s-zaizen's picture
Publish optimized vLLM recipe and Spark Arena results
7e17ce5 verified
|
Raw
History Blame Contribute Delete
2.07 kB

Optimized vLLM and Spark Arena results

The selected configuration is the non-speculative T5 recipe in ../recipes/vllm-optimized-sm121-tp2.yaml. It keeps the T3 FULL_DECODE_ONLY CUDA Graph configuration and extends the qualified operating envelope to a 262,144-token API ceiling with 12 GiB of explicit KV-cache reservation. Growing-prefix and isolated long-context gates passed 12/12 through 250,000 tokens.

Quality and fixed-workload selection

Comparison Quality delta Throughput delta
Upstream -> NVFP4 S0, GSM8K +0.0 pt +0.725 tok/s
NVFP4 S0 -> selected T3, GSM8K +0.0 pt +6.419 tok/s
NVFP4 S0 -> selected T3, fixed workload n/a +9.376 tok/s
T3 -> selected long-context T5, GSM8K -1.0 pt -0.405 tok/s
T3 -> selected long-context T5, fixed workload n/a -0.042 tok/s

T6 DSpark k=3 and T7 DSpark k=5 remained inside the three-point quality gate, but reduced fixed-workload throughput by 7.349 and 9.039 tok/s respectively. They were rejected; T5 remains selected.

Spark Arena submission

  • Submission: sub1788319379140
  • Status: completed
  • Profile: official/spark-arena-v2
  • Runtime: vLLM, tensor parallel 2, two DGX Spark nodes
  • Cells: 28/28 completed, 0 failed, three runs per cell
  • Model identity: s-zaizen/DeepSeek-V4-Flash-Vision-Exp-NVFP4

Representative results:

Depth Concurrency PP tok/s TG tok/s TTFR ms
0 1 1206.7 25.4 1699.1
0 2 483.6 21.5 5087.1
0 10 325.5 19.2 32509.2
16,384 1 130.2 25.2 15731.4
16,384 2 112.0 9.8 26150.1
16,384 10 100.9 6.6 109304.5
100,000 1 22.0 24.9 93032.5
100,000 2 21.4 2.4 142417.9
100,000 10 21.0 1.4 534753.7

The accompanying CSV and JSON contain all cells. Their model values were mechanically replaced with the canonical Hugging Face ID before publication; all numeric benchmark values remain byte-for-byte equal to the original fields.