[CI] BugFix Eval Small Models Distributed test for DiffusionGemma (#47980)
Signed-off-by: Markov Ilya <markovilya19@gmail.com> Co-authored-by: Markov Ilya <markovilya19@gmail.com>
Ilya Markov committed
Jul 8, 2026 at 19:00 UTC
b2cf70ea3a29870ab720dd3162cadec3d79a63fe
2 files changed
+6
tests/evals/gsm8k/configs/DiffusionGemma-26B-A4B-it-FP8-dynamic.yaml
+4
@@ -4,6 +4,10 @@ num_questions: 1319
4
num_fewshot: 5
5
startup_max_wait_seconds: 1200
6
use_chat_completions: true
7
+# Diffusion models use a fixed internal temperature schedule and do not
8
+# support per-request temperature or seed overrides.
9
+temperature: 1.0
10
+seed: null
11
server_args: >-
12
--enforce-eager
13
--max-model-len 4096
tests/evals/gsm8k/test_gsm8k_correctness.py
+2
@@ -69,6 +69,8 @@ def run_gsm8k_eval(eval_config: dict, server_url: str) -> dict:
69
use_chat_completions=eval_config.get("use_chat_completions", False),
70
host=host,
71
port=port,
72
+ temperature=eval_config.get("temperature", 0.0),
73
+ seed=eval_config.get("seed", 42),
74
request_timeout_seconds=request_timeout_seconds,
75
gen_prefix=eval_config.get("gen_prefix", ""),
76
)