MindPatternSubscribe

Entity trail

GB300 TP4

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

1

Findings

1

Edges

0

Sources

1

Corpus findings

2026-09-14 / github-pulse-researchervLLM swaps GLM-5.3-Flash's 15-kernel Triton prefill for FlashKDA, measuring 1.7-3.8x on the kernel and up to 13% off TTFTMerged 2026-09-14, PR #55737 reuses vllm._flashkda_C (already built for Kimi-K3) for GLM-5.3-Flash's KDA chunked prefill because FlashKDA implements the same bounded-gate recurrence, making it a drop-in. Per-layer on GB300: 8x2048 tokens 183us to 49us, 4x8192 555us to 146us, 1x16384 1371us to 787us. End-to-end on 4x GB300 TP4 with prefix caching off, mean TTFT fell 12.2% at 8x2048 and 7.9% at 2x65536; selection is automatic on SM90/SM10x/SM12x with bf16 and head_dim 128, and additional_config.kda_prefill_backend=triton keeps the old path.

Source trail

Graph sources

entity graphfindings textkg entitiesnewsletter issues