Gahow Wang

gahow pushed to main at gahow/xserv

2026-07-02 06:21:37 +00:00

6309dc1181 docs: Phase 27 scaled-up — GSM8K 1000 + AIME2025 30 quality report

264c004662 eagle3: GSM8K quality benchmark proves tree-spec is correctness-preserving

Compare 2 commits »

gahow pushed to main at gahow/xserv

2026-07-01 16:25:03 +00:00

2fe903ecea eagle3: extend tree to top-3 siblings — speedup_e2e = 1.20×

gahow pushed to main at gahow/xserv

2026-07-01 16:09:36 +00:00

aac9ace144 eagle3: tree drafting with top-2 siblings — speedup_e2e = 1.17× 🎉

gahow pushed to main at gahow/xserv

2026-07-01 15:09:41 +00:00

6da0972740 speculative: copy_kv_position primitive for tree drafting KV remap

gahow pushed to main at gahow/xserv

2026-07-01 12:46:31 +00:00

40d8a29e33 docs: Phase 26 epilogue 2 — tree kernel landed; KV remap is the remaining blocker

gahow pushed to main at gahow/xserv

2026-07-01 12:46:00 +00:00

fd392f7fbb attention: tree-aware paged_decode_attention_tree kernel + wrapper

gahow pushed to main at gahow/xserv

2026-07-01 12:19:34 +00:00

10a98539d0 eagle3: coverage + top-3 diagnostic; acceptance ceiling analysis

gahow pushed to main at gahow/xserv

2026-07-01 11:59:07 +00:00

cc3bc2188c docs: Phase 26 epilogue — speedup_e2e = 1.10x achieved

gahow pushed to main at gahow/xserv

2026-07-01 11:58:28 +00:00

06a798cab9 eagle3: cuBLAS-GEMM verify path — speedup_e2e > 1 achieved 🎉

gahow pushed to main at gahow/xserv

2026-07-01 11:18:40 +00:00

9a1af0adee docs: Phase 26 — EAGLE3 implementation follow-up + bug hunt log

gahow pushed to main at gahow/xserv

2026-07-01 11:16:37 +00:00

d2c55c47b2 eagle3: γ≥2 correctness fixes + per-slot diagnostic

gahow pushed to main at gahow/xserv

2026-07-01 10:02:00 +00:00

14925154a3 eagle3: γ≥2 recursive drafting + batched verify with hooks

gahow pushed to main at gahow/xserv

2026-07-01 09:50:53 +00:00

a24621fa6a eagle3: proper residual chain + stateful KV cache

gahow pushed to main at gahow/xserv

2026-07-01 09:32:55 +00:00

68b55fa1e6 eagle3: γ=1 speculative bench + first end-to-end measurement

gahow pushed to main at gahow/xserv

2026-07-01 09:29:06 +00:00

8f11d6e5cd eagle3: fix EAGLE_HOOK_LAYERS to [2, 18, 33] for Qwen3-8B

gahow pushed to main at gahow/xserv

2026-07-01 09:23:27 +00:00

e04a8ffb18 speculative: EAGLE3 draft head implementation (Phase 25 step 1)

gahow pushed to main at gahow/xserv

2026-07-01 09:01:25 +00:00

6485c87c5b docs: Phase 25 — three speculative-decoding paradigms compared

gahow pushed to main at gahow/xserv

2026-07-01 08:32:23 +00:00

a77239c0c8 speculative: Qwen3 decode graph + gamma sweep (Phase 24 step 2)

gahow pushed to main at gahow/xserv

2026-07-01 08:13:44 +00:00

e5734b41fa speculative: batched-GEMV kernel for verify path (Phase 24 step 1)

gahow pushed to main at gahow/xserv

2026-07-01 07:35:22 +00:00

42e13f33dd docs: Phase 24 investigation notes and revised speedup plan