This document details the implementation of multihead attention (MHA),
multiquery attention (MQA) and group-query attention (GQA) for auto-regressive
GPT-like models in TensorRT-LLM. As a quick reminder, the multihead attention
is the sequence of a batched matmul, a softmax and another batched matmul
described in the
Attention Is All You Need article.
Multi-query Attention (MQA) [https://arxiv.org/abs/1911.02150]
Group-query Attention (GQA) [https://arxiv.org/abs/2307.09288]
are variants of MHA that use fewer, so-called, K/V head than the number of
query heads. TensorRT-LLM, MHA, MQA and GQA are implemented by the operator
tensorrt_llm.functional.gpt_attention.
As discussed below, the current implementation supports two input modes: Padded and packed (non-padded). As the packed mode is always more memory-efficient and faster than the padded mode, support for padded mode may be removed in the future.
In TensorRT-LLM, the GPT attention operator supports two different types
of QKV inputs: Padded and packed (i.e. non padded) inputs. The mode is
determined by the global configuration parameter remove_input_padding defined
in tensorrt_llm.plugin.
When padding is enabled (i.e. remove_input_padding is False), the sequences
that are shorter than the max_sequence_length are padded to that maximum
length. It may result in excessive memory consumption as well as unneeded
computations on padding tokens (in the various matrix multiplications that
surround the MHA block).
To overcome that problem, TensorRT-LLM supports a mode without padding where the different tokens are packed together and the user provides the operator with a 1D tensor containing the lengths of the different sequences. It is recommended that users to always use packed mode (and support for the padded mode may be removed in the future).
The GPT attention operator encapsulates different implementations for both context and generation phases in auto-regressive models like GPT.
If the context_fmha_type is set to disabled (see
tensorrt_llm.plugin),
the implementation maps to a sequence of GPU kernels that will store the
intermediate Q*K^T tensor in memory before calling the softmax operator. It
is the slowest method and the memory footprint is significant (quadratically
depends on the sequence length).
Otherwise, if context_fmha_type is set to a enabled or
enabled_with_fp32_acc (accumulation in the first batched matmul is forced to
FP32), that function will trigger a kernel that performs the MHA/MQA block
using a single kernel. For short sequences, that kernel uses a vanilla
implementation of MHA/MQA. For larger sequences, this kernel uses the Flash
Attention algorithm as described in
https://arxiv.org/abs/2205.14135
and
https://arxiv.org/abs/2307.08691.
Currently, the implementation triggers extra kernels that apply pre-processing to the elements (like RoPE) and populate the KV cache (see below). In a future release, the number of such kernels is planned on being reduced in order to improve the overall performance.
The generation phase is implemented using a single kernel, called the masked multihead attention in TensorRT-LLM. That kernel is able to apply pre-processing on the Q, K and V elements on-the-fly: Add the QKV bias, apply RoPE, do dequantization/quantization. TensorRT-LLM will continue to add (or enable) additional features in future releases. For example, enable the support for ALiBi or IA3.
The masked MHA kernel has a special version that distributes the work across
multiple CUDA thread-blocks on the GPU for cases where the GPU occupancy is
low. That mode called multi-block can be enabled using the multi_block_mode
flag. Users are recommended to test that mode in scenarios where both the batch
size and the number of heads in the model are relatively small. The exact
definition of small in that context will depend on the model of the GPU and is
hard to predict but to provide with a rule of thumb, it is worth testing that
mode when batch_size * num_heads is less than the number of multi-processors
on the GPU (that suggestion may evolve in the future as more research is
conducted and the software improves).
Note that even if the multi-block mode is enabled, the attention operator will not immediately trigger the multi-block version of the GPU kernel. There is a minimum number of tokens (input + generated) that are required for the multi-block version to become more efficient than the "vanilla" implementation that uses a single CUDA thread-block per head. It is controlled by an internal heuristic.
Another note is that as the masked MHA kernels use shared memory size proportional to sequence length, so there can be some cases that GPU's shared memory is not enough when multi-block mode is not enabled. To get masked MHA kernel work in these cases, multi-block mode is forced on and a warning log is printed.
TensorRT-LLM supports a feature called in-flight batching. With that feature, sequences in context phase can be processed together with sequences in generation phase. The purpose of that technique is to better interleave requests to reduce latency as well as make better use the of the GPUs. For efficiency reasons (1), the support for inflight batching requires the input tensors to be packed (no padding).
In the current implementation, the sequences that are going through the
context phase must be before the sequences in the generation phase in the input
tensor. For example, for sequences S0, S1 and S2, if S0 and S2 are in
context phase (and S1 in generation), tokens from S0 and S2 must appear
before the tokens of S1 in the input tensor. The constraint may or may not
be relaxed in a future version.
(1) Padding sequences in the generation phase, that contain a single token, to the length of the maximum input sequence is inefficient use of resources.
In the generation phase, a common optimization is to provide the MHA kernel with a cache containing the values of the past K and V elements that have already been computed. That cache is known as the KV cache. TensorRT-LLM uses that technique to accelerate its generation phase. In TensorRT-LLM, there is one KV cache per Transformer layer, which means that there are as many KV caches as layers in a model. The current version of TensorRT-LLM supports two different types of KV caches: contiguous and paged KV caches.
The contiguous KV cache is a monolithic tensor. Its shape is:
[max_batch_size * max_beam_width, 2, num_heads, max_seqlen, hidden_dim_per_head].
That implementation uses a lot more memory than needed when the sequences are shorter than the maximum sequence length (even if they end up close to the limit after the generation of many output tokens, it may take a lot of steps to reach that point).
The paged KV cache decomposes the KV cache into blocks that are distributed to
the different requests by a cache manager during processing. That cache manager
keeps track of the sequences, allocate new blocks from a pool and recycle those
blocks when required. See the simplified implementation of
tensorrt_llm.runtime.KVCacheManager.
A more efficient C++ implementation is included in the
Batch Manager.
In its current implementation, even if the rest of the network runs in INT8 or
FP8, the GPT attention operator works with FP32, FP16, and BFloat16 inputs and
outputs. However, TensorRT-LLM supports INT8 and FP8
(kv_cache_quant_mode=QuantMode.INT8_KV_CACHE and
kv_cache_quant_mode=QuantMode.FP8_KV_CACHE) KV caches.
The GPT attention operator populates the KV cache. When INT8 or FP8 KV caches
are enabled, the input values have to be quantized to 8 bits using a scaling
factor. For quantization, the scaling factor is stored in the
kv_orig_quant_scale tensor. Its shape is [1] and only per-tensor
quantization is supported in the current version.
During generation, the values read from the cache are dequantized on-the-fly in
the MHA/MQA kernel. The scaling factor to dequantize those values is stored in
the kv_quant_orig_scale tensor. That tensor contains a single value (per
tensor scaling).
TensorRT-LLM has a feature called Cyclic KV Cache, which treats the kv cache
as a circular buffer. This means that it only stores the kv cache for the last N
tokens, where N is determined by the max_attention_window_size parameter in
GenerationSession.setup. You can see examples of this in the run.py or
summarize.py files. When the cache is full, new tokens’ kv cache will
overwrite the "least recently used" caches.
In the context phase, if the input length surpasses the max_attention_window_size,
Sliding Window Attention will be activated. This serves the same function as
the sliding window_size.
This feature helps to reduce the memory footprint of the kv cache when dealing with very long sequences.
_Note that the cyclic kv cache feature doesn't work with beam searching currently as the context kv cache are shared across beams.
The experimental feature, which allows different max_attention_window_size values
for each layer, is also supported. To utilize this feature, simply provide an
int32 torch.Tensor with a shape of [num_layers] to the GenerationSession.setup.
This tensor will serve as the buffer for max_attention_window_size,
setting unique values for each layer. However, it’s important to note that the
memory allocation for the kv cache still relies on the buffer’s maximum value.
The GPT attention operator supports beam-search. In the context phase, a single
beam is computed per input sequence. In the generation phase, the MHA/MQA/GQA
kernel uses an additional tensor to reconstruct the correct path for each beam.
That tensor is called the cache_indirection. Its shape is [batch_size, beam_width, max_seqlen].
For a sequence si, a beam bi and a token ti, the element
cache_indirection[si][bi][ti] is an integer between 0 and beam_width-1
that indicates which path in the beam to read the K and V elements from in the
KV cache. This tensor is populated in the sampling stage.
The input QKV tensor packs the Q, K and V tensors (concatenated along the last dimension) after the projection of the hidden states. It is a 3D tensor. RoPE and quantization to INT8 or FP8 (when needed) are performed by the GPT attention operator.
In padded mode, its shape is [batch_beam_size, max_seqlen, 3 * hidden_dim]
where batch_beam_size is the batch size (number of sequences) for the context
phase and the batch size multiplied by the beam width for the generation phase.
Having different beam widths per sequence in padded mode is not supported.
In packed mode, its shape is [num_tokens, 3 * hidden_dim] where
num_tokens is the total number of tokens in the batch. For the sequences in
context phase, the number of tokens of a sequence corresponds to its input
length (even if the beam width is greater than 1 for beam search). For the
sequences in generation phase, there are beam_width tokens per sequence. The
beam width can be different for each sequence.
In other words, the pseudo-code to compute the number of tokens is:
num_tokens = 0
# Add the length of each sequence in context phase.
for seq in context_phase:
num_tokens += seq.length
# Add the width of the beam for each sequence in generation phase.
for seq in generation_phase:
num_tokens += seq.beam_widthThe GPT attention operation can perform the computation of the Rotary
Positional Embedding (RoPE). When that operation is enabled,
rotary_embedding_dim is set to a value greater than 0, it is fused with other
operations. The GPT operator supports GPT-NeoX and GPT-J forms of RoPE by
setting position_embedding_type to PositionEmbeddingType.rope_gpt_neox
or PositionEmbeddingType.rope_gptj.
The GPT attention operator can apply ALiBi to the result of the Q*K^T
product. The bias is computed on-the-fly from the ALiBi slopes in the optimized
kernel.
In MHA, the output of the Q*K^T product is scaled by a constant value that
is computed as:
norm_factor = 1.f / (q_scaling * sqrt(head_size)).
On top of the MHA as self attention needed by GPT-style decoder-only models, gpt_attention also supports cross attention.
This enables using gpt_attention in a broader aspect as a generic decoder component. For example, the Encoder-Decoder model uses gpt_attention to issue both the self attention and cross attention modules in its Decoder.
Relative attention bias (RAB) is a kind of relative position modeling, adding an attention bias (Q*K^T+bias) according to relative positions. RAB is a lightweight method to include the information of relative positions, and is used in the popular Encoder-Decoder model T5 and also other models in the T5 family.
RAB is supported in two modes: i) regular mode which user passes in relative attention bias computed ahead of MHA. ii) implicit mode which computes the relative attention bias on the fly in MHA. The implicit mode suits the case when the relative attention bias is too large to fit in memory and can be turned on by passing in max_distance.