fix(test): remove dead num_worst_tokens test from internode file

[Syan-Lin]

Problem

test_internode.py included a copy-pasted num_worst_tokens != 0
test block from test_intranode.py. This block was dead code —
Buffer::dispatch asserts num_worst_tokens == 0 for multi-RDMA-rank
internode dispatch (legacy.py:378-379), so the test could never run.
When triggered, it crashed the test suite.

Fix

Remove the block entirely from test_internode.py. Coverage is
preserved in test_intranode.py where the feature is supported.

References

• Fixes [BUG] DeepEPv2 internode num_worst_tokens tests #670
• deep_ep/buffers/legacy.py:354-355: "this flag is for intranode only"
• docs/legacy.md:172: "this flag is for intranode only"

[ds-review-bot]

🔵 suggestion: The internode test always runs multi-RDMA-rank (num_ranks > 8), where this block is now skipped entirely, so it is dead code in this file. The same block in test_intranode.py already covers the supported single-RDMA case. Consider either deleting the block from the internode test, or narrowing it to a documented single-RDMA-rank scenario, and exposing a public buffer.num_rdma_ranks/buffer.is_internode instead of the private buffer.runtime.

_{🤖 v4}

[ds-review-bot]

🤖 ds-review-bot Code Review

v6

The guard matches Buffer.dispatch's existing restriction and prevents the unsupported internode test path. Static compilation and diff checks pass.

v4

Verification/cleanup complete for the num_worst_tokens test in test_internode.py. The internode (multi-RDMA-rank) dispatch hard-asserts num_worst_tokens == 0 (legacy.py:378-379) because the flag is intranode-only. The copy-pasted intranode-only test block was running unconditionally for with_topk, so it crashed CI in every multi-node/multi-RDMA internode run. The applied fix adds the guard with_topk and buffer.runtime.get_num_rdma_ranks() <= 1. I verified the guard exactly mirrors the dispatch's own internode criterion (get_num_rdma_ranks() > 1), the runtime call exists/exposes the same value the dispatch uses, and the intranode feature remains fully exercised by test_intranode.py. No residual/unsupported branch is left running into the assert; state is clean. Overall the change is correct and minimal.

v3

The change adds a buffer.runtime.get_num_rdma_ranks() <= 1 guard to the num_worst_tokens != 0 test block in tests/legacy/test_internode.py, preventing the test from exercising an intranode-only feature under multi-RDMA-rank internode dispatch. This is correct and aligns precisely with the runtime assertion in deep_ep/buffers/legacy.py:379 (assert num_worst_tokens == 0, 'Internode dispatch does not support num_worst_tokens > 0'), which is reached whenever get_num_rdma_ranks() > 1 (see legacy.py:378), and with the documented behavior in deep_ep/buffers/legacy.py:354-355 and docs/legacy.md:172 ("this flag is for intranode only"). The new condition is correctly combined with the existing with_topk gate so the intranode / single-RDMA-rank path continues to run the full test block unchanged, while multi-RDMA-rank internode correctly skips the unsupported feature. An explanatory comment referencing Buffer::dispatch was added, documenting the rationale for future maintainers. The fix is minimal, scoped to a single test file, and has no impact on production code. No correctness, safety, or style concerns were found; the change is approved.

Files reviewed: 1
Issues found: 🔵 1 suggestion
Inline comments posted: 1

[Syan-Lin]

Hi @sphish @LyricZhao, gentle ping on this PR. It's a small test-only cleanup: the num_worst_tokens block in test_internode.py is dead code, since the feature is documented as intranode-only and Buffer::dispatch asserts it to be 0 for multi-RDMA-rank runs. Coverage stays in test_intranode.py, and this would also close #670. The branch is up to date with main. Happy to rebase or adjust anything if needed. Thanks for your time!