10 Commits

Include renames

Author	SHA1	Message	Date
Tarjei Mandt	4f806387f7	Keep the DeepSeek V4.1 wide index scorer exact and off the GLM path	2026-10-05 11:21:26 +11:00
Tarjei Mandt	8be1005089	Fold the one-row DeepSeek V4.1 entry points into their rows form	2026-09-20 01:51:24 +10:00
Tarjei Mandt	0ac42d6fe0	Tidy the DeepSeek V4.1 comments and switches	2026-09-20 00:50:13 +10:00
Tarjei Mandt	66dff86949	Remove the DeepSeek V4.1 exact rows attention path	2026-09-20 00:50:12 +10:00
Tarjei Mandt	d8d1523b33	Speed up DeepSeek V4.1 Metal decode and run DSpark verify rows on its decode kernels	2026-09-17 08:54:55 +10:00
Tarjei Mandt	3b7f8f2242	Select every key for short DeepSeek V4.1 index rows	2026-09-17 08:54:55 +10:00
Tarjei Mandt	ce5a812fd3	Select DeepSeek V4.1 candidate blocks in batches	2026-09-17 08:54:55 +10:00
Tarjei Mandt	edceb7ab7f	Select DeepSeek V4.1 experts in one dispatch	2026-09-17 08:54:54 +10:00
Tarjei Mandt	d95f8b6101	Round DeepSeek V4.1 activations inside the producing kernels	2026-09-17 08:54:54 +10:00
Salvatore Sanfilippo	a04f46fa42	DeepSeek v4.1 Flash support for CUDA ... Add bounded SSD streaming, exact short prefills, accelerated large prefills, and native session batching. Support resident tensor parallel inference across two DGX Sparks over RoCE, with aligned expert shards, split vocabulary projection, decode graphs, and shared-expert overlap. Include primitive, state, transport, and model regression tests, plus usage documentation and measured QA results.	2026-09-13 17:32:48 +02:00