srivy-together/k3-dflash-sequence-compaction-validation-7cf713b

K3 DFlash Sequence-Compaction GB300 Validation

This repository is the evidence bundle for an independent correctness review of the TensorRT-LLM K3 DFlash sequence-compaction patch at commit 7cf713b08b6892aae44a12d41b1a67d029d4b234. The implementation was evaluated against the design contract at f7789542915749fc9e6cd9b165b3a271cbe92184.

Bottom line

No patch correctness defect was found within the implemented and supported envelope. The patch's padded-attention/compact-MLP design, strict verifier, rejection verifier, exact-profile CUDA graphs, graph padding, overlap handoff, KV/KDA accounting, and dynamic compact lengths all ran successfully on a TP8 Kimi-K3 workload using two GB300 nodes in one NVLink/MNNVL clique.

The recommendation is to approve the change as an opt-in prototype, not as a production-default feature. Production enablement should remain gated on exact-profile CUDA-graph memory economics and on resolving or explicitly waiving the inherited hybrid-Mamba block-reuse limitation.

Read VALIDATION_REPORT.md for the complete verdict, evidence matrix, correctness argument, qualifications, and performance observations. Read VALIDATION_LOG.md for the chronological test ledger.

Scope and results

• Hardware: TP8 on GB300 nodes r16 and r15 in one NVLink/MNNVL clique.
• Target: Kimi-K3 NVFP4 Traffic-Mix4.
• Drafter: five-layer K3 DFlash checkpoint, block size eight, production K=5 (W=6), YaRN 4.0.
• Native wheel comparison: all 113 entries byte-identical; no native entry missing.
• Runtime-source comparison: all 12 changed Python files matched their corresponding commits.
• Patch tests: 30 official and graph-replay CUDA tests passed.
• Runtime matrix: five successful servers, 42 responses, and zero scanner findings.
• Compact eager: 839 accepted of 1,265 eligible drafts.
• Compact CUDA graph: 908 accepted of 1,236 eligible drafts; all nine exact profiles captured and replayed.
• Compact rejection: 1,022 accepted of 1,644 eligible drafts; live temperature-1 requests passed.
• Every eligible compact length from one through five appeared in all compact-mode histograms.
• Cleanup completed; no model was deployed to production infrastructure.

The disabled-path byte-level text oracle was unstable on the unmodified base wheel. A serialized B1 follow-up produced three distinct base hashes for identical requests, while the patch produced one stable hash three times. Consequently, the report does not classify the observed 4/8 text match rate as a regression; disabled-path confidence comes from artifact identity, exact source matching, unchanged strict-verification algebra, stable-prompt parity, and clean runtime evidence.

The requested repeated-prompt block-reuse case was attempted but could not physically reuse target blocks. TensorRT-LLM disabled reuse for this hybrid linear model under the legacy Mamba manager, and metrics reported zero cached tokens. This is recorded as inherited backend coverage debt, not as a patch pass.

Repository layout

Path	Contents
VALIDATION_REPORT.md	Final review, evidence matrix, reasoning, limitations, and verdict.
VALIDATION_LOG.md	Timestamped work and test ledger.
harness/	Reproducibility scripts and K3 runtime configurations.
artifacts_final/status.json	Machine-readable validation stage history.
artifacts_final/summary.json	Case completion and text-comparison summary.
artifacts_final/analysis.json	Aggregate accounting and telemetry analysis.
artifacts_final/comparison.json	Record-level response comparisons.
artifacts_final/unit-tests.log	Unit and CUDA graph-replay test output.
artifacts_final/build.log	Local patch-wheel build output.
artifacts_final/*/server.log	Per-case TensorRT-LLM runtime logs.
artifacts_final/*/responses.json	Captured test responses.
artifacts_final/*/perf_metrics.json	Per-request and per-forward telemetry.
artifacts_final/serial_disabled_parity/	Serialized B1 nondeterminism follow-up.

The copied xp and infra reference repositories used to bootstrap the test environment are intentionally excluded. No model weights, container images, or production deployment resources are included in this bundle.

Artifact hashes

• Base wheel SHA-256: 18737092113842df8f922e087b7d3bd4d8973a92b68681b9536216de6b36816b
• Patch wheel SHA-256: 25d96576c030b9527a347232d8e79978312fd7e56c61b1fc8a1b4520f2f2353f
• Production-aligned drafter configuration SHA-256: cbffb34abb7caa6e7dd51568019e4682f6ae8052ed140d4ccec625ffc80763d3
• Final local build log SHA-256: 4e0c0614b6e0a89c8c217318138b589c892a034d6ad77955ee6d5c3a0d030611

Interpretation boundary

These artifacts establish correctness evidence for the tested topology and supported feature slice. They are not a production throughput benchmark, a proof for unsupported attention-DP/PP/CP topologies, or a live validation of guided decoding and the K3 thinking-guard state machine.