feat: agentic-replay (mooncake_trace) on official AIPerf — Qwen3.5-4B smoke#7

Closed

thangquang09 wants to merge 17 commits into

devfrom

feature/aiperf

Commits on Jun 3, 2026

test: narrow gemma4 bf16 h100 sweep to 1k1k conc=1 aiperf only

Thắng. Lý Quang (5)
and
claude
committed
fix: set max-num-batched-tokens=4096 for Gemma4 BF16 H100 to clear MM encoder floor

Thắng. Lý Quang (5)
and
claude
committed
ci: ignore compare-results failure on PR sweeps
Thắng. Lý Quang (5)
committed
feat: agentic-replay (mooncake_trace) on official AIPerf for Qwen3.5-4B smoke

Thắng. Lý Quang (5)
and
claude
committed
ci: trigger agentic-replay smoke sweep for qwen3.5-4b-bf16-h100-vllm

Thắng. Lý Quang (5)
and
claude
committed
fix: forward BENCHMARK_CLIENT into greennode container; route agentic-replay in e2e dispatch

Thắng. Lý Quang (5)
and
claude
committed
feat: add tokens/Watt (tok/s/W) energy-efficiency metric to results

Thắng. Lý Quang (5)
and
claude
committed
docs: document tokens/Watt energy-efficiency stage in AIPerf integration

Thắng. Lý Quang (5)
and
claude
committed
feat: commit 64k/128k agentic traces in-repo + #N subset; 64k conc=32 run

Thắng. Lý Quang (5)
and
claude
committed
Swap Qwen3.5-4B→Qwen3-4B-Instruct-2507; add latency percentiles + dual tok/W

Thắng. Lý Quang (5)
and
claude
committed
Fix 64k agentic-replay max-model-len 40960→73728 (AIPerf threads sessions)

Thắng. Lý Quang (5)
and
claude
committed
docs(aiperf): context-length per dataset + correct multi-turn note

Thắng. Lý Quang (5)
and
claude
committed
feat(aiperf): gemma4-31B agentic-replay vLLM vs SGLang pair (2xH100, TP=2, fp8)

Thắng. Lý Quang (5)
and
claude
committed
chore(aiperf): drop smoke leg from gemma4-agentic configs (run only the 64k pair)

Thắng. Lý Quang (5)
and
claude
committed
docs(aiperf): agentic-replay verification + engine-vs-engine + backend trap; SGLang fa3

Thắng. Lý Quang (5)
and
claude
committed

Commits on Jun 4, 2026

fix(aiperf): fail closed on partial replay results
Thắng. Lý Quang (5)
committed
fix(aiperf): keep replay CLI compatible with PyPI
Thắng. Lý Quang (5)
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: agentic-replay (mooncake_trace) on official AIPerf — Qwen3.5-4B smoke#7

feat: agentic-replay (mooncake_trace) on official AIPerf — Qwen3.5-4B smoke#7
thangquang09 wants to merge 17 commits into
devfrom
feature/aiperf

Commits on Jun 3, 2026

test: narrow gemma4 bf16 h100 sweep to 1k1k conc=1 aiperf only

fix: set max-num-batched-tokens=4096 for Gemma4 BF16 H100 to clear MM encoder floor

ci: ignore compare-results failure on PR sweeps

feat: agentic-replay (mooncake_trace) on official AIPerf for Qwen3.5-4B smoke

ci: trigger agentic-replay smoke sweep for qwen3.5-4b-bf16-h100-vllm

fix: forward BENCHMARK_CLIENT into greennode container; route agentic-replay in e2e dispatch

feat: add tokens/Watt (tok/s/W) energy-efficiency metric to results

docs: document tokens/Watt energy-efficiency stage in AIPerf integration

feat: commit 64k/128k agentic traces in-repo + #N subset; 64k conc=32 run

Swap Qwen3.5-4B→Qwen3-4B-Instruct-2507; add latency percentiles + dual tok/W

Fix 64k agentic-replay max-model-len 40960→73728 (AIPerf threads sessions)

docs(aiperf): context-length per dataset + correct multi-turn note

feat(aiperf): gemma4-31B agentic-replay vLLM vs SGLang pair (2xH100, TP=2, fp8)

chore(aiperf): drop smoke leg from gemma4-agentic configs (run only the 64k pair)

docs(aiperf): agentic-replay verification + engine-vs-engine + backend trap; SGLang fa3

Commits on Jun 4, 2026

fix(aiperf): fail closed on partial replay results

fix(aiperf): keep replay CLI compatible with PyPI