123cf9342a6733bd9fd2d578e99108c3b
+size 15984

rng_state_6.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b94120d8d88502ec8d8b623ec7550315caca003b44fcffbb5767ab0de91baefe
+size 15984

rng_state_7.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:332e4d901be380f740b5d8578f7b80ef1865c7fba83bc288c8a35852205cc668
+size 15984

rope_embeddings.py

@@ -0,0 +1,160 @@
+# --------------------------------------------------------
+# Adapted from EVA CLIP
+# https://github.com/baaivision/EVA/tree/master/EVA-CLIP/rei/eva_clip
+# --------------------------------------------------------
+
+from math import pi
+
+import torch
+from einops import rearrange, repeat
+from torch import nn
+
+
+def broadcast(tensors, dim=-1):
+    num_tensors = len(tensors)
+    shape_lens = set(list(map(lambda t: len(t.shape), tensors)))
+    assert len(shape_lens) == 1, 'tensors must all have the same number of dimensions'
+    shape_len = list(shape_lens)[0]
+    dim = (dim + shape_len) if dim < 0 else dim
+    dims = list(zip(*map(lambda t: list(t.shape), tensors)))
+    expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
+    assert all(
+        [*map(lambda t: len(set(t[1])) <= 2, expandable_dims)]
+    ), 'invalid dimensions for broadcastable concatentation'
+    max_dims = list(map(lambda t: (t[0], max(t[1])), expandable_dims))
+    expanded_dims = list(map(lambda t: (t[0], (t[1],) * num_tensors), max_dims))
+    expanded_dims.insert(dim, (dim, dims[dim]))
+    expandable_shapes = list(zip(*map(lambda t: t[1], expanded_dims)))
+    tensors = list(map(lambda t: t[0].expand(*t[1]), zip(tensors, expandable_shapes)))
+    return torch.cat(tensors, dim=dim)
+
+
+def rotate_half(x):
+    x = rearrange(x, '... (d r) -> ... d r', r=2)
+    x1, x2 = x.unbind(dim=-1)
+    x = torch.stack((-x2, x1), dim=-1)
+    return rearrange(x, '... d r -> ... (d r)')
+
+
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim,
+        pt_seq_len,
+        ft_seq_len=None,
+        custom_freqs=None,
+        freqs_for='lang',
+        theta=10000,
+        max_freq=10,
+        num_freqs=1,
+    ):
+        super().__init__()
+        if custom_freqs:
+            freqs = custom_freqs
+        elif freqs_for == 'lang':
+            freqs = 1.0 / (
+                theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)
+            )
+        elif freqs_for == 'pixel':
+            freqs = torch.linspace(1.0, max_freq / 2, dim // 2) * pi
+        elif freqs_for == 'constant':
+            freqs = torch.ones(num_freqs).float()
+        else:
+            raise ValueError(f'unknown modality {freqs_for}')
+
+        if ft_seq_len is None:
+            ft_seq_len = pt_seq_len
+        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len
+
+        freqs_h = torch.einsum('..., f -> ... f', t, freqs)
+        freqs_h = repeat(freqs_h, '... n -> ... (n r)', r=2)
+
+        freqs_w = torch.einsum('..., f -> ... f', t, freqs)
+        freqs_w = repeat(freqs_w, '... n -> ... (n r)', r=2)
+
+        freqs = broadcast((freqs_h[:, None, :], freqs_w[None, :, :]), dim=-1)
+
+        self.register_buffer('freqs_cos', freqs.cos(), persistent=False)
+        self.register_buffer('freqs_sin', freqs.sin(), persistent=False)
+
+    def forward(self, t, start_index=0):
+        rot_dim = self.freqs_cos.shape[-1]
+        end_index = start_index + rot_dim
+        assert rot_dim <= t.shape[-1], (
+            f'feature dimension {t.shape[-1]} is not of sufficient size to rotate in '
+            f'all the positions {rot_dim}'
+        )
+        t_left, t, t_right = (
+            t[..., :start_index],
+            t[..., start_index:end_index],
+            t[..., end_index:],
+        )
+        t = (t * self.freqs_cos) + (rotate_half(t) * self.freqs_sin)
+
+        return torch.cat((t_left, t, t_right), dim=-1)
+
+
+class VisionRotaryEmbeddingFast(nn.Module):
+    def __init__(
+        self,
+        dim,
+        pt_seq_len,
+        ft_seq_len=None,
+        custom_freqs=None,
+        freqs_for='lang',
+        theta=10000,
+        max_freq=10,
+        num_freqs=1,
+        patch_dropout=0.0,
+    ):
+        super().__init__()
+        if custom_freqs:
+            freqs = custom_freqs
+        elif freqs_for == 'lang':
+            freqs = 1.0 / (
+                theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)
+            )
+        elif freqs_for == 'pixel':
+            freqs = torch.linspace(1.0, max_freq / 2, dim // 2) * pi
+        elif freqs_for == 'constant':
+            freqs = torch.ones(num_freqs).float()
+        else:
+            raise ValueError(f'unknown modality {freqs_for}')
+
+        if ft_seq_len is None:
+            ft_seq_len = pt_seq_len
+        t = torch.arange(ft_seq_len) / ft_seq_len * pt_seq_len
+
+        freqs = torch.einsum('..., f -> ... f', t, freqs)
+        freqs = repeat(freqs, '... n -> ... (n r)', r=2)
+        freqs = broadcast((freqs[:, None, :], freqs[None, :, :]), dim=-1)
+
+        freqs_cos = freqs.cos().view(-1, freqs.shape[-1])
+        freqs_sin = freqs.sin().view(-1, freqs.shape[-1])
+
+        self.patch_dropout = patch_dropout
+
+        self.register_buffer('freqs_cos', freqs_cos, persistent=False)
+        self.register_buffer('freqs_sin', freqs_sin, persistent=False)
+
+    def forward(self, t, patch_indices_keep=None):
+        if patch_indices_keep is not None:
+            batch = t.size()[0]
+            batch_indices = torch.arange(batch)
+            batch_indices = batch_indices[..., None]
+
+            freqs_cos = repeat(
+                self.freqs_cos, 'i j -> n i m j', n=t.shape[0], m=t.shape[1]
+            )
+            freqs_sin = repeat(
+                self.freqs_sin, 'i j -> n i m j', n=t.shape[0], m=t.shape[1]
+            )
+
+            freqs_cos = freqs_cos[batch_indices, patch_indices_keep]
+            freqs_cos = rearrange(freqs_cos, 'n i m j -> n m i j')
+            freqs_sin = freqs_sin[batch_indices, patch_indices_keep]
+            freqs_sin = rearrange(freqs_sin, 'n i m j -> n m i j')
+
+            return t * freqs_cos + rotate_half(t) * freqs_sin
+
+        return t * self.freqs_cos + rotate_half(t) * self.freqs_sin

scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f05512bf38916e185cca93d0ada0f63479a3d982044c9a30eec1c58ba2ff27e3
+size 1064

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95a8c73db4abbe7f04d20fc397e5ca2a49e6027339b0470fc61067769542260c
+size 7032

vision_embedding.py

@@ -0,0 +1,70 @@
+import torch
+import torch.nn as nn
+
+from .configuration_vora import VoRAConfig
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class AIMv2PatchEmbed(nn.Module):
+    def __init__(self, config: VoRAConfig):
+        super().__init__()
+        self.proj = nn.Conv2d(
+            3,
+            config.vision_embedding_intermediate_size,
+            kernel_size=(config.patch_size, config.patch_size),
+            stride=(config.patch_size, config.patch_size),
+        )
+        self.norm = RMSNorm(config.vision_embedding_intermediate_size, eps=config.rms_norm_eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        x = self.norm(x)
+        return x
+
+
+class AIMv2Embedding(nn.Module):
+    def __init__(self,
+        config: VoRAConfig = None,
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        num_patches = (config.image_size // config.patch_size) ** 2
+        self.config = config
+
+        self.patchifier = AIMv2PatchEmbed(config)
+        self.pos_embed = nn.Parameter(torch.zeros((1, num_patches, config.vision_embedding_intermediate_size)))
+        self.out_proj = nn.Linear(config.vision_embedding_intermediate_size, hidden_size, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = x.shape
+        h_token = H // self.config.patch_size
+        w_token = W // self.config.patch_size
+        tokens = self.patchifier(x)
+        _, N, _ = tokens.shape
+        pos_embed = self.pos_embed.to(tokens.device)
+
+        if N <= pos_embed.size(1):
+            tokens = tokens + pos_embed[:, :N]
+        else:
+            pos_embed = pos_embed.view(1, int(pos_embed.size(1)**0.5), int(pos_embed.size(1)**0.5), -1).permute(0, 3, 1, 2)
+            pos_embed = nn.functional.interpolate(pos_embed, size=(h_token, w_token), mode='bilinear', align_corners=False).permute(0, 2, 3, 1)
+            pos_embed = pos_embed.view(1, N, pos_embed.size(-1))
+            tokens = tokens + pos_embed
+
+        return self.out_proj(tokens)

vora_generation_utils.py

@@ -0,0 +1,101 @@
+from typing import Any, Dict, Optional
+
+import torch
+from transformers import GenerationMixin
+from transformers.cache_utils import Cache
+from transformers.utils import ModelOutput
+
+
+class VoraGenerationMixin(GenerationMixin):
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Cache] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        if attention_mask is not None and attention_mask.ndim == 4:
+            attention_mask_2d = (attention_mask[:, 0, :, :] == 0).any(dim=1).long().to(attention_mask.device)
+            model_input = super().prepare_inputs_for_generation(
+                input_ids,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask_2d,
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                **kwargs,
+            )
+            model_input['attention_mask'] = attention_mask
+            return model_input
+        else:
+            return super().prepare_inputs_for_generation(
+                input_ids,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+    def _update_model_kwargs_for_generation(
+        self,
+        outputs: ModelOutput,
+        model_kwargs: Dict[str, Any],
+        is_encoder_decoder: bool = False,
+        num_new_tokens: int = 1,
+    ) -> Dict[str, Any]:
+        if "attention_mask" in model_kwargs and model_kwargs["attention_mask"].ndim == 4:
+            attention_mask = model_kwargs.pop("attention_mask")
+            model_kwargs = super()._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder, num_new_tokens=num_new_tokens
+            )
+            bs, _, seq_len, tgt_len = attention_mask.shape
+            dtype = attention_mask.dtype
+            min_dtype = torch.finfo(dtype).min
+            new_col = attention_mask.new_zeros((bs, 1, seq_len, 1)).fill_(min_dtype)
+            new_row = attention_mask.new_zeros((bs, 1, 1, tgt_len + 1))
+            model_kwargs["attention_mask"] = torch.cat([
+                torch.cat([attention_mask, new_col], dim=-1),
+                new_row
+            ], dim=2)
+            return model_kwargs
+        else:
+            return super()._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder, num_new_tokens=num_new_tokens
+            )
+
+
+def custom_prepare_4d_causal_attention_mask_with_cache_position(
+    attention_mask: torch.Tensor,
+    sequence_length: int,
+    target_length: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    cache_position: torch.Tensor,
+    batch_size: int,
+    **kwargs,
+):
+    if attention_mask is not None and attention_mask.dim() == 4:
+        # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+        causal_mask = attention_mask[:, :, -sequence_length:, -target_length:]
+    else:
+        min_dtype = torch.finfo(dtype).min
+        causal_mask = torch.full(
+            (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+        )
+        if sequence_length != 1:
+            causal_mask = torch.triu(causal_mask, diagonal=1)
+        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+        if attention_mask is not None:
+            causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+            mask_length = attention_mask.shape[-1]
+            padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+            padding_mask = padding_mask == 0
+            causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                padding_mask, min_dtype
+            )
+
+    return causal_mask

zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)