ling: Ling-3.1-flash on CUDA at TP=1 on #804's kernels (draft, stacked on #804) by jschmied · Pull Request #874 · ashhart/TensorFold

glm: group-limited top-k routing (Ling's 8 groups, keep 4) and a per-layer shared SwiGLU limit, GLM's bytes unchanged

jschmied

committed

f9fc1a7

ling: Ling-3.1-flash's config loader and weight map, its routing and SwiGLU limits as GLM's kernels take them

jschmied

committed

a655957

ling: weights in the checkpoint's own format (bf16, FP8 blocks, ModelOpt NVFP4) at TP=1, synthetic checkpoints in each

jschmied

committed

d46e372

docs: Ling-3.1-flash recipe: what exists, what is missing, testing without the checkpoint, memory per format

jschmied

committed

faf7af1

glm: MLA kernels take Ling-3.1's shape: 64 roped key columns, 128-wide heads, head-wise gate

jschmied

committed

7d184cb

glm: KDA serves Ling-3.1 as is (real layer 0 vs fla and float64); an opt-in 16-row WY chunked prompt path

jschmied

committed

ef8f12f

ling: the TP=1 layer graph in any loaded format: KDA and MLA mixers, dense and MoE MLPs, head, windows and commits

jschmied

committed

6f583ba

ling: SwiGLU clamp in Ling's order, min(silu(g), L) * clip(u, -L, L)

jschmied

committed

bfb44dd

tests: zig fmt ling_mla_real.zig

jschmied

committed

2a5e1b6

ling: served through the lane core at TP=1: engine, lane backend, server hooks and kept prompt states

jschmied

committed

a3f14d3

ling: the MTP layer (model.layers.72) drafts: enorm | hnorm through eh_proj, MLA + MoE on its own cache, chains of 3

jschmied

committed

577436b

tests: ling-pack times one MoE layer's expert load (read, pack, cold) and checks its bytes against the host packers; ling-layer-real can dump its rows

jschmied

committed

7bd7a40

ling: routed experts staged in batches through pinned memory and packed on the GPU (Packer.fp8Experts, fp4Experts), the file's pages dropped per batch

jschmied

committed

bbe360f

tests: ling-layer-speed times one real layer whole and op by op per phase; ling-layer-bytes writes every path's bytes for cmp

jschmied

committed

412f66f

ling: row copies 16 bytes a thread on aligned rows (prompt KDA and MLP column copies -50 %), the same bytes

jschmied

committed

99652ba

ling: a single decode row's projections write their columns in place, no copies (4 KDA, 2 per MLP), the same bytes

jschmied

committed

b3faea6

ling: FP8 prompt chunks under 16 routed pairs an expert run the decode expert kernels (same bits, -7 to -27 % expert time up to 768 rows)

jschmied

committed

5b00637

glm: group-limited select without divisions where groups are whole lane columns (Ling's 64), same picks and weights; ling-route-old

jschmied

committed

e78d23b

tests: ling-layer-speed takes a rows list, times prompt ops from an idle GPU, the expert kernels and plan tiles side by side

jschmied

committed

0ffce4f

tests: ling-layer-speed starts each rep after a random spin (tf_probe_spin), so other contexts' time slices miss the least

jschmied

committed

bd7c23e

glm: swiglu keeps GLM's limit as given; only Ling's swigluWith maps 0 to unclamped

jschmied

committed

32055e9

glm: swigluWith's doc comment within 120 columns

jschmied

committed

991383f

glm: drop the opt-in 16-row chunked KDA prompt path; GLM's cuda_kda.zig is 6380c8da's again

jschmied

committed

03171bf

ling: config refuses every geometry the kernels don't take, at load, with a named error

jschmied

committed

66579cb

ling: routing groups of one and a single pick refused at load; one group and every group kept checked vs float64

jschmied

committed

92afed6

ling: Store.alloc owns a buffer only once it is in both lists; a fake driver to fail every call in turn

jschmied

committed

674163f

ling: a new stream's slot is chosen before its map entry exists (lanes.zig read the entry's unset slot)

jschmied

committed

0ba510a

ling: rows past a sequence's capacity are refused before any layer runs, not at the first MLA layer

jschmied

committed

82b3817

ling: Engine.init frees its sequences when binding slot 0 or the last sync fails; every init failure checked

jschmied

committed

0ba2adf

ling: every tensor of the full checkpoint checked against the loader's names, dtypes and shapes, from headers alone

jschmied

committed

5649c63

oracle: the Ling oracles check their HF source against inclusionAI/Ling-3.1-flash@a6943f4b by sha256

jschmied

committed

e4c9bcc

ling: a segment the scratch cannot hold is refused before any layer runs, not at the first MLA layer

jschmied

committed

5e931c3

ling: staging and loaded weights freed only after the stream drains; a failed drain leaks them on purpose

jschmied

committed

1395e26

docs/recipes/ling-3.1: file names for hashes the rebase changed, current check counts, the recorder on main

jschmied

committed

e99e650

Navigation Menu

Repository navigation

ling: Ling-3.1-flash on CUDA at TP=1 on #804's kernels (draft, stacked on #804) - #874

ling: Ling-3.1-flash on CUDA at TP=1 on #804's kernels (draft, stacked on #804)#874
jschmied wants to merge 34 commits into
ashhart:mainfrom
jschmied:ling-family

Commits on Oct 16, 2026

glm: group-limited top-k routing (Ling's 8 groups, keep 4) and a per-layer shared SwiGLU limit, GLM's bytes unchanged

ling: Ling-3.1-flash's config loader and weight map, its routing and SwiGLU limits as GLM's kernels take them

ling: weights in the checkpoint's own format (bf16, FP8 blocks, ModelOpt NVFP4) at TP=1, synthetic checkpoints in each

docs: Ling-3.1-flash recipe: what exists, what is missing, testing without the checkpoint, memory per format

glm: MLA kernels take Ling-3.1's shape: 64 roped key columns, 128-wide heads, head-wise gate

glm: KDA serves Ling-3.1 as is (real layer 0 vs fla and float64); an opt-in 16-row WY chunked prompt path

ling: the TP=1 layer graph in any loaded format: KDA and MLA mixers, dense and MoE MLPs, head, windows and commits

ling: SwiGLU clamp in Ling's order, min(silu(g), L) * clip(u, -L, L)

tests: zig fmt ling_mla_real.zig

ling: served through the lane core at TP=1: engine, lane backend, server hooks and kept prompt states

ling: the MTP layer (model.layers.72) drafts: enorm | hnorm through eh_proj, MLA + MoE on its own cache, chains of 3

tests: ling-pack times one MoE layer's expert load (read, pack, cold) and checks its bytes against the host packers; ling-layer-real can dump its rows

ling: routed experts staged in batches through pinned memory and packed on the GPU (Packer.fp8Experts, fp4Experts), the file's pages dropped per batch

tests: ling-layer-speed times one real layer whole and op by op per phase; ling-layer-bytes writes every path's bytes for cmp

ling: row copies 16 bytes a thread on aligned rows (prompt KDA and MLP column copies -50 %), the same bytes

ling: a single decode row's projections write their columns in place, no copies (4 KDA, 2 per MLP), the same bytes

ling: FP8 prompt chunks under 16 routed pairs an expert run the decode expert kernels (same bits, -7 to -27 % expert time up to 768 rows)

glm: group-limited select without divisions where groups are whole lane columns (Ling's 64), same picks and weights; ling-route-old

tests: ling-layer-speed takes a rows list, times prompt ops from an idle GPU, the expert kernels and plan tiles side by side

tests: ling-layer-speed starts each rep after a random spin (tf_probe_spin), so other contexts' time slices miss the least

glm: swiglu keeps GLM's limit as given; only Ling's swigluWith maps 0 to unclamped

glm: swigluWith's doc comment within 120 columns

glm: drop the opt-in 16-row chunked KDA prompt path; GLM's cuda_kda.zig is 6380c8da's again

ling: config refuses every geometry the kernels don't take, at load, with a named error

ling: routing groups of one and a single pick refused at load; one group and every group kept checked vs float64

ling: Store.alloc owns a buffer only once it is in both lists; a fake driver to fail every call in turn

ling: a new stream's slot is chosen before its map entry exists (lanes.zig read the entry's unset slot)

ling: rows past a sequence's capacity are refused before any layer runs, not at the first MLA layer

ling: Engine.init frees its sequences when binding slot 0 or the last sync fails; every init failure checked

ling: every tensor of the full checkpoint checked against the loader's names, dtypes and shapes, from headers alone

oracle: the Ling oracles check their HF source against inclusionAI/Ling-3.1-flash@a6943f4b by sha256

ling: a segment the scratch cannot hold is refused before any layer runs, not at the first MLA layer

ling: staging and loaded weights freed only after the stream drains; a failed drain leaks them on purpose

docs/recipes/ling-3.1: file names for hashes the rebase changed, current check counts, the recorder on main