Add source-window collective materialization#192
Conversation
Address local-mapped materialization planning regressions and keep executable byte-range maps available for binding realization fallback paths. Add source-window collective design and validation coverage, propagate the new daemon/runtime options, and include focused C++ and Python tests for the planner and binding execution paths.
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
| "normalize_binding_realization_plan", | ||
| "plan", | ||
| "from_disk", | ||
| "from_resolved_public_disk_source", |
wolegechu
changed the title
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fa00f614fd
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if (!physical_source_table_or.ok()) { | ||
| return to_grpc_status(physical_source_table_or.status()); |
There was a problem hiding this comment.
When parsing physical_source_index_table_future fails, this early return leaves the key inserted in execution_template_inflight by the builder path and never signals cache.cv. Any concurrent or later request with the same execution-template cache key will miss the cache, see the stale inflight marker, and wait indefinitely in the loop above instead of retrying or surfacing the parse error; the later error paths already erase/signal, so this one needs the same cleanup.
Useful? React with 👍 / 👎.
Signed-off-by: wolegechu <ychu.luo@gmail.com>
Signed-off-by: wolegechu <ychu.luo@gmail.com>
| raise RuntimeError( | ||
| "TensorCast retained binding response has no mem_handle" | ||
| ) | ||
| except ValueError: |
| import pytest | ||
| import torch | ||
|
|
||
| import tensorcast.artifact_runtime.binding.retained as retained_binding_module |
1 participant
Summary
cudaMemcpy3DParmsidentifier so spelling checks accept the official CUDA type name.Impact
This makes source-window collective materialization a first-class automatic strategy for eligible disk materialization flows while preserving explicit config knobs for selection and fallback behavior.
Validation
Notes
MODULE.bazelandthird_party/libtorch/BUILDwere intentionally left out of the commits and remain local uncommitted changes.