################################################################################ # Copyright(c)2020-2025 Shanghai Biren Technology Co., Ltd. All rights reserved. # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ################################################################################ # SPDX-License-Identifier: Apache-2.0 # adapted from https://huggingface.co/OpenGVLab/InternVL2-4B/blob/main/modeling_intern_vit.py # -------------------------------------------------------- # InternVL # Copyright (c) 2023 OpenGVLab # Licensed under The MIT License [see LICENSE for details] # -------------------------------------------------------- from typing import Optional import torch import torch_br from fastcore.basics import patch_to from transformers import PretrainedConfig from vllm.model_executor.layers.quantization import QuantizationConfig # isort: off from vllm.model_executor.models.intern_vit import (InternMLP, InternVisionEmbeddings, InternVisionModel, InternVisionEncoder) from vllm.model_executor.models.intern_vit import InternParallelAttention from vllm.distributed.parallel_state import get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size from vllm.distributed.utils import divide from vllm.model_executor.layers.layernorm import RMSNorm # isort: on @patch_to(InternVisionModel) def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, *, num_hidden_layers_override: Optional[int] = None, num_dummy_heads: int = 0, prefix: str = "", use_data_parallel: bool = False, ) -> None: """ [Patch] enable data parallelism for InternVisionModel """ super(InternVisionModel, self).__init__() self.config = config self.use_data_parallel = use_data_parallel self.embeddings = InternVisionEmbeddings(config) self.encoder = InternVisionEncoder( config=config, quant_config=None, num_hidden_layers_override=num_hidden_layers_override, num_dummy_heads=num_dummy_heads, prefix=f"{prefix}.encoder", use_data_parallel=use_data_parallel, ) @patch_to(InternVisionEmbeddings) def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor: target_dtype = self.patch_embedding.weight.dtype if self.patch_size == 14: import torch_br.supa._debug as supa_debug supa_debug.set_disable_zero_ws(False) supa_debug.set_disable_zero_output_uma(False) supa_debug.set_disable_zero_output_numa(False) supa_debug.set_disable_reorder_zero(False) patch_embeds = torch_br.supa_conv2d_knxn_snxn_p0x0_fwd( pixel_values.to(dtype=target_dtype), self.patch_embedding.weight, self.patch_size, self.patch_size, 0) if self.patch_embedding.bias is not None: patch_embeds += self.patch_embedding.bias[None, :, None, None] supa_debug.set_disable_zero_ws(True) supa_debug.set_disable_zero_output_uma(True) supa_debug.set_disable_zero_output_numa(True) supa_debug.set_disable_reorder_zero(True) else: patch_embeds = self.patch_embedding(pixel_values.to( target_dtype)) # shape = [*, channel, width, height] batch_size, _, height, width = patch_embeds.shape patch_embeds = patch_embeds.flatten(2).transpose(1, 2) class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype) embeddings = torch.cat([class_embeds, patch_embeds], dim=1) if self.patch_embedding.bias is None: position_embedding = self._get_position_embedding(height, width) else: position_embedding = torch.cat([ self.position_embedding[:, :1, :], self._get_pos_embed(self.position_embedding[:, 1:, :], height, width) ], dim=1) embeddings = embeddings + position_embedding.to(target_dtype) return embeddings @patch_to(InternParallelAttention) def __init__( self, config: PretrainedConfig, quant_config: Optional[QuantizationConfig] = None, *, num_dummy_heads: int = 0, prefix: str = "", use_data_parallel: bool = False, ) -> None: super(InternParallelAttention, self).__init__() # [Patch] enable data parallelism self.use_data_parallel = True self.config = config self.embed_dim = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.embed_dim // self.num_heads if self.head_dim * self.num_heads != self.embed_dim: raise ValueError(f'embed_dim must be divisible by num_heads ' f'(got `embed_dim`: {self.embed_dim} and `num_heads`:' f' {self.num_heads}).') self.tp_size = (1 if use_data_parallel else get_tensor_model_parallel_world_size()) self.tp_rank = (0 if use_data_parallel else get_tensor_model_parallel_rank()) # Additional dummy heads are used to enable TP for common GPU counts. self.dummy_dim = (num_dummy_heads + self.num_heads) * self.head_dim self.num_heads_per_partition = divide(num_dummy_heads + self.num_heads, self.tp_size) assert self.tp_size == 1 self.scale = self.head_dim**-0.5 # self.qkv = QKVParallelLinear( # self.embed_dim, # self.head_dim, # num_dummy_heads + self.num_heads, # bias=config.qkv_bias, # quant_config=quant_config, # prefix=f"{prefix}.qkv", # disable_tp=use_data_parallel, # ) self.qkv = torch.nn.Linear(self.embed_dim, 3 * self.dummy_dim, bias=config.qkv_bias) self.qk_normalization = config.qk_normalization if self.qk_normalization: self.q_norm = RMSNorm(self.dummy_dim, eps=config.layer_norm_eps, var_hidden_size=self.embed_dim) self.k_norm = RMSNorm(self.dummy_dim, eps=config.layer_norm_eps, var_hidden_size=self.embed_dim) # self.proj = RowParallelLinear( # self.dummy_dim, # self.embed_dim, # quant_config=quant_config, # prefix=f"{prefix}.proj", # disable_tp=use_data_parallel, # ) self.proj = torch.nn.Linear(self.dummy_dim, self.embed_dim) # self.attn = MultiHeadAttention(self.num_heads_per_partition, # self.head_dim, self.scale) @patch_to(InternParallelAttention) def forward(self, x: torch.Tensor) -> torch.Tensor: B, N, C = x.shape x_tmp = [] for i in range(B): qkv = self.qkv(x[i:i + 1, :]).reshape(1, N, 3, self.num_heads, C // self.num_heads) q, k, v = qkv.unbind( 2) # make torchscript happy (cannot use tensor as tuple) if self.qk_normalization: q = self.q_norm(q.flatten(-2, -1)).view(1, N, self.num_heads, qkv.shape[4]) k = self.k_norm(k.flatten(-2, -1)).view(1, N, self.num_heads, qkv.shape[4]) q = q.permute(0, 2, 1, 3) k = k.permute(0, 2, 1, 3) v = v.permute(0, 2, 1, 3) attn = ((q * self.scale) @ k.transpose(-2, -1)) attn = attn.softmax(dim=-1) # x = (attn @ v).transpose(1, 2).reshape(B, N, C) x0 = attn[:, :, :, :512] @ v[:, :, :512, :] x1 = attn[:, :, :, 512:] @ v[:, :, 512:, :] x_tmp.append((x0 + x1).transpose(1, 2).reshape(1, N, C)) x = torch.cat(x_tmp, dim=0) x = self.proj(x) return x @patch_to(InternMLP) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: if hidden_states.shape[0] > 1: output = torch_br._empty_ut_only(hidden_states.shape, "COLMAJOR", is_numa=False, sbp="BB", axis=0, dtype=torch.bfloat16) for i in range(hidden_states.shape[0]): hidden_states_tmp, _ = self.fc1(hidden_states[i:i + 1, :, :]) hidden_states_tmp = self.activation_fn(hidden_states_tmp) hidden_states_tmp, _ = self.fc2(hidden_states_tmp) hidden_states_tmp += self.fc2.bias[None, None, :] output[i] = hidden_states_tmp[0] return output else: hidden_states, _ = self.fc1(hidden_states) hidden_states = self.activation_fn(hidden_states) hidden_states, _ = self.fc2(hidden_states) return hidden_states