[fank]

Scope

Narrow allocation-only change, not upstream #55575's auxiliary-compute optimization. With VLLM_PCAI_DSPARK_SKIP_MTP_BUFFER=1, V2 DSpark with valid configured auxiliary layers and PP1 does not allocate the target _mtp_hidden_buffer. Default, ordinary MTP, V1, PP and missing/invalid auxiliary configurations retain the existing buffer.

The DSpark auxiliary consumer ignores the MTP residual. At our 4096-token, 4096-hidden, hc_mult4 bf16 geometry this omits 128 MiB/GPU. No auxiliary arithmetic, target computation or workspace/graph-lifetime changes. No production changes.

Evidence

• 20 repository CPU tests pass, including opt-in, default, MTP/no-aux and V1/PP fallback cases.
• Patch applies with zero fuzz to the installed audit image and the resulting model source compiles.
• Installed consumer CPU probe passes ordinary, partial graph-staging and all-staged branches with a poison residual that fails on any read. Uses synthetic auxiliary tensors and a stub projection; NOT CUDA-graph execution or full-model equivalence.
• GPU graph capture/replay, output correctness, performance and actual profiled KV refund remain untested. Source-estimated 128 MiB is not a measured capacity gain.

Evaluation instructions: docs/dspark-mtp-buffer.md. Compare the SAME image flag off/on, without packed scales or other experiment flags.

[fank]

GPU follow-up on TP2 H100 NVL, same-image flag off/on:

• Off: 6.76 GiB available KV, 1,129,474 reported tokens.
• On: 6.89 GiB available KV, 1,150,350 reported tokens: +20,876 tokens (+1.85%). GiB log values are rounded; the source-derived omitted allocation is 128 MiB/GPU.
• Both ranks logged the omission only in the enabled run. Profile graph memory was 0.92 GiB and final graph memory 0.61 GiB in both runs.
• Serving reports both suites exited zero; vision checks were 3/3 in each. Full performance/acceptance comparison remains under review. Free-form golden text hashes differ; this is not bitwise output parity evidence.

This establishes an actual profiled KV capacity increase, not merely an allocation estimate. No production deployment or default-flag change.