[EanWang211123]

Motivation

ReplaySSM spec-verify support was introduced for MTP in:

• RFC: #28511 - Porting ReplaySSM to SGLang
• Implementation: #28695 - Support ReplaySSM Ring Spec-Verify

It avoids storing a full intermediate SSM state for every draft token. Instead, the verify stage stores compact raw inputs in ReplaySSM buffers and folds only the accepted prefix into the persistent SSM state after verification.

DFlash and DSPARK have their own post-verify state commit paths, so hybrid GDN models could not reuse the existing ReplaySSM implementation. Enabling --enable-linear-replayssm-spec with these algorithms was therefore rejected or could fall back to a commit path that expects intermediate_ssm.

This PR connects DFlash and DSPARK to the existing ReplaySSM fold-every-commit path while preserving their original KDA and non-ReplaySSM behavior.

Modifications

• dflash_worker_v2.py

  • Route DFlash GDN ReplaySSM states through the common post-verify commit path.

• dspark_worker_v2.py

  • Add GDN ReplaySSM state commit support for DSPARK.

• kv_cache_configurator.py

  • Allow DFlash and static-mode DSPARK to use ReplaySSM with hybrid GDN models.

Accuracy Tests

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ❌ Run #37716399910
Latest PR Test (Extra): ❌ Run #37716399565
Latest PR Test (AMD ROCm 10): ❌ Run #37716399934

[stefanskiasan]

Tested this PR as a Python-only overlay on the released v0.5.19 image (no rebuild): all helper symbols it relies on — commit_mamba_states_after_verify, replayssm_spec_fold/replayssm_is_kda, hybrid_gdn_config, and ragged_verify — already ship in 0.5.19, so the 3-file diff applies cleanly with patch -p1 --fuzz=5.

Ran DFLASH (jzinno/Ornith-1.5-35B-A3B-DFlash2) against Ornith-1.5-35B-A3B (FP8, W8A8) on a single H100 94GB with --enable-linear-replayssm-spec --mamba-radix-cache-strategy no_buffer --disable-overlap-schedule --attention-backend flashinfer:

• Before the patch: ValueError: --enable-linear-replayssm-spec with DSPARK/DFLASH requires a KDA (kimi_linear) model — i.e. DFLASH+GDN was hard-rejected.
• After the patch: the rejection is correctly lifted, GDN kernels + the DFLASH draft CUDA graphs init cleanly, and it serves. Accept-length up to 8.7 on structured/repetitive output, ~2.6–7.6 on code/JSON → 2.5–3.1× single-stream decode speedup on code/JSON, lossless (exact verify). Free-form prose lands ~2–3 accept as expected.

Confirms DFLASH + GDN + ReplaySSM works on Hopper for the Qwen3.5-A3B lineage, not just KDA. Would love to see this merged. Thanks for the work!