[valarLip]

Summary

Structure

• Model registry (support_model_arch_dict, mono_decode_installers) moves from model_runner.py to atom/models/registry.py; docs and skills point there.
• The DSA indexer op sparse_attn_indexer moves from models/deepseek_v2.py to model_ops/indexer/indexer_forward.py.
• model_ops/indexer/schedule.py is the one classification of which layers reuse an earlier layer's top-k. The model, the MLA layer (which layer owns a selection and its work plan) and the PP runner (which stage hands its selection on) all read it; the runner's hand-copied version is gone.
• The MLA metadata builder decides sparse prefill once a step (attn_metadata.run_sparse); MLA, the indexer, PCP and the sglang bridge follow it instead of each re-deriving max_seqlen_k > index_topk.
• flatten_dsa_config publishes a nested dsa_config as the flat index_* fields; index planes are allocated only for the layers it names.
• GDN causal-conv metadata is built from host-side sequence lengths (no device-to-host copy).
• atom.model_ops.indexer loads its names on first use, so its kernel-free submodules (schedule) import without AITER.
• Kimi-K3 image request inputs move to atom/models/kimi_k3/image_processing.py, so building them does not import the vision model or AITER.

Indexer and sparse MLA

• NoPE indexer support: index_rope_dim, a fused FP8 query-quant and key-cache writer (model_ops/indexer/nope.py), and keys/head weights projected by the owning attention. The unpooled glm5_next path now uses the shared indexer.
• A single-rank FP8 prefill subtracts request bases inside the slot conversion instead of rewriting the top-k tensor.
• Sparse-prefill work buffers are sized for prefill's own gathered head width (it can differ from decode's).
• A BF16 sparse-prefill row stays in one work unit, so its result does not depend on batch composition.
• Small BF16 persistent DCP prefills split 128 heads into two independent groups (attentions/sparse_mla_head_groups.py); the builder owns one buffer set and binds it to the layers that own a selection, and only when no layer shares top-k.
• BF16 ASM prefill pads 17..127 query heads to the 128 the kernel accepts.
• Stable top-k whenever tensor parallelism is on.

DCP

• The prefill query gather fuses rank-to-token reordering with head padding.
• zero_nonfinite_rows checks 32 rows per program and exits before touching output when all are finite; the prefill indptr prefix sum is one launch; the decode candidate exchange uses the custom all-gather.
• Ownership and slot mapping are shared helpers for the count and compact kernels.
• Fixes: the local logits width covers partial interleave groups; the prefill key gather ignores padded block-table rows.

Loading

• A layer the checkpoint stores in a plain float dtype is treated as unquantized, decided from the safetensors headers (no tensor is loaded).
• Loading warns when a checkpoint float tensor is converted to another float dtype.

Kimi-K3

• KDA control parameters, short convolutions and the output norm are FP32; e_score_correction_bias is FP32.
• When the config names DSA layers, those layers fuse the indexer key and head-weight projections into the q_a/kv_a projection (K3FusedIndexerProjection).
• The BF16 output gate uses the fused sigmoid-mul; BF16 gated norm and attention residual keep one reduction layout at every batch size.
• Vision projections run on the aiter BF16 GEMM.

Other

• MoE: pure tensor parallelism routes pad rows like real rows for now (TODO in _select_logical); a -1 route faults the TP a16w4 MoE path.
• The multimodal chat endpoint returns prompt_token_ids when requested.

Test plan

• ruff / black on every changed file
• Targeted unit and GPU tests (807 passed; the two test_row_view_* cases failed only because aiter could not compile module_gemm_a8w8_blockscale_bpreshuffle in that test environment): index top-k schedule, NoPE writer and FP8 MQA, sparse MLA prefill (batch-composition independence, head groups), DCP topk / merge / prefill query gather / LSE mask, MLA index cache, glm5_next kpool, FP4 indexer, DeepSeek-V4.1 indexer and paged scoring, Kimi-K3 activations and norm batch invariance
• Full CI-scope unit tests (tests/ minus tests/plugin and test_openai_server.py); the only failures should be ones that already fail on main

[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 2776 --add-label <label>