[WIP] support forge agent#52
Closed
jiaqiang-dot-liu wants to merge 20 commits into
Closed
Conversation
jiaqiang-dot-liu
requested review from
irvineoy,
sharareh-y and
sinarafati-amd
as code owners
July 2, 2026 02:16
…rnel resolution - New `image_kernel` task type: optimize a kernel that already ships inside the container image. setup_workspace seeds a per-run working copy from `image_repo_path` (reflink CoW, `.git` excluded) instead of git-cloning an upstream repo, so in-image submodules / prebuilt build cache / CK headers are reused directly (no clone, no submodule fetch, no network). - prompt_builder: wire image_kernel into the task-type chain, the repository_language knowledge selection, and the workspace-info section. - forge (agents/forge/launch_agent.py): resolve the kernel file robustly (workspace-relative -> repo subdir -> unique suffix match) so repository / image_kernel tasks whose sources live under a repo subdir work; fully backward compatible with legacy workspace-root snippets. - cheatsheet: map MI325X / MI325 -> gfx942 (reuse MI300X architecture doc). - tests: pure-python repro + regression tests (no GPU/network) for all of the above. Co-authored-by: Cursor <cursoragent@cursor.com>
…not capped at 8h bootstrap patches the run's timeout_seconds (KA_TIMEOUT_SECONDS) onto the forge agent_config but leaves max_hours at its static default (8), so the forge-loop's own time budget stopped a 32h run after ~8h (~25-110 iters) regardless of the configured max_iters/timeout. Derive --max-hours from timeout_seconds (with a 15-min margin under the hard process-wait kill) so the loop budget tracks the run's timeout. Default timeout (29700s) still yields ~8h (backward compatible). Adds a unit test for the derivation. Co-authored-by: Cursor <cursoragent@cursor.com>
jiaqiang-dot-liu
force-pushed
the
feat/add-forge-agent
branch
from
July 6, 2026 09:29
3a4d871 to
e020eb0
Compare
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…rrectness Co-authored-by: Cursor <cursoragent@cursor.com>
- add triton2flydsl softmax/rmsnorm/layernorm tasks (torch2hip-style 4-file layout) with a self-contained BYOD measurement driver that is the correctness oracle + performance baseline and defines the op I/O - route *2flydsl to KernelForge forge-rewrite (is_flydsl_rewrite): port the source to FlyDSL, then forge-loop optimizes it; shapes from the driver, candidate from target_file_path - score the FINAL kernel via the task driver like torch2hip (write_rewrite_task_result), not an agent result JSON - harden the forge subprocess: own process group + timeout group-kill + workspace orphan reap (shared _run_streamed) to avoid GPU leaks Co-authored-by: Cursor <cursoragent@cursor.com>
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
2 participants
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.
No description provided.