Gahow Wang

gahow created branch phase18-pipeline-parallelism in gahow/xserv

2026-05-29 10:58:01 +00:00

gahow pushed to phase18-pipeline-parallelism at gahow/xserv

2026-05-29 10:58:01 +00:00

11e0154e4d docs: Phase 18 pipeline parallelism — design + benchmark results

d5dcf1a5ab bench: PP harness (xserv --pp vs llama.cpp -sm layer)

824cc58daa server: pipeline-parallel HTTP engine (--pp N)

da3aaa134a model: pipeline-parallel Qwen3 (from_weights_pp + stage forward)

859c0cc0b6 distributed: NCCL P2P primitives (PpContext + send/recv)

Compare 5 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-29 10:19:47 +00:00

cf812b6264 Workload characterization C1-C3 on full production trace

847f52f03b PD-disagg crossover: regular synthetic trace + goodput sweep + figure

48ae72467a Replayer: closed-loop inter-turn think-time mode

657cd36f3d Gate evict_sent_blocks behind VLLM_EVICT_SENT_BLOCKS

Compare 4 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-29 09:07:28 +00:00

a0db3cbe77 Add leastwork_kappa decode-aware ablation (net-negative, documented)

gahow pushed to main at gahow/agentic-kvc

2026-05-29 08:08:46 +00:00

71b0747b3b 600s-truncated trace + LPWL 5-policy results

160c29133d Unified bench report: mean+TPS+per-worker GPU util, auto-captured

d9046322c6 Add parameter-free LPWL routing policy (--policy leastwork)

Compare 3 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-29 04:07:18 +00:00

8a876e90d1 traces/README: clarify w600 is the session-start window, not span

e532e83d3e mb5_run: scrape per-instance prefix-cache counters before teardown

d376d91fe1 Engine-state ablation: full sweep harness + results

08c3cf48aa Ship anonymized benchmark trace w600_r0.0015_st30 + provenance

8708b75520 Merge layerwise KV transfer + engine-state ablation onto main

Compare 22 commits »

gahow pushed to main at gahow/xserv

2026-05-29 03:36:58 +00:00

c2362df1f1 fix(xserv-chat): UTF-8/CJK-aware line input

gahow pushed to main at gahow/xserv

2026-05-29 03:11:08 +00:00

7b8b520cda docs: TP=1/2/4 xserv vs llama.cpp benchmark results

a4a171d425 bench: TP sweep harness (xserv --tp, llama row-split, concurrent groups)

95eb61d639 server: tensor-parallel HTTP engine (--tp N)

f17011129e model: tensor-parallel Qwen3 (sharded weights + AllReduce)

453520d622 distributed: NCCL tensor-parallel primitives (TpContext + AllReduce)

Compare 6 commits »

gahow pushed to main at gahow/xserv

2026-05-28 13:38:24 +00:00

14a44b503e docs: add Chinese README (overview + usage)

gahow pushed to main at gahow/xserv

2026-05-28 13:32:19 +00:00

80157e614a docs: update llama.cpp comparison with 8192 results (OOM fixed)

gahow pushed to main at gahow/xserv

2026-05-28 11:59:24 +00:00

fc1900a745 server: VRAM-sized KV pool + vLLM-style swap scheduler

d52baa0006 model: paged KV cache with CPU swap pool, decode graph, qwen3 updates

4c3f914459 kernels/cuda: paged-attention kernel, dispatch, pinned host memory

Compare 3 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-28 09:11:17 +00:00

4cd71b6631 Working-set figure: extend left panel to ~50 nodes

gahow pushed to main at gahow/agentic-kvc

2026-05-28 08:31:29 +00:00

2247d1de08 Working-set figure: right panel = W(t) time series

gahow pushed to main at gahow/agentic-kvc

2026-05-28 08:24:22 +00:00

c94b2e237a Working-set figure: linear node axes + benefit/cost split

gahow pushed to main at gahow/agentic-kvc

2026-05-28 08:16:05 +00:00

3b8be5bb61 Working-set figure: express footprint in node count, not GB

gahow pushed to main at gahow/agentic-kvc

2026-05-28 08:03:29 +00:00

dae98c6472 Working-set sizing tool + GLM-5.1-FP8/B300 result

2e6a369046 PD_DISAGG_RESULTS §5.1: D-pool pressure crashes consumers

3957c2df86 MB5 patch: clamp PD-consumer metrics counter underflow

8596135680 MB5 analysis: per-role KV split proves static-partition mismatch

e8980ce957 MB5 proxy: session-affinity P routing (MB5_P_ROUTING=session)

Compare 8 commits »

gahow pushed to main at gahow/xserv

2026-05-28 07:06:39 +00:00

3f1c3d429a docs: llama.cpp vs xserv benchmark results + summary

950ccf3822 bench: fix llama.cpp per-slot context (was 1/parallel of intended)

7cb9ee3870 bench: run one server at a time, match thinking mode, fix tools package

49c7653222 tools: add llama.cpp comparison baseline + standard benchmark suite

Compare 4 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-27 15:23:27 +00:00

e0d3b5150a MB5 driver fixes: bash env-prefix + replayer flag names + python date math

gahow pushed to main at gahow/agentic-kvc

2026-05-27 15:03:01 +00:00

e9abd70c8d MB5 driver: launcher, orchestrator, KV-pool timeline plotter

gahow pushed to main at gahow/agentic-kvc

2026-05-27 14:30:57 +00:00

a4f5dd56aa MB5 instrumentation: per-request KV-block snapshot from vLLM V1 scheduler