- New experiments/p800/dsv4_p800_sglang_tp_dp_matrix: TP8/DP1, TP4/DP2,
TP2/DP4 matrix with smoke results; TP2/DP4 documents the weight-loading
OOM root cause (274 GiB INT8 weights sharded only across TP group).
- Launch args drop --ep-size/--chunked-prefill-size/--max-prefill-tokens/
--max-running-requests; experts fall back to TP sharding.
- Move dsv4_p800_256k_4k_probe under experiments/p800/.
- scripts/common: jq-free parsing fixes in adaptive_bench_lib.sh and
parse_backend.py.
- .gitignore: cover raw_outputs under nested platform experiment layout.
- Add RESUME_RUN_ID env var for breakpoint resume in adaptive benchmarks.
When set, the script reuses an existing result directory and skips
already-tested (TP, DP, ISL, OSL) shapes based on adaptive_shapes.jsonl.
- Fix RUN_ID unbound variable in resume mode.
- Fix jq query to use -s (slurp) for jsonl files.
- Add resume skip logic to DRY_RUN mode as well.
- Rename DSL -> OSL across all adaptive benchmark files for consistency:
- scripts/common/adaptive_bench_lib.sh
- scripts/common/adaptive_concurrency.py
- experiments/dsv4_h200_vllm_tp_dp_matrix/adaptive_config.env
- experiments/dsv4_h200_vllm_tp_dp_matrix/run_adaptive_concurrency.sh
- experiments/dsv4_h200_sglang_tp_dp_matrix/adaptive_config.env
- experiments/dsv4_h200_sglang_tp_dp_matrix/run_adaptive_concurrency.sh
- experiments/ADAPTIVE_CONCURRENCY_USAGE.md
- 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.
- Add experiments/dsv4_p800_256k_4k_probe for testing 256k input + 4k output
on DeepSeek-V4-Flash-INT8 with Kunlun P800.
- Add SGLANG_EXTRA_LAUNCH_ARGS support to scripts/common/server_docker.sh
so probe can inject --context-length 270000.
- Add w8a8_int8_baseline server mode without EAGLE speculative decoding,
matching the long-context P800 tuning.
- Switch dsv4_p800_sglang and the new probe to use w8a8_int8_baseline by
default; original tuned mode remains available via SERVER_MODE=w8a8_int8.
- Remove USE_FAST_BFP16_MOE=1 which caused torch parameter assignment
crash in fresh containers for w8a8_int8 mode.
- Mount model path at its host absolute path inside the container so the
bench client tokenizer can load it regardless of server --model-path.
- Default WARMUP to 0 and NUM_PROMPTS to 10 for a fast smoke test.
- Default to a single scenario; document how to run a full grid.
- Update experiment README with PLATFORM env var and override examples.
