glm/cuda: KV pages from the core pool, kept states sized by their position (#1142) by Bizuayeu · Pull Request #1163 · ashhart/TensorFold

kv pool: a family without snapshot slots keeps its whole prompt cache as copied rows

Bizuayeu

committed

19123fa

glm cuda: the latent and pooled-key kernels take a stream's page table

Bizuayeu

committed

87cb6d5

glm cuda: a sequence's latent rows and pooled keys can live on a pool's pages

Bizuayeu

committed

1ea8acc

glm cuda: streams past the engine's own take their KV pages from the core pool

Bizuayeu

committed

26d5e5e

glm cuda tests: a paged stream against its whole-window twin, byte for byte

Bizuayeu

committed

00eda0d

glm cuda: a follower makes rank 0's pool through Pages.follow, under a host test

Bizuayeu

committed

a177519

glm cuda: on pages too the chunk kernels read keys in main's loop: a dense chunk's page once a block, a sparse key's row from the token list

Bizuayeu

committed

88c2b4d

glm cuda tests: the page kernels under the kernel contract, on a page table out of order

Bizuayeu

committed

3996c1e

glm cuda: gate 5's records for the six KV page launches, from nvcc 13.0 and 13.3 on every SM

Bizuayeu

committed

bc05317

glm cuda tests: a kept prefix's piece inside a page's span, and its size while a paged stream is current

Bizuayeu

committed

b94edf9

glm cuda: a stream takes its pages before its sequence is bound, so a slot past the paged sequences waits

Bizuayeu

committed

6f3c0fc

glm cuda: Lanes.seat's comment on one line, as the lean check asks

Bizuayeu

committed

1af8e24

docs: the KV pages receipt for 0b094b74, on one GPU and at TP=2 and TP=3

Bizuayeu

committed

9166d3b

docs: the KV pages receipt's two stack traces indented with spaces, as git diff --check asks

Bizuayeu

committed

0411056

glm cuda tests: the KV page kernels' contract cases in a file of their own, so contract_tests.zig stays under 600 lines

Bizuayeu

committed

b83359e

glm cuda tests: with TF_KV_PAGES=all the lane host admits a request by the pool alone, though the engine's own sequence is free

Bizuayeu

committed

50549a0

Navigation Menu

Repository navigation

glm/cuda: KV pages from the core pool, kept states sized by their position (#1142) - #1163

glm/cuda: KV pages from the core pool, kept states sized by their position (#1142)#1163
Bizuayeu wants to merge 16 commits into
ashhart:mainfrom
Bizuayeu:pr/glm-kv-pages

Commits on Oct 24, 2026

kv pool: a family without snapshot slots keeps its whole prompt cache as copied rows

glm cuda: the latent and pooled-key kernels take a stream's page table

glm cuda: a sequence's latent rows and pooled keys can live on a pool's pages

glm cuda: streams past the engine's own take their KV pages from the core pool

glm cuda tests: a paged stream against its whole-window twin, byte for byte

glm cuda: a follower makes rank 0's pool through Pages.follow, under a host test

glm cuda: on pages too the chunk kernels read keys in main's loop: a dense chunk's page once a block, a sparse key's row from the token list

glm cuda tests: the page kernels under the kernel contract, on a page table out of order

glm cuda: gate 5's records for the six KV page launches, from nvcc 13.0 and 13.3 on every SM

glm cuda tests: a kept prefix's piece inside a page's span, and its size while a paged stream is current

glm cuda: a stream takes its pages before its sequence is bound, so a slot past the paged sequences waits

glm cuda: Lanes.seat's comment on one line, as the lean check asks

docs: the KV pages receipt for 0b094b74, on one GPU and at TP=2 and TP=3

docs: the KV pages receipt's two stack traces indented with spaces, as git diff --check asks

glm cuda tests: the KV page kernels' contract cases in a file of their own, so contract_tests.zig stays under 600 lines

glm cuda tests: with TF_KV_PAGES=all the lane host admits a request by the pool alone, though the engine's own sequence is free