Pytorch glow bundle. But I’d like to have Int16 precision also.

Pytorch glow bundle produced by means of -emit-bundle and linked with your executable) is being executed? Aug 22, 2021 · Summary: **Summary** During #5146 the **bundle_with_multiple_entries** was modified to use multiple intermediate layers in order to have a meaningful testing of the activations allocations for multiple functions (previous test did not use activations at all). I tried with all the NLP models in onnx zoo but none of the onnx models compile. I successfully compiled GLOW (commit-id: 5bb41e1) using default triple settings (i. pb Convert tf graph to onnx Hi, I am trying to build Glow with GLOW_WITH_BUNDLES option to create standalone executable bundles and I am following the instructions here: Creating standalone executable bundles. weights file is saved using native endianness (native to the build platform, not necessarily the target platform). onnx -backend=CPU -emit-bundle=<bundle-dir> This generates a library. Jun 23, 2020 · pytorch / glow Public Notifications You must be signed in to change notification settings Fork 688 Star 3. May 1, 2019 · I have tried emitting bundles for targets other than x86_64 to no avail. Basically I can have a full recompilation of the bundle but I need the guarantee that the same model but with different weights and quantization parameters is compiled into the same bundle object (program) but with a different file of weights and quantization parameters such that the bundle object does not need to be re-integrated (re-linked Jan 24, 2021 · I used the lenet_mnist. In particular I’m looking at the GemmTest. Apr 14, 2019 · There are plenty of examples for using the image-classifier builder, but I have not found any examples on using the generic model builder. You could look at one of the main. To run inference on DSP standalone (with HIFI-NN optimizations), generate a Glow bundle using xt-clang compiler version with -use-hifi flag and insert the . Jun 18, 2023 · Hi, i'm having issue when building the glow bundle , as I am getting access denied when trying to download glow model from http://fb-glow-assets. Next, I made another project, and I changed lenet_mnist. It’s a binary blob of the weights. This application note demonstrates the process of creating a simple image recognition neural network written in PyTorch and trained on the CIFAR-10 dataset. the XModelRunner or ImageClassifier What I want is to have Glow emit calls for external functions and have a mechanism which links to the bundle pre-compiled object/binary code. I am following these build steps: https://github. pb model in your guide. The only thing close to it is something called “ @jitmain ”. Compiler for Neural Network hardware accelerators. Lastly the bundle would return errorCode Compiler for Neural Network hardware accelerators. That is, the x-inference-engine is a generic (for the most part) engine that is able to execute inference using bundles produced by Glow (e. weights file Compiler for Neural Network hardware accelerators. cpp, and try to inspect it somehow when loading it in. Use the following command (similar to how @ponnamsairam used it): model-compiler -model=model. When loading the model on Netron, the constant not found by the compiler seems to be the output of a "Where" node (precisely node "Where_186" in the graph), which along with other nodes previous to it, is the result of translating a gather operation from my original Pytorch model to an equivalent set of operations in the ONNX model. The x-inference-engine was conceived and developed as a complement to the XModelRunner -- the latter being a generic model compiler/runner distributed with Glow along with ImageClassifier and the generic ModelLoader. I made the quantization profile using the MNIST training dataset and used it to create quantized bundle Jun 20, 2018 · Hi, I'm trying to run the Resnet 50 example in examples/bundles. Nov 14, 2025 · This blog post will provide a detailed overview of Glow for PyTorch model benchmarks, including fundamental concepts, usage methods, common practices, and best practices. h”, “quantization. More details you can find Jul 16, 2020 · Hi, I compiled a TFLite Flatbuffer file into a bundle executable and got a major increase in inference time. In principle this could be well done manually by the user in the application code when integrating the bundle but linking those dependencies automatically make a better user experience. Compile a bundle for a floating-point model The model-compiler front-end tool is the main Glow tool used to compile ONNX, Caffe2 and TensorFlowLite models into bundles. Implement bundle saving for a partitioned network and the ability for the HostManager to save and load these bundles. pb to bundle using model-compiler tool in Pytorch Glow and put it in the project and ran it. Contribute to rosinality/glow-pytorch development by creating an account on GitHub. (In fact, I also encountered that error if I didn’t delete the line “bundle_with_multiple Oct 16, 2022 · Error message: No node under name gpu_0/data glow Yilinz (YiLin Zhang) October 16, 2022, 8:04am 1 I am a beginner and I was trying to use model-compiler to get the Apr 25, 2021 · Thanks for the reply. Normal build works. However, doing only that does not result in the test/GemmTest binary being built for RISC-V. I am loading the model by giving Glow an init_net. May 16, 2019 · Essentially my cross-compilation tools are further along in versioning than my target runtime environment. For example the TopK operator uses the qsort function which is not part of the LIBJIT/bundle and also not defined in some of the MCU toolchains. The first and the last options are feasible. 1 and cuDNN 7. The lower-level instruction-based address-only IR allows the compiler to perform memory-related optimizations, such as instruction scheduling, static memory allocation and Jun 18, 2020 · Hi, are you interested in ahead-of-time compilation (called creating a “bundle” in Glow terminology), or in JIT compilation? You can use image-classifier for both purposes, or model-compiler for ahead of time. Quantization is controlled for all bundles using the GLOW_QUANTIZE_BUNDLES cmake If I understood @bertmaher 's comments on the PR correctly, the resnet50 bundle is sufficiently recent and complex to serve as a good example of how to make a standalone bundle, and the lenet-mnist bundle is small enough to run in CI. 3k Apr 29, 2019 · nickgg commented on Apr 29, 2019 In the Runtime we may partition a network into multiple sub-graphs intended for different cards, or in the case of heterogeneous partitioning, even different backend kinds. An application code must be written afterwards to use the library code. Jan 6, 2023 · Hi, I am using glow to make an executable bundle for an arm aarch64 platform. model-compiler -model=. When integrated with PyTorch, it allows users to take advantage of the flexibility of PyTorch's dynamic Jan 21, 2022 · The files in question refer to the header (. Testing: This must have been missed because all our backends now create compiled functions with runtime bundles? But I have a private backend that does not use runtime bundle, and thus hit Oct 1, 2019 · now I have my own pytorch and onnx model. Jul 8, 2020 · If you want to run it driven via PyTorch, you’d need to still create a C++ wrapper (see here for how to use the bundle that’s generated) that PyTorch has a bridge to execute on. This I can do successfully and all stages the resulting models work as Compile a bundle for a floating-point model The model-compiler front-end tool is the main Glow tool used to compile ONNX and Caffe2 models into bundles. Both projects have the same code. It show up it give incorrect outputs after some investigation I realize that in in generated weights some layers are missing, After recompilation with NXP version of glow it start Apr 29, 2021 · pytorch / glow Public Notifications You must be signed in to change notification settings Fork 691 Star 3. In particular: What would be the input format to the generic builder? I would guess that I would have to write my own input handler. Mor Currently there is no unit tests or regression tests for Bundle saving and loading, we could probably just write a script that runs the four example bundles lenet_mnist and resnet50 (+ their quanti Aug 29, 2022 · luyaor commented on Aug 29, 2022 When compiling following ONNX model with Glow, the compiler fails with errors. pdf, I can not find the Batch Normalization in the pdf. Glow’s flow blocks consist of 3 components: act norm, 1x1 invertible convolutions and affine coupling layers. So, I think it is worthwhile to mention this somewhere. The motivation for this work is to remove the cost of compile time by allowing the users of Glow to compile the package ahead of time. Next, I made another project, and I changed lenet_mnist. Nov 14, 2025 · In this blog post, we will take a deep dive into a Glow PyTorch walk - through example, covering fundamental concepts, usage methods, common practices, and best practices. g. What could be the reason for such behavior? Device: Raspberry Pi 4 (Cortex-A72) Inference Time: (min, mean,&hellip; Apr 25, 2021 · Thanks for the reply. onnx -emit-bundle/ -bundle-api=dynamic It runs perfectly fine, but when I compiled it after creating the profile, by command . Both projects have the same code. But the inference time takes Mar 24, 2020 · pytorch / glow Public archive Notifications You must be signed in to change notification settings Fork 699 Star 3. (In fact, I also encountered that error if I didn’t delete the line “bundle_with_multiple Oct 16, 2022 · Error message: No node under name gpu_0/data glow Yilinz (YiLin Zhang) October 16, 2022, 8:04am 1 I am a beginner and I was trying to use model-compiler to get the May 9, 2020 · ptrblck May 10, 2020, 7:03am 2 What does model-compiler do and when is it used? Do you know, how kdim is defined and what it stands for? It seems that it’s another placeholder for the kernel and tries to compare the shapes with the weight tensor in the transposed convolution? sheh (Alexander Shekhovtsov) May 10, 2020, 9:34am 3 I’m testing glow compiler, model-compiler creates model bundle Mar 20, 2022 · I followed the Glow installation in GitHub - pytorch/glow: Compiler for Neural Network hardware accelerators and made a special buildwithbundles folder for the Glow build as it is mentioned here glow/AOT. 1 Introduction Glow is a machine learning compiler that accelerates the performance of neural network frameworks on different hardware platforms. weights. Apr 2, 2019 · Trying to use glow to emit bundle for image classifier using resnet50 model, which is converted from tf to onnx. If you're referring to the DeviceInfo passed into the partitioner it's the same idea. how can I quantize it using glow in python API, and then how can I inference it in glow? is there any clear doc? thanks. May 6, 2020 · I’m interested in using the Glow stand-alone bundles for microcontroller applications but I ran into the following memory issue. The user can generate bundles directly from the Python script, without having to first generate ONNX models. /201207_mnist_softmax. onnx -emit-bundle=bundles -backend=CPU However, the Apr 15, 2020 · The Resnet50 bundle for this example was obtained from a Caffe2 model. onnx -dump-profile Mar 22, 2019 · I don’t think there’s a good way to dump the values from the . If the output placeholder is "softmax" the user expects to find the symbol "softmax" in the bundle API. Feb 24, 2021 · I tried to compile the model (resnet50-v2-7. . How to convert the model into onnx format and what are the steps for running onnx model using Glow compiler? Mar 29, 2021 · Ubuntu 20. · Issue #3259 · pytorch/glow · GitHub pytorch / glow Public Notifications Fork 677 Star 3. This is inconvenient. onnx -emit-bundle/ -bundle-api=dynamic It runs perfectly fi Oct 21, 2018 · # Creating standalone executable bundles This document provides a short description about producing ahead-of-time compiled executable bundles. I have tried emitting bundles for targets other than x86_64 to no avail. For instance, other projects tend to allow the user's license to apply to generated code/files by their tool. I’ve been studying the “base. The compiler takes in machine learning frameworks such as PyTorch and Tensorflow and produces optimized code for accelerators. I tried exporting for opset 7 but pytorch is throwing this error: RuntimeError: ONNX export failed on arange, which is not implemented for opset 7. 1 Clang 8. Keep in mind Glow currently has no way to save training results, so you’re limited to training and performing inference in a single process. I believe it needs to be set in your environment if you’re not building the whole project via -DGLOW_WITH_BUNDLES=ON. Glow PyTorch is a powerful tool that plays a significant role in this process. In the document I referenced are also plenty of details in how to integrate the bundle. Additionally there are some examples located here. Jun 21, 2019 · This will require changes to the bundle saver (if, by some off chance, the build machine is big endian, then we have to explicitly save weights in little endian). May 15, 2019 · I’m trying to have tests be created for RISC-V. But I’d like to have Int16 precision also. Glow primarily targets neural network accelerators, though it does have a CPU backend and supports automatic profiling + quantization. Furthermore, I want to check memory usage or memory transfer. png -image-mode=0to1 -model=resnet50-v2-7. So I want to know that how glow solves the bn operator when we use glow to quantize the resnet50 model. Is there something else I should be doing? From the build log I can see that GemmTest gets Description This commit adds the capability to build the ResNet50 standalone bundle using cmake in addition to make. 2k Compiler for Neural Network hardware accelerators. Contribute to mbrukman/pytorch-glow development by creating an account on GitHub. 2k Sep 26, 2018 · Are you using a debug build of glow, by any chance? image-classifier is pretty much unusably slow unless you build in release mode (cmake -DCMAKE_BUILD_TYPE=Release when configuring). The documentation only tells about using the bundle for trained networks, or did I miss something? Oct 20, 2019 · It looks like GLOW_BINARY_DIR is not being set correctly. It might come with a surprise and a headache if the symbol name is actually "softmax__1" (especially in Sep 8, 2022 · pytorch / glow Public archive Notifications You must be signed in to change notification settings Fork 699 Star 3. What version of Opset can I specify in pytorch to make compatible with Glow? I am using Pytorch 1. When I just compiled the model, by command . Perhaps we could use an additional option, say -shared, along with -e When you don't legalize the bundleName_ in LLVMIRGen class then if passed name was wrong then the generated function would not be a valid C/C++ function name. When the test image is inputted, the inference time and the confidence can be confirmed to be output. It may be easier to just say that the . cpp drivers mentioned here, e. s3. To answer your questions, though: Jan 19, 2021 · Hi, I used the lenet_mnist. onnx) which I downloaded from ONNX-ZOO. 2k Mar 21, 2019 · Hi @r3krut, This category is for Glow, which is a different PyTorch backend from Caffe2 (which "natively integrates QNNPACK"). Ther 👍 1 rdzhabarov reacted with thumbs up emoji qcolombet merged commit ffe0e22 into pytorch:master Sep 10, 2018 qcolombet deleted the fix_resnet50_bundle branch Sep 10, 2018 Contributor Glow is used as a software back-end for the PyTorch machine learning framework, including support for the ONNX model format. Dec 12, 2019 · Here we’re using a LSTM model which predicts the sentence by giving input of string, it works fine with torchvision without using glow compiler. You cannot simply add a const keyword because the model entry function expects a non-constant format Jun 23, 2020 · pytorch / glow Public Notifications You must be signed in to change notification settings Fork 688 Star 3. Aug 16, 2021 · Thanks for your prompt reply. com/pytorch/glow Nov 29, 2020 · Hi, We currently develop a Glow backend for CEVA target DSP that supports AOT (offline) compilation mode. Jul 12, 2019 · …ngle bundle (pytorch#3775) Summary: Extend LLVMBackend and BundleSaver to support saving multiple Glow functions into a single bundle. Fixes pytorch#4214 facebook-github-bot closed this as completed in 85d49fd Feb 26, 2020 vdantu pushed a commit to vdantu/glow that referenced this issue Jul 12, 2020 Apr 29, 2021 · pytorch / glow Public Notifications You must be signed in to change notification settings Fork 691 Star 3. 3k Mar 14, 2019 · Do you need the quantized weights during the compilation of a network or when an AOT compiled network (i. o), weights (. Error message: Oct 15, 2019 · Glow Compiler Dependencies: LLVM 8. the Resnet50 bundle’s main. It also elaborates how to deploy it on an i. It has something to do with tensor types. Glow has multiple backends which can run models but not all of them are capable of saving a compiled model. bertmaher force-pushed the bundle branch from 157913c to d9da1d2 Compare Mar 28, 2019 bertmaher merged commit c5d2d7a into pytorch:master Mar 29, 2019 bertmaher deleted the bundle branch Mar 29, 2019 May 22, 2020 · Hi, After visualizing the graph. From what I understand because GemmTest depends on libjit_matmul_f I should be building libjit targeting RISC-V. txt The reason w Sep 26, 2018 · Are you using a debug build of glow, by any chance? image-classifier is pretty much unusably slow unless you build in release mode (cmake -DCMAKE_BUILD_TYPE=Release when configuring). cpp calls an extern function resnet50 (…), but I can’t find such function in the dumped llvm ir. when generati Compile a bundle for a floating-point model The model-compiler front-end tool is the main Glow tool used to compile ONNX, Caffe2 and TensorFlowLite models into bundles. Apr 7, 2021 · There is a section on the README for how to use a bundle here. The loader (perhaps in Loader::runInference) must guard against this to make sure that the Sep 25, 2020 · ONNX graph is bundle compiled. onnx. We need to access this data from the main application on the ARM host before it calls the bundle entry point to start the actual Glow inference process. x86_64). 2k I understand Glow might be more strict than the proto but Glow should try to maintain the placeholder names in the graph since it directly impacts the bundle API and the user experience. When declaring the constant weight array uint8_t constantWeight[CONSTANT_MEM_SIZE] = { #include "MODEL_NAME. 04. /image-classifier cat_285. From this example, you'll notice reshape and transpose nodes between convolution nodes are unnecessary, which will lead to extra computation. bin) files of a user-generated standalone glow-compiled static bundle. Dec 21, 2019 · For compiling an LSTM model try using the model-compiler tool and NOT the model-runner which is not intended for general purpose use. h files in your DSP project in "glow_bundle" folder. The high-level IR allows the optimizer to perform domain-specific optimizations. cpp” modules but I’m a little confused on how to progress. Fixes pytorch#4214 facebook-github-bot closed this as completed in 85d49fd Feb 26, 2020 vdantu pushed a commit to vdantu/glow that referenced this issue Jul 12, 2020 Sep 9, 2018 · Missed adding these in pytorch#1578 I cannot verify things are working as it seems we have a regression here (see pytorch#1604). Even i386 is failing. I verified that pytorch#1578 was not the cause of this regression. This report from NXP explains how to create an executable bundle using arm's cmsis-nn library using the -use-cmsis flag This is solved now by emitting all entry functions at once at the very end of the bundle emission process and thus this replacement is not needed anymore. pb to bundle using model-compiler tool in Pytorch Glow and put it in the project May 24, 2021 · I followed a solution in this post to avoid the “bundle error”. In the first phase, the optimizer performs domain-specific optimizations. So I should either upgrade the target, downgrade my cross-compilation tools (this probably won't work well with Glow), or link statically. Jul 13, 2021 · Hello I’m using Resnet18, Resnet50 as an AOT bundle now. Apr 8, 2021 · pytorch / glow Public Notifications You must be signed in to change notification settings Fork 688 Star 3. Mar 23, 2020 · The build is failing for the "DGLOW_WITH_BUNDLES=ON" option. h” & “quantization. 2 LTS build fails while creating standalone executable bundles (AOT) #5480 May 28, 2019 · I'm trying to emit a bundle for a Neural Network created using caffe2. Dec 26, 2018 · This category is for the Glow neural network accelerator compiler: https://github. o. By the end of this guide, you will have a solid understanding of how to use Glow in PyTorch for your own projects. 2k 6 days ago · PyTorch Foundation is the deep learning community home for the open source PyTorch framework and ecosystem. 3k Dec 7, 2020 · I tried to convert the onnx model with softmax into Bundles. And I changed the model to Bundles using model-compiler tool in eIQ Glow, and I made the project using eIQ SDK. Apr 10, 2022 · [AOT]Model Compiler shows error for tf2onnx converted onnx MobileNetV2 model #5935 We would like to show you a description here but the site won’t allow us. txt" }; the array is not defined as constant. The tool is generic in the sense that it can compile models with any number of inputs or outputs, without being limited to a particular application. o and . Glow is a framework developed by Facebook that focuses on optimizing deep learning models for efficient execution. However it would probably be better/easier to modify BundleSaver. 4 (nightly) version. 1 Anaconda 3 ` Pytorch if GPU is used need to install CUDA 10. Description This commit removes the zfnet512 and vgg19 bundles because they are not really necessary if there are two (resnet50 and lenet) working examples of how to create standalone bundles. e. So maybe we should get rid of the other two, and add cmake support for lenet-mnist? May 11, 2020 · Using a simple fully-connected network (784*128*10) on the MNIST dataset, I tested the quantization method proposed in this document. Is there an existing Compiler for Neural Network hardware accelerators. Glow requires the machine learning model to be in specific formats (that is, ONNX, Caffe2, and TFLite) for conversion Sep 2, 2020 · And with Glow as an extension of PyTorch, it will be even easier to generate bundles. MX RT1060-EVK board using the Glow Library provided by eIQ software environment. Is it possible to do such kind of partitioning with the current implementation of Partitioner? SplitInfinity force-pushed the disable-runtime-bundle-copy branch from f6b2fb3 to 9e4ec9f Compare Jul 23, 2019 SplitInfinity marked this pull request as ready for review Jul 23, 2019 Aug 3, 2021 · Hi, I am trying to use model-compiler to compile NLP model into Ahead Of Time (AOT) compiled executable bundles. Runtime Bundle should not be copiable. Only the bundle is different. pb and and predict_net. If you want to use QNNPACK, I believe all you need to do is make sure your model (e. 2k May 29, 2019 · Currently, there is no way to emit bundles as shared libraries without specifying external llvm compiler. The structure of the model is When running model-runner (or in fact any model builder that uses the loader) in non-bundle-emitting mode, specifying -backend=CPU causes SIGSEGV. As stated in the documentation if you build Glow with the CMake option -DGLOW_WITH_BUNDLES=ON it will build the Resnet50 example automatically (both floating-point and quantized version). weights file out of the box. Look at this example: #ifdef __cplusplus extern "C" { #endif // Bundle memory Feb 23, 2021 · I tried to compile the model (resnet50-v2-7. However, it looks like both the ExternalFunctionCall node and the enhancement of this feature, as described in the link you mentioned, add the possibility to use custom code during execution. My examples/bundles/resnet50/Makefile bundle has Summary Updated the AOT document: make the model-compiler the default tool for producing bundles since it is more generic than the image-classifier tool and also simpler to use (e. 1k Jul 6, 2021 · Glow gets input as a traditional neural network data flow graph from high level frameworks like Tensorflow , Pytorch and it lowers them into two phase strongly- typed Intermediate Representation. Mar 4, 2019 · The performance of Glow is quite impressive to me with a first very basic MNIST train run. So I cannot verify the full bundle creation, however this PR will still be required for the full fix. Glow is a normalizing flow model introduced by OpenAI that uses an invertible generative architecture. Dec 8, 2023 · http://fb-glow-assets. But the destructor reads constants_ to determine if it needs to free, and thus can free a garbage pointer. Try exporting with other opset versions Description: RuntimeBundle has a default constructor but no default initializers for its members. Any help or Nov 14, 2025 · In the realm of deep learning, optimizing model performance and deployment is crucial. predict_net) is using operators such as Jul 24, 2020 · pytorch / glow Public Notifications You must be signed in to change notification settings Fork 684 Star 3. com/pytorch/glow/blob/master/docs Compiler for Neural Network hardware accelerators. Apr 8, 2021 · pytorch / glow Public archive Notifications You must be signed in to change notification settings Fork 697 Star 3. And I changed the model to Bundles using model-compiler tool in eIQ Glow, and I made the project using eIQ SDK. 1 Compiler for Neural Network hardware accelerators. Overview A bundle is a self-contained compiled network model that can be used to execute the model in a standalone mode. md at master · pytorch/glow · GitHub so I’m not entirely sure what I am doing wrong. Jul 27, 2020 · Hello. amazonaws. It successfully generates ninja build file but fail dur Sep 6, 2019 · When creating bundles (AOT mode) we found useful to use an utility function to create calls to external functions (outside of libjit). I suppose different targets expect/support different tensor types; however, I made … Compiler for Neural Network hardware accelerators. I suppose different targets expect/support different tensor Jan 14, 2020 · When building bundles for some targets (e. I'm working with a simple example of taking the stock LSTM from within PyTorch, converting it to ONNX and then to Glow. Steps followed for conversion: Freeze the graph - sample. txt) and bin (. /model-compiler -backend=CPU -model=resnet50-v2-7. com/models/lenet_mnis Compiler for Neural Network hardware accelerators. Except for a few operators like LRN, this schema works fine. May 29, 2019 · In the standalone bundle, the main. Does PyTorch implementation of Glow. Nov 15, 2021 · A few days ago I tried to compile the model we worked on. Glow, or graph lowering, compiler derives its name because it lowers a neural network into a two-phase strongly typed intermediate representation. cpp when you’re generating the . h), object (. Where would I do that? Relating to 2, is there a high level class that I can subclass from to create my own builder How does it work? Glow lowers a traditional neural network dataflow graph into a two-phase strongly-typed intermediate representation (IR). But how can I trigger custom layer parsing during the parsing of the model? For example: I’m trying to compile a bundle from a tensorflow-lite model which contains the Jan 9, 2019 · Fair warning, the Glow team is pretty focused on inference right now, so you’ll likely encounter some rough edges in training. During the bundle creation stage, our backend generates metadata which is required by the target DSP during runtime. pb. 0. Mar 11, 2021 · You can provision them all to the single device and just use one Glow DeviceManager if that's what you mean and a single device is what you want. The code is attached here: createExternalCall. MCUs) I have problems with some of the operators from LIBJIT which are not self-contained. com/models Jun 22, 2021 · Hello everyone, I’m using the quantization-schema=symmetric_with_power2_scale and quantization-precision=Int8 for a new device. Contribute to pytorch/glow development by creating an account on GitHub. What should I do to check? Do I have to implement it directly on the code? Mar 5, 2020 · The main bundle function could have a local variable uint32_t errorCode initialized with 0 and maybe have an option in LLVMIRgen for createCall called errorCheck which does bitwise OR between errorCode and the uint32_t returned by the function (or this could be done unconditionally by createCall for all the libjit functions which return an uint32_t). AOT compilation for sample example (mnist) finished properly, emitting mnix. Now I wanted to do more advanced train runs using CNNs and am wondering, if I now need to implement the CNN by hand in Glow or simply use the image-classifier as described in the glow documentation. tdpfo vquq koa plupah pbtw amq hnfulga qqaojp njho uvaxn phld jjp hrg cqpm chllhjg