Repository navigation
feat: add dense prediction output foundation - #2545
Merged
Merged
Conversation
The latest MLX pin added backend-specific AsStrided temporary retention, but the ROCm core overlay kept the upstream CUDA-or-Metal dispatch and pulled Metal/Metal.hpp into Linux builds. Select the ROCm device header and command encoder explicitly, then refresh the overlay provenance records. The ROCm overlay gate and a syntax-only compile of the affected translation unit pass. Refs #2349
Add the `predict` CLI and shared dense predictor, interpolation, preprocessing, registry, and file-output contracts so upcoming model ports have one integrated path. Bound image and video decoding, stream safetensors tensor materialization, and safely replace only writer-owned output artifacts. Validate the ROCm compile, focused dense, CLI, and registry tests including real MP4 decoding, clippy with warnings denied, and formatting. Refs #2349
inureyes
added
status:review
Fix the dense family registration guide so it clearly tells operators that `KNOWN_MODEL_TYPES` redirects accidental `mlxcel generate` use to `mlxcel predict`. Validation: - `cargo fmt --all -- --check` - `git diff --check` - Python fixture syntax and referenced documentation paths checked Refs #2349
Keep the PyTorch reference decimals unchanged with a fixture-scoped precision lint allowance, matching the existing scheduler fixtures. CI checks test code in addition to the library and binary, so its wider scope caught these literal-only warnings. Refs #2349
13 tasks done
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
mlxcel predictwith bounded image and video loading, task selection, parsed model options, and a reusable dense-predictor registry.generate, and document the foundation status and model integration contract.65cb1fdd, which fixes ROCm AsStrided backend dispatch in the pinned MLX overlay so the implementation can compile and run on Linux.Test plan
cargo test ... vision::dense --no-runmlxcelbinary with warnings deniedcargo fmt --all -- --checkandgit diff --checkThe full ROCm workspace run completed with 13,299 passed, 9 failed, and 424 ignored tests across 170 targets. The failures are existing ROCm limitations: six native NVFP4 fixtures, two compiled complex-number tests, and one PhiMoE parity tolerance. The PhiMoE relative error is identical (
0.022405015) on this branch and the pre-feature baseline65cb1fdd; the other failures match pre-existing unsupported-backend checks and ROCm JIT implementations. The full suite is not green.Metal runtime validation was unavailable on this ROCm Linux host.
Closes #2349