[Test] Add get_swa_key_page_size to the Q8KV8 sparse-prefill fake KV pool

[nvpohanh]

[by Claude Code]

Motivation

#42128 changed DeepseekV4AttnBackend to read the SWA page size through token_to_kv_pool.get_swa_key_page_size() instead of token_to_kv_pool.swa_kv_pool.page_size. The fake _TokenToKVPool in test/registered/kernels/ops/attention/test_q8kv8_sparse_prefill_backend.py does not define that method. Since #42128 merged, four tests fail on main in the jit-kernel-unit-test-1-gpu-large shard:

AttributeError: '_TokenToKVPool' object has no attribute 'get_swa_key_page_size'. Did you mean: 'get_extra_key_page_size'?
  python/sglang/srt/layers/attention/deepseek_v4_backend.py:2423 (_build_sparse_prefill_chunk_cache)

• test_q8kv8_sparse_prefill_helper_builds_fp8_workspace_matching_bf16_path[0|4|128]
• test_q8kv8_sparse_prefill_real_kernel_matches_bf16_sparse_path

This breaks every PR that runs the jit-kernel suite, regardless of what the PR changes. Examples:

• https://github.com/sgl-project/sglang/actions/runs/36957704232/job/110725022663
• https://github.com/sgl-project/sglang/actions/runs/36972658154/job/110733999900
• https://github.com/sgl-project/sglang/actions/runs/36963048488/job/110700915381
• https://github.com/sgl-project/sglang/actions/runs/36959771934/job/110691037440

Also, the shard stops at the first failing file, so the tests scheduled after this one in that shard never run.

Modifications

Add get_swa_key_page_size() to the fake pool. It returns swa_kv_pool.page_size, matching the real DeepSeekV4TokenToKVPool when no request window is configured. This is a 3-line, test-only change with no production code changes.

This is the same change as commit 94d21e3 in #41251, split out so main is unblocked without waiting for that larger perf PR. The hunk is identical, so #41251 should merge cleanly on top of it.

Test Escape Postmortem

Summary: #42128 broke this test on main, but its own PR Test Base run passed. The suite that contains this test never ran on #42128, and the same gap let #38954 break the same fake pool in September.

Timeline (UTC)

Time	Event
2026-10-01 21:28	#42128's PR Test Base run passes (run 36928953138). Its check-changes step logs Filter jit_kernel = false, so call-jit-kernel-tests was skipped.
2026-10-01 23:10	The scheduled all-suites run on main passes, including both jit-kernel-unit-test-1-gpu-large shards (run 36939300008). It ran on f17f7705a5, which predates #42128.
2026-10-02 01:50	#42128 merges as ef867fa40d.
2026-10-02 02:24	First failure, on an unrelated PR that runs the jit-kernel suite (job 110676994618). The same signature then appears on other unrelated PRs (links in Motivation).
2026-10-02 07:47	This fix PR is opened.
2026-10-02 11:10	Next scheduled main run, the first one that would run this suite with #42128 included.

Why CI didn't catch it

• test_q8kv8_sparse_prefill_backend.py lives under test/registered/kernels/ and is registered in the base-b-kernel-unit stage. That stage runs only inside call-jit-kernel-tests.
• call-jit-kernel-tests runs only when check-changes sets jit_kernel=true.
  • The filter in .github/workflows/_pr-test-check-changes.yml matches only kernel and CI paths: test/registered/kernels/**, python/sglang/kernels/**, .github/workflows/pr-test.yml, .github/workflows/pr-test-jit-kernel.yml, and python/pyproject.toml.
• [Fix][DSV4.1] SWA page size with bounded replay #42128 changed none of those paths. It touched only python/sglang/srt/layers/attention/deepseek_v4_backend.py, python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py, and test/registered/unit/mem_cache/test_dsv4_compressed_pools.py.
• Despite its location, this is not a pure kernel test. It drives DeepseekV4AttnBackend._forward_prefill_sparse in srt through a hand-written fake _TokenToKVPool. Any change to the pool interface that the backend calls can therefore break the test without triggering it.
• What the gating covers ✅:
  • PRs that touch kernel or kernel-test paths run this test.
  • The scheduled main runs (11:10 and 23:10 UTC) run every suite.
• What it misses ❌:
  • PRs that change only the DSV4 srt backend or memory pool this test depends on.
  • The scheduled runs catch such breakage only after merge, and by then unrelated PRs are already failing.
• This is a repeat escape. [Refactor] Generalize DeepSeek V4 compressed pool management #38954 changed only srt files and unit tests, renamed a pool field, and broke the same fake pool with an AttributeError; [Test] Fix Q8KV8 sparse-prefill pool fixture after page-size rename #39019 fixed it. Both times the fake pool drifted from the real DeepSeekV4TokenToKVPool, and the path filter kept the breaking PR from running the test.

Remediation (follow-ups, not in this PR)

• Close the trigger gap. Two options:
  • Add the srt paths this test depends on to the jit_kernel filter: at least python/sglang/srt/layers/attention/deepseek_v4_backend.py, python/sglang/srt/layers/attention/dsv4/**, and python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py.
  • Or move the backend-level tests in this file into a suite that main-package changes already trigger.
• Reduce fake-pool drift. Derive the fake from the real DeepSeekV4TokenToKVPool interface, for example by subclassing it or using create_autospec, instead of hand-listing methods. Accessors added to the real pool are then inherited or fail at the fixture, rather than surfacing only when the suite happens to run.

Accuracy Tests

• No local GPU was available, so I did not rerun the CUDA test locally.
• The test file in this PR is byte-identical to the one at [Perf] Optimize DeepSeek V4.1 Flash Hopper paths and Blackwell prefill selection #41251's head. With that change, [Perf] Optimize DeepSeek V4.1 Flash Hopper paths and Blackwell prefill selection #41251's jit-kernel-unit-test-1-gpu-large (1) shard passed: https://github.com/sgl-project/sglang/actions/runs/36967456129/job/110714360608. Before the change, the same shard failed with the signature above: https://github.com/sgl-project/sglang/actions/runs/36959032864/job/110688647087.
• [Perf] Optimize DeepSeek V4.1 Flash Hopper paths and Blackwell prefill selection #41251's other changes do not touch the modules this test exercises (deepseek_v4_backend.py, dsv4/sparse_prefill_utils.py, the DSV4 KV-layout and accessor kernels).
• No commit on main since [Fix][DSV4.1] SWA page size with bounded replay #42128 touches those modules either.
• BLACK_NUM_WORKERS=1 SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure passes.

Speed Tests and Profiling

Not applicable: test-only change.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. (Fixes an existing test fixture; no new tests.)
• Update documentation according to Write documentations. (Not needed.)
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. (Not applicable.)
• Follow the SGLang code style guidance.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #36980423210
Latest PR Test (Extra): ❌ Run #36980422882
Latest PR Test (AMD ROCm 10): ❌ Run #36980423204

[nvpohanh]

/rerun-test test/registered/kernels/ops/attention/test_q8kv8_sparse_prefill_backend.py

[github-actions]

Results for /rerun-test test/registered/kernels/ops/attention/test_q8kv8_sparse_prefill_backend.py:

🚀 1-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/kernels/ops/attention/test_q8kv8_sparse_prefill_backend.py

[nvpohanh]

force-merging since this is a very local test change and the test has passed. and this is to fix a wide-spread CI failure.