[Feature Request] 4bit and 2bit and 1bit quantization support #14997

elephantpanda · 2023-03-10T07:37:02Z

Describe the feature request

Support for quantizing and running quantized models in 4bit, 2bit and 1bit. Also saving and loading these models in onnx format for lower file sizes.

The GPU doesn't necessarily have to support 4bit operations since it can just use gpu cores to convert them to float operations or int8 operations when needed.

Describe scenario use case

Some models such as Large Language Models are very big but run fairly well when quantized down to 8bit, 4bit, 2bit or even 1bit.

jchen351 · 2023-03-10T17:07:54Z

Hi, Pauldog thanks for reaching out. We have received your message and put these requests under consideration!

Thank you for your time,

Jian Chen (not a A.I.)

jchen351 · 2023-03-10T18:38:28Z

Also Could you please provide me more information about your scenarios, like: hardware to you wants to run on, and models you are interested in? Again, out currently priority is on fp16 support. And there isn't any hardware we have that supports the 4bit or lower.

elephantpanda · 2023-03-10T20:18:50Z

Sure here is a very recent example of a practical use case:

Llama 4bit

As far as I'm aware it doesn't require 4bit hardware it simply stores the weights on the GPU in 4bit, then uses GPU cores at runtime to convert them to int8 or float16 at runtime to do the calculations.

The main benefit is the ability to run larger models on the same hardware.

Use cases would be

Running very large language models on consumer hardware
Running large models on mobile hardware

Here are some papers

https://arxiv.org/abs/1810.05723
https://arxiv.org/abs/2202.05292

and articles
https://karanbirchahal.medium.com/aggressive-quantization-how-to-run-mnist-on-a-4-bit-neural-net-using-pytorch-5703f3faa599

Now, I don't know whether onnxruntime already can support this or not? Since technically say a 4bit quantized model would presumably appear like an 8bit quantized model as two 4bits are combined into one 8bit.

josephrocca · 2023-03-14T03:36:02Z

Hey @jchen351, I'm wondering why this is closed? Shouldn't it stay open if this is being considered?

The WebML ecosystem in particular could really do with a 4-bit quantization solution, since model size is such an important factor on the web.

xenova · 2023-03-14T03:47:18Z

100% agree with @josephrocca. 4-bit quantization would be massive for my Transformers.js library (and other WebML libraries)!

jchen351 · 2023-03-14T21:04:46Z

@xenova @josephrocca The only hardware we know that can support 4 bit quantization with performance gain is Nvidia A100, but we cannot get our hands on enough A100, and the newer H100 has dropped that support. We don't foresee any performance gain in doing 4 bit quantization on any other popular hardwares.
So, until them, I will keep this closed :)

xenova · 2023-03-14T21:40:11Z

This repo supports 4-bit quantization: https://github.com/ggerganov/llama.cpp
(And, as stated in the README, it runs on the CPU)

Also, considering that WASM uses a 32-bit address space (i.e., max 4GB), the only real way to get large models running on consumer hardware is quantization.

josephrocca · 2023-03-15T01:45:52Z

@jchen351, yes, as xenova pointed out, this is more about running large models on hardware that has a small amount of memory, rather than performance improvements.

For example, please see this demo of llama 7B running on a pixel 5 at 1 token/sec using 4 bit quantization: https://twitter.com/ggerganov/status/1635605532726681600

So this issue can probably be re-opened considering it is viable to gain this benefit without hardware support? llama.cpp has grown faster than the original stable diffusion repo (which was one of the fastest growing of all time) because it allows people to run big models on small hardware -- there's definitely demand for this! :)

skyne98 · 2023-04-24T09:19:42Z

@jchen351, can we have a second look at this? It's not really about performance, but rather allowing running models in places they couldn't before. I insist!

It just seems like the points that guys made, which are really valid, got seemingly plainly ignored.

tikikun · 2023-06-02T01:59:59Z

Re-open please, everyone is using 4-5bit quantization now

jywu-msft · 2023-07-06T20:02:50Z

re-opening this. this should not be closed.

jywu-msft · 2023-07-06T20:03:42Z

+@yufenglee FYI

ThisisBillhe · 2023-07-12T01:48:44Z

Hi everyone! I have successfully quantized a diffusion model to 2-bit and manually packed them into uint8 format (store 4x 2-bit weight in an uint8 variable) in pytorch. During inference, they are unpacked to float format for calculation. In this way, the model size has been reduced from 1545M to 150M, and the VRAM for loading the model is also greatly reduced (from 2500M to 1000M) in pytorch. However, when I export the model to onnx, only the model size is reduced (to around 190M), the VRAM for loading the model can still reach 3000M. I guess the uint8 parameters are cast to int32 or float32 during loading the onnx model.

Any ideas on how to lower the VRAM for loading this ONNX model? I have upload the model at googledrive.

elephantpanda · 2023-07-12T02:06:50Z

Hi everyone! I have successfully quantized a diffusion model to 2-bit and manually packed them into uint8 format (store 4x 2-bit weight in an uint8 variable) in pytorch. During inference, they are unpacked to float format for calculation. In this way, the model size has been reduced from 1545M to 150M, and the VRAM for loading the model is also greatly reduced (from 2500M to 1000M) in pytorch. However, when I export the model to onnx, only the model size is reduced (to around 190M), the VRAM for loading the model can still reach 3000M. I guess the uint8 parameters are cast to int32 or float32 during loading the onnx model.

Any ideas on how to lower the VRAM for loading this ONNX model? I have upload the model at googledrive.

2-bit diffusion model? Does it actually produce images?

Guess you could try packing 16 2-bits into an int32.

ThisisBillhe · 2023-07-12T02:19:50Z

Hi everyone! I have successfully quantized a diffusion model to 2-bit and manually packed them into uint8 format (store 4x 2-bit weight in an uint8 variable) in pytorch. During inference, they are unpacked to float format for calculation. In this way, the model size has been reduced from 1545M to 150M, and the VRAM for loading the model is also greatly reduced (from 2500M to 1000M) in pytorch. However, when I export the model to onnx, only the model size is reduced (to around 190M), the VRAM for loading the model can still reach 3000M. I guess the uint8 parameters are cast to int32 or float32 during loading the onnx model.
Any ideas on how to lower the VRAM for loading this ONNX model? I have upload the model at googledrive.

2-bit diffusion model? Does it actually produce images?

Guess you could try packing 16 2-bits into an int32.

The work is in progress..I guess you make a point, I will have a try.

dfiru · 2023-10-14T15:02:20Z

are there any branches or forks of the 2 x 4bit packing?

josephrocca · 2023-10-15T03:43:40Z

I noticed this point in the v1.16.0 release notes (3 weeks ago):

Support 4-bit quantization on CPU

I haven't tried it yet. @xenova I'm curious if you've tried this yet with the Web Wasm backend?

dfiru · 2023-10-15T03:47:10Z

https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/quantization/quant_utils.py#L71
QuantType still doesn't include it.

### Description This PR contains fusion-level and kernel-level optimizations for [Meta's LLaMA-2](https://blogs.microsoft.com/blog/2023/07/18/microsoft-and-meta-expand-their-ai-partnership-with-llama-2-on-azure-and-windows/). Some of the added optimizations include: - SimplifiedLayerNorm changes - Fusions for multiple variants - SkipSimplifiedLayerNorm changes - Kernel support for CPU - Rotary embeddings (previously did not exist) - Fusions for multiple variants - CPU and CUDA kernels - Supports interleaving and non-interleaving in the same kernels - Optimized cache that requires half of its originally exported sizes - Reduced from `(max_sequence_length, head_size)` to `(max_sequence_length, head_size / 2)` - Multi-head attention - Support for 2D and 3D attention masks - Group query attention (for FP16 CUDA and INT4 CUDA) - Integration with flash attention v2 and past-present buffer sharing - Removes need for `attention_mask` input as it is supported in the kernel - 4 bit quantization - `block_size` parameter is available for customizing - Support the new changes for [Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Support combinations of the below variants (ex: export ORT version and run with Optimum) Supported variants of LLaMA-2 include: - [ORT version](https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/python/tools/transformers/models/llama) - Produces one ONNX file that is already optimized (and quantized if requested) - Integrates with Optimum - [Another Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Already exported and available off-the-shelf - Faster versions of those models will be uploaded there soon - [Hugging Face version](https://huggingface.co/meta-llama) - Models that end with `-hf` - Some older and current versions of [`transformers`](https://github.com/huggingface/transformers) and [`optimum`](https://github.com/huggingface/optimum) that export the model to ONNX differently - Note that while some older versions are supported, it is recommended to use the latest package versions. ### Usage To use the optimizations, please see `README.md` for details. Please note the various `requirements.txt` files for the package versions recommended in order to use these changes. To run the ORT transformer optimizer separately, run the script as follows: ``` $ cd onnxruntime/onnxruntime/python/tools/transformers/ $ python3 optimizer.py --input <filename>.onnx --output <filename>.onnx --model_type gpt2 --num_heads <number of attention heads> --hidden_size <attention hidden size> --use_external_data_format --opt_level 0 ``` ### Motivation and Context This PR helps the following issues: - #14997 - #16254 - #17681 - #17925 - microsoft/onnxruntime-inference-examples#320 This PR uses changes from the following PRs: - pytorch/pytorch#104468 - pytorch/pytorch#109759 - #17020 - #17674 - #17890 - #17920 - huggingface/transformers#26162 - huggingface/optimum#1257 - huggingface/optimum#1289 - huggingface/optimum#1462 ### New TorchDynamo Exporter (experimental stage) This PR uses changes from the following issues and PRs to begin supporting the [new TorchDynamo exporter](https://pytorch.org/docs/stable/onnx.html#torchdynamo-based-onnx-exporter): - huggingface/transformers#26307 - pytorch/pytorch#104903 - pytorch/pytorch#105040 - microsoft/onnxscript#847 - microsoft/onnxscript#862 - microsoft/onnxscript#493

@skottmckay

commit 538e97c Author: Patrice Vignola <[email protected]> Date: Wed Oct 25 19:56:16 2023 -0700 [DML EP] Add dynamic graph compilation (#17876) Historically, DML was only able to fuse partitions when all sizes are known in advance or when we were overriding them at session creation time. But in practice, it should be possible to compile partitions at compute time if the caller knows that the dimensions won't be changed for every inference (e.g. resizing a webcam window, or padding the input to powers of 2). This graph will be cached and reused until the sizes change. This is an opt-in option gated under the `enable_dynamic_graph_fusion` option, which means that it will only be enabled when the caller requests it since they have more context on how their model will be called between inferences. This PR also adds the option to disable metacommands from the python API, which is an option for the C API but was lacking for python. commit d30d4d3 Author: Jambay Kinley <[email protected]> Date: Wed Oct 25 15:34:58 2023 -0700 Add MatMul FP4 and NF4 Support (#18066) Add a contrib op MatMulBnb4 (FP4 and NF4) and related toolchain to support quantization on weight. This PR adds: - schema for contrib op MatMulBnb4 which can support FP4 (4-bit floating point) and NF4 (4-bit NormalFloat) quantization on weight. - a naive implementation for MatMulBnb4 on CPU and GPU, i.e., implemented like MatMul(A, Dequantize(B)). - a special implementation for GemV for MatMulBnb4 and related benchmark tool. - tool to quantize model to FP4 or NF4. commit d88d52e Author: snadampal <[email protected]> Date: Wed Oct 25 13:34:57 2023 -0500 [aarch64] Remove mmla kernel support from apple (#18082)  The mmla kernels require additional ISA flags and are currently supported only on Linux  more context is in #15270 cc: @skottmckay , @chenfucn , @snnn commit 706e13e Author: liqun Fu <[email protected]> Date: Wed Oct 25 10:46:04 2023 -0700 implement affinegrid cpu kernel (#17777) commit 2c6b31c Author: pengwa <[email protected]> Date: Wed Oct 25 15:11:02 2023 +0800 FP16 optimizer automatically detect DeepSpeed compatibility (#18084) Optimum/Transformers are using accelerate lib to prepare models, so our FP16 optimizer wrapper does not work for long time. Because the namespace is `accelerate.utils.deepspeed.DeepSpeedOptimizerWrapper`, which underlying is still calling into DeepSpeed stage1and2 optimizer. This PR includes following changes: 1. Add `accelerate.utils.deepspeed.DeepSpeedOptimizerWrapper` in the modifier registry, plus a check on its contained `optimizer` property MUST be DeepSpeed stage 1 and 2 optimizer. (let's cover Stage 3 optimizer later) 2. For DeepSpeed version > 0.9.1, we will store the source code in a version list. As long as the related function in DeepSpeed remains unchanged during its new release, we won't need manually upgrade the version check any more. If some day, the source code did not match, a warning will be raised to users, to add a new version of source code in the list. With the above change, we will have our FP16 Optimizer working again in Optimum. ![image](https://github.com/microsoft/onnxruntime/assets/10530022/d35b4aa9-b371-46f1-98ae-73114f91179b) commit ae85619 Author: Sumit Agarwal <[email protected]> Date: Tue Oct 24 19:41:10 2023 -0700 Introduce new optimizer MatMul + BatchNormalization (#17915) Introduce new ORT L1 optimizer under RewriteRule category to fuse MatMul + BatchNormalization node. This optimizer look for a specific pattern observed in one of the impacting customer models and fuse the Matmul and Batchnormalization node into a Gemm node. For details on the pattern matching and fusion please refer to the comment section of `matmul_bn_fusion.cc`. To visualize, this optimizer will replace following subgraph to a Gemm node. <pre> MatMul GEMM | | Reshape ^ ---> Reshape ^ | | Transpose ^ Transpose ^ | BatchNormalization Note: ^ means there can be >=0 occurrence(s) of that node. Few example fusable pattern: * - MatMul -> Reshape -> Transpose -> BatchNormalization ---> GEMM -> Reshape -> Transpose * - MatMul -> Reshape -> BatchNormalization ---> GEMM -> Reshape * - MatMul -> Transpose -> BatchNormalization ---> GEMM -> Transpose * - MatMul -> Reshape -> Reshape -> BatchNormalization ---> GEMM -> Reshape -> Reshape * - MatMul -> Reshape -> Transpose -> Reshape -> BatchNormalization ---> GEMM -> Reshape -> Transpose -> Reshape * - MatMul -> BatchNormalization ---> GEMM </pre> Note: This optimizer may evolve in the future to be more generic in terms of the pattern matching. - Why is this change required? What problem does it solve? One of the user of ORT+DML ep needs this to better target the model to DML. But this transformation applies more broadly, so added L1 optimizer.  commit 76e275b Author: Jian Chen <[email protected]> Date: Tue Oct 24 15:17:36 2023 -0700 Merge Cuda docker files into a single one (#18020)   commit 6ec45f2 Author: Changming Sun <[email protected]> Date: Tue Oct 24 13:04:08 2023 -0700 Merge aiinfra-linux-ARM64-CPU-2019 and onnxruntime-linux-ARM64-CPU-2019 (#18069) Merge aiinfra-linux-ARM64-CPU-2019 and onnxruntime-linux-ARM64-CPU-2019 machines to a single one to ease management. commit efa0cc2 Author: liqun Fu <[email protected]> Date: Tue Oct 24 10:58:54 2023 -0700 implement isinf20 and isnan20 (#17874) commit abb3291 Author: Changming Sun <[email protected]> Date: Tue Oct 24 10:50:12 2023 -0700 Update win-wasm-ci.yml: increase the timeout value (#18023) commit e63ccd3 Author: Jian Chen <[email protected]> Date: Tue Oct 24 10:47:23 2023 -0700 Install CUDA 12.2 on Windows (#18044)   commit eb47008 Author: Jiajia Qin <[email protected]> Date: Tue Oct 24 13:56:56 2023 +0800 [js/webgpu] FP16 Cast, Resize (#18035)  Cast/Resize with f16 are missing in vae-decoder-f16. With this change, vae-decoder-f16 becomes 315 ms from over than 1 seconds. commit 688524a Author: Tianlei Wu <[email protected]> Date: Mon Oct 23 22:00:02 2023 -0700 [CUDA EP] Add warning logs when adding memcpy nodes (#18032) Memcpy nodes could have negative impact on performance, they also cause ORT unable to run CUDA graph. Here we add a warning log for CUDA EP when this happens. It could help trouble shooting. For example, when CUDA graph cannot run, we can see the logs to find out where the Memcpy nodes are inserted (Although it is also possible through saving optimized model, but that need more time and disk space). Note that the warning is per graph. When there are subgraphs, we might see multiple warnings if the issue happens in multiple graphs. Example logs: ``` 2023-10-19 20:58:10.678176531 [I:onnxruntime:, transformer_memcpy.cc:329 AddCopyNode] Add MemcpyFromHost after input_ids for CUDAExecutionProvider 2023-10-19 20:58:10.678198702 [I:onnxruntime:, transformer_memcpy.cc:329 AddCopyNode] Add MemcpyFromHost after /text_model/ArgMax_output_0 for CUDAExecutionProvider 2023-10-19 20:58:10.678211727 [I:onnxruntime:, transformer_memcpy.cc:329 AddCopyNode] Add MemcpyFromHost after /text_model/Gather_3_output_0 for CUDAExecutionProvider 2023-10-19 20:58:10.678257903 [W:onnxruntime:, transformer_memcpy.cc:74 ApplyImpl] 3 Memcpy nodes are added to the graph main_graph for CUDAExecutionProvider. It might have negative impact on performance (including unable to run CUDA graph). Set session_options.log_severity_level=1 to see the detail logs before this message. ``` commit 555b2af Author: Chi Lo <[email protected]> Date: Tue Oct 24 02:41:15 2023 +0000 [TensorRT EP] Add unit test for user provided cuda stream (#17974) Add a unit test for testing user provided CUDA stream commit 4ffd022 Author: Chi Lo <[email protected]> Date: Tue Oct 24 00:46:38 2023 +0000 [TensorRT EP] Refactor of TRT plugins support (#17946) Make sure "trt.plugins" custom op domain only being registered once. The bottom line is "trt.plugins" custom op domain needs to be registered before model load. `CreateTensorRTCustomOpDomainList()` is TRT EP's function to create "trt.plugins" custom op domain. Following are places where this function will be called. (This function only fetches all the TRT plugins from TRT plugin registry but not yet registered them to ORT custom op registry. The real registration happens in AddCustomOpDomains()) C/C++ APIs: - `OrtApis::SessionOptionsAppendExecutionProvider_TensorRT_XX`: This function will make session option object contain the "trt.plugins" custom op domain for ORT to register. So that later the session creation api can register the custom op domain accordingly and won't complain about invalid onnx node. - `InferenceSession::RegisterExecutionProvider`: In some cases, users might create the session object first and later call session_object.RegisterExecutionProvider(). This function will call p_exec_provider->GetCustomOpDomainList() which returns "trt.plugins" custom op domain. Otherwise, session_object.Load(model) will complain. Python APIs: - `RegisterTensorRTPluginsAsCustomOps`: Need to call this function so that session option object contains the "trt.plugins" custom op domain for ORT to register. Different language bindings have slightly different workflow of initializing the session. This might cause duplicate custom op domain in `session_option.custom_op_domains_` or `CreateTensorRTCustomOpDomainList()` being called more than once, but we put checks to make sure ep's custom op domain won't be registered twice. commit 2c50b75 Author: Dmitri Smirnov <[email protected]> Date: Mon Oct 23 17:42:20 2023 -0700 Functions Ahead Of Time inlininng (#17764) Inline functions in an EP aware fashion. The result of this PR is that models that are having been inlined by ONNX inliner and optimized and models that have been AOT inlined appear to be visually identical. For tests I used two models. The only difference is the resulting size because ONNX inliner removes local function definitions and AOT does not. Difference in sizes for `HF Mobile` model was 2.5 MB, and for `HF Bart` it was ~500K. It seems that the resuling model size affects the load time more than the actual optimizations. In general, the inlined models grow in size very fast and can easily exceed 2Gb limit. Q. Should we make AOT optional? `If` costant folding and the removal of local inlined models will be coming in other PRs. Some stats: ![image](https://github.com/microsoft/onnxruntime/assets/11303988/fcb4c815-2e06-4574-8d96-5a0a727d1ecf) commit f3cfe08 Author: satyajandhyala <[email protected]> Date: Mon Oct 23 16:02:50 2023 -0700 [JS/Web] Enabled 1d spacial input to GlobalAveragePool (#17973) Enable one-dim special input to GlobalAveragePoll input  Currently only 2D input is supported. commit 780ee18 Author: snadampal <[email protected]> Date: Mon Oct 23 16:49:04 2023 -0500 [aarch64] Implement QGEMM kernels with UMMLA/SMMLA instructions (#17160)  This PR adds UMMLA and SMMLA based QGEMM kernels for aarch64. This covers (i) symmetric quantization (zero point is Zero) (ii) asymmetric quantization (zero point is non zero) (iii) per channel as well as per tensor quantization (iv) Signed weights (U8S8 Gemm) (v) Unsigned weights (U8U8 Gemm) and (vi) Signed activations and weights (S8S8 Gemm) scenarios I've enabled the ummla/smmla kernels based on cpuinfo check for `I8MM` support MMLA QGEMM kernels are enabled for all the devices that support I8MM instructions.  This is to improve INT8 quantized MatMul performance on aarch64 platform. I have run the below benchmarking script (bert , roberta and gpt2 model inference) on AWS Graviton3 based c7g.4xl instance and observed up to 1.33x performance improvement compared to the optimized UDOT qgemm kernel performance. ``` cd onnxruntime/python/tools/transformers python3 benchmark.py ``` I have also run the unit tests, and made sure all are passing ``` ./build.sh --config RelWithDebInfo --build_shared_lib --parallel --compile_no_warning_as_error --skip_submodule_sync ``` commit 2a17d5c Author: kunal-vaishnavi <[email protected]> Date: Mon Oct 23 13:00:56 2023 -0700 LLaMA Model Optimization (#18021) This PR contains fusion-level and kernel-level optimizations for [Meta's LLaMA-2](https://blogs.microsoft.com/blog/2023/07/18/microsoft-and-meta-expand-their-ai-partnership-with-llama-2-on-azure-and-windows/). Some of the added optimizations include: - SimplifiedLayerNorm changes - Fusions for multiple variants - SkipSimplifiedLayerNorm changes - Kernel support for CPU - Rotary embeddings (previously did not exist) - Fusions for multiple variants - CPU and CUDA kernels - Supports interleaving and non-interleaving in the same kernels - Optimized cache that requires half of its originally exported sizes - Reduced from `(max_sequence_length, head_size)` to `(max_sequence_length, head_size / 2)` - Multi-head attention - Support for 2D and 3D attention masks - Group query attention (for FP16 CUDA and INT4 CUDA) - Integration with flash attention v2 and past-present buffer sharing - Removes need for `attention_mask` input as it is supported in the kernel - 4 bit quantization - `block_size` parameter is available for customizing - Support the new changes for [Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Support combinations of the below variants (ex: export ORT version and run with Optimum) Supported variants of LLaMA-2 include: - [ORT version](https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/python/tools/transformers/models/llama) - Produces one ONNX file that is already optimized (and quantized if requested) - Integrates with Optimum - [Another Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Already exported and available off-the-shelf - Faster versions of those models will be uploaded there soon - [Hugging Face version](https://huggingface.co/meta-llama) - Models that end with `-hf` - Some older and current versions of [`transformers`](https://github.com/huggingface/transformers) and [`optimum`](https://github.com/huggingface/optimum) that export the model to ONNX differently - Note that while some older versions are supported, it is recommended to use the latest package versions. To use the optimizations, please see `README.md` for details. Please note the various `requirements.txt` files for the package versions recommended in order to use these changes. To run the ORT transformer optimizer separately, run the script as follows: ``` $ cd onnxruntime/onnxruntime/python/tools/transformers/ $ python3 optimizer.py --input <filename>.onnx --output <filename>.onnx --model_type gpt2 --num_heads <number of attention heads> --hidden_size <attention hidden size> --use_external_data_format --opt_level 0 ``` This PR helps the following issues: - #14997 - #16254 - #17681 - #17925 - microsoft/onnxruntime-inference-examples#320 This PR uses changes from the following PRs: - pytorch/pytorch#104468 - pytorch/pytorch#109759 - #17020 - #17674 - #17890 - #17920 - huggingface/transformers#26162 - huggingface/optimum#1257 - huggingface/optimum#1289 - huggingface/optimum#1462 This PR uses changes from the following issues and PRs to begin supporting the [new TorchDynamo exporter](https://pytorch.org/docs/stable/onnx.html#torchdynamo-based-onnx-exporter): - huggingface/transformers#26307 - pytorch/pytorch#104903 - pytorch/pytorch#105040 - microsoft/onnxscript#847 - microsoft/onnxscript#862 - microsoft/onnxscript#493 commit 8a12b2c Author: Jiajia Qin <[email protected]> Date: Tue Oct 24 02:02:19 2023 +0800 [js/webgpu] Fix the transpose error when dims > 4D (#18027)  Currently, the uniform support has bugs when dims rank is larger than 4. See #17860 item 1. So this PR only enables shapes uniforms when shape rank is <= 4 for transpose. Otherwise, below compilation errors are thrown: ``` 1 error(s) generated while compiling the shader: :3:50 error: uniform storage requires that array elements are aligned to 16 bytes, but array element of type 'u32' has a stride of 4 bytes. Consider using a vector or struct as the element type instead. struct Uniforms { output_size:u32, a_shape:array<u32, 5>, a_strides:array<u32, 5>, output_shape:array<u32, 5>, output_strides:array<u32, 5> }; ^^^^^^^^^^^^^ :3:7 note: see layout of struct: /* align(4) size(84) */ struct Uniforms { /* offset( 0) align(4) size( 4) */ output_size : u32; /* offset( 4) align(4) size(20) */ a_shape : array<u32, 5>; /* offset(24) align(4) size(20) */ a_strides : array<u32, 5>; /* offset(44) align(4) size(20) */ output_shape : array<u32, 5>; /* offset(64) align(4) size(20) */ output_strides : array<u32, 5>; /* */ }; struct Uniforms { output_size:u32, a_shape:array<u32, 5>, a_strides:array<u32, 5>, output_shape:array<u32, 5>, output_strides:array<u32, 5> }; ^^^^^^ :4:42 note: 'Uniforms' used in address space 'uniform' here @group(0) @binding(2) var<uniform> uniforms: Uniforms; ^^^^^^^^ ``` commit f0d5ea5 Author: Hector Li <[email protected]> Date: Mon Oct 23 09:01:29 2023 -0700 [QNN EP] Disable flaky test QnnCPUBackendTests.MatMulOp_Broadcast (#18033) Disable flaky test QnnCPUBackendTests.MatMulOp_Broadcast. The test failed on Linux randomly. commit b7ae293 Author: JiCheng <[email protected]> Date: Sun Oct 22 23:33:29 2023 +0800 Support large model export using multi-gpu (#17990) This PR is to implemente a exporter which works for large language models(LLM). It works for models like Llama2-70b or gpt-175. The main idea is to utilize multiple-GPU and dispatch differnet layers to different GPU, in short, it symply implemented auto pipeline parallelism. For example : to export Llama2-70b, you need 8x V100-32GB or 4x A100-80G or More GPU memories. It would expect to export decoder-only models. For encoder-decoder arch-like models, we didn't test it yet.  --------- Co-authored-by: Justin Chu <[email protected]> commit 444a0ed Author: pengwa <[email protected]> Date: Sat Oct 21 19:45:45 2023 +0800 Avoid one time clone to save memory peak (#17934) commit 009cd4e Author: RandySheriffH <[email protected]> Date: Fri Oct 20 16:12:21 2023 -0700 Allow cuda custom ops allocate deferred cpu mem (#17893) Expose a new allocator from cuda stream. The allocator manages deferred cpu memory which only get recycled before stream destruction. --------- Co-authored-by: Randy Shuai <[email protected]> commit 2f57625 Author: Chi Lo <[email protected]> Date: Fri Oct 20 22:09:46 2023 +0000 [TensorRT EP] Add stream sync after enqueue (#18026) If the model is partitioned into TRT subgraphs and CUDA EP node, we observed cuda stream synchronization issue when multithreading. Calling stream sync API after enqueue can solve this issue without adding much performance overhead. commit 020824e Author: liqun Fu <[email protected]> Date: Fri Oct 20 15:08:25 2023 -0700 Update ONNX to 1.15.0rc1 (#17914) commit a43c57f Author: Baiju Meswani <[email protected]> Date: Fri Oct 20 11:39:57 2023 -0700 ResizeGrad CUDA/ROCM kernel implementation (#17772) commit cc7e8cc Author: Changming Sun <[email protected]> Date: Fri Oct 20 09:24:21 2023 -0700 Update dockerfiles/Dockerfile.source to avoid installing onnx (#17975) Update dockerfiles/Dockerfile.source to avoid installing onnx python package. ONNX is not listed in https://github.com/microsoft/onnxruntime/blob/main/requirements.txt.in. We do not have to install it. Especially when we do not run tests, the package provides no help when building onnxruntime from source. Resolve #17781 commit 99b8dca Author: Yi Zhang <[email protected]> Date: Fri Oct 20 23:41:40 2023 +0800 Disable dml stage in windows GPU pipeline temporarily. (#18034)

This PR contains fusion-level and kernel-level optimizations for [Meta's LLaMA-2](https://blogs.microsoft.com/blog/2023/07/18/microsoft-and-meta-expand-their-ai-partnership-with-llama-2-on-azure-and-windows/). Some of the added optimizations include: - SimplifiedLayerNorm changes - Fusions for multiple variants - SkipSimplifiedLayerNorm changes - Kernel support for CPU - Rotary embeddings (previously did not exist) - Fusions for multiple variants - CPU and CUDA kernels - Supports interleaving and non-interleaving in the same kernels - Optimized cache that requires half of its originally exported sizes - Reduced from `(max_sequence_length, head_size)` to `(max_sequence_length, head_size / 2)` - Multi-head attention - Support for 2D and 3D attention masks - Group query attention (for FP16 CUDA and INT4 CUDA) - Integration with flash attention v2 and past-present buffer sharing - Removes need for `attention_mask` input as it is supported in the kernel - 4 bit quantization - `block_size` parameter is available for customizing - Support the new changes for [Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Support combinations of the below variants (ex: export ORT version and run with Optimum) Supported variants of LLaMA-2 include: - [ORT version](https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/python/tools/transformers/models/llama) - Produces one ONNX file that is already optimized (and quantized if requested) - Integrates with Optimum - [Another Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Already exported and available off-the-shelf - Faster versions of those models will be uploaded there soon - [Hugging Face version](https://huggingface.co/meta-llama) - Models that end with `-hf` - Some older and current versions of [`transformers`](https://github.com/huggingface/transformers) and [`optimum`](https://github.com/huggingface/optimum) that export the model to ONNX differently - Note that while some older versions are supported, it is recommended to use the latest package versions. To use the optimizations, please see `README.md` for details. Please note the various `requirements.txt` files for the package versions recommended in order to use these changes. To run the ORT transformer optimizer separately, run the script as follows: ``` $ cd onnxruntime/onnxruntime/python/tools/transformers/ $ python3 optimizer.py --input <filename>.onnx --output <filename>.onnx --model_type gpt2 --num_heads <number of attention heads> --hidden_size <attention hidden size> --use_external_data_format --opt_level 0 ``` This PR helps the following issues: - #14997 - #16254 - #17681 - #17925 - microsoft/onnxruntime-inference-examples#320 This PR uses changes from the following PRs: - pytorch/pytorch#104468 - pytorch/pytorch#109759 - #17020 - #17674 - #17890 - #17920 - huggingface/transformers#26162 - huggingface/optimum#1257 - huggingface/optimum#1289 - huggingface/optimum#1462 This PR uses changes from the following issues and PRs to begin supporting the [new TorchDynamo exporter](https://pytorch.org/docs/stable/onnx.html#torchdynamo-based-onnx-exporter): - huggingface/transformers#26307 - pytorch/pytorch#104903 - pytorch/pytorch#105040 - microsoft/onnxscript#847 - microsoft/onnxscript#862 - microsoft/onnxscript#493

Fritskee · 2024-02-20T12:06:01Z

Any updates on this issue?

ogencoglu · 2024-02-24T11:26:36Z

4 bit would indeed be great. Any updates?

### Description This PR contains fusion-level and kernel-level optimizations for [Meta's LLaMA-2](https://blogs.microsoft.com/blog/2023/07/18/microsoft-and-meta-expand-their-ai-partnership-with-llama-2-on-azure-and-windows/). Some of the added optimizations include: - SimplifiedLayerNorm changes - Fusions for multiple variants - SkipSimplifiedLayerNorm changes - Kernel support for CPU - Rotary embeddings (previously did not exist) - Fusions for multiple variants - CPU and CUDA kernels - Supports interleaving and non-interleaving in the same kernels - Optimized cache that requires half of its originally exported sizes - Reduced from `(max_sequence_length, head_size)` to `(max_sequence_length, head_size / 2)` - Multi-head attention - Support for 2D and 3D attention masks - Group query attention (for FP16 CUDA and INT4 CUDA) - Integration with flash attention v2 and past-present buffer sharing - Removes need for `attention_mask` input as it is supported in the kernel - 4 bit quantization - `block_size` parameter is available for customizing - Support the new changes for [Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Support combinations of the below variants (ex: export ORT version and run with Optimum) Supported variants of LLaMA-2 include: - [ORT version](https://github.com/microsoft/onnxruntime/tree/main/onnxruntime/python/tools/transformers/models/llama) - Produces one ONNX file that is already optimized (and quantized if requested) - Integrates with Optimum - [Another Microsoft version](https://github.com/microsoft/Llama-2-Onnx) - Already exported and available off-the-shelf - Faster versions of those models will be uploaded there soon - [Hugging Face version](https://huggingface.co/meta-llama) - Models that end with `-hf` - Some older and current versions of [`transformers`](https://github.com/huggingface/transformers) and [`optimum`](https://github.com/huggingface/optimum) that export the model to ONNX differently - Note that while some older versions are supported, it is recommended to use the latest package versions. ### Usage To use the optimizations, please see `README.md` for details. Please note the various `requirements.txt` files for the package versions recommended in order to use these changes. To run the ORT transformer optimizer separately, run the script as follows: ``` $ cd onnxruntime/onnxruntime/python/tools/transformers/ $ python3 optimizer.py --input <filename>.onnx --output <filename>.onnx --model_type gpt2 --num_heads <number of attention heads> --hidden_size <attention hidden size> --use_external_data_format --opt_level 0 ``` ### Motivation and Context This PR helps the following issues: - microsoft#14997 - microsoft#16254 - microsoft#17681 - microsoft#17925 - microsoft/onnxruntime-inference-examples#320 This PR uses changes from the following PRs: - pytorch/pytorch#104468 - pytorch/pytorch#109759 - microsoft#17020 - microsoft#17674 - microsoft#17890 - microsoft#17920 - huggingface/transformers#26162 - huggingface/optimum#1257 - huggingface/optimum#1289 - huggingface/optimum#1462 ### New TorchDynamo Exporter (experimental stage) This PR uses changes from the following issues and PRs to begin supporting the [new TorchDynamo exporter](https://pytorch.org/docs/stable/onnx.html#torchdynamo-based-onnx-exporter): - huggingface/transformers#26307 - pytorch/pytorch#104903 - pytorch/pytorch#105040 - microsoft/onnxscript#847 - microsoft/onnxscript#862 - microsoft/onnxscript#493

ideasbyjin · 2024-06-17T14:56:36Z

Being able to convert a HF model for 4-bit quantization would be awesome!!

yufenglee · 2024-06-18T17:01:18Z

Being able to convert a HF model for 4-bit quantization would be awesome!!

The QLLM tool can convert a 4-bit HF model to ONNX: https://github.com/wejoncy/QLLM. And a tool from ORT Generate API can also convert it with this PR:microsoft/onnxruntime-genai#600

ideasbyjin · 2024-06-20T13:17:39Z

Thanks, I might be missing something but for my models (which are encoder-only models), I'm not sure how to get it to work. I was able to 4-bit quantize it using BitsAndBytes on HF, but not export it to ONNX

elephantpanda · 2024-07-24T02:09:50Z

Hi I see ONNX is now supporting 4bit data type. Is there any more information about how to make use of these and do quantization down to 4bit?

elephantpanda added the feature request request for unsupported feature or enhancement label Mar 10, 2023

github-actions bot added the quantization issues related to quantization label Mar 10, 2023

jchen351 self-assigned this Mar 10, 2023

jchen351 closed this as not planned Won't fix, can't repro, duplicate, stale Mar 10, 2023

elephantpanda mentioned this issue Mar 22, 2023

how to save 4bit/2bit models? onnx/onnx#4192

Open

jywu-msft reopened this Jul 6, 2023

jywu-msft unassigned jchen351 Jul 6, 2023

kunal-vaishnavi mentioned this issue Oct 18, 2023

LLaMA Model Optimization #18021

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Feature Request] 4bit and 2bit and 1bit quantization support #14997

[Feature Request] 4bit and 2bit and 1bit quantization support #14997

elephantpanda commented Mar 10, 2023

jchen351 commented Mar 10, 2023

jchen351 commented Mar 10, 2023

elephantpanda commented Mar 10, 2023 •

edited

Loading

josephrocca commented Mar 14, 2023

xenova commented Mar 14, 2023

jchen351 commented Mar 14, 2023 •

edited

Loading

xenova commented Mar 14, 2023 •

edited

Loading

josephrocca commented Mar 15, 2023 •

edited

Loading

skyne98 commented Apr 24, 2023

tikikun commented Jun 2, 2023

jywu-msft commented Jul 6, 2023

jywu-msft commented Jul 6, 2023

ThisisBillhe commented Jul 12, 2023

elephantpanda commented Jul 12, 2023

ThisisBillhe commented Jul 12, 2023

dfiru commented Oct 14, 2023

josephrocca commented Oct 15, 2023 •

edited

Loading

dfiru commented Oct 15, 2023

Fritskee commented Feb 20, 2024

ogencoglu commented Feb 24, 2024

ideasbyjin commented Jun 17, 2024

yufenglee commented Jun 18, 2024

ideasbyjin commented Jun 20, 2024

elephantpanda commented Jul 24, 2024

[Feature Request] 4bit and 2bit and 1bit quantization support #14997

[Feature Request] 4bit and 2bit and 1bit quantization support #14997

Comments

elephantpanda commented Mar 10, 2023

Describe the feature request

Describe scenario use case

jchen351 commented Mar 10, 2023

jchen351 commented Mar 10, 2023

elephantpanda commented Mar 10, 2023 • edited Loading

josephrocca commented Mar 14, 2023

xenova commented Mar 14, 2023

jchen351 commented Mar 14, 2023 • edited Loading

xenova commented Mar 14, 2023 • edited Loading

josephrocca commented Mar 15, 2023 • edited Loading

skyne98 commented Apr 24, 2023

tikikun commented Jun 2, 2023

jywu-msft commented Jul 6, 2023

jywu-msft commented Jul 6, 2023

ThisisBillhe commented Jul 12, 2023

elephantpanda commented Jul 12, 2023

ThisisBillhe commented Jul 12, 2023

dfiru commented Oct 14, 2023

josephrocca commented Oct 15, 2023 • edited Loading

dfiru commented Oct 15, 2023

Fritskee commented Feb 20, 2024

ogencoglu commented Feb 24, 2024

ideasbyjin commented Jun 17, 2024

yufenglee commented Jun 18, 2024

ideasbyjin commented Jun 20, 2024

elephantpanda commented Jul 24, 2024

elephantpanda commented Mar 10, 2023 •

edited

Loading

jchen351 commented Mar 14, 2023 •

edited

Loading

xenova commented Mar 14, 2023 •

edited

Loading

josephrocca commented Mar 15, 2023 •

edited

Loading

josephrocca commented Oct 15, 2023 •

edited

Loading