[CI/Build] Temporary fix to LM Eval Small Models (#28324)

Signed-off-by: zhewenli <zhewenli@meta.com>
This commit is contained in:
Zhewen Li
2025-11-09 21:08:38 +00:00
committed by GitHub
parent 4a8d6bd168
commit a65a934ebe
3 changed files with 8 additions and 3 deletions
@@ -2,4 +2,7 @@ model_name: "nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16"
accuracy_threshold: 0.45
num_questions: 1319
num_fewshot: 5
max_model_len: 4096
max_model_len: 4096
# Duo stream incompatabilbe with this model: https://github.com/vllm-project/vllm/issues/28220
env:
VLLM_DISABLE_SHARED_EXPERTS_STREAM: "1"
+3 -1
View File
@@ -62,9 +62,11 @@ def test_gsm8k_correctness_param(config_filename, tp_size):
str(tp_size),
]
env_dict = eval_config.get("env", None)
# Launch server and run evaluation
with RemoteOpenAIServer(
eval_config["model_name"], server_args, max_wait_seconds=480
eval_config["model_name"], server_args, env_dict=env_dict, max_wait_seconds=480
) as remote_server:
server_url = remote_server.url_for("v1")