From bc704886956c906619d63e609bb0cbb465d05520 Mon Sep 17 00:00:00 2001 From: jsheng_Linkedin <162757399+fortunecookiee@users.noreply.github.com> Date: Wed, 6 May 2026 14:28:39 -0700 Subject: [PATCH] [CI] Temporarily disable marco/mcdse-2b-v1 in test_embedding_models (#24279) Co-authored-by: Claude Opus 4.7 (1M context) --- test/registered/prefill_only/test_embedding_models.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/test/registered/prefill_only/test_embedding_models.py b/test/registered/prefill_only/test_embedding_models.py index 7dc4aa6c8..ad95a84f2 100644 --- a/test/registered/prefill_only/test_embedding_models.py +++ b/test/registered/prefill_only/test_embedding_models.py @@ -40,8 +40,14 @@ register_cuda_ci(est_time=136, suite="stage-b-test-1-gpu-small") MODEL_TO_CONFIG = { "Alibaba-NLP/gte-Qwen2-1.5B-instruct": (1, 1e-5), "intfloat/e5-mistral-7b-instruct": (1, 1e-5), - "marco/mcdse-2b-v1": (1, 1e-5), "Qwen/Qwen3-Embedding-8B": (1, 1e-5), + # Temporarily disable: HF reference path in runners.py runs this Qwen2-VL + # fine-tune with bidirectional attention (the non-sentence-transformers + # branch in _get_sentence_transformer_embedding_model does not pass + # is_causal=True), while SGLang's Qwen2-VL embedding is always causal — + # producing ~0.30 cosine diffs vs HF on short prompts. + # See https://github.com/sgl-project/sglang/actions/runs/25224929325/job/73966043206 + # "marco/mcdse-2b-v1": (1, 1e-5), # Temporarily disable before this model is fixed # "jason9693/Qwen2.5-1.5B-apeach": (1, 1e-5), }