Skip to the content.

Origin. PLI Labs research archive, file kernel-research/UPSCAN1105-SUPPLEMENT.md. Published 2026-11-18 in the Volta V100 research repository. Edited for publication. Machine names, private file paths, network addresses, session names and the operator’s name were replaced with generic terms (the V100 server, [private path], [address], the owner). Links to files that are not published were turned into plain text. No figure and no finding was changed. Dates inside the note are its own; later results are in the build pages of this repository.

This is a metadata watch alongside the three sources required by BRIEF.md, not a replacement for their full inventories. Public GitHub reads only, 2026-11-05. Raw evidence is in evidence/.

Primary change Potential use / KCAMP mapping Scope and exclusion
#61464 MRV2 DBO with speculative decoding C3: request-boundary microbatch split, full-batch metadata rebuild after target forward so speculator FULL graphs see valid buffers Engine design reference, not proof of PP2/DSpark/SM70 support or measured benefit here. DBO communication overlap is not automatically PP-stage interleaving.
#64458 MRV2 PP broadcast fixes C0/C2/C3/S6: avoid non-last-stage host sync; gather on main stream before broadcast stream reads; skip absent drafts; size broadcast per actual step Correctness/ownership audit before greater concurrency; compare with our already ported PP hooks. No extra speed credit for duplicate code.
#60618 fused prefill score + top-k New/S10-related prefill selection design: remove logits materialization, one persistent launch for <=16,384 query rows / <=1M keys; 296 MiB workspace stated vs 2.6 GiB LiteTopK Their gated 24-bit ranking permits boundary ties and has a rescanning path; our exactness must be preserved. Modern FP8 scoring implementation needs a Volta FP16/FP32 replacement. No portable whole-model multiplier.
#59131 gather compressed KV once S1/S2/S4 context/memory traffic: 38 compressed-layer gathers become 4 source-group gathers per chunk on the described path Mega-attention B200 path; our packed ring and fused sparse attention do not necessarily gather whole history. Audit actual calls first; no forecast without traffic count.
#63956 token-sharded sparse MLA prefill New/S1/S4: sequence-parallel prefill, preserve head-sharded decode in mixed steps Author claims GLM-5.3 NVFP4 TP4 GB200 TTFT 1.24x/1.29x/1.40x at 32K/64K/128K. Those are neither V4.1 nor Volta data. Would change communication and GEMM shapes; no exactness assumption.
#64442 DSpark inverse-CDF Markov fusion C1/S6 implementation pattern for fused sampler; author GB200 one-request Markov 111.9→71.6 µs, not whole round Our coupled keyed-Gumbel sampler is already fused/sharded. Switching random transforms would lose seeded byte identity/coupling unless redesigned, so do not adopt as a drop-in speed win.

All six selected PR descriptions and merge SHAs are captured as evidence/vllm-pr-*.json. None of these supplementary reads authorize deployment. The ranked report is grounded in the requested 1Cat/Maya/DSX inventories and our measured baseline.