Repository navigation
Conversation
…ication primitive
YAMY1234
marked this pull request as ready for review
September 28, 2026 00:05
YAMY1234
added a commit
that referenced
this pull request
Oct 15, 2026
…s, compressor to 16 full.env sets SGLANG_OPT_DSV41_ROUTED_EXPERT_GEMM_MAX_ROWS=48 (#14 and the #15 draft MoE) and SGLANG_OPT_DSV41_LOW_LATENCY_COMPRESSOR_MAX_ROWS=16 (#9 compressor). Section 4.3: same node, FULL / NEW / FULL, two rounds each, step -5.4 % at 12 rows (5.157 -> 4.879 ms), -4.5 % at 24, -4.1 % at 48, c=1 -0.2 %; the 12-row saving is the routing kernel and less MoE on the critical path. GSM8K-200 greedy NEW 196/200, FULL 197/200. The row-tiled low-latency MXFP8 path past 8 rows (92a4007) is documented and left off: correct, but the mHC prenorm + sinkhorn chain becomes the critical path. The item table and known limits list the new bounds.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PD 的恢复、因子化和状态发布有大量主机发射开销。本变更按真实末块长度选桶,将 P 层恢复、卷积、递推与因子化捕获为图,并在 P→D 读取前完成发布;末块状态、16/32 输出 tile、r8/W8、seed 与生产配方保持一致。最新 b42a773 让 radix 检查点复用 PrefillFactorGraph。
旧组件 A 臂 C1:P 384.0→178.996 ms,因子 host 29.7 ms,长度52/52通过。但同作业 C1/C8 的纯解码 P90 分别比 stock 低1.58%/1.12%,低并发性能门未闭合。最新 factor 补丁 C1 为217.513 ms,超过同作业 stock206.639+3.87 ms,因子 host59.7 ms;profile显示剩余开销分散,按负责人指令停止进一步猜测补丁。数值902101确定20/20及已定义交接逐位通过、19末块KL最大2.737774×,GSM197→194(−3)失败;固定冷集仅3条,不宣称固定35完整准入。所有失败及费用保留。
CPU与GPU逐位夹具覆盖真实长度、槽复用、首/中/末层、发布顺序和已定义交接字节;比较时掩码只用于未定义第N token及页尾,不改变生产线包,已定义区单字节反例必须失败。历史17-token离群、参照修订及后续补测均与原判值并列。
本 PR 是旧组件 P 图与负结果的可审实现,不代表所有性能门通过。k31 原生 PD 兼容移植在独立 PR #15:A/final 已通过 AGG→PD 数值,六格使用 k31 原生基线;k31 的本线 P 图另行准入,不与旧 v3 曲线直接相减。
C1分账、四格三列表、六格、逐位与数值回执、完整费用和UTC状态。未改 paper/,未提交权重或原始张量。