port(KERNEL-ATTN-MLA-SPARSE): MLA chunked context is the pre-b38e111d3e equal-share grid, not per-request packing

Row: KERNEL-ATTN-MLA-SPARSE

Found by #2647 (PORT-NOW wave
PORTQ-3, tranche 81-120). Nothing was executed: this is a static reading of the
tree at e24ec8bfd against upstream b38e111d3e (vllm#50613), read with
git show at that revision.

What is missing

Upstream b38e111d3e replaces MLA chunked-context planning. The old scheme gives
every with-context prefill an equal share of the workspace and lays chunks out on
a [num_chunks, num_prefills] grid, so a chunk can contain requests that
contribute zero tokens to it. The new plan_mla_context_chunks packs requests in
order and splits on an aligned boundary when the workspace fills, so a chunk
covers a contiguous run of requests and charges attention, up-projection and
merge only to them.

This tree carries the pre-commit scheme, line for line.
include/vllm/model_executor/layers/attention/mla_chunked_context.h:129-130:

int64_t max_context_chunk = workspace_size / num_prefills_with_context;
max_context_chunk = (max_context_chunk / page_size) * page_size;

That is the equal-share division the commit deletes. The header's own comment at
:35-40 records chunks containing zero-token requests as a live hazard, which is
exactly the condition upstream's packer removes by construction.

Size

Roughly 400-600 lines over 6-8 files: rewrite BuildMlaChunkedContext, change
the MlaChunkedContextMetadata / MlaChunkDeviceMetadata shape to carry a
request slice and a continuation flag, rework ComputeMlaPrefillContext
(mla_chunked_context.h:254-300) into a request-slice merge, update the three
call sites (src/vllm/model_executor/models/deepseek_v2.cpp:228,
minicpm3.cpp:119, kimi_linear_device.cpp:2028) and extend
tests/vt/test_ops_mla_chunked_context.cpp.

The DCP half of the upstream diff is inert here: dcp_world_size is 1 at
src/vllm/v1/core/sched/scheduler.cpp:280 and the header states it ports only
the non-DCP branch.

Expected observable effect

Workspace efficiency and chunk count, not tokens. Merging is order-associative
and vt::MergeAttnStates already handles the double -inf case, so token-exact
output is the expectation. That is an argument from the code, not a
measurement — nothing was run for this issue, and whoever ports it owes the
differential against the pinned oracle.

Owner

KERNEL-ATTN-MLA-SPARSE (.agents/kernel-matrix.md:140, state PARTIAL, claim
CLAIM-MLA-DEEPSEEK) explicitly owns "the PREFILL PATH and CHUNKED-CONTEXT LOOP
(W5)", which is this code.

Not in scope here

This issue does not advance the parity pin. b38e111d3e is inside the
5559679229..e126687a9a window that #2611
owns.

Start by comparing include/vllm/model_executor/layers/attention/mla_chunked_context.h with upstream b38e111d3e, focusing on BuildMlaChunkedContext and ComputeMlaPrefillContext. Then trace the call sites in deepseek_v2.cpp, minicpm3.cpp, and kimi_linear_device.cpp, and extend tests/vt/test_ops_mla_chunked_context.cpp; done means matching the pinned oracle's chunking while preserving token-exact output.