[ChangLi811]

Motivation

SGLANG_HIP_DSV41_INDEX_K_PAGE8=1 (sgl-project/sglang#44006) keeps the ratio-1 index-K pool in aiter's row-group page-8 byte layout, which SGLang's own indexer uses for prefill and wide decode. The monokernels require it off, so a server cannot have both; it is a server-wide switch, and on PD both peers must share it.

Modifications

• The pool's 8-row pages are K2i's virtual pages (8 × the 64-row page + the sub page), so the page tables, the candidate table and the selection's map back to slots are unchanged; only the bytes of a page differ. score_fp4._keys reads a page8 plane as aiter's page-8 walk does; the page-64 branch is unchanged.
• build_k2i(page8) builds that variant; each layer's plane carries its pool's low_ratio_index_k_page8, which picks the build (aiter without the row-group kernel keeps page-64).
• SPEC no longer requires the variable off; the README's PD note names it among the layout both peers must share, as SGLang's PD handshake does not compare it.
• test_deepseek_v41_k2i.py runs each case on a page8 plane too (the same keys packed by aiter's pack_kv_cache).
• Also in the README: the V4.1 usage note no longer says decode graph capture fails without AITER_BF16_FP8_MOE_BOUND=0; on the current image (aiter 285579b62) SGLang's decode starts and captures without it. The command keeps the variable, as the serving tests do.

Accuracy Tests

K2i kernel tests (24): FULL logits on a page8 plane bit for bit with the page-64 plane's and within k2i.logits.rel of SGLang's and fp64; a REINDEX consumer's top-512 as a set equal to SGLang's topk_within_candidate_blocks_page8; the two-level pipeline and a graph replay.

End to end: DeepSeek-V4.1-Flash, one MI355X node, rocm/sgl-dev:v0.5.22-rocm10-mi35x-20261025, SGLang aa6e51982 (a local merge of #43465, #44730, #44031 and #46036), PAGE8=1 with aiter's row-group kernel, TP4 + prefill CP4 + DSpark block 5; GSM8K: sgl-eval, 1319 questions, max tokens 4096; on and off on fresh servers:

KV	monokernels	GSM8K	accept length
FP8 unified	on	0.967	3.54
FP8 unified	off	0.961	3.51
paged	on	0.965	3.55
paged	off	0.961	3.51

TP2 + CP2 + FP8 unified: 0.965 on, 0.963 off. With PAGE8=0 the same server scores 0.964 (the page-64 path unchanged).

GSM8K above runs 64 threads, so most verify steps are wider than 8 requests and take SGLang's path. With AgentX's c1 recipe (max-running-requests 2, its 19 environment variables, SGLANG_HIP_DSV41_INDEX_K_PAGE8=1), every verify step runs on the monokernels; on #70 (now in main) plus this PR, #78 and #79 (DeepSeek-V4.1-Flash-1017 weights, fresh servers):

config	on	off
TP4, paged	0.967 / 3.41	0.967 / —
TP4 + prefill CP4, FP8 unified	0.967 / 3.41	0.963 / —
TP2, paged	0.967 / 3.39	0.968 / 3.40
TP2 + prefill CP2, FP8 unified	0.968 / 3.41	0.967 / 3.39

(GSM8K / DSpark accept length. TP4 "off" accept lengths are left out: with the recipe's --disable-shared-experts-fusion, SGLang's own path accepts less at batch 1–3, which --enforce-shared-experts-fusion, as the README uses, restores.)

PD

MoRI PD (enabled by #70) over two MI355X nodes: P = TP4 + prefill CP4 + DSpark without the monokernels, D = TP4 + DSpark; SGLANG_HIP_DSV41_INDEX_K_PAGE8=1 on both peers; P and D started afresh for every row; GSM8K through the PD router.

KV	D's monokernels	GSM8K, 64 threads	GSM8K, 8 threads	accept length, 8 threads
FP8 unified	on	0.964	0.966	3.54
FP8 unified	off	0.962	0.962	3.52
paged	on	0.967	0.964	3.55
paged	off	0.963	0.964	3.53

At 8 threads every verify step of the D runs on the monokernels when they are on. Greedy prompts (short, 15k–100k) and 50 long needles pass on every row; no GPU faults on either peer.

Speed Tests and Profiling

c1 TPOT on the FP8 unified server above: 1.17 ms on, 1.78 ms off.

Checklist

• pre-commit run --all-files passes (ruff, isort, codespell).
• Tests under test/ cover the change.
  • CPU, against SGLang aa6e51982: pytest test/unit passes 262, as on main, including the deepseek_v41 contract, which also passes against SGLang main (c6c753395). The 7 failures and 3 collection errors (GLM-5.3, MiniMax and DeepSeek-V4-Pro tests) are the same on main and unrelated to this PR. The K2i change is covered by test_deepseek_v41_k2i.py on the GPU (above).
• The SGLang commit or PR this was tested against is named above.
• Accuracy and speed results are attached when kernel or forward code changes.
• A new model adds one line to .github/CODEOWNERS that names its owner. (No new model.)