Gahow Wang

gahow pushed to main at gahow/xserv

2026-06-11 16:58:52 +00:00

24c49c31c2 tools: warm-server FP8 vs BF16 benchmark + results doc

5a16225c1f quantization: cache cuBLASLt FP8 plan per shape — fix per-expert heuristic churn

Compare 2 commits »

gahow pushed to main at gahow/xserv

2026-06-08 07:45:14 +00:00

3a530956af tools: add FP8 vs BF16 benchmark and GSM8K eval harness

gahow pushed to main at gahow/xserv

2026-06-07 16:25:58 +00:00

21caba0f06 tools: add FP8 vs BF16 benchmark and GSM8K eval harness

76487b7963 quantization: W8A8 FP8 compute via cuBLASLt tensor cores

Compare 2 commits »

gahow pushed to main at gahow/xserv

2026-06-07 11:33:18 +00:00

9f1fbbb98b quantization: add FP8 E4M3 W8A16 for gpt-oss MoE expert weights

gahow pushed to main at gahow/xserv

2026-06-02 17:01:47 +00:00

e1eb77baa4 xserv-chat: fix unclosed <think> on early termination and flush analysis tokens

gahow pushed to main at gahow/xserv

2026-06-02 13:37:32 +00:00

34e9bee375 xserv-chat: render gpt-oss analysis as a Qwen3-style <think> block

gahow pushed to main at gahow/xserv

2026-06-02 10:11:49 +00:00

3b9e32e6cd kernels: fix uninitialized shared-memory read in M=1 decode GEMV

5157b2cd30 kernels: fix NaN in flash-attention sinks on fully-masked window tiles

ea5d8ba7ea xserv-chat: render gpt-oss multi-turn as canonical harmony (drop CoT)

c0a81c84e7 server: canonical harmony system message in gpt-oss fallback

3d6bb1918e xserv-chat: fix gpt-oss harmony chat (canonical system prompt + routing)

Compare 5 commits »

gahow pushed to main at gahow/xserv

2026-06-02 06:19:15 +00:00

f2e60218b4 xserv-chat: harmony channel routing + repetition penalty for gpt-oss

gahow pushed to main at gahow/xserv

2026-06-02 04:05:15 +00:00

3ee8df2c0f xserv-chat: filter harmony control tokens + stop at <|end|> for gpt-oss

gahow pushed to main at gahow/xserv

2026-06-01 16:58:18 +00:00

ae08896f46 xserv-chat: support gpt-oss-20b with TP; fix GEMV precision bug

gahow pushed to main at gahow/agentic-kvc

2026-06-01 01:35:44 +00:00

54e1f5266a MB5 PD ablation v2 results: concurrency axis + reuse 3-way writeup

3f997fda14 MB5 PD ablation v2 tooling: conc completion-panel plot + gpu_monitor dep

Compare 2 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-31 17:04:48 +00:00

19c443e3bc paper f2a: reuse-topology decomposition + mixture-sensitivity sweep

9c105cf05a MB5 PD ablation: controlled-variable reuse/conc redo + campaign tooling

Compare 2 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-31 16:55:45 +00:00

32f7f55990 v2: linear (default cache-aware) baseline + 2x wall-cap on first600s

gahow pushed to main at gahow/agentic-kvc

2026-05-31 12:15:16 +00:00

7529284cee v2: LMetric PD-colo vs PD-disagg on the real agentic trace

fafc44da79 MB5 PD reuse-centric ablation: tooling, data, Fig 1-3

a2111b6e18 PD-disagg docs: annotated corrections for e13391e contamination

Compare 3 commits »

gahow pushed to main at gahow/xserv

2026-05-31 05:23:37 +00:00

1d0ec32e8d server: Jinja chat template rendering via minijinja

4368e79695 model: fused GPU MoE kernel — eliminate CPU roundtrip

377a04b81f tokenizer: read pre-tokenizer regex from tokenizer.json

Compare 3 commits »

gahow pushed to main at gahow/xserv

2026-05-30 17:06:32 +00:00

241009a96c docs: remove TO-BE-FIXED.md — all listed issues have been resolved

gahow pushed to main at gahow/xserv

2026-05-30 16:56:57 +00:00

0c6135aea3 bench-gpt-oss: teacher-forced diagnostics + --prompt flag

ffd90ce7fb server: emit harmony developer instructions block for gpt-oss

3c9d5e260e server: harmony termination via is_eos + TP repetition penalty

99b212e6c1 model/sampling: NaN-safe argmax + optional repetition penalty

e11f15e009 tokenizer: support multiple end-of-generation tokens

Compare 6 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-30 13:10:31 +00:00

0b180c191e v2 exp(d): expand figure to 6 panels (TTFT/E2E mean+p90, TPS, per-worker GPU util)

gahow pushed to main at gahow/agentic-kvc

2026-05-30 13:03:14 +00:00

9b6091fe6e v2 exp(d): 5-policy routing under tracets vs thinktime — ranking flip

68f21bef23 bench harness: env-tunable vLLM health timeout + both-modes 5-policy driver

075f5bbc22 trace: time_to_parent_chat annotation + thinktime trace variants

Compare 3 commits »

gahow pushed to main at gahow/xserv

2026-05-30 08:36:13 +00:00

5cb3cf28f9 server: add gpt-oss chat template for proper prompt formatting

15c51f143e server: support GptOss in TP engine + benchmark script

d29c39d74e fix: GEMV NaN bug — skip custom kernel for small N (<256)

9ad91a4a92 phase19: MoE support — gpt-oss-20b end-to-end inference with TP=2

Compare 4 commits »