Faster generation using AWQ + Fused modules

docker/transformers-all-latest-gpu/Dockerfile

@@ -56,7 +56,7 @@ RUN python3 -m pip install --no-cache-dir auto-gptq --extra-index-url https://hu
 RUN python3 -m pip install --no-cache-dir einops
 
 # Add autoawq for quantization testing
-RUN python3 -m pip install --no-cache-dir https://github.com/casper-hansen/AutoAWQ/releases/download/v0.1.6/autoawq-0.1.6+cu118-cp38-cp38-linux_x86_64.whl
+RUN python3 -m pip install --no-cache-dir https://github.com/casper-hansen/AutoAWQ/releases/download/v0.1.7/autoawq-0.1.7+cu118-cp38-cp38-linux_x86_64.whl
 
 # For bettertransformer + gptq 
 RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/optimum@main#egg=optimum

docs/source/en/quantization.md

@@ -85,6 +85,147 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 model = AutoModelForCausalLM.from_pretrained("TheBloke/zephyr-7B-alpha-AWQ", use_flash_attention_2=True, device_map="cuda:0")
 ```
 
+
+### Benchmarks
+
+We performed some speed, throughput and latency benchmarks using [`optimum-benchmark`](https://github.com/huggingface/optimum-benchmark) library. 
+
+Note at that time of writing this documentation section, the available quantization methods were: `awq`, `gptq` and `bitsandbytes`.
+
+The benchmark was run on a NVIDIA-A100 instance and the model used was [`TheBloke/Mistral-7B-v0.1-AWQ`](https://huggingface.co/TheBloke/Mistral-7B-v0.1-AWQ) for the AWQ model, [`TheBloke/Mistral-7B-v0.1-GPTQ`](https://huggingface.co/TheBloke/Mistral-7B-v0.1-GPTQ) for the GPTQ model. We also benchmarked it against `bitsandbytes` quantization methods and native `float16` model. Some results are shown below:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/forward_memory_plot.png">
+</div>
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/generate_memory_plot.png">
+</div>
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/generate_throughput_plot.png">
+</div>
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/forward_latency_plot.png">
+</div>
+
+You can find the full results together with packages versions in [this link](https://github.com/huggingface/optimum-benchmark/tree/main/examples/running-mistrals).
+
+From the results it appears that AWQ quantization method is the fastest quantization method for inference, text generation and among the lowest peak memory for text generation. However, AWQ seems to have the largest forward latency per batch size. 
+
+
+### Make use of fused modules
+
+You can benefit from fused modules by passing an `AwqConfig` with `fuse_modules=True` and your expected maximum sequence length for generation to `fuse_max_seq_len`. For architectures that do not support `do_fuse=True`, you can still fuse the modules, however you need to pass a custom `fusing_mapping` to `AwqConfig()`. Let's dive into these specific usecases.
+
+Note that you cannot combine fusing modules and other optimization techniques such as Flash Attention 2.
+
+#### Fusing modules for supported architectures
+
+Currently we support out of the box AWQ module fusing for `llama` and `mistral`. 
+
+To enable this feature for supported architectures simply create an `AwqConfig` and pass the arguments `fuse_max_seq_len` and `do_fuse=True`.
+
+For example to enable module fusing for the model `TheBloke/Mistral-7B-OpenOrca-AWQ`, run:
+
+```python
+import torch
+from transformers import AwqConfig, AutoModelForCausalLM
+
+model_id = "TheBloke/Mistral-7B-OpenOrca-AWQ"
+
+quantization_config = AwqConfig(
+    bits=4,
+    fuse_max_seq_len=512,
+    do_fuse=True,
+)
+
+model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config).to(0)
+```
+
+Note that you need to define `fuse_max_seq_len` to `AwqConfig`. That total sequence length should include the context length and the expected generation length. You can set it to a large value to be on the safe zone.
+
+You can also apply module fusing for other architectures that are not supported.
+
+#### Fusing modules for unsupported architectures
+
+For architectures that do not support out of the box module fusing, you can pass a custom fusing mapping; simply pass a dictionnary `modules_to_fuse` to `AwqConfig`, let's take an example with the Yi model:
+
+
+```python
+import torch
+from transformers import AwqConfig, AutoModelForCausalLM
+
+model_id = "TheBloke/Yi-34B-AWQ"
+
+quantization_config = AwqConfig(
+    bits=4,
+    fuse_max_seq_len=512,
+    modules_to_fuse={
+        "attention": ["q_proj", "k_proj", "v_proj", "o_proj"],
+        "layernorm": ["ln1", "ln2", "norm"],
+        "mlp": ["gate_proj", "up_proj", "down_proj"],
+        "use_alibi": False,
+        "num_attention_heads": 56,
+        "num_key_value_heads": 8,
+        "hidden_size": 7168
+    }
+)
+
+model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=quantization_config).to(0)
+```
+
+The parameter `modules_to_fuse` needs to have the following respective fields: 
+
+- `"attention"`: The names of the attention layers to fuse - in the order: query, key, value and output projection layer. In case you don't want to fuse the attention layers you can pass an empty list.
+- `"layernorm"`: The names of all the layernorm layers you want to replace with a custom fused layer norm. In case you don't want to fuse these layers you can also pass an empty list.
+- `"mlp"`: The names of the MLP layers you want to fuse into a single MLP layer in the order: (gate (dense layer post-attention) / up / down layers).
+- `"use_alibi"`: If you model uses alibi positional embedding
+- `"num_attention_heads"`: The number of attention heads
+- `"num_key_value_heads"`: This is the number of key value heads that should be used to implement Grouped Query Attention. If num_key_value_heads=num_attention_heads, the model will use Multi Head Attention (MHA), if num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. 
+- `"hidden_size"`: Dimension of the hidden representations.
+
+
+#### Benchmarks
+
+We benchmarked the model with and without fused modules first using only `batch_size=1` on the `TheBloke/Mistral-7B-OpenOrca-AWQ` model and below are the results:
+
+*unfused case*
+
+|   Batch Size |   Prefill Length |   Decode Length |   Prefill tokens/s |   Decode tokens/s | Memory (VRAM)   |
+|-------------:|-----------------:|----------------:|-------------------:|------------------:|:----------------|
+|            1 |               32 |              32 |            60.0984 |           38.4537 | 4.50 GB (5.68%) |
+|            1 |               64 |              64 |          1333.67   |           31.6604 | 4.50 GB (5.68%) |
+|            1 |              128 |             128 |          2434.06   |           31.6272 | 4.50 GB (5.68%) |
+|            1 |              256 |             256 |          3072.26   |           38.1731 | 4.50 GB (5.68%) |
+|            1 |              512 |             512 |          3184.74   |           31.6819 | 4.59 GB (5.80%) |
+|            1 |             1024 |            1024 |          3148.18   |           36.8031 | 4.81 GB (6.07%) |
+|            1 |             2048 |            2048 |          2927.33   |           35.2676 | 5.73 GB (7.23%) |
+
+*fused case*
+
+|   Batch Size |   Prefill Length |   Decode Length |   Prefill tokens/s |   Decode tokens/s | Memory (VRAM)   |
+|-------------:|-----------------:|----------------:|-------------------:|------------------:|:----------------|
+|            1 |               32 |              32 |            81.4899 |           80.2569 | 4.00 GB (5.05%) |
+|            1 |               64 |              64 |          1756.1    |          106.26   | 4.00 GB (5.05%) |
+|            1 |              128 |             128 |          2479.32   |          105.631  | 4.00 GB (5.06%) |
+|            1 |              256 |             256 |          1813.6    |           85.7485 | 4.01 GB (5.06%) |
+|            1 |              512 |             512 |          2848.9    |           97.701  | 4.11 GB (5.19%) |
+|            1 |             1024 |            1024 |          3044.35   |           87.7323 | 4.41 GB (5.57%) |
+|            1 |             2048 |            2048 |          2715.11   |           89.4709 | 5.57 GB (7.04%) |
+
+We also performed benchmarks with [`optimum-benchmark`](https://github.com/huggingface/optimum-benchmark) library. And below are the results:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/fused_forward_memory_plot.png">
+</div>
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/quantization/fused_generate_throughput.png">
+</div>
+
+
 ## AutoGPTQ
 
 <Tip>

src/transformers/integrations/__init__.py

@@ -17,7 +17,7 @@
 
 
 _import_structure = {
-    "awq": ["replace_with_awq_linear"],
+    "awq": ["fuse_awq_modules", "replace_with_awq_linear"],
     "bitsandbytes": [
         "get_keys_to_not_convert",
         "replace_8bit_linear",
@@ -80,7 +80,7 @@
 }
 
 if TYPE_CHECKING:
-    from .awq import replace_with_awq_linear
+    from .awq import fuse_awq_modules, replace_with_awq_linear
     from .bitsandbytes import (
         get_keys_to_not_convert,
         replace_8bit_linear,