[RFC] GDN hybrid decider kernels: close HF→vLLM/SGLang gap (causal_conv1d + FlashInfer/Triton GDN) #141

Activity

1. hsliuustc0106

   added

   kind:rfcProposal / design discussion; scope review before buildProposal / design discussion; scope review before build

   area:knowledgeKnowledge line: packs, skill library, knowledge admissionKnowledge line: packs, skill library, knowledge admission

   on Oct 25, 2026
2. Dong1017

   self-assigned this

   on Oct 27, 2026
3. Dong1017 commented on Oct 29, 2026

   

   Dong1017

   on Oct 29, 2026 – with ChatGPT Codex Connector

   Last edited by Dong1017

   Collaborator

   More actions

   Original issue-owner plan (retained)

   Issue-owner plan: finish the remaining GDN evidence chain

   The authorized Codex parent agent is resuming accountability for this whole issue at the user's request, through delivery and an evidence-based closure-readiness assessment. The existing causal-conv1d and vLLM source slices are already delivered by #164 and #169; those slices will be reused.

   Planned remaining slices:

   1. HF Transformers caller contract (one source PR): pin the actual Qwen3.5 implementations used by the workers (Transformers 5.10.2 for Clef and 5.17.0 for PPLX), trace convolution/GDN binding and cache/forward behavior, and document the classification-prefill transfer boundary.
   2. SGLang caller/backend contract (a separate source PR): qualify the model-to-GDN/convolution dispatch, backend selection and state layout at a fixed upstream revision, preserving the existing skill owners and distinguishing prefill from decode.
   3. Whole-issue evidence reconciliation: independently inspect the existing system1-omni fix: compact Skill Eval comments while retaining evidence boundaries #173 and related measurement records for the required sm_89 installation/ready binding, Public231 C1 p50/p95 delta and quality controls. Newer hardware or changed workloads will not be substituted for the stated acceptance conditions.

   Each delivery PR will retain immutable source provenance and licenses, source-derived candidate limits, independent revision-bound review and all applicable CI. Existing records and negative results remain intact. This plan does not start new device/model/provider execution or modify the serving engine.

   Closure readiness will be assessed against all four success criteria, with a final Delivered / Verified / Limits card and exact remaining blockers if any. Source delivery or passing software CI alone will not mark the performance criteria complete. The issue remains open during that assessment.

   Source delivery complete; original empirical acceptance pending

   Issue141 remains open. Source criteria 1/4 are delivered; empirical criteria 2/3 are not yet independently verified.

   flowchart LR
       A["Pinned Dao, vLLM, HF and SGLang sources"] --> B["HF worker/operator map and C1 limits"]
       B --> C["Candidate references and source cases"]
       C --> D["Independent review and software checks passed"]
       D --> F["Whole-Issue acceptance pending"]
       E["Original install, ready and kernel A/B receipts"] -. "still needed" .-> F
   

   Loading

   • Delivered: reuse the merged Dao knowledge: qualify causal-conv1d forward and state contracts #164 and vLLM knowledge: trace vLLM GDN convolution phases and state #169 slices. HF knowledge: qualify HF Qwen3.5 conv/GDN bindings #250 is merged at a8215b3ec314b8bc5cee152cdebbf4bb76ba4188; SGLang docs: qualify SGLang GDN and causal-conv caller contracts #251 is merged at 7d1d82f55cda89842bd23ee2da91b521ee60eb08. The map includes the actual eligible default PPLX GVA forward/adapter, independent convolution/GDN binding, grouped heads/gates, state orientation/initialization and classification-prefill transfer boundaries. See the HF binding contract and integrated operator map.
   • Verified: both new slices have genuine final-head independent Sol/ultra LGTM and all applicable current CI. HF review / Ready checks: each full lane 2,465 passed / 11 skipped / 259 subtests, both isolated wheels 124 records. SGLang review / Ready checks: each full lane 2,509 passed / 11 skipped / 1 warning / 284 subtests, both isolated wheels 124 records. Both actual squash trees match their assessed merge results. Post-merge CPU smoke (8 tests each) and declaration checks passed; SGLang post-merge software / declaration.
   • Limits: source knowledge remains candidate/not_run. No new kernel/device/model/provider execution, quality run, original latency remeasurement or agent-efficacy experiment occurred. The positive upstream Clef/PPLX reports remain author evidence; their matching original raw records were not accessible for independent verification. Completed source PRs and software CI do not replace those empirical gates.

   Full acceptance and evidence record

   Original criterion	Disposition and evidence	Remaining action
   1. Document vLLM/SGLang GDN+conv operators against the actual HF worker	Source scope delivered: #164/#169/#250/#251, pinned primary readsets and the concrete five-row worker/serving cross-map. The default eligible PPLX GVA body and FI adapter are covered explicitly.	Retain source-derived limits; no runtime/kernel equivalence is implied.
   2. Reproducible sm_89 causal-conv1d installation with no fallback on ready	Not independently accepted: installation recipes and source guards are documented, but matching original successful build/install, actual capability and ready/first-forward receipts were not recovered.	Provide the original revision/environment/build/ready records and reconcile the public hardware correction.
   3. At least one verified C1 p50/p95 delta versus the issue's baseline table	Positive author reports preserved; original comparison not independently verified here. System1 #173 reports Clef 9030065 → 594528a9: 247.518/413.025 → 164.969/216.755 ms, and first PPLX 1ba07d15 → f968e576: 98.777/341.202 → 98.037/302.569 ms, each 231/231. Later settled/kept tips remain distinct author records, not replacements for the first-run figures.	Durable original baseline/tip full revisions, workload/protocol hashes, warmup5/concurrency1, A/B order or remeasure-gate, request/outcome, quality and matching environment records.
   4. Explain classification-C1 transfer and non-transfer	Source scope delivered: complete-prompt prefill versus decode/continuation, call/alias/architecture predicates, state initialization/axes and unsupported transfer conditions are documented.	No numerical, model-quality or latency acceptance follows from the source note.

   Final source/software identities. HF reviewed head f0dfee58054c2be8d17afd75d78eb5782a861ff8 generated from committed source b24d3525755376d2b0a5184ce66526b120729e6e. SGLang reviewed head 1799639b95ec99e320346fb85963d4a207004131 is a test-only child of projection witness 5672d0075e214e4000b5d20f580ff0451ce9a270; canonical source is 2f72a0ccabd87eefdc76223c9b82f7091c379da1. All 517 committed inputs and all 78 native outputs were verified. The final one-line exact inventory expectation accounts for the genuine sixth typed case while retaining every per-case check; source/generator/projection bytes did not change. The complete 11,346-byte Apache license is packaged through the existing Markdown contract, byte-identical to the pinned upstream license. No package/test policy was relaxed.

   Historical failures retained. Earlier SGLang head 2515627 omitted the unsupported .txt license asset from its wheels; 5672d00 then exposed a stale exact case count5. Their results remain tied to those heads; the final head above passes actual full and isolated-package checks. The older automatic Codex review failed without a readable finding and supplies no current-head approval. Independent review checked the complete final change, submissions, comments and threads. SkillEval retained zero comparisons/zero calls and 90 gaps with an explicit source/software-only disposition, not an efficacy pass.

   Read-only archive recovery. The existing 29,114,507-byte migration archive's whole-file SHA-256 was verified as af09735124e0630284740e0ac42884ab0176f2fb7985918efab28e8a40628e6e. Its inspected protocol compares Python HTTP 504b272c46dc1963f981c6bed9ef0c7ebdd40c69 with Rust HTTP 3364dbd7f8b2cb414c24e73f24b74dd785888ecc. Selected protocol, metadata and summaries were independently reconciled; not every manifest member was rehashed. This is a different comparison and cannot replace the original kernel A/B evidence. The original run path remains unknown in the granted readable scope.

   Hardware scope. The original issue names L20X/sm_89; the engineering owner corrected the target to H200/sm_90, followed by a later L20X-reported8.9 startup statement. Original run properties must reconcile this discrepancy; a driver label or current server profile does not establish historical capability.

   Next acceptance action: make the original experiment receipts readable, or record an explicit accepted scope decision from the responsible owner. Keep the issue open while criteria2/3 lack matching evidence. New device/model/provider experiments require their own concrete execution authorization and are not started by this source delivery.