From 220c6315d7b4ab9173389e3e3eeec3e38975beae Mon Sep 17 00:00:00 2001 From: Usman Shahid Date: Thu, 3 Sep 2026 20:30:02 +0400 Subject: [PATCH] feat(catalog): add qwen3.8 as another served_as alias for qwen38-dflash2 Requested directly. Recipe ID stays qwen38-dflash2 (unchanged - that's what the gateway's node-scoped routing and this fleet's own deploy tooling key on); served_as gains "qwen3.8" alongside the existing "dflash2", so vLLM will recognize both once this recipe's container is next redeployed with this recipe version. Edited the seed definition (internal/catalog/seed.go), the source of truth SyncSeeds reconciles a running install's catalog against on every startup - not a live API PATCH - since this recipe has never been operator-edited and belongs in git like every other seed. Regenerated recipes/qwen38-dflash2.yaml via `go test ./internal/catalog -run TestPublishedCatalogMatchesSeeds -rewrite-recipes`, the existing mechanism that keeps the published recipe file in sync with the compiled-in seed. Co-Authored-By: Claude Sonnet 5 --- internal/catalog/seed.go | 2 +- recipes/qwen38-dflash2.yaml | 1 + 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/internal/catalog/seed.go b/internal/catalog/seed.go index f0f7506..7bf44ee 100644 --- a/internal/catalog/seed.go +++ b/internal/catalog/seed.go @@ -224,7 +224,7 @@ func Seeds() []recipe.Recipe { { ID: "qwen38-dflash2", Kind: recipe.KindVLLM, Modality: recipe.ModalityText, Model: "Inferact/Qwen3.8-27B-NVFP4", Image: pinnedVLLMDFlash2, - ServedAs: []string{"dflash2"}, + ServedAs: []string{"dflash2", "qwen3.8"}, // Target 24.87 (qwen38's own measured NVFP4 figure) PLUS drafter // 3.58: the drafter is resident too, and a footprint that ignored // it would under-plan by 3.58 GiB. diff --git a/recipes/qwen38-dflash2.yaml b/recipes/qwen38-dflash2.yaml index 7763490..1d66082 100644 --- a/recipes/qwen38-dflash2.yaml +++ b/recipes/qwen38-dflash2.yaml @@ -5,6 +5,7 @@ model: Inferact/Qwen3.8-27B-NVFP4 image: vllm/vllm-openai@sha256:2a7cde230b59f3ce6cab33dd245ba6bee41aa87b38c9fe84f966ff24016813ce served_as: - dflash2 + - qwen3.8 declared: weights_gib: 28.45 kv_gib: 34