GPU plotting via wgpu (Vulkan/Metal), backed by abundance's proof-of-space - #3883
GPU plotting via wgpu (Vulkan/Metal), backed by abundance's proof-of-space#3883vedhavyas wants to merge 19 commits into
Conversation
|
@BugBot run |
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
✅ Bugbot reviewed your changes and found no new issues!
Comment @cursor review or bugbot run to trigger another review on this PR
Reviewed by Cursor Bugbot for commit 64a7290. Configure here.
Latest nightly that still supports generic_const_exprs.
64a7290 to
816db7f
Compare
|
@BugBot run |
There was a problem hiding this comment.
✅ Bugbot reviewed your changes and found no new issues!
Comment @cursor review or bugbot run to trigger another review on this PR
Reviewed by Cursor Bugbot for commit 816db7f. Configure here.
🛡️ Immunefi PR ReviewsWe noticed that your project isn't set up for automatic code reviews. If you'd like this PR reviewed by the Immunefi team, you can request it manually using the link below: Once submitted, we'll take care of assigning a reviewer and follow up here. |
Vendor abundance's chiapos proof-of-space engine (fc8da99) into shared/ab-proof-of-space as a path dependency, re-pointed from ab_core_primitives to subspace_core_primitives. Only the chiapos tables + PosProofs that ChiaV2Table needs are kept; abundance's Chia/Shim table-trait layer is left out. Adds ab-blake3 + rclite as workspace deps. Binning is abundance's raw first_k_bits; Subspace's little-endian s-bucket convention lands in a later commit.
Vendor abundance's GPU proof-of-space (fc8da99) into shared/ab-proof-of-space-gpu as a path dependency, re-pointed to subspace_core_primitives. Only the rust-gpu shader kernels + the proofs-only host (GpuRecordsEncoderInstance/create_proofs) are kept; abundance's Reed-Solomon RecordsEncoder host, the ab-erasure-coding/ab-farmer-components deps, and the shader kernel tests are left out. subspace-proof-of-space-wgpu is host-only, consuming the vendored shader + proofs-only API and encoding with subspace's KZG scheme, so there is one cargo-gpu shader compile. f7 binning is abundance's raw first_k_bits; Subspace's little-endian convention lands next. Adds spirv-std + wgpu + cargo-gpu-install workspace deps.
Subspace bins proofs into s-buckets little-endian (a bucket's little-endian bytes are the challenge, matching the ChiaTable verifier's challenge_index.to_le_bytes()). Bake it directly into the vendored CPU chiapos create_proof_targets and the GPU find_matches_and_compute_f7 shader, replacing abundance's raw first_k_bits binning. No generic map; the convention is fixed. chiapos self_verification checks the dense binning against the little-endian search (for_s_bucket(cs) == find_proof(cs.to_le_bytes())).
chia_v2::ChiaV2Table wraps the vendored ab-proof-of-space chiapos Tables and builds the full proof table with create_proofs, so the proofs verify under the existing ChiaTable verifier; verification delegates to ChiaTable. The alloc feature gates it so the no_std runtime never links it.
816db7f to
98bd5aa
Compare
|
@BugBot run |
There was a problem hiding this comment.
✅ Bugbot reviewed your changes and found no new issues!
Comment @cursor review or bugbot run to trigger another review on this PR
Reviewed by Cursor Bugbot for commit 98bd5aa. Configure here.
|
@BugBot run |
There was a problem hiding this comment.
✅ Bugbot reviewed your changes and found no new issues!
Comment @cursor review or bugbot run to trigger another review on this PR
Reviewed by Cursor Bugbot for commit 6f0a848. Configure here.
0e1f9ca to
d3f0043
Compare
|
@BugBot run |
PlotMetadataHeader gains a cutover field (Option<HistorySize>) and a typed PlotMetadataVersion (V0/V1); a version-aware Decode reads old V0 headers as cutover = None and rejects unknown versions (the old-binary lockout). Opening a V0 plot records the cutover as the highest history size among existing sectors and rewrites the header as V1. New-sector plotting then waits until history advances past the cutover, so no sector is stamped at history_size <= cutover and later misread with the old proof-of-space. Sectors at or below the cutover use the old proof-of-space, newer ones the new; nothing dispatches on it yet.
PosTable (subspace-proof-of-space, alloc-gated) dispatches between the old ChiaTable and the new ChiaV2Table; both proofs verify under ChiaTable, so the choice is a farmer-side read/prove concern. The Table trait gains generator_for(is_post_cutover) and SolutionCandidates a history_size() accessor; the farmer picks per sector by comparing history_size against the farm's cutover at reading, proving, and auditing. New sectors always use the new implementation.
Cross-verify abundance-backed proofs against the existing ChiaTable verifier, a plot->read round-trip, and the per-sector V1/V2 dispatch selection. Point the benches at the new table.
WgpuRecordsEncoder wires the vendored wgpu proof-of-space into the farmer behind a wgpu feature (Vulkan/Metal, macOS included): the host-side WgpuDevice runs the little-endian f7 shader and does subspace KZG encoding, with the plotting setup shared by the farm and cluster-plotter commands.
Plot the same sector with wgpu and with CpuRecordsEncoder<ChiaV2Table> and assert the sector bytes are identical, keeping the GPU path consensus-compatible with the CPU read/verify path.
wgpu covers all GPU vendors (SPIR-V + Metal/Vulkan) with CPU fallback, so the vendor-specific CUDA (sppark) and ROCm plotters are redundant. Drops the subspace-proof-of-space-gpu crate, the cuda/rocm farmer plotters, their features, and the CUDA/ROCm CI, release, and Docker steps; wgpu becomes a default feature so one farmer binary per platform ships with GPU built in. CI moves over to the wgpu build: the macOS CI and release runners go to 26, and lavapipe (software Vulkan) is installed on the Linux and Windows runners so the byte-identity test runs there instead of skipping. There is a single cargo-gpu shader compile, so no rust-gpu toolchain pre-install step is needed.
Pick plotting devices by type instead of using every enumerated GPU: discrete GPUs, else integrated, else a virtual GPU as a last resort; unknown and CPU-emulated adapters are skipped. --wgpu-gpus still selects exact GPUs by device id and bypasses the tiering, and the chosen GPUs are logged.
c2c8ae0 to
3021641
Compare
|
@BugBot run |
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.
Comment @cursor review or bugbot run to trigger another review on this PR
Reviewed by Cursor Bugbot for commit 3021641. Configure here.
|
|
||
| if let Some(error) = self.mapping_error.lock().take() { | ||
| return Err(RecordEncodingError::BufferMapping(error)); | ||
| } |
There was a problem hiding this comment.
GPU error permanently stalls plotting
High Severity
A transient failure in create_proofs leaves tainted set, so that queue returns DeviceBroken forever. WgpuRecordsEncoder runs every queue via spawn_broadcast and fails the whole sector if any queue errors, so one dead queue breaks all later encodes. With GPU present, CPU plotting is auto-disabled, leaving no fallback and retries that cannot recover.
Additional Locations (2)
Reviewed by Cursor Bugbot for commit 3021641. Configure here.
nazar-pc
left a comment
nazar-pc
left a comment
There was a problem hiding this comment.
You should update subspace.plotter.*.plot-sector to something like subspace.plotter.*.plot-sector-v2 in
Resolve Cargo.toml and Cargo.lock against the prometheus-client 0.25.0 bump and the advanced libp2p and polkadot-sdk fork pins, keeping the wgpu dependencies from this branch.
The plot-sector payload is unchanged, so during a partial cluster upgrade an old plotter would decode a new farmer's request and plot the sector with the old proof-of-space. Versioning the subject makes that routing impossible, and versioning free-instance keeps a new farmer from discovering old plotters at all.
2 participants
This brings GPU plotting to the farmer through wgpu: one farmer binary plots on Vulkan and Metal (macOS included) with a CPU fallback. It retires the vendor-specific CUDA and ROCm plotters, which wgpu replaces.
The proof-of-space comes from abundance, vendored into the tree rather than pulled as a git dependency - its CPU and GPU proof-of-space crates are copied into
shared/, re-pointed at our own core primitives, and Subspace's little-endian s-bucket convention is baked into the binning so new proofs verify under the existingChiaTableverifier with no consensus change. This is the vendoring route we agreed on over the git-dep in nazar-pc/abundance#724 (now closed). Supersedes #3878, the earlier CPU-only approach.Commits in order, worth reading one by one:
Bump nightly to 2026-05-03: the toolchain bump the rest needs.vendor abundance ab-proof-of-space (cpu) at fc8da99: abundance's CPU chiapos re-pointed to our primitives, trimmed to what we use.vendor abundance ab-proof-of-space-gpu (gpu) at fc8da99: the rust-gpu shaders plus a proofs-only host, same re-point, one shader crate so there's a single cargo-gpu compile.bake little-endian s-bucket binning into vendored proof of space: our s-bucket convention baked into the CPU binning and the f7 shader so proofs land whereChiaTableexpects them.add abundance-backed proof of space: a denseChiaV2Tableon top of the vendored crate, verified by the existingChiaTableverifier.add farm cutover point with in-place v0 upgrade: an existing farm upgrades in place at a cutover point; old sectors keep working, new ones plot with the new table.dispatch proof-of-space per sector by history size: pick the old or new table per sector, at the cutover boundary.add tests and benches for the abundance proof of space.add wgpu GPU plotter backed by abundance proof of space: the wgpu plotter wired into the farmer behind awgpufeature, encoding records with our KZG scheme.add wgpu byte-identity test against CPU ChiaV2Table: a GPU-plotted sector is byte-for-byte identical to a CPU-plotted one, so both read and verify under the same path.remove CUDA/ROCm + sppark; make wgpu the default GPU plotter: drop the vendor plotters and their CI, release, and Docker steps; wgpu becomes the default. CI moves over to the wgpu build, with the macOS runners (CI and release) on 26 and lavapipe on Linux and Windows so the byte-identity test runs there.add --cpu-only flag to skip GPU plotting.GPU plotting is verified on Apple Silicon M1; M2/M3/M4 aren't validated yet and may need further work.
Code contributor checklist: