From 1e1dd106a2268807298dc8b9a725d989b6e1fc53 Mon Sep 17 00:00:00 2001 From: adibarra <93070681+adibarra@users.noreply.github.com> Date: Mon, 4 May 2026 01:16:41 -0500 Subject: [PATCH] fix(db): drop tp/ep/num_gpu >= 1 checks to allow aggregated multinode configs --- .../004_relax_decode_zero_constraints.sql | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) create mode 100644 packages/db/migrations/004_relax_decode_zero_constraints.sql diff --git a/packages/db/migrations/004_relax_decode_zero_constraints.sql b/packages/db/migrations/004_relax_decode_zero_constraints.sql new file mode 100644 index 00000000..612a5fcd --- /dev/null +++ b/packages/db/migrations/004_relax_decode_zero_constraints.sql @@ -0,0 +1,17 @@ +-- ============================================================ +-- CONFIGS — allow prefill/decode tp / ep / num_gpu = 0 +-- ============================================================ +-- +-- Aggregated multinode disagg runs can produce a config with no decode +-- workers for low-conc sweeps (e.g. conc=1 on dsv4 run 25296668638 had +-- decode_tp=0, decode_ep=0, num_decode_gpu=0 with all 8 GPUs on prefill). +-- The original `>= 1` checks rejected these rows during ingestion, so the +-- conc=1 point silently disappeared from the official chart. Drop the +-- prefill-side mirrors too for symmetry. + +alter table configs drop constraint configs_decode_tp_positive; +alter table configs drop constraint configs_decode_ep_positive; +alter table configs drop constraint configs_num_decode_gpu_positive; +alter table configs drop constraint configs_prefill_tp_positive; +alter table configs drop constraint configs_prefill_ep_positive; +alter table configs drop constraint configs_num_prefill_gpu_positive;