add eval config for Qwen3-235B-A22B-Thinking-2507-FP8

Signed-off-by: Huamin Li <[email protected]>

Author: hl475

.buildkite/lm-eval-harness/configs/Qwen3-235B-A22B-Instruct-2507-FP8.yaml

@@ -0,0 +1,13 @@
+model_name: "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8"
+tasks:
+  - name: "mmlu_pro"
+    metrics:
+      - name: "exact_match,custom-extract"
+        value: 0.82
+limit: 250 # will run on 250 * 14 subjects = 3500 samples
+num_fewshot: 5
+enforce_eager: false
+max_model_len: 49152
+apply_chat_template: true
+fewshot_as_multiturn: true
+gen_kwargs: "temperature=0,top_p=1,top_k=0,max_gen_toks=6144,until=<|ENDANSWER|>"

.buildkite/lm-eval-harness/configs/models-large-h100.txt

@@ -0,0 +1 @@
+Qwen3-235B-A22B-Instruct-2507-FP8.yaml

.buildkite/lm-eval-harness/configs/models-large-hopper.txt

@@ -21,10 +21,11 @@ def launch_lm_eval(eval_config, tp_size):
     max_model_len = eval_config.get("max_model_len", 4096)
     batch_size = eval_config.get("batch_size", "auto")
     backend = eval_config.get("backend", "vllm")
+    enforce_eager = eval_config.get("enforce_eager", "true")
     model_args = (
         f"pretrained={eval_config['model_name']},"
         f"tensor_parallel_size={tp_size},"
-        f"enforce_eager=true,"
+        f"enforce_eager={enforce_eager},"
         f"add_bos_token=true,"
         f"trust_remote_code={trust_remote_code},"
         f"max_model_len={max_model_len},"
@@ -37,8 +38,13 @@ def launch_lm_eval(eval_config, tp_size):
         limit=eval_config["limit"],
         # TODO(yeq): using chat template w/ fewshot_as_multiturn is supposed help
         # text models. however, this is regressing measured strict-match for
-        # existing text models in CI, so only apply it for mm.
-        apply_chat_template=backend == "vllm-vlm",
+        # existing text models in CI, so only apply it for mm, or explicitly set
+        apply_chat_template=eval_config.get(
+            "apply_chat_template", backend == "vllm-vlm"
+        ),
+        fewshot_as_multiturn=eval_config.get("fewshot_as_multiturn", False),
+        # Forward decoding and early-stop controls (e.g., max_gen_toks, until=...)
+        gen_kwargs=eval_config.get("gen_kwargs"),
         batch_size=batch_size,
     )
     return results

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -392,7 +392,7 @@ steps:
       --ignore=lora/test_deepseekv2_tp.py \
       --ignore=lora/test_gptoss.py \
       --ignore=lora/test_qwen3moe_tp.py
-      
+
   parallelism: 4
 
 - label: PyTorch Compilation Unit Tests # 15min
@@ -1115,7 +1115,7 @@ steps:
   - tests/weight_loading
   commands:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt
-  
+
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
@@ -1157,6 +1157,19 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
+##### H100 test #####
+- label: LM Eval Large Models (H100) # optional
+  gpu: h100
+  optional: true
+  num_gpus: 4
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  commands:
+    - export VLLM_USE_DEEP_GEMM=0
+    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
+
 ##### H200 test #####
 - label: Distributed Tests (H200) # optional
   gpu: h200