ceb1170969
- Created adaptive_summary.md to summarize adaptive concurrency search results, including metrics such as TP, DP, and TPS. - Added run_manifest.json to document experiment configuration details, including hardware specifications and search parameters. - Introduced shapes.tsv to define input and output lengths for various configurations in the adaptive experiment.
DSV4 DSpark H20 SGLang TP×DP Matrix Benchmark
NVIDIA H20 + SGLang (Docker) + DeepSeek-V4-Flash-DSpark TP×DP matrix benchmark.
Adapted from experiments/h20/dsv4_h20_sglang_tp_dp_matrix, with the model
swapped to /data1/hf_models/DeepSeek-V4-Flash-DSpark. All other parameters
(context length 1M, max-running-requests 128, mem-fraction-static 0.88,
moe-runner-backend marlin, ISL/OSL matrix, concurrency rules) are identical to
the non-DSpark H20 SGLang matrix experiment.
Note: DSpark speculative decoding is a vLLM-only feature in this repository; SGLang serves the DSpark checkpoint as a standard model.
Quick Start
# Dry run: print server args and scenario plan only
DRY_RUN=1 bash experiments/h20/dsv4_dspark_h20_sglang_tp_dp_matrix/run_bench.sh
# Full TP×DP matrix run
bash experiments/h20/dsv4_dspark_h20_sglang_tp_dp_matrix/run_bench.sh
# Adaptive concurrency search (+16 per step, aligned with
# dsv4_h20_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh)
bash experiments/h20/dsv4_dspark_h20_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh
Results land in experiments/h20/dsv4_dspark_h20_sglang_tp_dp_matrix/results/<RUN_ID>/.
Configuration
Edit config.env or override via environment variables:
MODEL_PATH=/data1/hf_models/DeepSeek-V4-Flash-DSpark \
SGLANG_PORT=30033 \
bash experiments/h20/dsv4_dspark_h20_sglang_tp_dp_matrix/run_bench.sh
Files
Same layout as dsv4_h20_sglang_tp_dp_matrix; see that experiment for details.
Platform
This experiment targets platforms/nvidia_h20.env.