test: cover Compressor overlap and non-overlap windowing (PRPUNDIT-25)

[jiagaoxiang]

This PR was created by dougljia via Test Gap Resolver.

Summary

• Adds a dedicated CPU behavioral test for Compressor.forward in tests/unit_tests/backends/megatron/test_compressor_forward.py, not the FP32-policy suite (test_v4_keep_in_fp32.py) and not the fused-pool kernel suite (test_compressor_pool.py).
• Independent manual oracle for CSA overlap (ratio=4) and HCA non-overlap (ratio=128): window i is [half_a[i], half_b[i-1]] with zeros on window 0, or a contiguous ratio slice, then eager softmax-pool + RMSNorm.
• Also covers fused vs unfused projections (PRIMUS_COMPRESS_FUSE_PROJ) and the sequence-not-divisible-by-ratio guard, per linxwang's PRPUNDIT-25 note.

Closes test gap PRPUNDIT-25.

Test plan

• PYTHONPATH=. pytest tests/unit_tests/backends/megatron/test_compressor_forward.py (8 passed, CPU)
• black --check on the new file

[Copilot]

🟢 Approval recommended

The PR is test-only and the new tests accurately mirror the production windowing/pooling/RMSNorm logic while adding concrete coverage for overlap vs non-overlap behavior and a key input-shape guard.

Pull request overview

Adds a focused CPU unit test suite that validates primus.backends.megatron.core.transformer.compressor.Compressor.forward behavior for both CSA overlap (ratio=4) and HCA non-overlap (ratio=128) by comparing against an independent, hand-derived windowing + softmax-pool + RMSNorm oracle.

Changes:

• Introduces tests/unit_tests/backends/megatron/test_compressor_forward.py with a manual reference implementation for overlap and non-overlap window construction.
• Adds behavioral assertions that (a) forward matches the independent oracle, (b) overlap window 0 uses causal zero padding (no predecessor leakage), and (c) contiguous-mode forward asserts when S % ratio != 0.

File summaries

File	Description
tests/unit_tests/backends/megatron/test_compressor_forward.py	New CPU behavioral tests that independently validate overlap/non-overlap windowing semantics and forward output against a manual oracle.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟢 Approval recommended

The new tests are consistent with the current Compressor implementation and meaningfully close the stated behavioral test gap without introducing risky production changes.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0 new
• Review effort level: Lite

[jiagaoxiang]

TestIntel PR Steward: Checked the build-docker (3.12) failure on head 971757f (run 33779247129).

The image build itself completes fine (all layers cached, image tagged docker.io/tasimage/primus:pr-1085). The job then fails on:

flag needs an argument: 'p' in -p
Usage:  docker login [OPTIONS] [SERVER]
##[error]Process completed with exit code 125.

This comes from the docker login -u rocmshared -p ${{ secrets.ROCM_DOCKER_HUB_TOKEN }} step in .github/workflows/ci.yaml. Since this PR's head branch runs from a fork (jiagaoxiang/Primus), ROCM_DOCKER_HUB_TOKEN isn't available to the pull_request workflow run (GitHub withholds repo secrets from fork PRs), so -p gets an empty value and docker login errors out.

This is a repo-wide CI/secrets-config issue, not something introduced by this test-only change (the diff only adds tests/unit_tests/backends/megatron/test_compressor_forward.py), and it isn't a transient flake — rerunning would fail the same way for any fork PR. I'm leaving it as-is rather than touching the workflow (out of scope for a test PR) or repeatedly rerunning a job that can't succeed under these permissions. A maintainer will need to either scope that login step to non-fork runs or make the Docker Hub push conditional.

The two Copilot reviews are both "approval recommended" with zero comments, so there's nothing to address there. Still waiting on a human approval and a passing/waived build-docker check before this is mergeable.

[jiagaoxiang]

TestIntel PR Steward: Checked in on this PR after the poll flagged failed CI + reviews.

• CI (build-docker (3.12), run 33779247129): the job fails at the docker login -u tasimage -p step with flag needs an argument: 'p' in -p — the password argument is empty, so the Docker Hub credential/secret isn't reaching this step. This happens after the image itself builds successfully (all layers CACHED/built fine), so it's a CI credential/secret plumbing issue on the runner side, not something caused by this PR's change (a single new test file, tests/unit_tests/backends/megatron/test_compressor_forward.py). This looks like repo-wide CI infrastructure breakage rather than a PR-caused failure or a simple flake, so I'm not rerunning it or touching CI config — a maintainer with access to the Docker Hub secret/runner config should look at it.
• Reviews: both reviews so far are from copilot-pull-request-reviewer[bot] (🟢 approval recommended, 0 line comments on either pass) — no findings to address, and bot approval doesn't count toward mergeability. Still waiting on a human review.

No code changes needed on my end right now. Will keep watching for a human review and for the build-docker check to be fixed/rerun by a maintainer.

[jiagaoxiang]

Superseded by same-repo PR #1098 (head AMD-AGI:testgap/compressor-forward) so Primus-CI-TAS receives Docker Hub credentials and can run torch unit tests. Closing this fork PR.