Gahow Wang

gahow pushed to main at gahow/xserv

2026-05-30 04:51:02 +00:00

9a01c60100 server: GPU argmax fast path for greedy decode

c679f618fd model: fuse QKV/gate_up projections, batched decode ops

cc4bd4cfe5 paged-kv: kernel-based scatter + fix data_ptr offset bug

13ae3de69e kernels: reshape_and_cache, GPU argmax, single-launch GEMV

6ce21345be cuda: add cached_trim() to release pooled GPU buffers

Compare 6 commits »

gahow pushed to main at gahow/agentic-kvc

2026-05-30 04:48:40 +00:00

dc8e6dd5a8 v2 exp(a): add remote KV-store (RDMA) tier

gahow pushed to main at gahow/agentic-kvc

2026-05-30 03:23:46 +00:00

ad754cfe0b v2 exp(b): GPU KV-capacity APC/latency knee + writeup

837df6bc9e v2 exp(a): three-tier KV-hit latency microbench (GPU >> CPU >> miss)

Compare 2 commits »

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

2026-05-29 14:35:01 +00:00

b8868d59ac moe: FIX decode — route all gpt-oss matmuls through dense cuBLAS

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

2026-05-29 14:29:57 +00:00

fcd7fa62b7 moe(wip): decode-path debugging tools — decode STILL broken

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

2026-05-29 14:28:50 +00:00

6fdfb1b9d9 moe(wip): zero dequant output — decode STILL broken (not fixed)

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

2026-05-29 14:24:59 +00:00

0b76207d21 moe: fix decode correctness — zero dequant output (was uninitialized-read)

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

2026-05-29 14:21:38 +00:00

afe7cc6645 moe(wip): KV-cached gpt-oss decode — NOT yet correct

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

2026-05-29 14:21:17 +00:00

603b6eb270 moe: correct + deterministic KV-cached gpt-oss decode (single card)

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

2026-05-29 13:54:09 +00:00

7e7d077ff1 moe: KV-cached GPU decode for gpt-oss (O(1)/token)

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

2026-05-29 13:51:02 +00:00

94957c5727 moe: MXFP4-resident experts on GPU (single-card gpt-oss)

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

2026-05-29 13:43:00 +00:00

f59ba73938 docs: MoE progress snapshot before dash5 reboot

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

2026-05-29 13:40:25 +00:00

58062bd326 kernels: fix MXFP4 build — wire quant module + mxfp4.cu (7ebdd7c didn't build)

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

2026-05-29 13:38:56 +00:00

7ebdd7c552 kernels: MXFP4 -> BF16 dequant kernel (verified vs numpy)

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

2026-05-29 13:31:05 +00:00

403879959a kernels: fix sink-decode build (b4db953 didn't compile)

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

2026-05-29 13:28:49 +00:00

1e8091a111 docs: MoE perf-path progress + remaining work (#7-#10)

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

2026-05-29 13:28:32 +00:00

b4db9535db kernels: attention sinks + sliding window in decode attention

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

2026-05-29 13:17:43 +00:00

2a515de7df moe: fix missing pub mod gptoss + record REAL verification

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

2026-05-29 13:13:48 +00:00

d0af97a2bf docs: MoE progress — gpt-oss forward verified correct (Paris); remaining perf work

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

2026-05-29 13:13:21 +00:00

0dd8851e88 moe: gpt-oss-20b forward verified correct (predicts "Paris")

05534611ca moe(wip): gptoss.rs first correctness-first forward + logit-dump bin

c7d0750c32 moe(wip): gpt-oss-20b groundwork — config fields, arch doc, MXFP4 tools

057a3c68a3 docs: Phase 19 MoE (gpt-oss-20b) design + progress snapshot

Compare 4 commits »