Gahow Wang

gahow pushed to main at gahow/agentic-kvc

2026-05-26 06:21:47 +00:00

06dd175441 Microbench 1 plots: prefill-decode interference heatmap + lines

gahow pushed to main at gahow/agentic-kvc

2026-05-26 06:06:24 +00:00

72790ae6c1 PD-sep server-side profiling: vLLM patches + per-request breakdown

gahow pushed to main at gahow/agentic-kvc

2026-05-26 05:28:46 +00:00

d76eb02637 Elastic migration v2 section: PD-sep on agentic workload is net negative

95c8ef853c Fix proxy shadow drift: actively reconcile against vLLM /metrics

4b833d33b7 unified_v2.1: relax gates + add unified_kv_both isolation control

19f69a9d2e unified_v2: selective per-request PD-sep via Mooncake (E3+E4)

Compare 4 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-25 17:12:49 +00:00

c63dc151a0 Agentic PD / Unified routing story plan draft

gahow pushed to main at gahow/agentic-kvc

2026-05-25 17:09:52 +00:00

0881942cf3 Window 1 results: recompute with fixed metrics + reframe limitations

0e82612100 Fix B3 analysis bugs from subagent audit (median + percentile + sweep)

8ac41a8684 Agentic dispatch coupling: trace-replay session-sequentiality is realistic

Compare 3 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-25 16:59:03 +00:00

f784e49c07 Microbench: prefill-decode interference + PD transfer lifecycle

gahow pushed to main at gahow/agentic-kvc

2026-05-25 16:37:29 +00:00

559faa1e26 B2 finding: TPOT idx peaks at 32k, not 65k — cost migrates to TTFT

gahow pushed to main at gahow/agentic-kvc

2026-05-25 16:22:20 +00:00

4722883903 Audit package refresh: Window 1 supported claims + risk register

0c3220cbb8 Window 1 results: combined B1' + B2 + B3 report and artifacts

b7902061d1 Window 1 analysis: APC upper bound, B2 window-overlap, figure renderer

b9f324f2e6 B2 interference driver: request return_token_ids + text fallback

df3249925b B3 analyze: prefer per-policy engine_state over slicing shared dir

Compare 6 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-25 11:58:43 +00:00

08530b3915 B3 policies: pseudocode reference for the five-policy sweep

123a74a4b9 B3 report renderer: incremental markdown table from comparison JSON

92db1c4370 B3 post-run helpers: engine_state slicer + per-policy aggregator

e23128ad65 B2: PD-colo interference microbench harness + sweep aggregator

c6b7c3471b B3: load_only + sticky policies, capped-trace builder, sweep driver

Compare 6 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-25 08:23:56 +00:00

cd82b8c2a2 PD-sep matrix results: C2/C3/C4 figures + empirical mechanism refined

25445e3d18 A5: joined analysis with reuse decomp, interference, hot-spot, labels

f42c715ec1 A4: open-loop session-causal SRR loadgen

5816aad731 A3: vLLM scheduler patch for step-level JSONL log

fe556b5d98 A2: proxy worker-state snapshot and request-id passthrough

Compare 7 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-25 07:15:25 +00:00

5ed6f6fe5b Add characterization result figures

gahow pushed to main at gahow/agentic-kvc

2026-05-25 07:01:42 +00:00

0f64fb3261 Add agentic workload characterization audit scaffold

gahow pushed to main at gahow/agentic-kvc

2026-05-25 03:47:36 +00:00

21ffb3d4f7 PD-sep matrix infrastructure: bench.sh pdsep mode + matrix driver

gahow pushed to main at gahow/agentic-kvc

2026-05-25 03:41:35 +00:00

4028c587b1 Paper section: system analysis + workload figures + KV-wall model

gahow pushed to main at gahow/agentic-kvc

2026-05-25 03:24:26 +00:00

d71a111099 Paper section: PD-sep scaffold + drop --enforce-eager from launch scripts

gahow pushed to main at gahow/agentic-kvc

2026-05-25 02:50:54 +00:00

6a27f75337 Docs: reconcile routing docs with current hybrid direction

ac6534c3ff Cleanup: retire dead PUSH path + extract hybrid picker

255c8e6884 Hybrid routing: LMetric for LB + explicit affinity for high-cache sessions

448361cf83 Update design doc: final results + review findings

4c583f2f1c Revert relaxed gate + push_cost fix: 134 offloads destroyed performance

Compare 10 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-24 14:45:15 +00:00

5772149d36 Approach B v2: TTFT-based migration trigger

45b82272c3 Add migration policy design doc with A/B experiment results

e9919605af Approach B: session-level lazy migration trigger

e06de5144b Approach A: contention-aware cost model with migration discount

e13391eeab Evict migrated blocks from prefix cache after KV send completes

Compare 61 commits »

gahow pushed to main at gahow/xserv

2026-05-23 06:14:00 +00:00

9bb5c5c328 tools: add correctness + performance test scripts for Qwen3-8B

986a289616 fix: 12 bug fixes from comprehensive review — 51 tok/s verified on RTX 5090

Compare 2 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-22 23:16:32 +00:00

8e0c6e78b0 Add comprehensive research findings document

gahow pushed to main at gahow/agentic-kvc

2026-05-22 23:15:03 +00:00

080a8fa138 Chunk-size ablation + comprehensive synthesis