From e832f820f21527ff68edd98cb527ca4f892c0c0a Mon Sep 17 00:00:00 2001
From: tensorrt-cicd <90828364+tensorrt-cicd@users.noreply.github.com>
Date: Mon, 15 Jun 2026 09:04:03 -0700
Subject: [PATCH] [nvbugs/6278377][fix] Add BF16 baseline accuracy reference
 for Step-3.7-Flash GSM8K

The TestStep3_7::test_auto_dtype[tp_size=8-ep_size=8] case loads the
unquantized Step-3.7-Flash checkpoint (no quant_algo, no spec_dec_algo),
but references/gsm8k.yaml only contained quantized entries
(FP8_BLOCK_SCALES, NVFP4) for stepfun-ai/Step-3.7-Flash. The lookup in
AccuracyTask.get_hypothesis_testing_params raised
ValueError("Not registered specs: ...").

Add a no-quant baseline entry (accuracy: 88) so the auto-dtype lookup
matches. Verified on B300 8xGPU: measured 90.26, threshold 84.797.

Signed-off-by: tensorrt-cicd <90828364+tensorrt-cicd@users.noreply.github.com>
---
 tests/integration/defs/accuracy/references/gsm8k.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tests/integration/defs/accuracy/references/gsm8k.yaml b/tests/integration/defs/accuracy/references/gsm8k.yaml
index 489beae5acfd..947f9b892207 100644
--- a/tests/integration/defs/accuracy/references/gsm8k.yaml
+++ b/tests/integration/defs/accuracy/references/gsm8k.yaml
@@ -488,6 +488,7 @@ zai-org/GLM-5-FP8:
 # (mtp_nextn=3) is lossless for both FP8 and NVFP4 and reuses the non-spec
 # baseline accuracy.
 stepfun-ai/Step-3.7-Flash:
+  - accuracy: 88
   - quant_algo: FP8_BLOCK_SCALES
     accuracy: 88
   - quant_algo: FP8_BLOCK_SCALES