feat(glm): add cuDNN DSA backend for GLM-5.2 (#3127)

* feat(glm): add cuDNN DSA kernels for GLM-5.2

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(moe): use reentrant checkpoint ordering

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(loss): bound long-context cross entropy memory

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(glm): bound TileLang top-k workspace

Restore the established masked cross-entropy path after the recompute loss changed the CP8 cuDNN optimizer trajectory. Chunk TileLang top-k across query rows only for large score matrices so CP8 fits without changing CP1 selection semantics.

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(thd): derive packed padding masks from lengths

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* fix(glm): exclude packed padding from cuDNN DSA

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* chore(distributed): drop unrelated comment change

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* docs(glm): add Tulu3 cuDNN recipe

Replace the validation-only mock benchmark recipes with a public Tulu3 4K
cuDNN example matching the validated topology. Document the recipe and
exclude it from release execution until the standard image includes FlashMLA.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* chore(glm): drop mock validation plumbing

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* refactor(glm): narrow PP mask handling

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* chore(glm): drop mock collater support

Remove the pre-batched THD collater branch after dropping the validation-only
mock recipes and dataset plumbing from this PR.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* chore(glm): defer lock updates to CI

Restore the checked-in lockfiles and let the repository lock-generation bot
produce the dependency updates from pyproject.toml.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* Update uv lock

Signed-off-by: NeMo Bot <nemo-bot@nvidia.com>

* chore(glm): note PP mask capability follow-up

Document that the model-type exceptions should move to a shared mask-ownership
capability once that contract is standardized.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* Update docs/model-coverage/llm/thudm/glm5-moe-dsa.mdx

Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>

* test(glm): configure IndexShare for cuDNN PP carry

The pipeline metadata contract only includes the top-k carry for models
with shared indexers. Make the cuDNN fixed-width carry test opt into that
configuration explicitly, matching the TileLang coverage.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: HuiyingLi <willwin.lee@gmail.com>

---------

Signed-off-by: HuiyingLi <willwin.lee@gmail.com>
Signed-off-by: NeMo Bot <nemo-bot@nvidia.com>
Co-authored-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: NeMo Bot <nemo-bot@nvidia.com>
Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com>