gpu, gpu-gates, model, refset: Nemotron 3 Nano 4B on the card and bloomery-serve --model nemotron by midagedev · Pull Request #259 · midagedev/bloomery

placement, model: Program::NemotronHBody, a dense whole-card program, and coverage rows for its Mamba-2 and squared-ReLU layers

midagedev

and

claude

committed

2fc3a55

gpu, gpu-gates: the nemotron_h decode body on the card, and gate_nemotron_e2e's self-consistency clauses

midagedev

and

claude

committed

1c6f94f

model: the nemotron_h program covers the pack-5 GQA row by its own entry

midagedev

and

claude

committed

be15476

refset: read nemotron logits as whole f32 chunks

midagedev

committed

d4082d6

gpu-gates, model, runtime: bloomery-serve --model nemotron, one slot on the card body, and gate_nemotron_serve

midagedev

and

claude

committed

583bf9f

placement, model: PIN(2026-10-15) the two test clauses the nemotron_h program retired

midagedev

and

claude

committed

79290ad

gpu-gates, refset: gate_nemotron_e2e holds the body to mainline's last-position logits

midagedev

and

claude

committed

816bbaf

gpu-gates: the group-5 flash check ends a tile on four live keys

midagedev

committed

052e3d3

tools: gate-paths after the rebase onto main: nemotron_h's row as settled, and gate_nemotron_e2e on the refset rows it reads

midagedev

and

claude

committed

9ea9da2

model: PIN(2026-10-15) the nemotron meta clause the program retired

midagedev

and

claude

committed

dc1b956

gpu: the nemotron_h attention takes main's one head-128 launch at group 5

midagedev

and

claude

committed

92d182d

gpu: relu_sqr returns with the nemotron_h body that calls it

midagedev

and

claude

committed

29fb061

gpu-gates: gate_nemotron_e2e takes the shared runner's argmax

midagedev

and

claude

committed

9d54d6b

gpu-gates: gate_nemotron_serve reads the trained context through generate::trained_ctx

midagedev

and

claude

committed

34111a1

tools: gate-paths' act_rule.rs and flip.rs rows name gate_nemotron_e2e, which reaches them through the shared e2e runner

midagedev

and

claude

committed

cb26dcb

tools: gate-paths' emb2.rs and gelu.rs rows name the nemotron gates that reach them on main

midagedev

and

claude

committed

1f9a93c

model, gpu, gpu-gates: review fold: the nemotron type pins, the shared blk name, one-slot plan, the keep rule in its seat, no copied hparams, the seat records the model's held context

midagedev

and

claude

committed

1b6550c

model, gpu, gpu-gates: review fold: the nemotron pins read every type of their tensors, an absent conv bias is a zero bias, the seat's context is the model's, the default short refusal

midagedev

and

claude

committed

6906329

gpu, tools: a file tensor never takes a derived weight's name; gate-paths' nemotron_h row names the gates coverage now reaches it through

midagedev

and

claude

committed

a25cbe0

model, gpu, gpu-gates: review fold: the nemotron vector and gain pins (f32, bf16), the serve gate's clauses in one loop, no blk literal outside arch/

midagedev

and

claude

committed

2b2b19d

Navigation Menu

Repository navigation

gpu, gpu-gates, model, refset: Nemotron 3 Nano 4B on the card and bloomery-serve --model nemotron - #259

gpu, gpu-gates, model, refset: Nemotron 3 Nano 4B on the card and bloomery-serve --model nemotron#259
midagedev merged 20 commits into
mainfrom
nemo7stack

Commits on Oct 15, 2026

placement, model: Program::NemotronHBody, a dense whole-card program, and coverage rows for its Mamba-2 and squared-ReLU layers

gpu, gpu-gates: the nemotron_h decode body on the card, and gate_nemotron_e2e's self-consistency clauses

model: the nemotron_h program covers the pack-5 GQA row by its own entry

refset: read nemotron logits as whole f32 chunks

gpu-gates, model, runtime: bloomery-serve --model nemotron, one slot on the card body, and gate_nemotron_serve

placement, model: PIN(2026-10-15) the two test clauses the nemotron_h program retired

gpu-gates, refset: gate_nemotron_e2e holds the body to mainline's last-position logits

gpu-gates: the group-5 flash check ends a tile on four live keys

tools: gate-paths after the rebase onto main: nemotron_h's row as settled, and gate_nemotron_e2e on the refset rows it reads

model: PIN(2026-10-15) the nemotron meta clause the program retired

gpu: the nemotron_h attention takes main's one head-128 launch at group 5

gpu: relu_sqr returns with the nemotron_h body that calls it

gpu-gates: gate_nemotron_e2e takes the shared runner's argmax

gpu-gates: gate_nemotron_serve reads the trained context through generate::trained_ctx

tools: gate-paths' act_rule.rs and flip.rs rows name gate_nemotron_e2e, which reaches them through the shared e2e runner

tools: gate-paths' emb2.rs and gelu.rs rows name the nemotron gates that reach them on main

model, gpu, gpu-gates: review fold: the nemotron type pins, the shared blk name, one-slot plan, the keep rule in its seat, no copied hparams, the seat records the model's held context

model, gpu, gpu-gates: review fold: the nemotron pins read every type of their tensors, an absent conv bias is a zero bias, the seat's context is the model's, the default short refusal

gpu, tools: a file tensor never takes a derived weight's name; gate-paths' nemotron_h row names the gates coverage now reaches it through

model, gpu, gpu-gates: review fold: the nemotron vector and gain pins (f32, bf16), the serve gate's clauses in one loop, no blk literal outside arch/