Add: integrate DeepSeek V4 DSpark serving

[hashiqiqixian]

Summary

Integrate the DeepSeek V4 DSpark model-side pipeline into PyPTO Serving with the target DP4 x TP4 dense topology and EP16 MoE topology.

The serving data flow covered by this change is:

terminal prefill
  -> target token sampling
  -> initialize three drafter KV layers
  -> produce seven draft tokens
  -> target verification decode
  -> commit the accepted prefix per request
  -> update request-local context and drafter KV
  -> produce the next seven-token proposal

Changes

• Admit the DSpark DP4 x TP4, EP16 topology and select the DSpark PyPTO-lib kernel family.
• Add DSpark checkpoint validation, weight staging, task arguments, and device-buffer ownership.
• Wire terminal Prefill to the three-layer drafter, TP4 LM head, seven-step Markov sampler, and confidence output.
• Keep stable TP4 request/query ownership through Decode and support different accepted-prefix lengths within one batch.
• Bound verification by remaining context capacity and fall back to target-only Decode when another full proposal cannot fit.
• Carry request sampling parameters and integrate the device-sampling ABI needed by the target path.
• Document the required full DSpark W8A8 checkpoint instead of the base DeepSeek-V4-Flash checkpoint.

Dependencies

• Depends on Refactor: align DSpark drafter with TP4 output flow pypto-lib#1057 for the TP4 DSpark model-side contract. The submodule currently points to that reviewed revision.
• Includes the DSpark adaptations on top of Add: integrate DeepSeek V4 device sampling #207. After Add: integrate DeepSeek V4 device sampling #207 merges, this branch should be rebased so the shared device-sampling changes disappear from this PR diff.

Validation

On ci-69:

• task_20260828_152804_31283327088: Ruff passed; 173 targeted Serving tests passed and 1 was deselected.
• PyPTO-lib TP4 Markov B4: task_20260828_145457_72515728721 passed.
• PyPTO-lib TP4 + EP8 drafter B4: task_20260828_145528_78912424701 passed.
• Checkpoint startup-contract inspection: task_20260828_153100_6586995834 correctly rejected the available base W8A8 checkpoint because 85 of 123 DSpark startup weights are missing.

Local review:

• Complete origin/main...HEAD diff reviewed.
• git diff --check passed.
• Copyright header check passed.

Current limitations

• A real DSpark checkpoint end-to-end Serving run is not yet complete. The available /data/models/dsv4-flash-w8a8 checkpoint contains only part of mtp.0 and lacks mtp.1, mtp.2, Markov, and confidence-head weights.
• DP4 x TP4, EP16 is the implemented/static serving contract, but it has not yet been physically validated as a complete 16-device Serving deployment.
• Confidence values are produced by the model pipeline, but confidence-driven adaptive verification scheduling is not included in this PR.
• Prefix caching is intentionally disabled for DSpark in this revision.

Related model work: hw-native-sys/pypto-lib#935.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}