test: stabilize Gemma4 26B-A4B MTP GSM8K test with deterministic inference + tuned threshold (#26653)

Co-authored-by: Cursor <cursoragent@cursor.com>

test/registered/spec/test_gemma4_mtp_26b_a4b_extra.py

@@ -31,10 +31,11 @@ GSM8K_NUM_THREADS = 128
 GSM8K_SCORE_MARGIN = 0.03
 SERVER_LAUNCH_TIMEOUT = DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 3
 
-# Initial values are seeded from current Gemma4 GSM8K observations in the
-# cookbook. Replace each top-k entry with exact MTP first-200-sample scores as
-# CI calibration data becomes available.
-OBSERVED_GSM8K_SCORES = {1: 0.450, 3: 0.450}
+# Calibrated from deterministic-inference GSM8K runs (200 examples, 5-shot,
+# greedy, triton, TP=2). With --enable-deterministic-inference the per-topk
+# score is reproducible run-to-run (std=0 over N=20): topk=1 -> 0.445,
+# topk=3 -> 0.440.
+OBSERVED_GSM8K_SCORES = {1: 0.445, 3: 0.440}
 GSM8K_SCORE_THRESHOLD = min(OBSERVED_GSM8K_SCORES.values()) - GSM8K_SCORE_MARGIN
 ACCEPT_LENGTH_THRESHOLD = 1.5
 
@@ -84,6 +85,9 @@ class TestGemma4MTP26BA4B(CustomTestCase):
             "--max-total-tokens",
             "32768",
             "--skip-server-warmup",
+            # Batch-invariant kernels make the GSM8K score reproducible
+            # run-to-run; without this the topk=3 score swings ~0.33-0.50.
+            "--enable-deterministic-inference",
         ]
         if TENSOR_PARALLEL_SIZE > 1:
             args += ["--tp-size", str(TENSOR_PARALLEL_SIZE)]