[gbyu-amd]

Stacked on #2647 (base gyu/tp-across-nodes); retarget to main once #2647 merges. This change needs #2647's dcp_persistent_supported(), which is true on gfx1250 when AITER_MLA_DECODE_PS1_FLYDSL=1. On main it is still gfx950-only.

Problem

Speculative decode with DCP is gated to gfx950. A q > 1 verify pass over a round-robin DCP KV shard needs a persistent MLA kernel that applies the causal mask on global positions. On gfx1250, --method dspark with --decode-context-parallel-size 8 stops at startup with Speculative decode + DCP is only supported on gfx950.

gfx1250's PS1 kernel (AITER_MLA_DECODE_PS1_FLYDSL=1) takes g_kv_indptr and applies that mask (mla_a8w8_qh96_1tg_16mx6_64nx1_lse_cprr_ps, ROCm/aiter#5844). But it takes a multi-token query only at 16 or 96 gathered heads, and aiter does not fold other widths onto it. Under TP8 + DCP8, Kimi-K3's DSpark draft gathers 8 × 8 = 64 heads, so its block pass fell through to the page-64 path.

Changes

• atom/config.py: gate speculative decode + DCP on dcp_persistent_supported() instead of gfx == "gfx950".
• atom/model_ops/attention_mla.py: mla_dcp_kernel_num_heads(..., multi_token=True) pads a multi-token decode on gfx1250 to the next PS1 width, 16 or 96. The draft's 64 becomes 96; the target's 96 is unchanged. MLAAttention passes multi_token whenever a speculative config is set, for the target's verify pass and the draft's block pass alike.
• atom/model_ops/attentions/aiter_mla.py: the metadata builder plans the same width.

Validation

Kimi-K3 on two MI455X nodes (4 GPUs each): TP8 + DCP8, MoE TP8, DSpark with 7 draft tokens (Inferact/Kimi-K3-DSpark), no forced acceptance. GSM8K 5-shot, all 1319:

ATOM under test	All-reduce between the nodes	strict-match	Draft tokens accepted	Tokens per forward
this commit on gyu/k3-455-tp8 (2026-10-19)	RCCL	0.9545 ± 0.0057	53.6%	4.75
this commit on the ATOM of image rocm/fw-bringup:gfx1250-atom-k3-20261027 (main 481c2bf + #2647 + #2737 + #2557 + #2554; 2026-10-27)	aiter custom all-reduce (symm_mem)	0.9530 ± 0.0058	53.2%	4.73

• Without DSpark, the first configuration scores 0.9621 ± 0.0053.
• aiter: the first run used cc8d4ff12 (the grouped-MoE K-tail fit of [FlyDSL] [MoE] Fit the gfx1250 grouped GEMM k-tile to K, and refuse one that does not divide it aiter#6706); the second used main 7061cbd88, which has [FlyDSL] [gfx1250] Fix the dropped K tail and the bias column offset in the grouped MoE TDM GEMM aiter#6663.
• Capturing the q_len 8 graphs with RCCL over the scale-up fabric needs NCCL_GRAPH_REGISTER=0.
• On 2026-10-27, a 28-minute MLPerf Endpoints agentic run with this commit (prefix caching, DSpark 4, MoE TP8) ended in a GPU hang on one node. MoE TP8 runs without DSpark have also brought nodes down (known issues of the Kimi-K3-455-TP8 recipe), so the cause is open.