Add Nemotron-Labs-Diffusion (93.3% GSM8K @ 2461 tok/s, 2.9× SOL) - #1

Draft

supercoolgreatcoder wants to merge 16 commits into

mainfrom

nemotron-labs-diffusion

Commits on Jun 16, 2026

[Model] Scaffold Nemotron Labs Diffusion onboarding

hutm
and
claude
committed
cf575f4
[Model][Diffusion] Make self_conditioning optional in shared ModelState

hutm
and
claude
committed
37e5975
[Model] Expose embed_input_ids on NemotronLabsDiffusionForBlockDiffusion

hutm
and
claude
committed
e48c1d7
[Model] Apply hf_to_vllm_mapper in NemotronLabsDiffusion.load_weights

hutm
and
claude
committed
53575d8
[Engine] Auto-populate diffusion_config from HF config when present

hutm
and
claude
committed
8be9ff7
[Model] Custom NemotronLabsDiffusion transformer with Llama-4 Q scaling

hutm
and
claude
committed
b62d22b
[Model] Wire diffusion-mode runtime for Nemotron Labs Diffusion

hutm
and
claude
committed
d87df7a
[Model] Add NemotronDiffusionSampler (FastDiffuser-style top-k unmasking)

hutm
and
claude
committed
ebce8e2
[Model] FastDiffuser EOS-freeze; 82.5% GSM8K via TRITON_ATTN

hutm
and
claude
committed
84dbba1
[Model] Route ar_mode=true through V1 model runner

hutm
and
claude
committed
3e8736a
[Model] FastDiffuser sampler: SGLang-exact softmax + next-block seeding

hutm
and
claude
committed
94e73ff
[Model] Revert next-block seeding (no net accuracy lift on full GSM8K)

hutm
and
claude
committed
97aecc9
[Model] FIX: disable vLLM YARN's auto-mscale for Nemotron (closes accuracy gap)

hutm
and
claude
committed
132c2a0
[Model] Add LinearSpec sampler (2 forwards/block, draft+verify)

hutm
and
claude
committed
fb5ccb5
[Model] LoRA-on-draft for LinearSpec (eager)

hutm
and
claude
committed
2eec881
[Attention] Pass FA4-only kwargs conditionally in flash_attn_varlen_func

hutm
and
claude
committed
f0e4be5

Navigation Menu

Add Nemotron-Labs-Diffusion (93.3% GSM8K @ 2461 tok/s, 2.9× SOL) - #1

Add Nemotron-Labs-Diffusion (93.3% GSM8K @ 2461 tok/s, 2.9× SOL)#1
supercoolgreatcoder wants to merge 16 commits into
mainfrom
nemotron-labs-diffusion

Commits on Jun 16, 2026

[Model] Scaffold Nemotron Labs Diffusion onboarding

[Model][Diffusion] Make self_conditioning optional in shared ModelState

[Model] Expose embed_input_ids on NemotronLabsDiffusionForBlockDiffusion

[Model] Apply hf_to_vllm_mapper in NemotronLabsDiffusion.load_weights

[Engine] Auto-populate diffusion_config from HF config when present

[Model] Custom NemotronLabsDiffusion transformer with Llama-4 Q scaling

[Model] Wire diffusion-mode runtime for Nemotron Labs Diffusion

[Model] Add NemotronDiffusionSampler (FastDiffuser-style top-k unmasking)

[Model] FastDiffuser EOS-freeze; 82.5% GSM8K via TRITON_ATTN

[Model] Route ar_mode=true through V1 model runner

[Model] FastDiffuser sampler: SGLang-exact softmax + next-block seeding

[Model] Revert next-block seeding (no net accuracy lift on full GSM8K)

[Model] FIX: disable vLLM YARN's auto-mscale for Nemotron (closes accuracy gap)

[Model] Add LinearSpec sampler (2 forwards/block, draft+verify)

[Model] LoRA-on-draft for LinearSpec (eager)

[Attention] Pass FA4-only kwargs conditionally in flash_attn_varlen_func