Skip to content

Instantly share code, notes, and snippets.

View KastanDay's full-sized avatar

Kastan Day KastanDay

36 followers · 9 following
View GitHub Profile
@KastanDay
KastanDay / sglang_pd_page_reuse_repro.py
Created July 27, 2026 15:47
Deterministic CPU reproducer: SGLang PD releases KV pages while a Mooncake transfer is still writing to them (sgl-project/sglang)
#!/usr/bin/env python3
"""Deterministic reproducer: SGLang PD releases KV pages while a Mooncake
transfer is still writing to them.
Run against an UNPATCHED sglang checkout. CPU only -- no GPU, no network, no
second process, ~1 second.
cd /path/to/sglang && PYTHONPATH=python python sglang_pd_page_reuse_repro.py
Invariant under test
@KastanDay
KastanDay / evidence.md
Created July 27, 2026 18:33
SGLang PR #25 multi-node Mooncake transfer-quiescence validation

PR #25 multi-node Mooncake validation

Revision and deployment

  • Current PR head: 7db762ddf260df8ff6b102a9f20616a09c0d0457
  • Deployed source snapshot: 8361cb5705493de9489cc57cc7536279ae3b4de0
  • python/ and test/ are byte-for-byte identical between those revisions. The only full-tree differences are two unrelated Kimi-K3 documentation files introduced by the newer base.
  • Runtime: lmsysorg/sglang:latest@sha256:984699c298a95b73c469b2191403ddc85fd780506e13c39c4afff3845e27bc6c
  • Topology: two TP8 prefill engines and one TP8 decode engine on three separate H200 nodes, 24 GPUs total.
  • Barrier policy: SGLANG_DISAGGREGATION_TRANSFER_BARRIER=2 (STRICT).