ci: do full build [skip tests]

Author: avik-pal

.buildkite/cuda_tutorials.yml

@@ -24,6 +24,5 @@ steps:
         timeout_in_minutes: 120
 
 env:
-  LUX_DOCS_DRAFT_BUILD: true # FIXME: remove before merging
   DATADEPS_ALWAYS_ACCEPT: true
   GKSwstype: "100" # https://discourse.julialang.org/t/generation-of-documentation-fails-qt-qpa-xcb-could-not-connect-to-display/60988

.github/workflows/Documentation.yml

@@ -96,7 +96,6 @@ jobs:
       - name: Run Tutorials
         run: julia --color=yes --project=docs --threads=auto docs/tutorials.jl
         env:
-          LUX_DOCS_DRAFT_BUILD: true # FIXME: remove before merging
           TUTORIAL_BACKEND_GROUP: "CPU"
           BUILDKITE_PARALLEL_JOB_COUNT: 4
           BUILDKITE_PARALLEL_JOB: ${{ matrix.group }}
@@ -122,6 +121,8 @@ jobs:
     needs: [tutorial-cpu]
     steps:
       - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
       - name: Collect Workflow Telemetry
         uses: catchpoint/workflow-telemetry-action@v2
         with:

docs/make.jl

@@ -31,7 +31,7 @@ makedocs(;
         repo="github.com/LuxDL/Lux.jl",
         devbranch="main",
         devurl="dev",
-        deploy_url="https://lux.csail.mit.edu"
+        deploy_url="https://lux.csail.mit.edu",
     ),
     plugins=[CitationBibliography(joinpath(@__DIR__, "references.bib"))],
     draft=DRAFT_MODE,

docs/package.json

@@ -5,7 +5,7 @@
     "@types/node": "^22.13.4",
     "markdown-it": "^14.1.0",
     "markdown-it-mathjax3": "^4.3.2",
-    "vitepress": "^1.6.3",
+    "vitepress": "^1.6.4",
     "vitepress-plugin-tabs": "^0.6.0"
   },
   "scripts": {

docs/src/.vitepress/config.mts

@@ -492,5 +492,7 @@ export default defineConfig({
         timeStyle: "medium",
       },
     },
+    metaChunk: true,
+    mpa: true,
   },
 });

docs/src/introduction/overview.md

@@ -12,61 +12,59 @@ it both compiler and autodiff friendly.
 
 Lux.jl takes a **Reactant-first approach** to deliver exceptional performance and seamless deployment capabilities:
 
-* **XLA Compilation** -- Lux models compile to highly optimized XLA code via [Reactant.jl](https://github.com/EnzymeAD/Reactant.jl), delivering significant speedups on CPU, GPU, and TPU.
+- **XLA Compilation** -- Lux models compile to highly optimized XLA code via [Reactant.jl](https://github.com/EnzymeAD/Reactant.jl), delivering significant speedups on CPU, GPU, and TPU.
 
-* **Cross-Platform Performance** -- Run the same Lux model with optimal performance across different hardware backends (CPU, NVIDIA GPUs, AMD GPUs, TPUs) without code changes, simply by switching the Reactant backend.
+- **Cross-Platform Performance** -- Run the same Lux model with optimal performance across different hardware backends (CPU, NVIDIA GPUs, AMD GPUs, TPUs) without code changes, simply by switching the Reactant backend.
 
-* **Production Deployment** -- Compiled models can be exported and deployed to production servers and edge devices by leveraging the rich TensorFlow ecosystem, making Lux suitable for real-world applications.
+- **Production Deployment** -- Compiled models can be exported and deployed to production servers and edge devices by leveraging the rich TensorFlow ecosystem, making Lux suitable for real-world applications.
 
-* **Large Model Support** -- With Reactant compilation, Lux now excels at training very large models that were previously challenging, making it competitive with other frameworks for large-scale deep learning.
+- **Large Model Support** -- With Reactant compilation, Lux now excels at training very large models that were previously challenging, making it competitive with other frameworks for large-scale deep learning.
 
 ## Design Principles
 
-* **Layers must be immutable** -- cannot store any parameter/state but rather store the
+- **Layers must be immutable** -- cannot store any parameter/state but rather store the
   information to construct them
-* **Layers are pure functions**
-* **Layers return a Tuple containing the result and the updated state**
-* **Given same inputs the outputs must be same** -- yes this must hold true even for
+- **Layers are pure functions**
+- **Layers return a Tuple containing the result and the updated state**
+- **Given same inputs the outputs must be same** -- yes this must hold true even for
   stochastic functions. Randomness must be controlled using `rng`s passed in the state.
-* **Easily extensible**
-* **Extensive Testing** -- All layers and features are tested across all supported AD
+- **Easily extensible**
+- **Extensive Testing** -- All layers and features are tested across all supported AD
   backends across all supported hardware backends.
 
 ## Why use Lux over Flux?
 
-* **High-Performance XLA Compilation** -- Lux's Reactant-first approach enables XLA compilation for dramatic performance improvements across CPU, GPU, and TPU. Models compile to highly optimized code that eliminates Julia overhead and leverages hardware-specific optimizations.
+- **High-Performance XLA Compilation** -- Lux's Reactant-first approach enables XLA compilation for dramatic performance improvements across CPU, GPU, and TPU. Models compile to highly optimized code that eliminates Julia overhead and leverages hardware-specific optimizations.
 
-* **Production-Ready Deployment** -- Deploy Lux models to production environments using the mature TensorFlow ecosystem. Compiled models can be exported and run on servers, edge devices, and mobile platforms.
+- **Production-Ready Deployment** -- Deploy Lux models to production environments using the mature TensorFlow ecosystem. Compiled models can be exported and run on servers, edge devices, and mobile platforms.
 
-* **Neural Networks for SciML**: For SciML Applications (Neural ODEs, Deep Equilibrium
+- **Neural Networks for SciML**: For SciML Applications (Neural ODEs, Deep Equilibrium
   Models) solvers typically expect a monolithic parameter vector. Flux enables this via its
   `destructure` mechanism, but `destructure` comes with various
   [edge cases and limitations](https://fluxml.ai/Optimisers.jl/dev/api/#Optimisers.destructure). Lux
   forces users to make an explicit distinction between state variables and parameter
   variables to avoid these issues. Also, it comes battery-included for distributed training.
 
-* **Sensible display of Custom Layers** -- Ever wanted to see Pytorch like Network printouts
+- **Sensible display of Custom Layers** -- Ever wanted to see Pytorch like Network printouts
   or wondered how to extend the pretty printing of Flux's layers? Lux handles all of that
   by default.
 
-* **Truly immutable models** - No *unexpected internal mutations* since all layers are
-  implemented as pure functions. All layers are also *deterministic* given the parameters
+- **Truly immutable models** - No _unexpected internal mutations_ since all layers are
+  implemented as pure functions. All layers are also _deterministic_ given the parameters
   and state: if a layer is supposed to be stochastic (say [`Lux.Dropout`](@ref)), the state
   must contain a seed which is then updated after the function call.
 
-* **Easy Parameter Manipulation** -- By separating parameter data and layer structures,
+- **Easy Parameter Manipulation** -- By separating parameter data and layer structures,
   Lux makes implementing [`WeightNorm`](@ref), `SpectralNorm`, etc. downright trivial.
   Without this separation, it is much harder to pass such parameters around without
   mutations which AD systems don't like.
 
-* **Wider AD Support** -- Lux has extensive support for most
+- **Wider AD Support** -- Lux has extensive support for most
   [AD systems in julia](@ref autodiff-lux), while Flux is mostly tied to Zygote (with some
   initial support for Enzyme).
 
-* **Optimized for All Model Sizes** -- Whether you're working with small prototypes or large production models, Lux delivers optimal performance. For the smallest networks where minimal overhead is critical, you can use [`ToSimpleChainsAdaptor`](@ref) to leverage SimpleChains.jl's specialized CPU optimizations.
+- **Optimized for All Model Sizes** -- Whether you're working with small prototypes or large production models, Lux delivers optimal performance. For the smallest networks where minimal overhead is critical, you can use [`ToSimpleChainsAdaptor`](@ref) to leverage SimpleChains.jl's specialized CPU optimizations.
 
-* **Reliability** -- We have learned from the mistakes of the past with Flux and everything
+- **Reliability** -- We have learned from the mistakes of the past with Flux and everything
   in our core framework is extensively tested, along with downstream CI to ensure that
   everything works as expected.
-
-

docs/src/introduction/resources.md

@@ -1,10 +1,10 @@
 # Resources to Get Started
 
-* Go through the [Quickstart Example](@ref Quickstart).
-* Read the introductory tutorials on
+- Go through the [Quickstart Example](@ref Quickstart).
+- Read the introductory tutorials on
   [Julia](https://jump.dev/JuMP.jl/stable/tutorials/getting_started/getting_started_with_julia)
   and Lux.
-* Go through the examples sorted based on their complexity in the documentation.
+- Go through the examples sorted based on their complexity in the documentation.
 
 !!! tip "Have More Questions?"
 

docs/src/tutorials/index.md

@@ -35,7 +35,7 @@ const beginner = [
   },
   {
     href: "https://luxdl.github.io/Boltz.jl/stable/tutorials/1_GettingStarted",
-    src: "https://production-media.paperswithcode.com/datasets/ImageNet-0000000008-f2e87edd_Y0fT5zg.jpg",
+    src: "https://blog.roboflow.com/content/images/2021/06/image-18.png",
     caption: "Pre-Built Deep Learning Models",
     desc: "Use Boltz.jl to load pre-built DL and SciML models."
   }