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 9, 02:30 UTC
[CI] Add DeepSeek V3.2 nightly eval (#33566)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Matthew Bonanni committed
Feb 2, 2026 at 11:10 UTC
9f8cb81b44ce2433facc60dec709dc8bf116c315
3 files changed
+24
tests/evals/gsm8k/configs/DeepSeek-V3.2-DP.yaml
new
+11
@@ -0,0 +1,11 @@
1
+model_name: "deepseek-ai/DeepSeek-V3.2"
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-V3.2-TP.yaml
new
+11
@@ -0,0 +1,11 @@
1
+model_name: "deepseek-ai/DeepSeek-V3.2"
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
+2
@@ -1,2 +1,4 @@
1
DeepSeek-R1-TP.yaml
2
DeepSeek-R1-DP.yaml
3
+DeepSeek-V3.2-TP.yaml
4
+DeepSeek-V3.2-DP.yaml