init

config.json

@@ -0,0 +1,40 @@
+{
+  "_name_or_path": "cogvlm-grounding-generalist",
+  "architectures": [
+    "CogVLMForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_cogvlm.CogVLMConfig",
+    "AutoModelForCausalLM": "modeling_cogvlm.CogVLMForCausalLM"
+  },
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_position_embeddings": 2048,
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-05,
+  "template_version": "base",
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.35.0",
+  "use_cache": true,
+  "vision_config": {
+    "dropout_prob": 0.0,
+    "hidden_act": "gelu",
+    "hidden_size": 1792,
+    "image_size": 490,
+    "in_channels": 3,
+    "intermediate_size": 15360,
+    "layer_norm_eps": 1e-06,
+    "num_heads": 16,
+    "num_hidden_layers": 63,
+    "num_positions": 1226,
+    "patch_size": 14
+  },
+  "vocab_size": 32000
+}

configuration_cogvlm.py

@@ -0,0 +1,45 @@
+from typing import Literal
+from transformers import PretrainedConfig
+
+
+class CogVLMConfig(PretrainedConfig):
+    _auto_class = "AutoConfig"
+
+    def __init__(
+            self,
+            vocab_size=32000,
+            hidden_size=4096,
+            intermediate_size=11008,
+            num_hidden_layers=32,
+            num_attention_heads=32,
+            hidden_act='silu',
+            max_position_embeddings=2048,
+            initializer_range=0.02,
+            rms_norm_eps=1e-06,
+            template_version: Literal["base", "chat"] = "chat",
+
+            pad_token_id=0,
+            bos_token_id=1,
+            eos_token_id=2,
+            tie_word_embeddings=False,
+            use_cache=True,
+            **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.rms_norm_eps = rms_norm_eps
+        self.initializer_range = initializer_range
+        self.vocab_size = vocab_size
+        self.num_hidden_layers = num_hidden_layers
+        self.hidden_act = hidden_act
+        self.template_version = template_version
+        self.use_cache = use_cache
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 0,
+  "transformers_version": "4.35.0"
+}

model-00001-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab245c9b171545099d652eefcd863aa20f95b7ec2c18dea754ffa661ddcdeebd
+size 4938885184

model-00002-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4c34bd99e8317404ef58c253a01648c120b356b3b8f95933da46dd02fbb73ba
+size 4947290688

model-00003-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0affb830c00b94b9fc5ecfa41d7ae0d62fa42c73300df2537cd7c4496f947014
+size 4947307592

model-00004-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed1d44e236c4af9263f8a32406769d4e8dcf2fefad205a17c2c2191d948c0e07
+size 4991331080

model-00005-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9dc0499401747dd885bd9e212cec57296fad9b0c2d59c8c1984063f9c9d22eeb
+size 4991331088

model-00006-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7196fa96d96992a65ae041f2a2921afef149003e059bec884520d386c814ce0a
+size 4970162920

model-00007-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:340afeec8059f2d6fb2b890041611a448bb54440ae185d8a5433bf39e2711d9b
+size 4960543792

model-00008-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9122674b65c1b3c9c8d498ff684591c106081a906f0c85e9fcb8e1a2cc0a270e
+size 532677104

modeling_cogvlm.py

@@ -0,0 +1,785 @@
+"""largely copy from llama and adapt for cogvlm"""
+import warnings
+from typing import TYPE_CHECKING, Optional, Tuple, List, Union, Literal, Dict, Any
+
+import math
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from torchvision import transforms
+from einops import rearrange
+
+from transformers import PreTrainedModel, PreTrainedTokenizer
+from transformers.utils.logging import get_logger
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+
+from .configuration_cogvlm import CogVLMConfig
+from .util import FastRotaryEmbedding
+from .visual import EVA2CLIPModel
+
+if TYPE_CHECKING:
+    from transformers.utils import ModelOutput
+
+logger = get_logger(__name__)
+
+LANGUAGE_TOKEN_TYPE = 0
+VISION_TOKEN_TYPE = 1
+
+
+# Copied from transformers.models.bart.modeling_bart._make_causal_mask
+def _make_causal_mask(
+        input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+    mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, 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)
+
+    if past_key_values_length > 0:
+        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
+
+
+# Copied from transformers.models.bart.modeling_bart._expand_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)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return (self.weight * hidden_states).to(input_dtype)
+
+
+class MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def get_expert_mask(token_type_ids: "torch.LongTensor(B, L)") -> "[torch.BoolTensor(B, L), torch.BoolTensor(B, L)]":
+    vision_token_mask = torch.zeros_like(token_type_ids, dtype=torch.bool)
+    vision_token_mask[:, :-1] = (token_type_ids[:, :-1] == VISION_TOKEN_TYPE) & (token_type_ids[:, 1:] == VISION_TOKEN_TYPE)
+    language_token_mask = ~vision_token_mask
+    return vision_token_mask, language_token_mask
+
+
+class VisionExpertMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.language_mlp = MLP(config)
+        self.vision_mlp = MLP(config)
+
+    def forward(self, hidden_states: "torch.Tensor(B, L, D)", token_type_ids: "torch.LongTensor(B, L)"):
+        output = torch.empty(hidden_states.shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
+        output[vision_token_mask] = self.vision_mlp(hidden_states[vision_token_mask])
+        output[language_token_mask] = self.language_mlp(hidden_states[language_token_mask])
+        return output
+
+
+def attention_fn(
+        query_layer: "torch.tensor(B, H, L, HD)",
+        key_layer: "torch.tensor(B, H, L, HD)",
+        value_layer: "torch.tensor(B, H, L, HD)",
+        attention_mask: "torch.tensor(B, H, L, HD)",
+        *,
+        scaling_attention_score: bool = True,
+        attention_dropout: nn.Module = None
+):
+    attention_mask_bool = (attention_mask == 0)
+    is_low_triangle = (attention_mask_bool == torch.ones_like(attention_mask_bool, dtype=torch.float).tril()).all()
+    is_full = (attention_mask_bool > 0).all()
+    if not (int(torch.__version__.split('.')[0]) >= 2):
+        warnings.warn("It's recommended to use torch2.0 or higher.")
+    if int(torch.__version__.split('.')[0]) >= 2 and scaling_attention_score and (is_full or is_low_triangle):
+        dropout_p = 0. if attention_dropout is None or not attention_dropout.training else attention_dropout.p
+        return torch.nn.functional.scaled_dot_product_attention(
+            query_layer, key_layer, value_layer,
+            attn_mask=None,
+            dropout_p=dropout_p,
+            is_causal=not is_full
+        )
+    else:
+        if scaling_attention_score:
+            query_layer = query_layer / math.sqrt(query_layer.shape[-1])
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores + attention_mask
+        attention_scores = nn.functional.softmax(attention_scores, dim=-1, dtype=torch.float32).to(query_layer.dtype)
+        if attention_dropout is not None:
+            attention_scores = attention_dropout(attention_scores)
+        context_layer = torch.matmul(attention_scores, value_layer)
+        return context_layer
+
+
+class VisionExpertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+
+        # self.rotary_emb = RotaryEmbedding(self.hidden_size // self.num_heads)
+        self.rotary_emb = FastRotaryEmbedding(dim=self.head_dim, pos_idx_in_fp32=False)
+        self.vision_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False)
+        self.vision_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.language_expert_query_key_value = nn.Linear(self.hidden_size, self.hidden_size * 3, bias=False)
+        self.language_expert_dense = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+
+    def _transpose_for_scores(self, tensor):
+        """Transpose a 3D tensor [B, L, H*HD] into a 4D tensor with size [B H L HD]."""
+        new_tensor_shape = tensor.size()[:-1] + (self.num_heads, self.head_dim)
+        tensor = tensor.view(*new_tensor_shape)
+        return tensor.permute(0, 2, 1, 3)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            token_type_ids: torch.LongTensor,
+            position_ids: torch.LongTensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+        vision_token_mask, language_token_mask = get_expert_mask(token_type_ids)
+
+        shape = list(hidden_states.shape)
+        shape[-1] = shape[-1] * 3
+        mixed_raw_layer = torch.empty(shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        mixed_raw_layer[vision_token_mask] = self.vision_expert_query_key_value(hidden_states[vision_token_mask])
+        mixed_raw_layer[language_token_mask] = self.language_expert_query_key_value(hidden_states[language_token_mask])
+
+        query_states, key_states, value_states = torch.split(mixed_raw_layer, self.hidden_size, dim=-1)
+        query_states = self._transpose_for_scores(query_states)  # B, H, L, HD
+        key_states = self._transpose_for_scores(key_states)  # B, H, L, HD
+        value_states = self._transpose_for_scores(value_states)  # B, H, L, HD
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        query_states, key_states = self.rotary_emb(query_states, key_states, position_ids=position_ids, max_seqlen=position_ids.max() + 1)
+
+        if past_key_value is not None:
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        context_layer = attention_fn(
+            query_layer=query_states, key_layer=key_states, value_layer=value_states, attention_mask=attention_mask,
+            scaling_attention_score=True, attention_dropout=None)
+        if context_layer.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {context_layer.size()}"
+            )
+        context_layer = context_layer.transpose(1, 2).contiguous().reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = torch.empty(context_layer.shape, dtype=hidden_states.dtype, device=hidden_states.device)
+        attn_output[vision_token_mask] = self.vision_expert_dense(context_layer[vision_token_mask])
+        attn_output[language_token_mask] = self.language_expert_dense(context_layer[language_token_mask])
+
+        if output_attentions:
+            warnings.warn("output_attentions is not implemented.")
+
+        return attn_output, None, past_key_value
+
+
+class CogVLMDecoderLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = VisionExpertAttention(config=config)
+        self.mlp = VisionExpertMLP(config)
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            token_type_ids: torch.LongTensor,
+            position_ids: torch.LongTensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: Optional[bool] = False,
+            use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states, token_type_ids=token_type_ids)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs  # type: ignore
+
+
+class CogVLMPreTrainedModel(PreTrainedModel):
+    config_class = CogVLMConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["CogVLMDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+def is_empty(images_list: Optional[List[List[torch.Tensor]]]):
+    if images_list is None or len(images_list) == 0:
+        return True
+    for image_list in images_list:
+        if len(image_list):
+            return False
+    return True
+
+
+def build_position_ids(x: "torch.BoolTensor(B, L)", attention_mask: Optional["torch.BoolTensor(B, L)"] = None) -> "torch.LongTensor(B, L)":
+    if attention_mask is not None:
+        tmp = x.clone()
+        tmp[~(attention_mask.bool())] = -1
+    else:
+        tmp = x.clone()
+    # image boi eoi token as LANGUAGE_TOKEN_TYPE
+    is_boi_eoi = torch.zeros_like(x, dtype=torch.bool)
+    is_boi_eoi[:, 1:] |= (tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE)
+    is_boi_eoi[:, 0] |= (tmp[:, 0] == VISION_TOKEN_TYPE)
+    is_boi_eoi[:, :-1] |= (tmp[:, :-1] == VISION_TOKEN_TYPE) & (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE)
+    is_boi_eoi[:, -1] |= (tmp[:, -1] == VISION_TOKEN_TYPE)
+    tmp[is_boi_eoi] = LANGUAGE_TOKEN_TYPE
+    # final position ids
+    y = torch.zeros_like(x, dtype=torch.long)
+    y[:, 1:] = (tmp[:, 1:] == LANGUAGE_TOKEN_TYPE) | ((tmp[:, 1:] == VISION_TOKEN_TYPE) & (tmp[:, :-1] == LANGUAGE_TOKEN_TYPE))
+    y = y.cumsum(dim=-1)
+    return y
+
+
+class CogVLMModel(CogVLMPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList([CogVLMDecoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.vision = EVA2CLIPModel(config)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def encode_images(self, images: List[List[torch.Tensor]]) -> torch.Tensor:
+        images_list, images = images, []
+
+        images = []
+        for image_list in images_list:
+            for image in image_list:
+                images.append(image)
+
+        images = torch.stack(images)
+        images_features = self.vision(images)
+        return images_features
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            images: List[List[torch.Tensor]] = None,
+            token_type_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """take care of image_encode, token_type_ids, position_ids and (attention_mask = None is fine)"""
+
+        if past_key_values is not None:
+            pass  # generate mode with past_key_values. the image features are already mapped
+        else:
+            # not allow for inputs_embeds, because we want to process image feature
+            assert input_ids is not None and inputs_embeds is None, f"{input_ids} {inputs_embeds}"
+            if not is_empty(images):  # multi-modality
+                assert token_type_ids is not None, f"multi-modality requires `token_type_ids`!"
+                assert len(input_ids) == len(images), f"{len(input_ids)} {len(images)}"
+                inputs_embeds = self.embed_tokens(input_ids)
+                images_features = self.encode_images(images)
+                images_features = rearrange(images_features, 'b n d -> (b n) d')
+                images_features = images_features.to(dtype=inputs_embeds.dtype, device=inputs_embeds.device)
+                inputs_embeds = inputs_embeds.index_put([token_type_ids == VISION_TOKEN_TYPE], images_features)
+            else:  # single-modality
+                if token_type_ids is None:
+                    token_type_ids = torch.ones_like(input_ids, dtype=torch.long, device=input_ids.device) * LANGUAGE_TOKEN_TYPE
+                assert not (token_type_ids == VISION_TOKEN_TYPE).any(), f"{(token_type_ids == VISION_TOKEN_TYPE).sum()}"
+                inputs_embeds = self.embed_tokens(input_ids)
+
+            if position_ids is None:
+                position_ids = build_position_ids(token_type_ids, attention_mask)
+            input_ids = None
+
+        return self.llm_forward(
+            input_ids=input_ids,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+    def llm_forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            token_type_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        """largely copy from llama forward and adapt for cogvlm with `token_type_ids`"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        # embed positions
+        if attention_mask is None:
+            attention_mask = torch.ones(
+                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
+            )
+        attention_mask = self._prepare_decoder_attention_mask(
+            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
+        )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+            layer_outputs = decoder_layer(
+                hidden_states,
+                token_type_ids=token_type_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    # noinspection PyMethodMayBeStatic
+    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        combined_attention_mask = None
+        if input_shape[-1] > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape,
+                inputs_embeds.dtype,
+                device=inputs_embeds.device,
+                past_key_values_length=past_key_values_length,
+            )
+
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
+                inputs_embeds.device
+            )
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+
+        return combined_attention_mask
+
+
+def chat_history_to_prompt(history, query):
+    prompt = " [INST] "
+    for i, (old_query, response) in enumerate(history):
+        prompt += old_query + " [/INST] " + response + " [INST] "
+    prompt += query + " [/INST] "
+    return prompt
+
+
+def base_history_to_prompt(history, query):
+    prompt = query
+    return prompt
+
+
+_history_to_prompt = {
+    "base": base_history_to_prompt,
+    "chat": chat_history_to_prompt
+}
+
+
+class CogVLMForCausalLM(CogVLMPreTrainedModel):
+    _auto_class = "AutoModelForCausalLM"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = CogVLMModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            images: List[List[torch.Tensor]] = None,
+            token_type_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+            labels: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            images=images,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def _prepare_attention_mask_for_generation(
+            self,
+            inputs: torch.Tensor,
+            pad_token_id: Optional[int],
+            eos_token_id: Optional[Union[int, List[int]]],
+    ) -> torch.LongTensor:
+        return torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device)  # type: ignore
+
+    def prepare_inputs_for_generation(
+            self, input_ids, token_type_ids, images=None, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        # build position_ids if needed
+        position_ids = kwargs.get("position_ids", None)
+        if position_ids is None:
+            position_ids = build_position_ids(token_type_ids, attention_mask)
+
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+            token_type_ids = token_type_ids[:, -1:]
+            position_ids = position_ids[:, -1:]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "token_type_ids": token_type_ids,
+                "images": images,
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    def _update_model_kwargs_for_generation(
+            self,
+            outputs: "ModelOutput",
+            model_kwargs: Dict[str, Any],
+            is_encoder_decoder: bool = False,
+            standardize_cache_format: bool = False,
+    ) -> Dict[str, Any]:
+        # update past_key_values
+        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
+            outputs, standardize_cache_format=standardize_cache_format
+        )
+        if getattr(outputs, "state", None) is not None:
+            model_kwargs["state"] = outputs.state
+
+        # update token_type_ids with last value
+        if "token_type_ids" in model_kwargs:
+            token_type_ids = model_kwargs["token_type_ids"]
+            new_token_type_ids = torch.ones(size=(token_type_ids.shape[0], 1), dtype=token_type_ids.dtype, device=token_type_ids.device) * LANGUAGE_TOKEN_TYPE
+            model_kwargs["token_type_ids"] = torch.cat([token_type_ids, new_token_type_ids], dim=-1)
+
+        if not is_encoder_decoder:
+            # update attention mask
+            if "attention_mask" in model_kwargs:
+                attention_mask = model_kwargs["attention_mask"]
+                model_kwargs["attention_mask"] = torch.cat(
+                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+                )
+        else:
+            # update decoder attention mask
+            if "decoder_attention_mask" in model_kwargs:
+                decoder_attention_mask = model_kwargs["decoder_attention_mask"]
+                model_kwargs["decoder_attention_mask"] = torch.cat(
+                    [decoder_attention_mask, decoder_attention_mask.new_ones((decoder_attention_mask.shape[0], 1))],
+                    dim=-1,
+                )
+
+        return model_kwargs
+
+    def _reorder_cache(self, past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+    def build_conversation_input_ids(
+            self,
+            tokenizer: "PreTrainedTokenizer",
+            *,
+            query: str,
+            history: Optional[List[Tuple[str, str]]] = None,
+            images: Optional[List["PIL.Image"]] = None,
+            template_version: Optional[Literal["base", "chat"]] = None,
+    ):
+        image_size: int = self.config.vision_config['image_size']
+        patch_size: int = self.config.vision_config['patch_size']
+        template_version = template_version or self.config.template_version
+        assert images is None or len(images) <= 1, f"not support multi images by now."
+        history = history or []
+        text = _history_to_prompt[template_version](history, query)
+
+        input_ids = [tokenizer.bos_token_id]
+        token_type_ids = [LANGUAGE_TOKEN_TYPE]
+        if images is not None and len(images) == 1:
+            # vision
+            transform = transforms.Compose(
+                [
+                    transforms.Resize(
+                        (image_size, image_size), interpolation=transforms.InterpolationMode.BICUBIC
+                    ),
+                    transforms.ToTensor(),
+                    transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
+                ]
+            )
+            images = [transform(images[0])]
+            # language
+            vision_token_num = (image_size // patch_size) * (image_size // patch_size) + 2
+            input_ids += [tokenizer.pad_token_id] * vision_token_num
+            token_type_ids += [VISION_TOKEN_TYPE] * vision_token_num
+        text_ids = tokenizer.encode(text, add_special_tokens=False)
+
+        input_ids += text_ids
+        token_type_ids += [LANGUAGE_TOKEN_TYPE] * len(text_ids)
+        attention_mask = [1] * len(input_ids)
+
+        return {
+            'input_ids': torch.tensor(input_ids, dtype=torch.long),
+            'token_type_ids': torch.tensor(token_type_ids, dtype=torch.long),
+            'attention_mask': torch.tensor(attention_mask, dtype=torch.long),
+            'images': images,
+        }

util.py

@@ -0,0 +1,483 @@
+from typing import Optional, Tuple, Union
+
+import torch
+from einops import rearrange, repeat
+import torch.nn.functional as F
+
+import triton
+import triton.language as tl
+
+
+# @triton.autotune(
+#     configs=[
+#         triton.Config({"BLOCK_M": 2}),
+#         triton.Config({"BLOCK_M": 4}),
+#         triton.Config({"BLOCK_M": 8}),
+#         triton.Config({"BLOCK_M": 16}),
+#     ],
+#     key=["CACHE_KEY_SEQLEN", "BLOCK_K", "INTERLEAVED"],
+# )
+@triton.jit
+def rotary_kernel(
+        OUT,  # Pointers to matrices
+        X,
+        COS,
+        SIN,
+        CU_SEQLENS,
+        SEQLEN_OFFSETS,  # this could be int or a pointer
+        # Matrix dimensions
+        seqlen,
+        nheads,
+        rotary_dim,
+        seqlen_ro,
+        CACHE_KEY_SEQLEN,
+        # strides
+        stride_out_batch,
+        stride_out_nheads,
+        stride_out_seqlen,
+        stride_out_headdim,
+        stride_x_batch,
+        stride_x_nheads,
+        stride_x_seqlen,
+        stride_x_headdim,
+        # Meta-parameters
+        BLOCK_K: tl.constexpr,
+        IS_SEQLEN_OFFSETS_TENSOR: tl.constexpr,
+        IS_VARLEN: tl.constexpr,
+        INTERLEAVED: tl.constexpr,
+        CONJUGATE: tl.constexpr,
+        BLOCK_M: tl.constexpr,
+):
+    pid_m = tl.program_id(axis=0)
+    pid_batch = tl.program_id(axis=1)
+    pid_head = tl.program_id(axis=2)
+    rotary_dim_half = rotary_dim // 2
+
+    if not IS_VARLEN:
+        X = X + pid_batch * stride_x_batch + pid_head * stride_x_nheads
+        OUT = OUT + pid_batch * stride_out_batch + pid_head * stride_out_nheads
+        COS = COS + pid_batch * seqlen_ro * rotary_dim_half
+        SIN = SIN + pid_batch * seqlen_ro * rotary_dim_half
+    else:
+        start_idx = tl.load(CU_SEQLENS + pid_batch)
+        seqlen = tl.load(CU_SEQLENS + pid_batch + 1) - start_idx
+        X = X + start_idx * stride_x_seqlen + pid_head * stride_x_nheads
+        OUT = OUT + start_idx * stride_out_seqlen + pid_head * stride_out_nheads
+
+    if pid_m * BLOCK_M >= seqlen:
+        return
+    rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    if not IS_SEQLEN_OFFSETS_TENSOR:
+        rm_cs = rm + SEQLEN_OFFSETS
+    else:
+        rm_cs = rm + tl.load(SEQLEN_OFFSETS + pid_batch)
+    rk = tl.arange(0, BLOCK_K)
+    rk_half = tl.arange(0, BLOCK_K // 2)
+
+    if not INTERLEAVED:
+        # Load the 1st and 2nd halves of X, do calculation, then store to 1st and 2nd halves of OUT
+        X = X + (rm[:, None] * stride_x_seqlen + rk_half[None, :] * stride_x_headdim)
+        COS = COS + (rm_cs[:, None] * rotary_dim_half + rk_half[None, :])
+        SIN = SIN + (rm_cs[:, None] * rotary_dim_half + rk_half[None, :])
+        cos = tl.load(
+            COS, mask=(rm_cs[:, None] < seqlen_ro) & (rk_half[None, :] < rotary_dim_half), other=1.0
+        )
+        sin = tl.load(
+            SIN, mask=(rm_cs[:, None] < seqlen_ro) & (rk_half[None, :] < rotary_dim_half), other=0.0
+        )
+        x0 = tl.load(
+            X, mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half), other=0.0
+        )
+        x1 = tl.load(
+            X + rotary_dim_half * stride_x_headdim,
+            mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half),
+            other=0.0,
+        )
+        if CONJUGATE:
+            sin = -sin
+        o0 = x0 * cos - x1 * sin
+        o1 = x0 * sin + x1 * cos
+        # write back result
+        OUT = OUT + (rm[:, None] * stride_out_seqlen + rk_half[None, :] * stride_out_headdim)
+        tl.store(OUT, o0, mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half))
+        tl.store(
+            OUT + rotary_dim_half * stride_out_headdim,
+            o1,
+            mask=(rm[:, None] < seqlen) & (rk_half[None, :] < rotary_dim_half),
+        )
+    else:
+        # We don't want to load X[0, 2, 4, ...] and X[1, 3, 5, ...] separately since both are slow.
+        # Instead, we load x0 = X[0, 1, 2, 3, ...] and x1 = X[1, 0, 3, 2, ...].
+        # Loading x0 will be fast but x1 will be slow.
+        # Then we load cos = COS[0, 0, 1, 1, ...] and sin = SIN[0, 0, 1, 1, ...].
+        # Then we do the calculation and use tl.where to pick put the right outputs for the even
+        # and for the odd indices.
+        rk_swap = rk + ((rk + 1) % 2) * 2 - 1  # 1, 0, 3, 2, 5, 4, ...
+        rk_repeat = tl.arange(0, BLOCK_K) // 2
+        X0 = X + (rm[:, None] * stride_x_seqlen + rk[None, :] * stride_x_headdim)
+        X1 = X + (rm[:, None] * stride_x_seqlen + rk_swap[None, :] * stride_x_headdim)
+        COS = COS + (rm_cs[:, None] * rotary_dim_half + rk_repeat[None, :])
+        SIN = SIN + (rm_cs[:, None] * rotary_dim_half + rk_repeat[None, :])
+        cos = tl.load(
+            COS,
+            mask=(rm_cs[:, None] < seqlen_ro) & (rk_repeat[None, :] < rotary_dim_half),
+            other=1.0,
+        ).to(tl.float32)
+        sin = tl.load(
+            SIN,
+            mask=(rm_cs[:, None] < seqlen_ro) & (rk_repeat[None, :] < rotary_dim_half),
+            other=0.0,
+        ).to(tl.float32)
+        x0 = tl.load(X0, mask=(rm[:, None] < seqlen) & (rk[None, :] < rotary_dim), other=0.0).to(
+            tl.float32
+        )
+        x1 = tl.load(
+            X1, mask=(rm[:, None] < seqlen) & (rk_swap[None, :] < rotary_dim), other=0.0
+        ).to(tl.float32)
+        if CONJUGATE:
+            sin = -sin
+        x0_cos = x0 * cos
+        x1_sin = x1 * sin
+        out = tl.where(rk[None, :] % 2 == 0, x0_cos - x1_sin, x0_cos + x1_sin)
+        OUT = OUT + (rm[:, None] * stride_out_seqlen + rk[None, :] * stride_out_headdim)
+        tl.store(OUT, out, mask=(rm[:, None] < seqlen) & (rk[None, :] < rotary_dim))
+
+
+def apply_rotary(
+        x: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+        seqlen_offsets: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+        interleaved=False,
+        inplace=False,
+        conjugate=False,
+) -> torch.Tensor:
+    """
+    Arguments:
+        x: (batch, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim).
+        cos: (seqlen_ro, rotary_dim / 2)
+        sin: (seqlen_ro, rotary_dim / 2)
+        seqlen_offsets: integer or integer tensor of size (batch,)
+        cu_seqlens: (batch + 1,) or None
+        max_seqlen: int
+    Returns:
+        y: (batch, seqlen, nheads, headdim)
+    """
+
+    batch, nheads, seqlen, headdim = x.shape
+
+    batch_ro, seqlen_ro, rotary_dim = cos.shape
+
+    assert batch == batch_ro
+    assert sin.shape == cos.shape
+    rotary_dim *= 2
+    assert rotary_dim <= headdim, "rotary_dim must be <= headdim"
+    assert headdim <= 256, "Only support headdim <= 256"
+
+    assert seqlen_ro >= seqlen, "seqlen_ro must be >= seqlen"
+
+    assert (
+            cos.dtype == sin.dtype
+    ), f"cos and sin must have the same dtype, got {cos.dtype} and {sin.dtype}"
+    assert (
+            x.dtype == cos.dtype
+    ), f"Input and cos/sin must have the same dtype, got {x.dtype} and {cos.dtype}"
+
+    cos, sin = cos.contiguous(), sin.contiguous()
+    if isinstance(seqlen_offsets, torch.Tensor):
+        assert seqlen_offsets.shape == (batch,)
+        assert seqlen_offsets.dtype in [torch.int32, torch.int64]
+        seqlen_offsets = seqlen_offsets.contiguous()
+    else:
+        assert seqlen_offsets + seqlen <= seqlen_ro
+
+    output = torch.empty_like(x) if not inplace else x
+    if rotary_dim < headdim and not inplace:
+        output[..., rotary_dim:].copy_(x[..., rotary_dim:])
+
+    BLOCK_K = (
+        32
+        if rotary_dim <= 32
+        else (64 if rotary_dim <= 64 else (128 if rotary_dim <= 128 else 256))
+    )
+    grid = lambda META: (triton.cdiv(seqlen, META["BLOCK_M"]), batch, nheads)  # noqa
+    BLOCK_M = 4 if interleaved else (8 if rotary_dim <= 64 else 4)
+
+    # Need this, otherwise Triton tries to launch from cuda:0 and we get
+    # ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)
+    with torch.cuda.device(x.device.index):
+        rotary_kernel[grid](
+            output,  # data ptrs
+            x,
+            cos,
+            sin,
+            cu_seqlens,
+            seqlen_offsets,
+            seqlen,  # shapes
+            nheads,
+            rotary_dim,
+            seqlen_ro,
+            seqlen // 128,  # key for triton cache (limit number of compilations)
+            output.stride(0),  # batch_strides
+            output.stride(-3),  # nheads_stride
+            output.stride(-2),  # seqlen_stride
+            output.stride(-1),  # headdim_stride
+            x.stride(0),  # batch_strides
+            x.stride(-3),  # nheads stride
+            x.stride(-2),  # seqlen stride
+            x.stride(-1),  # headdim stride
+            BLOCK_K,
+            isinstance(seqlen_offsets, torch.Tensor),
+            False,
+            interleaved,
+            conjugate,
+            BLOCK_M,
+        )
+    return output
+
+
+class ApplyRotaryEmb(torch.autograd.Function):
+    @staticmethod
+    def forward(
+            ctx,
+            x,
+            cos,
+            sin,
+            interleaved=False,
+            inplace=False,
+            seqlen_offsets: Union[int, torch.Tensor] = 0,
+            cu_seqlens: Optional[torch.Tensor] = None,
+            max_seqlen: Optional[int] = None,
+    ):
+        out = apply_rotary(
+            x,
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            interleaved=interleaved,
+            inplace=inplace,
+        )
+        if isinstance(seqlen_offsets, int):
+            ctx.save_for_backward(cos, sin, cu_seqlens)  # Can't save int with save_for_backward
+            ctx.seqlen_offsets = seqlen_offsets
+        else:
+            ctx.save_for_backward(cos, sin, cu_seqlens, seqlen_offsets)
+            ctx.seqlen_offsets = None
+        ctx.interleaved = interleaved
+        ctx.inplace = inplace
+        ctx.max_seqlen = max_seqlen
+        return out if not inplace else x
+
+    @staticmethod
+    def backward(ctx, do):
+        seqlen_offsets = ctx.seqlen_offsets
+        if seqlen_offsets is None:
+            cos, sin, cu_seqlens, seqlen_offsets = ctx.saved_tensors
+        else:
+            cos, sin, cu_seqlens = ctx.saved_tensors
+        # TD [2023-09-02]: For some reason Triton (2.0.0.post1) errors with
+        # "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
+        if not ctx.interleaved and not ctx.inplace:
+            do = do.clone()
+        dx = apply_rotary(
+            do,
+            cos,
+            sin,
+            seqlen_offsets=seqlen_offsets,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=ctx.max_seqlen,
+            interleaved=ctx.interleaved,
+            inplace=ctx.inplace,
+            conjugate=True,
+        )
+        return dx, None, None, None, None, None, None, None
+
+
+def apply_rotary_emb(
+        x,
+        cos,
+        sin,
+        interleaved=False,
+        inplace=False,
+        seqlen_offsets: Union[int, torch.Tensor] = 0,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[int] = None,
+):
+    """
+    Arguments:
+        x: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim)
+        cos, sin: (seqlen_rotary, rotary_dim / 2)
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
+            of 1st half and 2nd half (GPT-NeoX style).
+        inplace: if True, apply rotary embedding in-place.
+        seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
+            Most commonly used in inference when we have KV cache.
+        cu_seqlens: (batch + 1,) or None
+        max_seqlen: int
+    Return:
+        out: (batch_size, seqlen, nheads, headdim) if cu_seqlens is None
+            else (total_seqlen, nheads, headdim)
+    rotary_dim must be <= headdim
+    Apply rotary embedding to the first rotary_dim of x.
+    """
+    return ApplyRotaryEmb.apply(
+        x, cos, sin, interleaved, inplace, seqlen_offsets, cu_seqlens, max_seqlen
+    )
+
+
+# For backward compatibility
+apply_rotary_emb_func = apply_rotary_emb
+
+
+class FastRotaryEmbedding(torch.nn.Module):
+    """
+    The rotary position embeddings from RoFormer_ (Su et. al).
+    A crucial insight from the method is that the query and keys are
+    transformed by rotation matrices which depend on the relative positions.
+
+    Other implementations are available in the Rotary Transformer repo_ and in
+    GPT-NeoX_, GPT-NeoX was an inspiration
+
+    .. _RoFormer: https://arxiv.org/abs/2104.09864
+    .. _repo: https://github.com/ZhuiyiTechnology/roformer
+    .. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
+
+    If scale_base is not None, this implements XPos (Sun et al., https://arxiv.org/abs/2212.10554).
+    A recommended value for scale_base is 512: https://github.com/HazyResearch/flash-attention/issues/96
+    Reference: https://github.com/sunyt32/torchscale/blob/main/torchscale/component/xpos_relative_position.py
+    """
+
+    def __init__(
+            self,
+            dim: int,
+            base=10000,
+            interleaved=False,
+            scale_base=None,
+            pos_idx_in_fp32=True,
+            device=None,
+    ):
+        """
+        interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
+            of 1st half and 2nd half (GPT-NeoX style).
+        pos_idx_in_fp32: if True, the position indices [0.0, ..., seqlen - 1] are in fp32,
+            otherwise they might be in lower precision.
+            This option was added because previously (before 2023-07-02), when we construct
+            the position indices, we use the dtype of self.inv_freq. In most cases this would
+            be fp32, but if the model is trained in pure bf16 (not mixed precision), then
+            self.inv_freq would be bf16, and the position indices are also in bf16.
+            Because of the limited precision of bf16 (e.g. 1995.0 is rounded to 2000.0), the
+            embeddings for some positions will coincide.
+            To maintain compatibility with models previously trained in pure bf16,
+            we add this option.
+        """
+        super().__init__()
+        self.dim = dim
+        self.base = base
+        self.pos_idx_in_fp32 = pos_idx_in_fp32
+        # Generate and save the inverse frequency buffer (non trainable)
+        inv_freq = self._compute_inv_freq(device)
+        self.register_buffer("inv_freq", inv_freq)
+        self.interleaved = interleaved
+        self.scale_base = scale_base
+        scale = (
+            (torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
+            if scale_base is not None
+            else None
+        )
+        self.register_buffer("scale", scale, persistent=False)
+
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+        self._cos_k_cached = None
+        self._sin_k_cached = None
+        self.cos = None
+        self.sin = None
+
+    def _compute_inv_freq(self, device=None):
+        return 1.0 / (
+                self.base
+                ** (torch.arange(0, self.dim, 2, device=device) / self.dim)
+                # ** (torch.arange(0, self.dim, 2, device=device).float() / self.dim)
+        )
+
+    def _update_cos_sin_cache(self, seqlen, position_id, device=None, dtype=None):
+
+        if (
+                seqlen > self._seq_len_cached
+        ):
+            self._seq_len_cached = seqlen
+            # We want fp32 here, not self.inv_freq.dtype, since the model could be loaded in bf16
+            # And the output of arange can be quite large, so bf16 would lose a lot of precision.
+            # However, for compatibility reason, we add an option to use the dtype of self.inv_freq.
+            if self.pos_idx_in_fp32:
+                t = torch.arange(seqlen, device=device, dtype=torch.float32)
+                # We want fp32 here as well since inv_freq will be multiplied with t, and the output
+                # will be large. Having it in bf16 will lose a lot of precision and cause the
+                # cos & sin output to change significantly.
+                # We want to recompute self.inv_freq if it was not loaded in fp32
+                if self.inv_freq.dtype != torch.float32:
+                    inv_freq = self._compute_inv_freq(device=device)
+                else:
+                    inv_freq = self.inv_freq
+            else:
+                t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
+                inv_freq = self.inv_freq
+            freqs = torch.einsum("i,j->ij", t, inv_freq)
+            if self.scale is None:
+                self._cos_cached = torch.cos(freqs).to(dtype)
+                self._sin_cached = torch.sin(freqs).to(dtype)
+
+            else:
+                power = (
+                                torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device)
+                                - seqlen // 2
+                        ) / self.scale_base
+                scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
+                # We want the multiplication by scale to happen in fp32
+                self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
+                self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
+                self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
+                self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
+
+    def forward(
+            self,
+            q: torch.Tensor,
+            k: torch.Tensor,
+            position_ids: torch.Tensor,
+            max_seqlen,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        q: (batch, nheads, seqlen, headdim)
+        k: (batch, nheads, seqlen, headdim)
+        position_id: (batch, seqlen)
+        max_seqlen: int
+        layer_id: int
+            only if layer_id == 0, then update cons and sin
+        Apply rotary embedding *inplace* to q k.
+        """
+
+        self._update_cos_sin_cache(max_seqlen, position_ids, device=q.device, dtype=q.dtype)
+        cos, sin = F.embedding(position_ids, self._cos_cached), F.embedding(position_ids, self._sin_cached)
+
+        q = apply_rotary_emb_func(
+            q,
+            cos,
+            sin,
+            interleaved=self.interleaved,
+            inplace=True
+        )
+        k = apply_rotary_emb_func(
+            k,
+            cos,
+            sin,
+            interleaved=self.interleaved,
+            inplace=True
+        )
+        return q, k

visual.py

@@ -0,0 +1,135 @@
+import torch
+from torch import nn
+from argparse import Namespace
+import xformers.ops as xops
+from transformers.activations import ACT2FN
+
+
+class PatchEmbedding(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.proj = nn.Conv2d(config.in_channels, config.hidden_size, kernel_size=config.patch_size, stride=config.patch_size)
+        self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size))
+        self.position_embedding = nn.Embedding(config.num_positions, config.hidden_size)
+
+    def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
+        x = self.proj(images)
+        x = x.flatten(2).transpose(1, 2)
+        cls_token = self.cls_embedding.expand(x.shape[0], -1, -1)
+        x = torch.cat((cls_token, x), dim=1)
+        x += self.position_embedding.weight.unsqueeze(0)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.num_heads = config.num_heads
+        head_dim = config.hidden_size // config.num_heads
+        self.scale = head_dim ** -0.5
+        self.query_key_value = nn.Linear(config.hidden_size, config.hidden_size * 3)
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.output_dropout = torch.nn.Dropout(config.dropout_prob)
+
+    def forward(self, x: "tensor(B, L, D)") -> "tensor(B, L, D)":
+        B, L, _ = x.shape
+        qkv = self.query_key_value(x)
+        qkv = qkv.reshape(B, L, 3, self.num_heads, -1).permute(2, 0, 1, 3, 4)  # 3, B, L, H, D
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        out = xops.memory_efficient_attention(
+            q, k, v, scale=self.scale,
+        )
+        output = self.dense(out.view(B, L, -1))
+        output = self.output_dropout(output)
+        return output
+
+    def attention(self, q, k, v):
+        attn_weights = torch.matmul(q * self.scale, k.transpose(-2, -1))
+        attn_weights = attn_weights.softmax(dim=-1)
+        output = torch.matmul(attn_weights, v)
+        return output
+
+
+class MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.fc1(x)
+        x = self.activation_fn(x)
+        x = self.fc2(x)
+        return x
+
+
+class TransformerLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attention = Attention(config)
+        self.mlp = MLP(config)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        attention_input = hidden_states
+        attention_output = self.input_layernorm(self.attention(attention_input))
+        hidden_states = attention_input + attention_output
+        mlp_input = hidden_states
+        mlp_output = self.post_attention_layernorm(self.mlp(mlp_input))
+        output = mlp_input + mlp_output
+        return output
+
+
+class Transformer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.layers = nn.ModuleList([TransformerLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(self, hidden_states):
+        for layer_module in self.layers:
+            hidden_states = layer_module(hidden_states)
+        return hidden_states
+
+
+class GLU(nn.Module):
+    def __init__(self, config, in_features):
+        super().__init__()
+        self.linear_proj = nn.Linear(in_features, config.hidden_size, bias=False)
+        self.norm1 = nn.LayerNorm(config.hidden_size)
+        self.act1 = nn.GELU()
+        self.act2 = nn.functional.silu
+        self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
+        self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+
+    def forward(self, x):
+        x = self.linear_proj(x)
+        x = self.act1(self.norm1(x))
+        x = self.act2(self.gate_proj(x)) * self.dense_h_to_4h(x)
+        x = self.dense_4h_to_h(x)
+        return x
+
+
+class EVA2CLIPModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        vision_config = Namespace(**config.vision_config)
+        self.patch_embedding = PatchEmbedding(vision_config)
+        self.transformer = Transformer(vision_config)
+        self.linear_proj = GLU(config, in_features=vision_config.hidden_size)
+        self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+        self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
+
+    def forward(self, images: "tensor(B, C, H, W)") -> "tensor(B, L, D)":
+        x = self.patch_embedding(images)
+        x = self.transformer(x)
+        x = x[:, 1:]
+        x = self.linear_proj(x)
+        boi = self.boi.expand(x.shape[0], -1, -1)
+        eoi = self.eoi.expand(x.shape[0], -1, -1)
+        x = torch.cat((boi, x, eoi), dim=1)
+        return x