Source code for tensorrt_llm.layers.embedding

# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# 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.
import math
from typing import Optional

import numpy as np
import torch

from .._utils import set_obj_attrs, str_dtype_to_torch, trt_dtype_to_np
from ..functional import concat, constant, embedding, unsqueeze, where
from ..mapping import Mapping
from ..module import Module
from ..parameter import Parameter


[docs] class Embedding(Module): """ The embedding layer takes input indices (x) and the embedding lookup table (weight) as input. And output the corresponding embeddings according to input indices. The size of weight is [num_embeddings, embedding_dim] Four parameters (tp_size, tp_group, sharding_dim, tp_rank) are involved in tensor parallelism. Only when "tp_size > 1 and tp_group is not None", tensor parallelism is enabled. When "sharding_dim == 0", the weight is shared in the vocabulary dimension. tp_rank must be set when sharding_dim == 0. When "sharding_dim == 1", the weight is shard in the hidden dimension. """ def __init__(self, num_embeddings: int, embedding_dim: int, dtype: Optional[str] = None, tp_size: int = 1, tp_group: Optional[list] = None, sharding_dim: int = 0, tp_rank: Optional[int] = None): super().__init__() # num_embeddings records the total vocab size no matter using TP or not self.num_embeddings = num_embeddings self.embedding_dim = embedding_dim self.tp_size = tp_size self.tp_group = tp_group self.sharding_dim = sharding_dim self.tp_rank = tp_rank self.dtype = dtype self.tp_dim = sharding_dim if sharding_dim == 1: shape = (self.num_embeddings, self.embedding_dim // self.tp_size) elif sharding_dim == 0: shape = (math.ceil(self.num_embeddings / self.tp_size), self.embedding_dim) self.weight = Parameter(shape=shape, dtype=dtype) self.weight_padding_size = ((8 - shape[0] % 8) % 8, shape[1]) set_obj_attrs(self.weight, { "weight_loader": self.weight_loader, })
[docs] def forward(self, x): # The embedding weight is padded to the multiple of 8. # The reason is that when lm_head and vocab_embedding are using the same embedding weight, # previously weights can't be depulicated in the engine because gemm will pad the weight to the multiple of 8. # If we also pad the embedding weight to the multiple of 8, the weights can be successfully deduplicated. # This will not affect the input and output of the gather op and perf impact is negligible. if self.weight_padding_size[0] != 0: padding_values = np.zeros(self.weight_padding_size, dtype=trt_dtype_to_np( self.weight.value.dtype)) padding = constant(padding_values) padded_weight = concat([self.weight.value, padding], dim=0) else: padded_weight = self.weight.value return embedding(x, padded_weight, tp_size=self.tp_size, tp_group=self.tp_group, sharding_dim=self.sharding_dim, tp_rank=self.tp_rank)
[docs] def weight_loader(self, mapping: Mapping, param: Parameter, loaded_weight: torch.Tensor): # use_parallel_embedding tp_rank = mapping.tp_rank if self.tp_size > 1: sharding_dim = self.sharding_dim shard_size = param._shape[sharding_dim] start_idx = tp_rank * shard_size loaded_weight = loaded_weight.narrow(sharding_dim, start_idx, shard_size) param.value = loaded_weight
[docs] def postprocess(self, tllm_key, weights, **kwargs): if weights is None: return {} weights = weights.to(str_dtype_to_torch(self.dtype)) return {tllm_key: weights}
[docs] class PromptTuningEmbedding(Embedding): """ PromptTuningEmbedding handles fine-tuned prompts with virtual tokens. At runtime, a supplementary embedding dictionary is passed. Tokens whose ids are >= vocab_size are embedded with that additional dictionary. The prompt tuning dictionary holds multiple tasks, and each sequence is assigned a given task. Prompt-tuned tokens from a given sequence use the adequate task dictionary, as defined by the `tasks` input. """ def __init__(self, num_embeddings, embedding_dim, vocab_size=None, dtype=None, tp_size=1, tp_group=None, sharding_dim=0, tp_rank=0): super().__init__(num_embeddings, embedding_dim, dtype, tp_size, tp_group, sharding_dim, tp_rank) if vocab_size is None: vocab_size = num_embeddings self.vocab_size = vocab_size
[docs] def forward(self, tokens, prompt_embedding_table, tasks, task_vocab_size): """ Pass all tokens through both normal and prompt embedding tables. Tokens are masked so that "normal" embedding only see "normal" tokens. Same logic for "prompt" embedding. After those two embedding, combine results based on whether the token was "normal" or "prompt-tuned". Parameters: tokens : Tensor the ids to embed, size [batch_size, seq_len] prompt_embedding_table : Tensor the additional embedding table for prompt-tuned tokens, size [num_tasks * num_tokens_per_task, hidden_size] tasks: Tensor the task required by each token, size [batch_size, seq_len] task_vocab_size: Tensor the number of tokens used for each task, should be equal to prompt_embedding_table's num_tokens_per_task, size [1] Returns: Tokens' embedding """ # do not use ">=" because internally the layer works with floating points prompt_tokens_mask = tokens > (self.vocab_size - 1) # clip tokens in the [0, vocab_size) range normal_tokens = where(prompt_tokens_mask, self.vocab_size - 1, tokens) normal_embeddings = embedding(normal_tokens, self.weight.value, self.tp_size, self.tp_group, self.sharding_dim, self.tp_rank) # put virtual tokens in the [0, max_prompt_vocab_size) range prompt_tokens = where(prompt_tokens_mask, tokens - self.vocab_size, 0) # add offsets to match the concatenated embedding tables tasks = tasks * task_vocab_size # tasks: [batch_size, seq_len] # prompt_tokens: [batch_size, seq_len] prompt_tokens = prompt_tokens + tasks prompt_embeddings = embedding(prompt_tokens, prompt_embedding_table) # prompt_tokens_mask: [batch_size, seq_len] -> [batch_size, seq_len, 1] # combine the correct sources of embedding: normal/prompt return where(unsqueeze(prompt_tokens_mask, -1), prompt_embeddings, normal_embeddings)