Skip to content

Commit 05fccac

Browse files
cquil11claude
andcommitted
test: DSv4-Pro B300 vLLM - bump to v0.20.2, fine conc sweep, no cpu offload
- Image v0.20.0-cu130 -> v0.20.2 (cu129; no -cu130 build exists for v0.20.2 on Docker Hub). Same image B200 Kimi runs cleanly on. - Strip all offloading=cpu entries (G+H runs showed both crash and hang on the cpu path; v0.20.2 may or may not fix the underlying DSv4 issue, isolate to none first). - Replace conc-lists with fine sweep [64, 70, 76, 82] across all 4 parallelism configs to find the cliff at 6-conc resolution. 16 jobs total. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
1 parent bcc838d commit 05fccac

1 file changed

Lines changed: 5 additions & 9 deletions

File tree

.github/configs/nvidia-master.yaml

Lines changed: 5 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -2811,7 +2811,7 @@ dsv4-fp8-h200-sglang-mtp:
28112811
# field, so dp-attn=true is used as the existing vLLM script switch for DP4
28122812
# layouts on 4 allocated GPUs.
28132813
dsv4-fp4-b300-vllm:
2814-
image: vllm/vllm-openai:v0.20.0-cu130
2814+
image: vllm/vllm-openai:v0.20.2
28152815
model: deepseek-ai/DeepSeek-V4-Pro
28162816
model-prefix: dsv4
28172817
runner: b300
@@ -2844,14 +2844,10 @@ dsv4-fp4-b300-vllm:
28442844
agentic-coding:
28452845
- duration: 1800
28462846
search-space:
2847-
- { tp: 4, offloading: none, conc-list: [16, 32, 64] }
2848-
- { tp: 8, offloading: none, conc-list: [16, 32, 64] }
2849-
- { tp: 4, ep: 4, dp-attn: true, offloading: none, conc-list: [64, 128, 256] }
2850-
- { tp: 8, ep: 8, dp-attn: true, offloading: none, conc-list: [128, 256, 512] }
2851-
- { tp: 4, offloading: cpu, conc-list: [16, 32, 64] }
2852-
- { tp: 8, offloading: cpu, conc-list: [16, 32, 64] }
2853-
- { tp: 4, ep: 4, dp-attn: true, offloading: cpu, conc-list: [64, 128, 256] }
2854-
- { tp: 8, ep: 8, dp-attn: true, offloading: cpu, conc-list: [128, 256, 512] }
2847+
- { tp: 4, offloading: none, conc-list: [64, 70, 76, 82] }
2848+
- { tp: 8, offloading: none, conc-list: [64, 70, 76, 82] }
2849+
- { tp: 4, ep: 4, dp-attn: true, offloading: none, conc-list: [64, 70, 76, 82] }
2850+
- { tp: 8, ep: 8, dp-attn: true, offloading: none, conc-list: [64, 70, 76, 82] }
28552851

28562852
dsv4-fp4-b300-trt:
28572853
image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-9aa3715

0 commit comments

Comments
 (0)