[v0.26.0rc][BugFix] resolve shape mismatch in DCP PD-disaggregated recomputation (#16487)

[lllyys]

What this PR does / why we need it?

Cherry-pick of #16487 to releases/v0.26.0rc.

Fix shape mismatches when DCP is enabled with PD-disaggregated recomputation:
short recomputation requests (including last-token recomputation) are now
correctly classified as decode requests in the attention, MLA, and SFA
metadata builders. DCP state is cached during builder initialization and the
builder's own configuration is used, so metadata construction works outside
the current-config context.

Original PR: #16487 (main commit cab3719)

Backport conflicts and how they were resolved

The cherry-pick conflicted in 4 files (this branch carries 165 backports of
its own; test_attention_cp.py / test_mla_v1.py live under
tests/ut/attention/a2/ here and were applied via rename detection).

1. vllm_ascend/attention/mla_v1.py — 3 hunks:
   • Imports: kept this branch's imports; added only
     is_pd_decode_recompute_scheduler_enabled (already present in
     utils.py from an earlier backport).
   • Builder __init__: main's hunk carried the whole PCP block
     (pcp_size/pcp_enabled/pcp_rank), which does not exist on this
     branch; added only self.dcp_enabled = enable_dcp().
   • Decode/prefill split gate: main changed
     not (pcp_enabled or dcp_size > 1) to add the DCP + PD-recompute
     override. This branch's gate is
     context_parallel_metadata is None; resolved as
     context_parallel_metadata is None or (self.dcp_enabled and is_pd_decode_recompute_scheduler_enabled(self.vllm_config))
     — same intent, expressed on this branch's gate. Verified
     AscendMlaDCPMetadataBuilder (mla_cp.py) inherits build() from
     AscendMLAMetadataBuilder, so the DCP MLA path is covered.
2. attention_cp.py — import hunk only: kept this branch's
   memcache_comm_fence import path (main moved that symbol elsewhere) and
   added the helper import. All functional hunks applied cleanly.
3. sfa_cp.py — import hunk only: main's side carried its whole
   utils/weight_switch/pcp import section, none of which exists here; added
   only from vllm_ascend.utils import is_pd_decode_recompute_scheduler_enabled.
   Functional hunks applied cleanly.
4. tests/ut/attention/test_sfa_cp.py — main's version of
   test_sfa_dcp_builder_sizes_replicated_view_from_padded_block_table is
   parametrized against buffers/patches this branch does not have; applied
   only the PR's intent onto this branch's shape of the test (patch
   enable_dcp, assert it is called exactly once at init and cached).

Adaptations of cleanly-applied test hunks:

• patch("vllm.config.get_current_vllm_config_or_none", ...) →
  patch("vllm.config.get_current_vllm_config", side_effect=AssertionError):
  this branch's helper fallback calls the latter, so patching the former
  would never simulate "no current config context".
• patch.object(builder, "_update_parallel_slot_mapping") →
  _update_dsa_cp_slot_mapping_for_dcp (this branch's method name).
• The MLA "without DCP" sub-test looped over pcp_size in (1, 2) — reduced
  to the single no-DCP case (no PCP in mla_v1 on this branch); it still
  asserts the DCP-only override is never evaluated when DCP is off.

Independent review (Codex)

Reviewed pre-submission with Codex against both branches and both commits.
Verdict: APPROVE WITH NITS; both findings addressed before submission:

1. (minor) The adapted MLA regression test left
   context_parallel_metadata=None, so the is None gate short-circuited
   and the new override was never exercised. Fixed: the test now populates
   context_parallel_metadata, so classification hinges on the override
   itself, and a recompute-off case asserts short extends stay prefills.
2. (nit) A documentation claim about the vLLM pin was wrong (vLLM v0.26.0
   exports both config getters; the patch-target change is needed because
   this branch's helper fallback calls get_current_vllm_config). Corrected.

Codex also confirmed: no functional hunks dropped, no other
split_decodes_and_prefills call sites on this branch need the same fix,
and no main-only code was pulled in.

Does this PR introduce any user-facing change?

No. Fixes failures in DCP-enabled PD-disaggregated recomputation.

How was this patch tested?

• Regression tests from the original PR, adapted to this branch's code shape
  (see above), in tests/ut/attention/a2/test_attention_cp.py,
  tests/ut/attention/a2/test_mla_v1.py, tests/ut/attention/test_sfa_cp.py.
• CI unit tests on this branch.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

---

Tip

💡 Consider Linking a Related Issue or RFC

Your PR title contains the [BugFix] tag, indicating a bug fix or new feature.

Linking a related issue or RFC in the PR description is strongly encouraged — it gives reviewers helpful context and speeds up the review. You can use any of these keywords:

• Fixes #<issue_number>
• Closes #<issue_number>
• Resolves #<issue_number>
• Refs #<rfc_or_issue_number> (for RFCs)

🙏 Thanks for helping us keep the project well-organized!

[gemini-code-assist]

Code Review

Suggested PR Title:

[Attention][Feature] Cache DCP state in metadata builders for PD decode recompute scheduler

Suggested PR Summary:

### What this PR does / why we need it?
This PR caches the DCP (Distributed Context Parallel) enabled state during the initialization of metadata builders (`AscendAttentionDCPMetadataBuilder`, `AscendSFADCPMetadataBuilder`, and `AscendMLAMetadataBuilder`). This cached state, along with the builder's own `vllm_config`, is used to determine whether to treat short extends as decodes when splitting decodes and prefills. This avoids relying on `get_current_vllm_config()` during metadata building, which can run outside the active vLLM config context, and ensures correct behavior when the PD decode recompute scheduler is enabled.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
Added unit tests in `test_attention_cp.py`, `test_mla_v1.py`, and `test_sfa_cp.py` to verify that the decodes/prefills split correctly uses the builder's config and cached DCP state without requiring an active vLLM config context.