[WIP] Enable GLM-5.2 DSpark compact verify functionality

[tanth47]

Summary

This is a draft/WIP GLM-5.2 DSpark functionality patch.

• Fixes GLM-5.2 DSpark draft block layout to verify anchor + gamma and sample/confidence over draft slots only.
• Enables compact/ragged verify plumbing for GLM-5.2, including DSA eager ragged support and CUDA-graph gating.
• Adds seeded non-greedy DSpark draft/accept paths for PyTorch sampling backend, with debug reference assertions.
• Adds request validation so seeded non-greedy DFlash/DSpark requests do not silently run unseeded on unsupported backends.
• Adds seeded DFlash standalone sampling uniforms.
• Adds trace/accuracy harness and focused tests for compact layout, target probs, DSA graph gating, request validation, and seeded sampling metadata.

Runtime validation

GLM-5.2 target: zai-org/GLM-5.2-FP8
Draft: RedHatAI/GLM-5.2-speculator.dspark
TP: 4
Backend: ROCm MI350, DSA tilelang, --sampling-backend pytorch, SGLANG_RAGGED_VERIFY_MODE=compact

• Greedy/default compact 64x128: 64/64 OK, aggregate AR 0.5126, AL 4.5536, trace 1098 compact records, 0 failed.
• Non-greedy seeded p19 64x128: 64/64 OK, aggregate AR 0.4962, AL 4.4353, trace 2021 compact/non-greedy records, 0 failed.
• Forced-budget non-uniform p20 8x32 concurrency 2: 8/8 OK, trace 103 compact/non-greedy records, 33 non-uniform verify-lens records, 103 seeded sampling records, 0 failed.
• p20 seeded mini 8x32: 8/8 OK, aggregate AR 0.6825, AL 5.6889, trace backend pytorch, 53 seeded records, 0 failed.

Unit/focused tests

python3 -m pytest \
  test/registered/unit/spec/test_dflash_request_validation.py \
  test/registered/unit/spec/test_dspark_softmax_temp.py \
  test/registered/unit/spec/test_dflash_verify_target_probs.py \
  test/registered/unit/spec/test_dspark_compact_layout.py \
  test/registered/unit/spec/test_dspark_dsa_ragged_graph_gate.py \
  test/registered/unit/sampling/test_sampling_batch_info.py \
  test/registered/spec/dspark/test_trtllm_mha_ragged_verify.py \
  -q -p no:cacheprovider

Result: 94 passed, 5 warnings.

Known limitations

• This is not yet a full performance-stack implementation comparable to the DeepSeek-V4-specific DSpark path.
• Exact A/B seeded replay is not used as the primary correctness oracle because GLM-5.2 DSA showed baseline nondeterminism; correctness is currently checked via verifier trace + runtime non-greedy accept reference asserts.
• CUDA graph folded accept/commit and profiled SPS/STS tables are still future work for GLM-5.2 performance.

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hnyls2002]

Heads up - #30261 has just merged, so the DSpark stack this PR builds on is now on main. A rebase/merge of origin/main should shrink the diff considerably (46 of the 100 changed files were part of #30261), leaving just the GLM-5.2-specific changes for review.

[yiakwy-xpu-ml-framework-team]

@tanth47 I guess this commit is targeted to ROCM 7.0. But in the tests, you have

 test/registered/spec/dspark/test_trtllm_mha_ragged_verify.py \

and passed the test. Can I know the reason ?

[tanth47]

Good catch. That line in the draft PR body is confusing.

test/registered/spec/dspark/test_trtllm_mha_ragged_verify.py was not meant as ROCm 7.0 runtime validation of the TRTLLM MHA backend. It is a focused ragged-verify layout / graph-capability test that uses CPU tensors for the layout invariants, and the TRTLLM MHA-specific cases are guarded by an import skip:

• if TRTLLMHAAttnBackend cannot be imported in the current environment, those cases are skipped;
• the actual GLM-5.2 runtime validation for this PR was on ROCm/MI350 with the DSA backend, not TRTLLM MHA.

I included it in the draft validation list because the old monolithic patch touched generic ragged verify semantics shared by graph-capable attention backends. After #30261 merged, I am splitting this draft into smaller PRs; in the cleaned GLM-5.2/ROCm stack I will not present this test as ROCm TRTLLM evidence. It should either be listed only as generic CPU/focused coverage, or removed from the ROCm validation section.