[unified-memory] Remove the kernel-page multiplier plumbing (5/7)

[caihuali95]

This is one of seven PRs split out of #38592. That PR was +2105/−1366 across 41
files, twelve of them attention backends on the serving hot path, and it interleaved
hardening, a layout change and dead-code deletion in a single diff — a reviewer could not
tell which hunk was safe-by-construction and which changed behaviour. #38592 is retained as the token-major
layout PR (3/7), which is the core of the series.

Stack — each PR is based on the one above it, and every one is green on its own:

#	PR	what it does	size
1	#40326	derive KV row addresses from strides, not shapes	21 files, +733/−118
2	#40327	build paged KV views through one helper	13 files, +704/−47
3	#38592	token-major layout	36 files, +1215/−1108
4	#40328	mark the write loc physical and check it at every write door	45 files, +935/−132
5	#40329	remove the kernel-page multiplier plumbing	17 files, +117/−230
6	#40330	stride the KV translate kernel and route every translate through it	11 files, +361/−159
7	#40331	name the fused KV translate for what it computes	28 files, +119/−121

Review order is the table order. 1 and 2 are no-ops on today's contiguous pools (2 also
fixes how HND pools with more than one KV head are read); 3 is where the behaviour changes
and deserves the most attention; 4 adds the physical write-loc check; 5 is deletion; 6
consolidates the translate; 7 is naming only.

5/7 — remove the kernel-page multiplier plumbing

Base: #40328 (4/7). 5 commits, 17 files, +117/−230.

What and why

3/7 pinned kernel_page_multiplier to 1: under the token-major entry a kernel-facing id
is the physical token id, so there is no per-page block scale left for an id to carry.
This PR deletes the plumbing that existed only to carry it.

Pure deletion. Nothing left behind computes differently — every site either multiplied
by a value that is now always 1, or passed it along.

What goes: the multiplier's keyword arguments and properties across the unified
allocators and the index translator, the * mult at every use site, the stride arguments
that existed only to re-derive the scaled id, and the corresponding parameters in the
kernels that took them. The read-table build loses its scaling step and reduces to a
clamped table lookup. The aiter MLA backend's DCP verify path passed the multiplier to
create_mla_kv_page_table_for_dcp too; it now passes the same arguments as trtllm_mla's.
The write-loc translate kernel's stride argument goes as well: its one caller passed the
page size, so the kernel computes phys * page_size + offset itself. The translator's module
doc now describes two id spaces, virtual and physical.

Both translate implementations are left untouched here. With the multiplier gone,
translate_kv_loc and translate_kv_loc_for_kernel compute the same id for any in-range
non-negative input, differing only in implementation (several torch ops vs one Triton
launch) and in how they treat a negative loc. Folding them is 6/7 — kept separate so
this PR carries no behaviour change at all.

Why it is not merged into 3/7

They are different kinds of change, and the split is deliberate:

• 3/7 is the feature, and its middle commit is where a reviewer must think hard. Seventeen
  files of deletion would dilute exactly the focus the split was made to create.
• They are independently revertible. If this removal turns out to break a consumer we did
  not find, reverting it alone leaves the layout intact — and the layout is what 4/7, 6/7,
  7/7 and downstream work all stand on. Merged, that revert would take the layout with it.

Review shape

Net −113 lines across 17 files, eight of them tests whose edits follow the removal; test
cases that only exercised a multiplier above 1 are converted to the page-size form or
dropped, and the two single-full-attention-layer tests (the multiplier-1 case), which the
removal turned into duplicates, are folded into the general ones. Checking this PR is
mostly confirming that every deleted argument was provably 1.

Validation

Model families exercised — a hybrid of every pool shape the unified pool supports, plus a
non-hybrid control. Perf covers the gated-deltanet hybrid (Qwen3.5-9B), the MLA hybrid
(Kimi-Linear-48B-A3B), the sliding-window hybrid (gpt-oss-20b), the tri-pool hybrid
(Inkling and Inkling-Small), and dense Llama-3.1-8B as a control. Accuracy covers the same
set and additionally the Mamba hybrid, Falcon-H1-7B. Each model is run unified against a
baseline on the same model, backend and configuration, so every delta isolates the pool.

Median inter-token latency, unified vs baseline: +1.30% (measured run-to-run noise floor
for this metric: 1.9%). Median GSM8K delta at n=200: +0.00 pt (binomial delta-sigma ≈ 3.0 pt).

Re-validated on the previous base (0318a8d) after review feedback: the stack tip (7/7) gives +0.95% median inter-token latency over 40 paired configurations (median per-configuration shift against the previous run of the same 40: +0.12 pp) and −0.25 pt median GSM8K over 26, with the full unit scope green.

The median per-model perf shift against the preceding PR's run is −0.09 pp.

End to end

The admitted SM100 backends were run with unified memory at the stack tip; see #40331.

---

CI States

Latest PR Test (Base): ⏳ Run #36944701062
Latest PR Test (Extra): ⏳ Run #36944700814
Latest PR Test (AMD ROCm 10): ⏳ Run #36944701099