[PerryZhang01]

Ready for review. No blocking issues.

A P/D producer with paged state (DeepSeek-V4, MLA, MHA) kept deferred output, so it held each request until its rank's next forward that samples, and forwards of middle chunks sample nothing. Every P/D producer now reports its first token from the forward that samples it, as recurrent-state producers already did. Port of jgong5#3165 (fork branch task-2655 at 4ced162) for jgong5#2655, without its simulator parts.

Dev record

• Found: since [feature] support pp>1 for Kimi-K3 #2365 the scheduler and tokenIDProcessor both take the rule from scheduler.defers_output, and recurrent-state producers and PP already take the non-deferred path, so only the producer branch changes.
• Decided: defers_output is False for every producer and drops its family argument, with the config lookup and imports only that branch used. Consumers and servers without P/D still defer; PP still does not. A producer schedules no drafts, so ScheduledBatch stages the same ids either way.
• Decided: a producer's sampling forward now waits on the host for its own ids (prepare_sampled_ids reads them with .tolist()), as recurrent-state producers' already do. Its throughput cost belongs to the served measurement.
• Surprised: the CPU gate skips tests/test_spec_deferred_status_queue.py, since model_runner imports only with aiter and a device arch. Red and green ran with Triton's target declared gfx950.
• Left undone: the served measurement.

Named result

The 14 s median hold was measured on Compass records in the fork; the served measurement is pending (jgong5#3173).

Gates

• acc5dee: tests/test_spec_deferred_status_queue.py 33 passed with gfx950 declared; 5 fail on main's source.
• acc5dee: tests/ without plugins, 13802 passed; its 125 failed and 99 errors are main's, test for test.
• acc5dee: P/D, PP, scheduler and runner tests with gfx950 declared: 1218 passed, and the same 30 failures as main.
• ruff, black clean.

Evidence

Revert-red: atom/model_engine/{scheduler,model_runner}.py and atom/kv_transfer/disaggregation/types.py from main (0189b35), this branch's test file, gfx950 declared:

tree	result	failing
acc5dee	33 passed
main's source	5 failed, 28 passed	test_every_producer_gives_up_deferred_output[v4], [mla], [mha]: assert True is False; test_a_producer_reports_t0_from_the_forward_that_samples_it: the deferred branch's send_to_cpu_async waits on sync_event=None; test_the_scheduler_resolves_the_runners_deferral[1]: main looks up the model's family (no attribute 'hf_config')

Declared target: A fake rocminfo on PATH and a Triton driver whose only answer is GPUTarget("hip", "gfx950", 64), as the fork's capture tooling does. It is a scratch wrapper around pytest.main, not part of this change.

Regression set, gfx950 declared: test_scheduler.py, test_scheduler_mtp_max_tokens.py, test_scheduler_spec_preempt.py, test_pp.py, test_pp_draft_input_ids.py, test_pd_{chunked_transfer,mla_landing,mla_staging,pp,producer_config,source_claim}.py, test_kv_connector_scheduler.py, test_kv_change.py, test_pause_generation.py, test_postprocess_width.py, test_output_transfer_runner.py, test_h2d_runner_publication.py, test_model_runner_staging_fence.py, test_mtp_draft_kv_allocation.py, test_dsv4_state_replay.py, test_deepseek_v4_transfer_regions.py, test_attn_family.py, test_spec_deferred_status_queue.py. The 30 failures are identical on main and here: 22 in test_postprocess_width.py and 8 in test_output_transfer_runner.py, Current atom config is not set or no _served on a stub runner. The CPU gate skips both files. Without the declared target the set gives 1163 passed, 39 skipped.

Full suite: in a container with aiter and no GPU, unlike the CPU gate's runner, main's tree gives 13801 passed and the same 224 failing or erroring tests.

Overlap: No open ROCm/ATOM PR changes defers_output: of 363 open PRs, the 94 touching scheduler.py, model_runner.py or disaggregation/types.py were read for it.

Fork measurement, jgong5#2655: in two DeepSeek-V4-Pro DPA-TP8 1P1D runs at c=256, output prefill to release p50 14.7 s and 14.4 s, and no first token reached the client before its release.

🤖 Generated with Claude Code

[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 3203 --add-label <label>