From 8b0f7d4e60e6582efd45fa1f1d3472e615e75644 Mon Sep 17 00:00:00 2001 From: chenfeiz0326 <203214996+chenfeiz0326@users.noreply.github.com> Date: Fri, 18 Sep 2026 08:38:26 -0700 Subject: [PATCH] [None][test] Disable ctx overlap scheduler for deepseek-v4-pro con8 disagg config The ctx_only perf-sanity case generated from this config measures 25% higher throughput with the ctx worker's overlap scheduler disabled (7.82/7.75 vs 6.20/6.25 req/s, 22 vs 33 iterations, 2 reps per arm on GB300/oci-aga). --max-concurrency 8 makes this a closed loop, so cadence is set by the permit round-trip. Deferred teardown keeps a request resident as GENERATION_TO_COMPLETE for one extra iteration before its response is sent, making the round trip 3 iterations instead of 2. Sustained cadence is then 8/3 = 2.67 forwards per iteration against a 4/iteration cap, and matching overlap-off would require 12 concurrent requests where the config specifies 8. #18457 is merged and active in the measured overlap-on arm; it removes the phantom admission charge (summed charge 86 -> 3, exactly the overlap-off value) but not the extra iteration of residency, so it recovers 1 iteration of 34. #18346 removed the unconditional ctx_only override on the rationale that the overlap scheduler works fine for a context-only run. That holds at high concurrency; this config is below the concurrency where it can. A per-config yaml value is used rather than restoring the override so the high-concurrency ctx_only configs keep overlap enabled. Co-Authored-By: Claude Signed-off-by: chenfeiz0326 <203214996+chenfeiz0326@users.noreply.github.com> --- ...o-fp4_8k1k_con8_ctx1_dep4_gen4_tep8_eplb0_mtp3_ccb-NIXL.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-fp4_8k1k_con8_ctx1_dep4_gen4_tep8_eplb0_mtp3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-fp4_8k1k_con8_ctx1_dep4_gen4_tep8_eplb0_mtp3_ccb-NIXL.yaml index 5977b36bcb71..b54449034d31 100644 --- a/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-fp4_8k1k_con8_ctx1_dep4_gen4_tep8_eplb0_mtp3_ccb-NIXL.yaml +++ b/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-fp4_8k1k_con8_ctx1_dep4_gen4_tep8_eplb0_mtp3_ccb-NIXL.yaml @@ -103,5 +103,5 @@ worker_config: kv_transfer_timeout_ms: 600000 backend: NIXL transceiver_runtime: PYTHON - disable_overlap_scheduler: false + disable_overlap_scheduler: true speculative_config: *id001