qwen3_5: reduce Gated DeltaNet preparation shared memory

[hsliuustc0106]

Purpose

Reduce Qwen3.5/3.8 Gated DeltaNet preparation shared memory from 92 KiB to 72 KiB. Convert normalized Q/K to TF32 once, store the meaningful bits in three-byte component planes, retain the original four-term accumulation, and write U/W fragments directly as bfloat16. The float32 inverse and existing workspace/C ABI contracts are preserved.

Extend the recurrent float64 reference under tests/qwen3_5/ to 20 configurations covering tile/chunk boundaries, grouped heads, long inputs, weak decay and tiny/zero Q/K, without changing its tolerance. Merge current main after #55, resolve the README/test conflicts, preserve both branches' test cases, and use the shared omni-qwen3-5-native test registration.

Test Plan

System1-Omni Version / Commit: 33df9d75a575833b04bd3292e9b6ff32c7f6b760, incorporating main 579fa8f2c4f75d7099179daf95695b6aef8766b9.

Passed on the integrated branch:

• cargo fmt --all --check
• cargo clippy --workspace --locked --all-targets -- -D warnings
• cargo test -p omni-jev --test frontend --locked — 12 frontend API tests.
• cargo test --workspace --locked
• cargo build --workspace --release --locked
• cargo test --release --locked -p omni-qwen3-5-native --test kernels --no-run
• CUDA 13.0 backend builds for SM80, SM89 and SM90.
• Benchmark smoke-manifest validation and all 8 benchmark unit tests.
• mkdocs build --strict using the existing prepared docs environment.

One scheduler-reserved H200 GPU 2 job, NUMA 0 / CPUs 0–15: all six current GPU tests passed, including 20 recurrent GDN configurations. Newly built merged ABI4 worker/frontend binaries passed genuine readiness, first inference and one 74-request Open-Jev fidelity pass. Every probability, decision and token-usage field matches the pinned native reference exactly. Task-owned processes exited; the reservation was released with 0 MB remaining on GPU 2.

Integration controls, build hashes, commands and results, GPU test log, and all request results.

Test Result

GitHub CI passed for the updated head 33df9d7: Rust, benchmark tests and the strict documentation build. GitHub reports this head as conflict-free (MERGEABLE).

Historical measurements from 2026-10-03, complete eager GDN wall time (two passes of 100 calls per shape, feasibility/warmups excluded):

Tokens	Baseline pass 1 / 2	Candidate pass 1 / 2	Reduction pass 1 / 2
107	0.051291 / 0.050976 ms	0.050914 / 0.050957 ms	0.74% / 0.04%
936	0.247565 / 0.247337 ms	0.214176 / 0.214178 ms	13.49% / 13.41%
3,399	0.800775 / 0.798935 ms	0.696768 / 0.697262 ms	12.99% / 12.73%

The ≥10% medium/long kernel gate passed. Inputs are seeded synthetic tensors at real workload shapes; wall time includes host submission. Sampled U/W/QD/KD/PB/decay representations and final outputs matched the native baseline bitwise.

Mean warm HTTP latency, 74 single-candidate requests per pass: baseline 48.375 / 48.441 ms, candidate 48.105 / 48.305 ms. Aggregate 48.408 → 48.205 ms (0.42%), below the predeclared 2% target. These two repetitions do not establish a material full-model latency improvement.

The latency comparison used baseline main 58b8cbe, the same frozen #55 worker/frontend (202c0e1) for both arms, and ABI4 library sources 7b93572 except for candidate GDN preparation. Graph capture and shared-prefix caching were disabled. The integrated head's GDN source matches that measured candidate exactly; the 2026-10-04 run above validates integration correctness and makes no new performance comparison. HF Transformers and open-jev-fast were not remeasured here.

Benchmark report, raw records and reproduction controls also document the separate installed FlashInfer comparison and rejected precision/accumulation variants. Modified-kernel execution on SM80/SM89 and hardware profiling counters remain unverified.

Self-review

Agent precheck completed with no actionable code correctness findings. Conflicts with merged #55 are resolved, test registration is shared and unique, and archived performance evidence is preserved. Contributor review/checklist is pending.

Before marking this PR ready for review or requesting maintainer review, complete the self-review checklist. Keep the PR in draft while this work is incomplete. The optional precheck-pr skill does not replace contributor or maintainer review.

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.