qwen27: score and decide passes queued onto the lane host's idle engine thread by grantoverton · Pull Request #1189 · ashhart/TensorFold

qwen27: serve qwen3_5_moe (Ornith-1.5-122B) on Metal: routed experts, 32-over-2 attention

BobClawblaw

committed

b117a8b

qwen27: register qwen3_5_moe on Metal; DFlash2 drafters of up to eight taps and a block of 16 (Ornith's)

BobClawblaw

committed

68fa105

qwen27: drop the port's debug switches; fmt

BobClawblaw

committed

1559553

qwen27: prompt rows route on the tensor units; the profile times the routed block (route, experts) and TF_PROFILE_PROMPT

BobClawblaw

committed

5ef0dbc

qwen27: routed checkpoints prefill in 1,024-row chunks

BobClawblaw

committed

6469810

qwen27: one-line comments for the lean check

BobClawblaw

committed

1cf5b96

qwen27: a config test for the routed shape

BobClawblaw

committed

5eeb412

qwen27: routed models let plain rounds compete with drafted depths

BobClawblaw

committed

4acc699

qwen27: the router matrix is padded to whole 64-row tiles for the tensor units' product

BobClawblaw

committed

32dad79

qwen27: tap commits gather every tap plane (the drafter's context held five of Ornith's eight); selector options by env

BobClawblaw

committed

e077860

qwen27: one-line comments for the lean check

BobClawblaw

committed

2f995f3

qwen27: decode rows run the shared expert as one more slot of the experts' kernel (core: indexed_extra row kernel)

BobClawblaw

committed

0d37dcb

qwen27: one-line comments for the lean check

BobClawblaw

committed

f11bef4

qwen27: an absent norm_topk_prob means renormalized top-k weights, as the reference config and the CUDA family read it

BobClawblaw

committed

30e5dcb

Navigation Menu

Repository navigation

qwen27: score and decide passes queued onto the lane host's idle engine thread - #1189

qwen27: score and decide passes queued onto the lane host's idle engine thread#1189
grantoverton wants to merge 16 commits into
ashhart:mainfrom
grantoverton:ornith-pr1-decide-queue

Commits on Oct 20, 2026

qwen27: serve qwen3_5_moe (Ornith-1.5-122B) on Metal: routed experts, 32-over-2 attention

qwen27: register qwen3_5_moe on Metal; DFlash2 drafters of up to eight taps and a block of 16 (Ornith's)

qwen27: drop the port's debug switches; fmt

qwen27: prompt rows route on the tensor units; the profile times the routed block (route, experts) and TF_PROFILE_PROMPT

qwen27: routed checkpoints prefill in 1,024-row chunks

qwen27: one-line comments for the lean check

qwen27: a config test for the routed shape

qwen27: routed models let plain rounds compete with drafted depths

qwen27: the router matrix is padded to whole 64-row tiles for the tensor units' product

qwen27: tap commits gather every tap plane (the drafter's context held five of Ornith's eight); selector options by env

qwen27: one-line comments for the lean check

qwen27: decode rows run the shared expert as one more slot of the experts' kernel (core: indexed_extra row kernel)

qwen27: one-line comments for the lean check

qwen27: an absent norm_topk_prob means renormalized top-k weights, as the reference config and the CUDA family read it

Commits on Oct 21, 2026

qwen27: score and decide passes queued onto the lane host's idle engine thread

Commits on Oct 22, 2026

qwen27: one-line comments for the lean check