[yidingcheng0206]

[Feature] GLM-5.3 internal prefill checkpoints

What changes

• Enforce prefill_save_alignment for both internal and prompt-end local
  prefill checkpoints. Default 8192 is unchanged; smaller endpoint saves
  previously admitted are now rejected, including with IC OFF.
• Add an optional native FlashKDA adapter using its existing
  checkpoint_state / checkpoint_offsets outputs.
• Add opt-in internal checkpoint capture for GLM-5.3 AR and deepseek_mtp,
  preserving reservation, restore/pinning, publication and cleanup.
• Retain MTP lookahead checks and absolute-position conv/recurrent ring writes.
  No FlashKDA, KPool, convolution, or TileLang decode kernel sources change.

IC OFF: forward [start,H) -> save S_H -> forward [H,E)
IC ON:  forward [start,E), exporting S_H; runtime advances to S_E

The same checkpoint H is preserved, rather than discarded for a faster run.
PR #5037 is already in main and is not reapplied.

Enablement and scope

export LMDEPLOY_KDA_PREFILL_BACKEND=flash_kda
export LMDEPLOY_INTERNAL_PREFILL_CHECKPOINT=1

Requires local prefix caching, BF16 inputs, FP32 recurrent state and
FlashKDA with the native checkpoint ABI. Validated FlashKDA:
ee0be888cd0e972f9409bf53756f8c38c6652173.

IC defaults to OFF; FLA remains the default prefill entry. Current scope is
exclusive single-sequence text prefill, one internal checkpoint, within the
existing prefill budget. Required endpoint saves and ineligible requests keep
the original path. IC does not support connectors/Mooncake, multimodal,
scoring/routed-expert output, or mixed local decode/prefill.
Intervals within the PCG cap are not merged; IC currently executes outside
prefill PCG while preserving decode graphs. Extra state export and larger
forwards have memory/time costs.

Validation

Submission head: a447f3768c5a7818563f85a40b9e4d5e55a0ec78.
Tested implementation: 170bc8b33b38412d0cbdb56b0e08ed041d804640 on
main 52e57d8f0d3a00fc53010f34bb5fe9109e73c560. The only change since
validation is documentation cleanup; implementation and tests are identical.

Check	Result
Targeted CPU regression	366 passed, 9 skipped, 1 CUDA-dependent test deselected
CUDA restoration/ring/routing regression	13 passed, none skipped; non-MTP and K=1/3/5
TP1/DP8/EP8 model check, K=3	IC OFF/ON: 11 requests each; 22/22 final answers correct
Checkpoint evidence	Expected split/merged shapes, save/restore and lookahead rejection passed
Decode graph	Eight workers per service logged capture and use_graph=True

Model checks used identical data, native FlashKDA, BF16 KV, prefix ON, prefill16384,
alignment8192, PCG2048, C1, greedy/seed42 and natural EOS with a 512-token cap.
Full responses matched verbatim in only 4/11 pairs; this is not bitwise
equivalence or full IC accuracy acceptance. No separate replay trace was added.
Existing Transformers-version and MP-shutdown warnings were retained.
Both services stopped and GPUs were released.

No new full GSM8K, performance matrix or production-stress claim is made.
Historical performance results are not substituted for this submission-base check.

Review order

Four commits: alignment fix → native backend → non-MTP IC → MTP IC.
User-facing details: docs/en/advance/internal_checkpoint.md.

[RunningLeon]

for this type hint, should it be 'PrefillStateCapture' ?

[caikun-pjlab]

please add zh_cn docs