lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) by jschmied · Pull Request #1634 · ashhart/TensorFold

cuda: the paged shared-window attention at head dim 64

jschmied

committed

f264ece

lfm2_moe: CUDA glue kernels under the contract

jschmied

committed

ac8f6e2

lfm2_moe: serve LFM2.5-24B-A2B on CUDA

jschmied

committed

692cf20

cuda: Liquid's short conv and its commit in liquid_conv.cuh, by tap count

jschmied

committed

173e069

kernel contract: the limits check reads lfm2_moe's fatbin

jschmied

committed

064089d

kernel contract: the head-64 paged attention, decode and prompt rows, against float64

jschmied

committed

1b48934

tools: LFM2.5-24B-A2B's float64 reference on the CPU, and its receipt of every top-1 miss

jschmied

committed

68b48aa

docs: LFM2.5-24B-A2B on CUDA, its receipt and the weights' licence

jschmied

committed

0a9de29

lfm2_moe: the attention caches in pages of the core pool, every stream leasing

jschmied

committed

b4f4f49

lfm2_moe: the prompt cache on CUDA, a kept state the conv states and the pages' rows through its position

jschmied

committed

abae447

lfm2_moe: keep a mark's state as the prompt pass goes by, so a keep adds no pass

jschmied

committed

928034a

docs: LFM2.5 on CUDA leases pool pages and keeps chat states in the prompt cache

jschmied

committed

10f3274

feat(lfm3): add strict MLX checkpoint admission checks

BlivionIaG

authored and

jschmied

committed

c17be81

build(lfm3): isolate admission in the standard build graph

BlivionIaG

authored and

jschmied

committed

7be5515

cuda: the paged prompt attention at head dim 128, four query heads a KV head

jschmied

committed

d0dbb68

lfm3: admit the ModelOpt NVFP4 text export beside the MLX layouts

jschmied

committed

18a6183

cuda: the 7-tap liquid conv and commit for LFM3, under the kernel contract

jschmied

committed

a8faad1

lfm3: serve the ModelOpt NVFP4 text export on CUDA, weight-only

jschmied

committed

3d3759f

liquid: one lane backend, kept-state walk and native glue for LFM2.5 and LFM3 on CUDA

jschmied

committed

0a9f78e

kernel contract: pattn_kernel<128, 8, 4, 8> over pages against float64, with a control

jschmied

committed

1709436

lfm3: Metal refuses the ModelOpt NVFP4 export by name

jschmied

committed

c311238

lfm3: say at load that NVFP4 on CUDA is weight-only

jschmied

committed

d6efbac

tools: LFM3's float64 reference on the CPU, and its receipt of every top-1 miss

jschmied

committed

6bb371c

docs: LFM3-1.7B NVFP4 on CUDA, weight-only, its receipt and the weights' licence

jschmied

committed

fad7e14

Navigation Menu

Repository navigation

lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) - #1634

lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508)#1634
jschmied wants to merge 29 commits into
ashhart:mainfrom
jschmied:lfm3-cuda-fp8

Commits on Oct 29, 2026

cuda: the paged shared-window attention at head dim 64

lfm2_moe: CUDA glue kernels under the contract

lfm2_moe: serve LFM2.5-24B-A2B on CUDA

cuda: Liquid's short conv and its commit in liquid_conv.cuh, by tap count

kernel contract: the limits check reads lfm2_moe's fatbin

kernel contract: the head-64 paged attention, decode and prompt rows, against float64

tools: LFM2.5-24B-A2B's float64 reference on the CPU, and its receipt of every top-1 miss

docs: LFM2.5-24B-A2B on CUDA, its receipt and the weights' licence

lfm2_moe: the attention caches in pages of the core pool, every stream leasing

lfm2_moe: the prompt cache on CUDA, a kept state the conv states and the pages' rows through its position

lfm2_moe: keep a mark's state as the prompt pass goes by, so a keep adds no pass

docs: LFM2.5 on CUDA leases pool pages and keeps chat states in the prompt cache

feat(lfm3): add strict MLX checkpoint admission checks

build(lfm3): isolate admission in the standard build graph

cuda: the paged prompt attention at head dim 128, four query heads a KV head

lfm3: admit the ModelOpt NVFP4 text export beside the MLX layouts

cuda: the 7-tap liquid conv and commit for LFM3, under the kernel contract

lfm3: serve the ModelOpt NVFP4 text export on CUDA, weight-only

liquid: one lane backend, kept-state walk and native glue for LFM2.5 and LFM3 on CUDA

kernel contract: pattn_kernel<128, 8, 4, 8> over pages against float64, with a control

lfm3: Metal refuses the ModelOpt NVFP4 export by name

lfm3: say at load that NVFP4 on CUDA is weight-only

tools: LFM3's float64 reference on the CPU, and its receipt of every top-1 miss

docs: LFM3-1.7B NVFP4 on CUDA, weight-only, its receipt and the weights' licence

Commits on Oct 30, 2026

receipts: SASS identity of nemotron_attention and prefill_attention against main, every function's hash

lfm3: admit the block FP8 text export beside NVFP4 and the MLX layouts

lfm3 CUDA: serve the FP8 export weight-only and the MLX 4-bit export as stored

docs, tools: LFM3's FP8 and MLX 4-bit on CUDA, and their float64 receipts

tools/lfm3: ref64.py runs on numpy without torch, and a bf16-points pass