Use AITER DSv4 Indexer top-k kernel

[Oseltamivir]

Motivation

Wire the DeepSeek-V4 CSA Indexer path in ATOM's PR #650 branch to the DSv4 Indexer scorer/top-k implementation from ROCm/aiter#2998 instead of materializing fp8_mqa_logits and running PyTorch topk in ATOM.

This is intended to make ATOM exercise the AITER PR #2998 Indexer implementation in downstream InferenceX DSv4 runs.

Technical Details

• Imports aiter.ops.triton.attention.dsv4_indexer.dsv4_indexer_topk.
• Applies the reference Indexer FP4 simulation to Q before scoring: RoPE, Hadamard rotate, FP4 round-trip back to BF16.
• Gathers ATOM's paged Indexer KV cache, converts it into the same rotated FP4-dequantized basis, and builds the dense [num_seqs, max_committed, head_dim] view expected by AITER PR #2998.
• Calls dsv4_indexer_topk(..., offset=0, seq_ids=..., kv_lens=...) so AITER handles batched per-sequence KV lengths and DSv4 causal compressed-entry visibility.
• Adds ATOM's per-token sparse-attention offset afterward to preserve the existing [window || compressed] sparse-attention layout.

Note: this PR intentionally wires the Indexer half of ROCm/aiter#2998 first. The AITER sparse_mqa_sink ABI is paged KV + uniform top-k, while PR #650 currently uses ATOM's ragged packed [window || compressed] sparse-attention layout. Routing sparse attention through AITER should be a follow-up ABI/layout change rather than a per-layer repack.

Test Plan

• Local syntax check: python3 -m py_compile atom/models/deepseek_v4.py atom/model_ops/attentions/deepseek_v4_attn.py
• Downstream InferenceX DSv4 ATOM job will overlay this branch plus Dsv4 sparse indexer aiter#2998.

Test Result

Local syntax check passes. GPU correctness/performance validation is pending in InferenceX.