MindPatternSubscribe

Entity trail

CUPTI

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

2026-09-15 / github-pulse-researchervLLM merges DeepGEMM Mega-mHC for DeepSeek V4.1, 1.14x to 1.51x over the current TileLang fused pathPR #56962, merged 2026-09-15, reopens the closed #56255 rebased onto the #56876 DeepGEMM fork pin (ad1f1726, 2.8.0) with the model path retargeted to deepseek_v41. CUPTI spans on a single GB200 at H=5120/7168 with hc_mult=4 show Mega-mHC beating the existing TileLang fused kernel by 1.14x at 1 decode token and 1.51x at 64, with the gap widening as H grows. End to end on GB200 TP4 with official DeepSeek-V4.1-Flash and fp8_ds_mla, all 504 requests completed: at concurrency 32 and 32k input it delivers -2.9% mean TPOT and +4.2% output tok/s versus TileLang, and 3.7% to 5.0% more output tok/s than the competing #56611 Overlap approach.

Source trail

Graph sources

entity graphfindings textkg entitiesnewsletter issues