sskj/experiments/dsv4_p800_long_context_matrix
yy-fighting f6a45f1ed8 Add P800 long-context matrix and tune benchmark/health settings
- Add experiments/dsv4_p800_long_context_matrix for 64k/128k/256k context
  grid on DeepSeek-V4-Flash-INT8 with Kunlun P800.
- Change bench_client_docker.sh to use --concurrency instead of
  --max-concurrency to match the user's long-context benchmark script.
- Add SGLANG_HEALTH_CHECK_TIMEOUT override to server_docker.sh for slow
  long-context server startups.
- Tune dsv4_p800_256k_4k_probe with lower max-running-requests and
  mem-fraction-static to avoid OOM on P800.
2026-07-08 15:02:22 +00:00
..

DSV4 P800 Long-Context Matrix

Kunlun P800 XPU + SGLang + DeepSeek-V4-Flash-INT8 long-context matrix benchmark.

What it does

Runs a grid of input/output lengths to measure SGLang long-context serving performance on P800:

  • Input lengths: 64k, 128k, 256k
  • Output lengths: 256, 1k, 4k
  • Concurrency: 25 for 64k; 10, 5, 2 for 128k and 256k

Each context group (64k / 128k / 256k) is started with its own server instance using the appropriate --context-length and --max-running-requests.

Quick Start

# Run the full matrix (starts server per group, runs all scenarios, stops server)
PLATFORM=kunlun_p800 bash experiments/dsv4_p800_long_context_matrix/run_bench.sh

Results land in experiments/dsv4_p800_long_context_matrix/results/<RUN_ID>/.

Configuration

Edit config.env or override via environment variables:

# Run with the tuned EAGLE speculative-decoding config instead of the baseline
SERVER_MODE=w8a8_int8 PLATFORM=kunlun_p800 \
bash experiments/dsv4_p800_long_context_matrix/run_bench.sh

# Reduce concurrency for the 256k group if OOM occurs
CONTEXT_GROUPS=("64k 80000 25" "128k 140000 10" "256k 270000 5") \
PLATFORM=kunlun_p800 \
bash experiments/dsv4_p800_long_context_matrix/run_bench.sh

Files

File Purpose
config.env Experiment-level configuration (model, scenarios, context groups)
start_server.sh Start the P800 SGLang Docker container for a context group
run_bench.sh Orchestrator: server per group → warmup → scenarios → parse

Platform

This experiment targets platforms/kunlun_p800.env and uses scripts/common/server_docker.sh / scripts/common/bench_client_docker.sh.