[DCP][Kimi-k3] Decode on the gathered head width

[whx-sjtu]

Summary

After ROCm/aiter#4964, we can run q_head 96 mla cprr directly without padding to 128.

Two changes, the first two on the Kimi-K3 tp8 + dcp8 decode path.

Drop the MLA decode head padding (attention_mla.py). mla_dcp_kernel_num_heads rounded the gathered width (num_heads * dcp) up to one of (16, 32, 64, 128), on the premise that aiter's fold onto the 16-head kernel desynchronises the DCP round-robin causal mask. That premise no longer holds — aiter now rebuilds qo_indptr and the kv indptrs as part of the fold, so every multiple of 16 is correct. A persistent decode now dispatches on the gathered width as-is, which is what lets K3 tp8+dcp8 (12 × 8 = 96) reach the 96-head kernel instead of padding to 128: 183.6us vs 190.8us at fp8 / ctx 32k / bs 64 / qlen 4.

Fuse add + RMSNorm in the K3 DSpark draft. Thread residual through the draft layers so the norms take their fused add form, in both the native and vLLM model files.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2060 --add-label <label>