feat(dllm): add SCDD training and sampling (#3354)

* feat(dllm): add mixed mask+uniform corruption kernel

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* feat(dllm): add SCDD schedule and NELBO loss

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* feat(dllm): add scdd training strategy

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* feat(dllm): add SCDD sampler

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* docs(dllm): document SCDD fine-tuning

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* test(dllm): add cpu smoke for scdd training path

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* test(dllm): add single-gpu scdd smoke test

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* docs(dllm): use openreview link, fix scdd train command

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* test(dllm): use llama fixture for scdd smoke

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* perf(dllm): chunk the SCDD vocab reduction

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* feat(dllm): make llada_scdd the go-to SCDD recipe

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* test(dllm): cover scdd sampler edge branches

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* docs(dllm): use uv run for dllm commands

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* Update docs/guides/dllm/finetune.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>
Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* Update docs/guides/dllm/finetune.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>
Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* Update docs/guides/dllm/finetune.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>
Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* Update docs/guides/dllm/finetune.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>
Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* Update docs/guides/dllm/finetune.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>
Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* Update docs/guides/dllm/finetune.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>
Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* Update docs/guides/dllm/finetune.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>
Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* fix(dllm): validate scdd corruption bounds and keep t on the grid

check mask_prob + uniform_prob <= 1 in corrupt_mix, clamp the uniform
draw so a clean [MASK] cannot produce an out-of-range id, and snap the
top diffusion time to the previous grid point instead of 1 - 1e-4.

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* docs(dllm): fix scdd sampler wording on remasking

visible tokens are never sent back to [MASK]; the reverse posterior puts
no mass there. reported by the scdd authors.

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* docs(dllm): drop external repo paths from scdd provenance notes

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* test(dllm): align scdd smoke script and config with repo conventions

shebang on the first line, wrap pytest in coverage run like the other L2
scripts, and use distributed.strategy instead of the FSDP2Manager target.

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* docs(dllm): drop duplicated idlm config line

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* fix(dllm): scdd final denoise removes uniform-channel noise, not just masks

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

* pin sdpa in the scdd smoke, fp32 breaks flash attn

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>

---------

Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>
Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>