[valarLip]

DP attention: one prefill token-graph bucket for the whole group

Under DP attention, a rank with no prefill of its own runs a dummy prefill while a peer prefills. Before this change, each prefilling rank padded to its own bucket, and the dummy sat on the smallest one. A recording's collectives are sized by its bucket, so a dummy replaying 256 beside a peer replaying 2048 hung the group.

ForwardMode._pad_to_token_graph now picks the bucket that holds the group's largest token count and raises every prefilling rank to it, dummy prefill included. The function no longer takes this rank's scheduled_tokens. test_every_prefilling_rank_replays_the_groups_bucket covers the dummy case.

Docs updated to match:

• the dummy_execution docstring;
• ATOM_TOKEN_GRAPH_SIZES in environment_variables.md;
• the token-graph section of compilation_cudagraph_guide.md;
• the dummy-batch section of distributed_guide.md, which named a _execute_dummy_prefill that no longer exists.

Known limitation (unchanged by this PR): without a prefill delayer (ATOM_ENABLE_PREFILL_DELAYER=0, or RapidServe), no rank learns before scheduling that a peer will prefill, so a decoding rank keeps decoding. It then runs eagerly beside a peer replaying a prefill token graph, and the MoE collectives' sizes can disagree. Making decode rows ride the prefill path when any rank prefills would remove the dummy prefill and this case together; that is follow-up work.

CI

• Fix main: tests/test_otel_scheduler.py::test_engine_initializes_tracing_before_ready compared the whole EngineCore startup call list. refactor: one graph mechanism per region, settled DP step shapes, cheaper V4.1 decode prep #2688 added a prefill_token_graphs call, and feat(observability): add opt-in standalone OpenTelemetry tracing #2644 merged on an older base, which turned main red (job). The test now asserts only allocate_kv_cache → initialize → freeze → ready.
• Accuracy: remove DeepSeek-V4-Pro DEP RCCL, DeepSeek-V4-Pro MTP, MiniMax-M2.7, GLM-5-FP8, GLM-5.2-FP8, GLM-5.2-MXFP4 and GLM-5.2-MXFP4 MTP. The atomesh accuracy whitelist drops DeepSeek-V4-Pro MTP too.
• Benchmark: remove GLM-5.2-FP8 and GLM-5.2-MXFP4, along with their atom-benchmark.yaml dispatch inputs, and DeepSeek-V4-Pro's MTP3 / DPA MTP3 variants. test_benchmark_catalog.py no longer names them.
• Agentic nightly: add DeepSeek-V4.1-Flash DSpark5 DTP2 (-tp 2 --enable-dp-attention) and DEP2 (-tp 2 --enable-dp-attention --enable-expert-parallel), both at concurrency 128 on GPUs 0,1.

Test plan

• 18 affected test files (forward mode, scheduler, DP metadata, DSpark, runner padding, benchmark / agentic catalogs, atomesh workflow, OTel): 672 passed, 45 skipped
• The old OTel assertion fails on this base; the new one passes
• validate_catalog.py; the agentic builder expands both new variants to the expected server args
• Agentic nightly DTP2 / DEP2 on GPU

[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 2696 --add-label <label>