mirror of
https://github.com/GraphiteEditor/Graphite.git
synced 2026-10-11 03:50:56 +08:00
Keavon Chambers
a708a54492
* Add rank polymorphism node audit classifying all 271 nodes
* Implement StaticType for Item<T>
* Generate Item and mapped List wire variants for nodes declaring an Item<T> primary input
* Migrate nine nodes to Item element-wise kernels, dissolving the blending trait boilerplate
* Document the Item kernel implementation and staging plan
* Route Item<Vector> through TaggedValue::TypeDefault
* Add executor integration tests covering the Item and List wire variants
* Collapse element-wise Item/List wire pairs to the List form for conversion insertion
* Migrate sixteen vector modifier nodes to Item element-wise kernels
* Migrate Sample Image, Extend Image to Bounds, and Dehaze to Item element-wise kernels
* Fix bevel_with_transform test to actually exercise the transform attribute
* Implement From<T> for Item<T>
* Register PromoteNode rank adapters wrapping bare values into Item wires
* Insert PromoteNode adapters for Item/List wire pair fields in the preprocessor
* Define a real promote node backing the PromoteNode registry identifiers
* Zip ranked Item connectors by frame slot in the mapped element-wise variant
* Register ItemToListNode singleton raise adapters
* Resolve Item wires against List connectors by inserting promotion adapters at construction
* Rank the Offset Points distance connector and prove mixed-rank resolution end-to-end
* Implement Clampable for Item and List wires with per-variant clamp bounds
* Rank the Round Corners radius connector, exercising hard bounds on a ranked wire
* Implement ApplyTransform for Item
* Add Item wire implementations to the Transform node, keeping rank-0 chains rank 0
* Detect element-wise nodes by lazy primary connectors declaring Output = Item
* Convert Transform to an Item kernel with ranked parameters, delivering the broadcast milestone
* Rename Apply Transform to Bake Transform, baking item transforms on Vector, DAffine2, and DVec2
* Promote bare wires onto Item connectors at resolution via WrapItemNode adapters
* Rank the numeric, vector, and boolean parameters across the migrated element-wise nodes
* Rank the enum, integer, and seed parameters, registering their rank adapters via a consolidated macro
* Amend the audit with the DashPattern value type resolution
* Migrate the string family to Item element-wise kernels
* Unwrap Item wires into bare legacy connectors at resolution via UnwrapItemNode adapters
* Shadow owned node parameters in bodies instead of mut in signatures
* Migrate the math family and string measure nodes to Item element-wise kernels
* Convert the comparison and clamp nodes to Item kernels, dropping unreachable &str rows
* Flat-map expander kernels returning List under the mapped variant's frame
* Migrate the expander nodes to Item kernels flat-mapping under the frame
* Remove the unused peel_list helper
* Rank the raster adjustment and blending kernels, recontextualizing shader nodes onto an Item stand-in
Migrate the 16 adjustment nodes, Mix, Color Overlay, and Gradient Map from whole-List kernels to rank-0 Item kernels, letting the macro derive the List-mapped (zip) variants. Move the Adjust and Blend per-element seams off List onto the element types (add the Raster<CPU> impls, drop the now-dead List impls).
Shader nodes keep their bodies verbatim: PerPixelAdjust re-emits the identical kernel against a transparent no_std Item stand-in, so every Item<T> connector and .element() call resolves to a zero-cost identity on the GPU while the uniform buffer stays bare repr(C). The macro peels Item off ranked uniform params, wraps the fetched texel and uniforms at the entry point, and unwraps the result. This drops the shader_node/Item incompatibility guard. Register rank adapters for the adjustment enums.
* Update the rank polymorphism roadmap for the landed shader-node and adjustments chunk
* Rename the GPU Item stand-in to ShaderItem, aliased as Item at its shader-node import sites
* Flip the vector shape generators to emit rank-0 Item<Vector>
The shape generators (Rectangle, Circle, Ellipse, Arc, Spiral, Polygon, Star, Arrow, Line, Grid, QR Code) each produced exactly one shape wrapped in a singleton List<Vector>. Emit Item<Vector> directly so they connect to the rank-0 content connector of the migrated Transform node. Downstream List consumers receive the value through the existing Item to List promotion.
Relax the element-wise validation so a `()` (generator) primary may return Item<T> without being element-wise. Adapt the Repeat on Points test, which still takes a List content connector, by raising the generator's Item output through a singleton wrapper node.
* Parse ranked Item<T> parameter defaults against the bare element type
A ranked `Item<T>` parameter's default value is a bare, unranked `T` (promoted to the wire at resolution), but the preprocessor was handed the wrapped `Item<T>` type and could not parse the literal, flooding the console with warnings and dropping the defaults. Key the field's default_type metadata off the peeled element type for concrete ranked parameters, leaving generic `Item<T>` primaries and skip_impl nodes untouched.
* Parse an element-wise primary's scalar default against the bare element type
An element-wise node's primary reports its default_type as the List wire form so an unconnected primary defaults to an empty list. But when the primary carries a scalar `#[default]` (such as Root's radicand), that literal must parse as a bare element, not a List. Key the primary's default_type off the bare element type when it has a Default value source, keeping the List form otherwise.
* Add the DashPattern value type for stroke dash sequences
Introduce a rank-0 DashPattern value type (a Vec<f64> of alternating dash and gap lengths) so a stroke's dash pattern is a single frameable value rather than a rank-1 List<f64>. Register it as an auto-generated TaggedValue variant, parse its default from a comma or space separated string, and register its rank adapters. Not yet wired into the Stroke node.
* Rank the Fill and Stroke nodes element-wise and give Stroke a DashPattern connector
Migrate Fill and Stroke to element-wise Item<V> primaries (over Vector and Graphic element types) via a new element-level VectorItemMut trait, so styling one shape yields one shape and rank is preserved instead of promoting the input to a singleton List and emitting a List. The macro derives the List-mapped variant for genuine collections.
Wire the Stroke dash sequence to the new rank-0 DashPattern value type, collapsing the old content x paint x dash cartesian and dropping the IntoF64Vec trait. Update the stroke properties dash widget, the drawing tool, and graph-operation plumbing to read and write DashPattern, and migrate legacy F64Array, F64, and String dash inputs on document open.
Assign Colors stays a whole-collection node: each element's gradient position depends on its index among all siblings, which the element frame does not expose, so it keeps its List primary and the VectorListIterMut trait.
* Register rank adapters for the ranked Stroke enum parameters
The element-wise Stroke node ranks its align, cap, and paint order parameters as Item<StrokeAlign>, Item<StrokeCap>, and Item<PaintOrder>, but those enums lacked promotion adapters, so a bare default enum value could not be promoted to its Item wire and no Stroke variant resolved ("No construct found for node"). Register their rank adapters alongside StrokeJoin.
* Display Item wires in the Data panel without a List's ID column
Add a TableItemLayout impl for Item<T> and recognize Item wire types when introspecting graph data. An Item holds a single element, so it renders as a one-row table of the element plus its attributes with no leading index column, and it labels as its element type T rather than a List's T[]. Add ItemAttributeValues::get_any for the attribute widget dispatch.
* Register MonitorNode for Item wire types so the Data panel introspects them directly
Graph introspection wraps the inspected output in a generic MonitorNode typed to the wire. Without Item<T> monitor registrations, an Item<Vector> output could only be monitored after an Item to List promotion, so the Data panel captured and displayed a List<Vector> despite the connector being Item<Vector>. Register monitors for the Item types the element-wise nodes emit, and add the matching Data panel downcast entries.
* Color and double Item/List wires and cleave layer-stack connectors in the node graph
* Route wire color and rank through hidden nodes and refresh them on type changes
* Rework the DashPattern connector conversions with element-wise promotion and an explicit reducer node
* Rank the remaining value, context, aggregation, and transform nodes onto Item<T> wires
* Back DashPattern with a List<f64> so the Data panel can introspect its lengths
* Carry a single Item<T> through varargs so the Read context nodes emit Item<T> not List<T>
* Relax rank validation for aggregation shapes, add element adapters, and match variants by fewest promotions
* Rank the remaining bare and unnecessarily-List connectors across the node catalog
* Add Graphic::None and the FillChoice paint value, making colors and gradients plain values
* Rename GradientStops to Gradient and the legacy Gradient/Fill structs to LegacyGradient/LegacyFill
* Restore generator frame-from-params ranking to the roadmap as a planned stage
* Rename the ranked-field adapter identifier from PromoteNode to FieldAdapterNode to reflect its full contract
* Unload only the wires whose displayed style changed when types update
* Peel wire rank in the editor's semantic type checks so rank-0 layers are recognized
* Restore the whole-List Transform variant so rank-1 content wires resolve again
* Register the Item wire forms for the Memoize and Context Modification infrastructure nodes
* Give every ranked connector a field adapter and add numeric cast variants for legacy wires
* Key a ranked param's type default off its Item wire form when no literal default exists
* Inherit the layer's content value when splicing a node into an empty chain
* Migrate stale List-form TypeDefault inputs to the definition's current default
* Generate the mapped wire variant only when the element-wise node has a frame source
* Let a bare wire feed a List connector via a wrap-raise adapter, costed as two rank steps
* Add a zip companion to the whole-List Transform so ranked List parameters pair per slot
* Add the Sum, Average, Minimum, Maximum, Any, and All list reducers
* Convert the measure family to element-wise Item kernels per the audit classification
* Prefer the bare element value over the Item type default so ranked params keep their widgets
* Rename GradientStopsUI to GradientUI
* Split Fill's optional transform into a _has_transform bool and a ranked _transform matrix
* Rename the migration-only OptionalDAffine2 TaggedValue to LegacyOptionalDAffine2
* Flow byte buffers as Item<Resource> instead of List<u8> across the byte nodes
* Macro-generate the list-content wire variant, retiring the hand-written Transform-zip, Area, and Centroid companions
* Let ()-primary generators take ranked params and frame over them via the mapped variant, ranking Circle's radius
* Rank the vector shape generators' params to Item, adding a rank-aware input grab to the introspection harness
* Rank the value, color, and text generator params to Item
* Rank the raster, web-request, and context-reader generator params to Item
* Fix the repeat and brush test wirings left behind by the param-ranking sweeps
* Delete the vestigial Some, Unwrap Option, and Size Of debug nodes
* Delete the Attach Attribute node, folding its role into Write Attribute
* Add the Filter and Sort list companion nodes
* Guard the removed-definition migration swap target with a test
* Add the Box Corners value type in place of the rectangle corner radius list
* Split Text to Vector's per-glyph mode into a Text to Vector Glyphs node
* Rank the Combine Channels node's channel connectors to Item
* Make Map Points an element-wise node
* Delete the deprecated Upload Texture node
* Update the implementation roadmap to reflect the landed stages
* Let monitor introspection read rank-0 wires, locking in the layer coercion promotion path
* Prefer the rank-0 default when disconnecting a rank-capable input
* Make Path Modify an element-wise node
* Wrap node paths in a NodeIdPath newtype so they flow as a single Item
* Give Item<Raster<CPU>> a default so an unconnected Brush background resolves
* Stop the Brush node from setting layer attributes its paint operation doesn't produce
* Present-gate Flatten Path's adopted layer path like its fill and stroke
* Gate carried layer attributes on static column presence, not runtime values
* Give the remaining graphic Item<T> types a default so unconnected primaries resolve
* Dispatch a ranked param's Properties widget from its rank-0 element type
* Make Extract Transform an element-wise node, restoring the Origins to Polyline body
* Rename Flatten Path to Combine Paths
* Stamp Legacy Layer Extend's adopted layer path as a readable NodeIdPath
* Drop the dead List<u8> and List<NodeId> wire rows
* Rank Flatten Graphic's Fully Flatten toggle to Item
* Update the implementation roadmap with the endgame scope
* Make Combine Paths a reducer that collapses the whole frame into one path
* Stop type-converter nodes from carrying the source's unrelated attributes
* Format the Origins to Polyline regression test
* Wrap the Brush node's trace in a BrushTrace newtype so it flows as one value
* Make Switch a framed element-wise select, bundling whole collections
* Widen and align element-type coverage across the list and graphic nodes
* Register the compiler's cache chain pair for every ranked enum and newtype wire
* Fix wire colors for Passthrough outputs, bundled lists, and bools, and widen list wires
* Represent List wire types structurally with Type::List, replacing name-parsed rank promotion
* Treat scope and data fields as environment, rank scope wires as Item, and feed the render boundary through a context vararg
* Delete the vestigial Clone debug node
* Reinstate Upload Texture as an element-wise node and fix the GPU variants' scope executor and rank adapters
* Rename Combine Paths back to Flatten Path, deferring that rename to its own PR
* Deduplicate the promotion adapter registrations into the field adapter macro
* Rank Write Attribute's value connector to Item<AttributeValueDyn>, retiring the UnwrapItem bridge
* Vertical wire styling
* Store the editor layer path attribute as a bare NodeIdPath, not an Item<NodeIdPath>
* Rank Context Modification's features connector to Item<ContextFeatures>, dropping the dead memoize row
* Rank Path Modify's modification parameter to Item<Box<VectorModification>>
* Rename the field adapter node family to input adapter
* Drop the dead bare scalar rows from Context Modification's implementations list
* Move the dynamic executor's test module into its own file
* Drop the registry's unreachable bare rows for Memoize, the cache chain, and ConvertNode
* Materialize stored TaggedValues as ranked Item wires at the source
* Remove the bare-wire promotion and adapter machinery made dead by ranked value materialization
* Plant the input adapter for List-only inputs, composing position conversion from standard rows
* Consolidate Into/Convert conversions into the input adapter umbrella and rename the rank adapter identifiers
* Fix grouped layers gaining a phantom None stack element from the FillChoice default hijacking every List<Graphic> disconnect
* Enforce ranked node inputs in the macro, rejecting bare wire declarations
* Remove the unit Context => () machinery rows, leaving () purely as the no-primary sentinel
* Add a --signatures rank-audit mode to node-docs for the ranked-wire migration
* Remove the node-docs --signatures rank-audit mode now that ranked wires are enforced
* Migrate legacy no-color values on the Black & White, Color Overlay, and Empty Image color inputs
* Rewrite the element-wise accessor wire type at the primary input, not raw index 0
* Register the cache chain for Resource wires, replacing the lone hand-written Monitor row
* Gate the remaining Raster<GPU> registry rows behind the gpu feature
* Let List<DVec2> wires erase to ListDyn for the attribute reader and element counter
* Rename Extract Element to Item at Index, Count Elements to List Length, and Omit Element to Remove at Index
* Store paint picks as plain color/gradient values, removing the FillChoice value type
* Code review restructuring
* Sort by the consumed sort_key attribute or natural element order, adding the Sort Key node
* Remove the new list-combinator and reducer nodes to defer them to a follow-up PR
* Parse Fill and Stroke color defaults through the paint wire's Graphic element
* Emit ranked implementation-row default types structurally so their element TypeIds survive to default-literal parsing
* Exempt the deliberate no-paint choice from the stale List-form TypeDefault migration
* Migrate the legacy 4-input Fill directly to the split has-transform shape
* Upgrade the demo artwork
* Fix the valid AI review findings: Item eq/hash contract, table-era no-paint migration, quantize List rows, and other smaller issues
* Remove the rank polymorphism working documents
* Hash Item attribute values directly instead of debug-formatting them, speeding up cached evaluation
* Replace the data panel's dead bare-wire downcast arms with full coverage of the ranked monitor row types
* Derive PartialEq for Item now that attributes participate in equality
* Extend the data panel's attribute dispatchers with the newly supported scalar and choice enum types
* Add List monitor rows for the framed numeric conversion outputs so inspecting them resolves, with matching data panel arms
85 lines
11 KiB
Markdown
85 lines
11 KiB
Markdown
+++
|
|
title = "Distributed computing in the Graphene runtime"
|
|
date = 2022-05-12
|
|
|
|
[extra]
|
|
banner = "https://static.graphite.art/content/blog/2022-05-12-distributed-computing-in-the-graphene-runtime__2.avif"
|
|
banner_png = "https://static.graphite.art/content/blog/2022-05-12-distributed-computing-in-the-graphene-runtime__2.png"
|
|
author = "Keavon Chambers"
|
|
summary = "Graphite's 2D editor is built upon Graphene, a node-based editing system for nondestructive design across various data types designed to render artwork faster using multiple machines. The system optimizes execution paths, minimizes latency, and uses a distributed runtime for quick data processing."
|
|
reddit = "https://www.reddit.com/r/graphite/comments/unw45k/blog_post_distributed_computing_in_the_graphene/"
|
|
twitter = "https://twitter.com/GraphiteEditor/status/1524664083554791424"
|
|
+++
|
|
|
|
Graphite is a professional 2D graphics editor for photo editing, image manipulation, graphic design, illustration, data visualization, batch processing, and technical art. It is designed to run on a variety of machines, from mobile hardware like iPads or web browsers on midrange laptops up to beefy workstations with dozens of CPU cores and multiple GPUs.
|
|
|
|
To provide a responsive user experience, its architecture is made to support the use of distributed computation to make up for deficiencies in local compute power. Resulting productivity benefits will scale for users on hardware ranging from low-end mobile devices up to high-end workstations because documents and use cases can grow to great complexity.
|
|
|
|
This article explores the current thinking about the problems and potential engineering solutions involved in Graphite and Graphene for building a high-performance distributed computing runtime environment. We are only just embarking on the graph engine implementation, meaning this post describes theoretical approaches to theoretical challenges. The aim is to shed light on what will need to be built and what we currently believe is the trajectory for this work. We hope this post prompts discussions that evolve the concepts and approaches described herein. If this topic sounds interesting and you have feedback, please [get in touch](/contact).
|
|
|
|
<!-- more -->
|
|
|
|
## Node-based editing
|
|
|
|
A core feature is Graphite's reliance on procedural content generation using a node graph engine called Graphene. In traditional editors like Photoshop and Gimp, certain operations like "blur the image" modify the image pixels permanently, destroying the original (unblurred) image information.
|
|
|
|
Graphite is a *nondestructive* editor. Its approach is to represent the "blur" operation as a step in the creation process. All editing steps made by the user are encoded as operations, such as: import or resize an image, draw with a paintbrush, select an area with a certain color, combine two geometric shapes, etc.
|
|
|
|
Operations are functions that process information and are called *nodes*. For example, the "Blur" node takes an image and a strength value and outputs a blurred version of the image. On the advanced end, machine learning-powered nodes may do things like image synthesis or style transfer. Many nodes perform a wide variety of image editing operations and these are connected together into a directed acyclic graph where the final output is the pixels drawn to the screen or saved to an image file.
|
|
|
|
Many nodes process raster image data, but others work on data types like vector shapes, numbers, strings, colors, and large data tables (like imports from a spreadsheet, database, or CSV file). Some nodes perform visual operations like blur while others modify data, like performing regex matching on strings or sorting tables of information. For example, a CSV file might be imported, cleaned up, processed, then fed into the visual nodes which render it in the form of a chart.
|
|
|
|
Different nodes may take microseconds, milliseconds, seconds, or occasionally even minutes to run. Most should not take more than a few seconds, and those which take appreciable time should run infrequently. During normal operations, hundreds of milliseconds should be the worst case for ordinary nodes that run frequently. Caching is used heavily to minimize the need for recomputation.
|
|
|
|
### Node authorship
|
|
|
|
The nodes that process the node graph data can be computationally expensive. The goal is for the editor to ordinarily run and render (mostly) in real-time in order to be interactive. Because operations are arranged in a directed acyclic graph rather than a sequential list, there is opportunity to run many stages of the computation and rendering in parallel.
|
|
|
|
Nodes are implemented by us as part of a built-in library, and by some users who may choose to write code using a built-in development environment. Nodes can be written in Rust to target the CPU with the Rust compilation toolchain (made conveniently accessible for users). The same CPU Rust code can be reused (with modifications where necessary) for authoring GPU compute shaders via the [rust-gpu](https://github.com/EmbarkStudios/rust-gpu) compiler. This makes it easier to maintain both versions without using separate code files and even languages between targets.
|
|
|
|
### Sandboxing
|
|
|
|
For security and portability, user-authored nodes are compiled into WebAssembly (Wasm) modules and run in a sandbox. Built-in nodes provided with Graphite run natively to avoid the nominal performance penalty of the sandbox. When the entire editor is running in a web browser, all nodes use the browser's Wasm executor. When running in a distributed compute cluster on cloud machines, the infrastructure provider may be able to offer sandboxing to sufficiently address the security concerns of running untrusted code natively.
|
|
|
|
## The Graphene distributed runtime
|
|
|
|
In the product architecture, Graphene is a distributed runtime environment for quickly processing data in the node graph by utilizing a pool of CPU and GPU compute resources available on local and networked machines. Jobs are run where latency, speed, and bandwidth availability will be most likely to provide a responsive user experience.
|
|
|
|
<img src="https://static.graphite.art/content/blog/2022-05-12-distributed-computing-in-the-graphene-runtime/local-and-cloud.avif" onerror="this.onerror = null; this.src = this.src.replace('.avif', '.png')" alt="Graphene in a local and cloud environment" />
|
|
|
|
### Scheduler
|
|
|
|
If users are running offline, their CPU threads and GPU (or multiple GPUs) are assigned jobs by the local Graphene scheduler. If running online, some jobs are performed locally while others are run in the cloud, an on-prem compute cluster, or just a spare computer on the same network. The schedulers generally prioritize keeping quicker, latency-sensitive jobs local to the client machine or LAN while allowing slower, compute-intensive jobs to usually run on the cloud.
|
|
|
|
Each cluster in a locality—such as the local machine, an on-prem render farm, and the cloud data center—runs a scheduler that commands the available compute resources. The multiple schedulers in different localities must cooperatively plan the high-level allocation of resources in their respective domains while avoiding excessive chatter over the network about the minutiae of each job invocation.
|
|
|
|
### Cache manager
|
|
|
|
Working together with the scheduler, the Graphene cache manager stores intermediate node evaluation results and intelligently evicts them when space is limited. If the user changes the node graph, it can reuse the upstream cached data but will need to recompute downstream nodes where the data changed. Memoization can be used to avoid recomputation when an input to a node is the same as received in the past, and that result is available in the cache.
|
|
|
|
### Progressive enhancement
|
|
|
|
The scheduler and cache manager work in lockstep to utilize available compute and cache storage resources on the local machine or cluster in order to minimize latency for the user. Immediate feedback is needed when, for example, drawing with a paintbrush tool.
|
|
|
|
Sometimes, nodes can be run with quality flags. Many nodes (like the paintbrush rasterizer) are implemented using several different algorithms that produce different results trading off speed for quality. This means it runs once with a quick and ugly result, then runs again later to render a higher quality version, and potentially several times subsequently to improve the visual fidelity when the scheduler has spare compute resources and time. Anti-aliasing, for example, will usually pop in to replace aliased renders after a few seconds of waiting.
|
|
|
|
### Batched execution
|
|
|
|
It is important to reduce the overhead between executions. Sometimes, Graphene will predict, or observe during runtime, the frequent execution of code paths (or rather, *node paths*) with significant sequential (not parallel) execution steps. These are good candidates for optimization by reducing the overhead between each execution.
|
|
|
|
In these cases, Graphene will batch multiple sequentially-run nodes by recompiling them such that they are inlined as one execution unit. This can happen for both CPU-based programs and GPU-based compute shaders. This is conceptually similar to how just-in-time (JIT) compilers can predict, or observe at runtime, frequently-run code paths in order to apply compiler optimizations where they are most impactful.
|
|
|
|
### Data locality
|
|
|
|
When dealing with sequential chains of nodes, if they haven't already been recompiled as a batched execution unit, the Graphene scheduler may also frequently prioritize grouping together a set of related nodes to be run together on the same machine for memory and cache locality.
|
|
|
|
Sequential nodes allocated to the same physical hardware can avoid the overhead of copying data over the internet, or between machines in the cluster, or between RAM and VRAM, or from RAM to the CPU cache.
|
|
|
|
Graphene should recognize when a certain intermediate result already lives in RAM or VRAM and prioritize using that CPU or GPU to compute the tasks which rely on that data. But when that's not possible, we need a fast architecture to transfer data between machines. For GPUs, it might be possible to use a DMA (Direct Memory Access) strategy like Microsoft's new DirectStorage API for DirectX and transfer data into or out of VRAM straight to SSDs or networked file systems. Efficiently transferring the final result, and maybe occasionally intermediate cache results, between networks (like the cloud and client) with latency and bandwidth considerations is also important.
|
|
|
|
## Conclusion
|
|
|
|
Presently, we are experimenting with CPU and GPU node composition for the beginnings of the Graphene visual programming language. Most of what was described in this post will likely evolve as we get further into the implementation stage and when we learn more from experts in the fields of computer science that Graphene overlaps with.
|
|
|
|
If you have a background or interest in programming language design, functional programming, ECS and data-oriented design, scheduling, distributed computing, general-purpose GPU compute (GPGPU), or high-performance computing (HPC), we'd love to have your ideas steer our work. Or better yet, join the team to make this dream a reality. We discuss most of our architecture and designs on our [Discord server](https://discord.graphite.art) through text and sometimes voice. Please come say hi!
|