SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Use HF papers (#38184)

Browse Source 
* Use hf papers

* Hugging Face papers

* doi to hf papers

* style
This commit is contained in:
Quentin Gallouédec
2025-06-13 13:07:09 +02:00
committed by GitHub
parent 1031ed5166
commit de24fb63ed
811 changed files with 2622 additions and 2617 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
docs/source/en/llm_tutorial_optimization.md
View File

@@ -16,7 +16,7 @@ rendered properly in your Markdown viewer.
Large Language Models (LLMs) such as GPT3/4, [Falcon](https://huggingface.co/tiiuae/falcon-40b), and [Llama](https://huggingface.co/meta-llama/Llama-2-70b-hf) are rapidly advancing in their ability to tackle human-centric tasks, establishing themselves as essential tools in modern knowledge-based industries.
Deploying these models in real-world tasks remains challenging, however:
- To exhibit near-human text understanding and generation capabilities, LLMs currently require to be composed of billions of parameters (see [Kaplan et al](https://arxiv.org/abs/2001.08361), [Wei et. al](https://arxiv.org/abs/2206.07682)). This consequently amplifies the memory demands for inference.
- To exhibit near-human text understanding and generation capabilities, LLMs currently require to be composed of billions of parameters (see [Kaplan et al](https://huggingface.co/papers/2001.08361), [Wei et. al](https://huggingface.co/papers/2206.07682)). This consequently amplifies the memory demands for inference.
- In many real-world tasks, LLMs need to be given extensive contextual information. This necessitates the model's capability to manage very long input sequences during inference.
The crux of these challenges lies in augmenting the computational and memory capabilities of LLMs, especially when handling expansive input sequences.
@@ -27,7 +27,7 @@ In this guide, we will go over the effective techniques for efficient LLM deploy
2. **Flash Attention:** Flash Attention is a variation of the attention algorithm that not only provides a more memory-efficient approach but also realizes increased efficiency due to optimized GPU memory utilization.
3. **Architectural Innovations:** Considering that LLMs are always deployed in the same way during inference, namely autoregressive text generation with a long input context, specialized model architectures have been proposed that allow for more efficient inference. The most important advancement in model architectures hereby are [Alibi](https://arxiv.org/abs/2108.12409), [Rotary embeddings](https://arxiv.org/abs/2104.09864), [Multi-Query Attention (MQA)](https://arxiv.org/abs/1911.02150) and [Grouped-Query-Attention (GQA)]((https://arxiv.org/abs/2305.13245)).
3. **Architectural Innovations:** Considering that LLMs are always deployed in the same way during inference, namely autoregressive text generation with a long input context, specialized model architectures have been proposed that allow for more efficient inference. The most important advancement in model architectures hereby are [Alibi](https://huggingface.co/papers/2108.12409), [Rotary embeddings](https://huggingface.co/papers/2104.09864), [Multi-Query Attention (MQA)](https://huggingface.co/papers/1911.02150) and [Grouped-Query-Attention (GQA)]((https://huggingface.co/papers/2305.13245)).
Throughout this guide, we will offer an analysis of auto-regressive generation from a tensor's perspective. We delve into the pros and cons of adopting lower precision, provide a comprehensive exploration of the latest attention algorithms, and discuss improved LLM architectures. While doing so, we run practical examples showcasing each of the feature improvements.
@@ -157,8 +157,8 @@ from accelerate.utils import release_memory
release_memory(model)
```
Now what if your GPU does not have 32 GB of VRAM? It has been found that model weights can be quantized to 8-bit or 4-bits without a significant loss in performance (see [Dettmers et al.](https://arxiv.org/abs/2208.07339)).
Model can be quantized to even 3 or 2 bits with an acceptable loss in performance as shown in the recent [GPTQ paper](https://arxiv.org/abs/2210.17323) 🤯.
Now what if your GPU does not have 32 GB of VRAM? It has been found that model weights can be quantized to 8-bit or 4-bits without a significant loss in performance (see [Dettmers et al.](https://huggingface.co/papers/2208.07339)).
Model can be quantized to even 3 or 2 bits with an acceptable loss in performance as shown in the recent [GPTQ paper](https://huggingface.co/papers/2210.17323) 🤯.
Without going into too many details, quantization schemes aim at reducing the precision of weights while trying to keep the model's inference results as accurate as possible (*a.k.a* as close as possible to bfloat16).
Note that quantization works especially well for text generation since all we care about is choosing the *set of most likely next tokens* and don't really care about the exact values of the next token *logit* distribution.
@@ -308,7 +308,7 @@ Long story short, the default self-attention algorithm quickly becomes prohibiti
As LLMs improve in text comprehension and generation, they are applied to increasingly complex tasks. While models once handled the translation or summarization of a few sentences, they now manage entire pages, demanding the capability to process extensive input lengths.
How can we get rid of the exorbitant memory requirements for large input lengths? We need a new way to compute the self-attention mechanism that gets rid of the \\( QK^T \\) matrix. [Tri Dao et al.](https://arxiv.org/abs/2205.14135) developed exactly such a new algorithm and called it **Flash Attention**.
How can we get rid of the exorbitant memory requirements for large input lengths? We need a new way to compute the self-attention mechanism that gets rid of the \\( QK^T \\) matrix. [Tri Dao et al.](https://huggingface.co/papers/2205.14135) developed exactly such a new algorithm and called it **Flash Attention**.
In a nutshell, Flash Attention breaks the \\(\mathbf{V} \times \text{Softmax}(\mathbf{QK}^T\\)) computation apart and instead computes smaller chunks of the output by iterating over multiple softmax computation steps:
@@ -316,13 +316,13 @@ $$ \textbf{O}_i \leftarrow s^a_{ij} * \textbf{O}_i + s^b_{ij} * \mathbf{V}_{j} \
with \\( s^a_{ij} \\) and \\( s^b_{ij} \\) being some softmax normalization statistics that need to be recomputed for every \\( i \\) and \\( j \\) .
Please note that the whole Flash Attention is a bit more complex and is greatly simplified here as going in too much depth is out of scope for this guide. The reader is invited to take a look at the well-written [Flash Attention paper](https://arxiv.org/abs/2205.14135) for more details.
Please note that the whole Flash Attention is a bit more complex and is greatly simplified here as going in too much depth is out of scope for this guide. The reader is invited to take a look at the well-written [Flash Attention paper](https://huggingface.co/papers/2205.14135) for more details.
The main takeaway here is:
> By keeping track of softmax normalization statistics and by using some smart mathematics, Flash Attention gives **numerical identical** outputs compared to the default self-attention layer at a memory cost that only increases linearly with \\( N \\) .
Looking at the formula, one would intuitively say that Flash Attention must be much slower compared to the default self-attention formula as more computation needs to be done. Indeed Flash Attention requires more FLOPs compared to normal attention as the softmax normalization statistics have to constantly be recomputed (see [paper](https://arxiv.org/abs/2205.14135) for more details if interested)
Looking at the formula, one would intuitively say that Flash Attention must be much slower compared to the default self-attention formula as more computation needs to be done. Indeed Flash Attention requires more FLOPs compared to normal attention as the softmax normalization statistics have to constantly be recomputed (see [paper](https://huggingface.co/papers/2205.14135) for more details if interested)
> However, Flash Attention is much faster in inference compared to default attention which comes from its ability to significantly reduce the demands on the slower, high-bandwidth memory of the GPU (VRAM), focusing instead on the faster on-chip memory (SRAM).
@@ -526,22 +526,22 @@ Therefore, for the LLM without position embeddings each token appears to have th
For the LLM to understand sentence order, an additional *cue* is needed and is usually applied in the form of *positional encodings* (or also called *positional embeddings*).
Positional encodings, encode the position of each token into a numerical presentation that the LLM can leverage to better understand sentence order.
The authors of the [*Attention Is All You Need*](https://arxiv.org/abs/1706.03762) paper introduced sinusoidal positional embeddings \\( \mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N \\) .
The authors of the [*Attention Is All You Need*](https://huggingface.co/papers/1706.03762) paper introduced sinusoidal positional embeddings \\( \mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N \\) .
where each vector \\( \mathbf{p}_i \\) is computed as a sinusoidal function of its position \\( i \\) .
The positional encodings are then simply added to the input sequence vectors \\( \mathbf{\hat{X}} = \mathbf{\hat{x}}_1, \ldots, \mathbf{\hat{x}}_N \\) = \\( \mathbf{x}_1 + \mathbf{p}_1, \ldots, \mathbf{x}_N + \mathbf{p}_N \\) thereby cueing the model to better learn sentence order.
Instead of using fixed position embeddings, others (such as [Devlin et al.](https://arxiv.org/abs/1810.04805)) used learned positional encodings for which the positional embeddings
Instead of using fixed position embeddings, others (such as [Devlin et al.](https://huggingface.co/papers/1810.04805)) used learned positional encodings for which the positional embeddings
\\( \mathbf{P} \\) are learned during training.
Sinusoidal and learned position embeddings used to be the predominant methods to encode sentence order into LLMs, but a couple of problems related to these positional encodings were found:
1. Sinusoidal and learned position embeddings are both absolute positional embeddings, *i.e.* encoding a unique embedding for each position id: \\( 0, \ldots, N \\) . As shown by [Huang et al.](https://arxiv.org/abs/2009.13658) and [Su et al.](https://arxiv.org/abs/2104.09864), absolute positional embeddings lead to poor LLM performance for long text inputs. For long text inputs, it is advantageous if the model learns the relative positional distance input tokens have to each other instead of their absolute position.
1. Sinusoidal and learned position embeddings are both absolute positional embeddings, *i.e.* encoding a unique embedding for each position id: \\( 0, \ldots, N \\) . As shown by [Huang et al.](https://huggingface.co/papers/2009.13658) and [Su et al.](https://huggingface.co/papers/2104.09864), absolute positional embeddings lead to poor LLM performance for long text inputs. For long text inputs, it is advantageous if the model learns the relative positional distance input tokens have to each other instead of their absolute position.
2. When using learned position embeddings, the LLM has to be trained on a fixed input length \\( N \\), which makes it difficult to extrapolate to an input length longer than what it was trained on.
Recently, relative positional embeddings that can tackle the above mentioned problems have become more popular, most notably:
- [Rotary Position Embedding (RoPE)](https://arxiv.org/abs/2104.09864)
- [ALiBi](https://arxiv.org/abs/2108.12409)
- [Rotary Position Embedding (RoPE)](https://huggingface.co/papers/2104.09864)
- [ALiBi](https://huggingface.co/papers/2108.12409)
Both *RoPE* and *ALiBi* argue that it's best to cue the LLM about sentence order directly in the self-attention algorithm as it's there that word tokens are put into relation with each other. More specifically, sentence order should be cued by modifying the \\( \mathbf{QK}^T \\) computation.
@@ -556,14 +556,14 @@ $$ \mathbf{\hat{q}}_i^T \mathbf{\hat{x}}_j = \mathbf{{q}}_i^T \mathbf{R}_{\theta
*RoPE* is used in multiple of today's most important LLMs, such as:
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**Llama**](https://arxiv.org/abs/2302.13971)
- [**PaLM**](https://arxiv.org/abs/2204.02311)
- [**Llama**](https://huggingface.co/papers/2302.13971)
- [**PaLM**](https://huggingface.co/papers/2204.02311)
As an alternative, *ALiBi* proposes a much simpler relative position encoding scheme. The relative distance that input tokens have to each other is added as a negative integer scaled by a pre-defined value `m` to each query-key entry of the \\( \mathbf{QK}^T \\) matrix right before the softmax computation.
![](/blog/assets/163_optimize_llm/alibi.png)
As shown in the [ALiBi](https://arxiv.org/abs/2108.12409) paper, this simple relative positional encoding allows the model to retain a high performance even at very long text input sequences.
As shown in the [ALiBi](https://huggingface.co/papers/2108.12409) paper, this simple relative positional encoding allows the model to retain a high performance even at very long text input sequences.
*ALiBi* is used in multiple of today's most important LLMs, such as:
@@ -572,7 +572,7 @@ As shown in the [ALiBi](https://arxiv.org/abs/2108.12409) paper, this simple rel
Both *RoPE* and *ALiBi* position encodings can extrapolate to input lengths not seen during training whereas it has been shown that extrapolation works much better out-of-the-box for *ALiBi* as compared to *RoPE*.
For ALiBi, one simply increases the values of the lower triangular position matrix to match the length of the input sequence.
For *RoPE*, keeping the same \\( \theta \\) that was used during training leads to poor results when passing text inputs much longer than those seen during training, *c.f* [Press et al.](https://arxiv.org/abs/2108.12409). However, the community has found a couple of effective tricks that adapt \\( \theta \\), thereby allowing *RoPE* position embeddings to work well for extrapolated text input sequences (see [here](https://github.com/huggingface/transformers/pull/24653)).
For *RoPE*, keeping the same \\( \theta \\) that was used during training leads to poor results when passing text inputs much longer than those seen during training, *c.f* [Press et al.](https://huggingface.co/papers/2108.12409). However, the community has found a couple of effective tricks that adapt \\( \theta \\), thereby allowing *RoPE* position embeddings to work well for extrapolated text input sequences (see [here](https://github.com/huggingface/transformers/pull/24653)).
> Both RoPE and ALiBi are relative positional embeddings that are *not* learned during training, but instead are based on the following intuitions:
- Positional cues about the text inputs should be given directly to the \\( QK^T \\) matrix of the self-attention layer
@@ -742,21 +742,21 @@ Researchers have proposed two methods that allow to significantly reduce the mem
#### 3.2.2 Multi-Query-Attention (MQA)
[Multi-Query-Attention](https://arxiv.org/abs/1911.02150) was proposed in Noam Shazeer's *Fast Transformer Decoding: One Write-Head is All You Need* paper. As the title says, Noam found out that instead of using `n_head` key-value projections weights, one can use a single head-value projection weight pair that is shared across all attention heads without that the model's performance significantly degrades.
[Multi-Query-Attention](https://huggingface.co/papers/1911.02150) was proposed in Noam Shazeer's *Fast Transformer Decoding: One Write-Head is All You Need* paper. As the title says, Noam found out that instead of using `n_head` key-value projections weights, one can use a single head-value projection weight pair that is shared across all attention heads without that the model's performance significantly degrades.
> By using a single head-value projection weight pair, the key value vectors \\( \mathbf{k}_i, \mathbf{v}_i \\) have to be identical across all attention heads which in turn means that we only need to store 1 key-value projection pair in the cache instead of `n_head` ones.
As most LLMs use between 20 and 100 attention heads, MQA significantly reduces the memory consumption of the key-value cache. For the LLM used in this notebook we could therefore reduce the required memory consumption from 15 GB to less than 400 MB at an input sequence length of 16000.
In addition to memory savings, MQA also leads to improved computational efficiency as explained in the following.
In auto-regressive decoding, large key-value vectors need to be reloaded, concatenated with the current key-value vector pair to be then fed into the \\( \mathbf{q}_c\mathbf{K}^T \\) computation at every step. For auto-regressive decoding, the required memory bandwidth for the constant reloading can become a serious time bottleneck. By reducing the size of the key-value vectors less memory needs to be accessed, thus reducing the memory bandwidth bottleneck. For more detail, please have a look at [Noam's paper](https://arxiv.org/abs/1911.02150).
In auto-regressive decoding, large key-value vectors need to be reloaded, concatenated with the current key-value vector pair to be then fed into the \\( \mathbf{q}_c\mathbf{K}^T \\) computation at every step. For auto-regressive decoding, the required memory bandwidth for the constant reloading can become a serious time bottleneck. By reducing the size of the key-value vectors less memory needs to be accessed, thus reducing the memory bandwidth bottleneck. For more detail, please have a look at [Noam's paper](https://huggingface.co/papers/1911.02150).
The important part to understand here is that reducing the number of key-value attention heads to 1 only makes sense if a key-value cache is used. The peak memory consumption of the model for a single forward pass without key-value cache stays unchanged as every attention head still has a unique query vector so that each attention head still has a different \\( \mathbf{QK}^T \\) matrix.
MQA has seen wide adoption by the community and is now used by many of the most popular LLMs:
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**PaLM**](https://arxiv.org/abs/2204.02311)
- [**PaLM**](https://huggingface.co/papers/2204.02311)
- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
- [**BLOOM**](https://huggingface.co/bigscience/bloom)
@@ -764,7 +764,7 @@ Also, the checkpoint used in this notebook - `bigcode/octocoder` - makes use of
#### 3.2.3 Grouped-Query-Attention (GQA)
[Grouped-Query-Attention](https://arxiv.org/abs/2305.13245), as proposed by Ainslie et al. from Google, found that using MQA can often lead to quality degradation compared to using vanilla multi-key-value head projections. The paper argues that more model performance can be kept by less drastically reducing the number of query head projection weights. Instead of using just a single key-value projection weight, `n < n_head` key-value projection weights should be used. By choosing `n` to a significantly smaller value than `n_head`, such as 2,4 or 8 almost all of the memory and speed gains from MQA can be kept while sacrificing less model capacity and thus arguably less performance.
[Grouped-Query-Attention](https://huggingface.co/papers/2305.13245), as proposed by Ainslie et al. from Google, found that using MQA can often lead to quality degradation compared to using vanilla multi-key-value head projections. The paper argues that more model performance can be kept by less drastically reducing the number of query head projection weights. Instead of using just a single key-value projection weight, `n < n_head` key-value projection weights should be used. By choosing `n` to a significantly smaller value than `n_head`, such as 2,4 or 8 almost all of the memory and speed gains from MQA can be kept while sacrificing less model capacity and thus arguably less performance.
Moreover, the authors of GQA found out that existing model checkpoints can be *uptrained* to have a GQA architecture with as little as 5% of the original pre-training compute. While 5% of the original pre-training compute can still be a massive amount, GQA *uptraining* allows existing checkpoints to be useful for longer input sequences.
@@ -776,7 +776,7 @@ The most notable application of GQA is [Llama-v2](https://huggingface.co/meta-ll
## Conclusion
The research community is constantly coming up with new, nifty ways to speed up inference time for ever-larger LLMs. As an example, one such promising research direction is [speculative decoding](https://arxiv.org/abs/2211.17192) where "easy tokens" are generated by smaller, faster language models and only "hard tokens" are generated by the LLM itself. Going into more detail is out of the scope of this notebook, but can be read upon in this [nice blog post](https://huggingface.co/blog/assisted-generation).
The research community is constantly coming up with new, nifty ways to speed up inference time for ever-larger LLMs. As an example, one such promising research direction is [speculative decoding](https://huggingface.co/papers/2211.17192) where "easy tokens" are generated by smaller, faster language models and only "hard tokens" are generated by the LLM itself. Going into more detail is out of the scope of this notebook, but can be read upon in this [nice blog post](https://huggingface.co/blog/assisted-generation).
The reason massive LLMs such as GPT3/4, Llama-2-70b, Claude, PaLM can run so quickly in chat-interfaces such as [Hugging Face Chat](https://huggingface.co/chat/) or ChatGPT is to a big part thanks to the above-mentioned improvements in precision, algorithms, and architecture.
Going forward, accelerators such as GPUs, TPUs, etc... will only get faster and allow for more memory, but one should nevertheless always make sure to use the best available algorithms and architectures to get the most bang for your buck 🤗