A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 17:45 UTC
[CI][DeepSeek] Add nightly DeepSeek R1 `lm_eval` tests on H200 (#30356)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>
Matthew Bonanni committed
Jan 5, 2026 at 17:17 UTC
276e03b92c168f9457f922d8d957900dd30e5d89
5 files changed
+33
-1
.buildkite/test-pipeline.yaml
+8
@@ -1351,6 +1351,14 @@ steps:
1351
- CUDA_VISIBLE_DEVICES=1,2 VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
1352
- pytest -v -s tests/v1/distributed/test_dbo.py
1353
1354
+- label: LM Eval Large Models (H200) # optional
1355
+ timeout_in_minutes: 60
1356
+ gpu: h200
1357
+ optional: true
1358
+ num_gpus: 8
1359
+ commands:
1360
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
1361
+
1362
##### B200 test #####
1363
- label: Distributed Tests (B200) # optional
1364
gpu: b200
tests/evals/gsm8k/configs/DeepSeek-R1-DP.yaml
new
+11
@@ -0,0 +1,11 @@
1
+model_name: "deepseek-ai/DeepSeek-R1"
2
+accuracy_threshold: 0.95
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+server_args: >-
7
+ --enforce-eager
8
+ --max-model-len 4096
9
+ --data-parallel-size 8
10
+ --enable-expert-parallel
11
+ --speculative-config '{"method":"mtp","num_speculative_tokens":1}'
tests/evals/gsm8k/configs/DeepSeek-R1-TP.yaml
new
+11
@@ -0,0 +1,11 @@
1
+model_name: "deepseek-ai/DeepSeek-R1"
2
+accuracy_threshold: 0.95
3
+num_questions: 1319
4
+num_fewshot: 5
5
+startup_max_wait_seconds: 1200
6
+server_args: >-
7
+ --enforce-eager
8
+ --max-model-len 4096
9
+ --tensor-parallel-size 8
10
+ --enable-expert-parallel
11
+ --speculative-config '{"method":"mtp","num_speculative_tokens":1}'
tests/evals/gsm8k/configs/models-h200.txt
new
+2
@@ -0,0 +1,2 @@
1
+DeepSeek-R1-TP.yaml
2
+DeepSeek-R1-DP.yaml
tests/evals/gsm8k/test_gsm8k_correctness.py
+1
-1
@@ -78,7 +78,7 @@ def test_gsm8k_correctness(config_filename):
78
eval_config["model_name"],
79
server_args,
80
env_dict=env_dict,
81
- max_wait_seconds=600,
81
+ max_wait_seconds=eval_config.get("startup_max_wait_seconds", 600),
82
) as remote_server:
83
server_url = remote_server.url_for("v1")
84
print(f"Server started at: {server_url}")