Repository navigation
[docs] Kimi-K3: widen the H200 High-Throughput recipe to 4x8 TP32/EP32 - #32834
Merged
Merged
Conversation
The H200 Unified High-Throughput cell was a copy of Balanced (2x8 TP16/EP16) with mem-frac 0.90 and extra_buffer_lazy. Replace it with the shape that was actually run: TP32/EP32 across 4 nodes, plus the cross-node fabric env the H100 4x8 cells already carry (SGLANG_HOST_IP, NCCL/GLOO_SOCKET_IFNAME, PYTORCH_CUDA_ALLOC_CONF, SGLANG_K3_ATTN_RES_MODE, SGLANG_MOE_FUSED_GATE_RADIX) and --dist-timeout 3600. As run — DSPARK and HiCache L1+L2 layered on, --mamba-full-memory-ratio 0.058 — the static allocation leaves 12.54 GB free per GPU and 1940352 KV tokens. Node count is a cell property, so the prose that states H200 is 2x8 is updated in three places: the hardware-recipe line, the Strategy list, and the per-platform table. The h200 multi-node hint said "BOTH ranks", which is no longer true for this cell. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
zijiexia
requested review from
JustinTong0323,
sogalin and
wisclmy0611
as code owners
July 30, 2026 00:09
Contributor
|
Caution The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased. |
github-actions
Bot
added
the
documentation
Contributor
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Workflows to automatically generate PRs for you. |
wisclmy0611
approved these changes
Jul 30, 2026
Contributor
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Workflows to automatically generate PRs for you. |
saturn-acc
pushed a commit
to saturn-acc/sglang
that referenced
this pull request
Aug 16, 2026
sgl-project#32834) Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Atituiset
pushed a commit
to Atituiset/sglang
that referenced
this pull request
Sep 10, 2026
sgl-project#32834) Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This branch was successfully deployed
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
2 participants
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.
Motivation
The Kimi-K3 cookbook's H200 / Unified / High-Throughput cell was a copy of Balanced (2×8 TP16/EP16) with
--mem-fraction-static 0.90andextra_buffer_lazy. The shape actually run for that operating point is TP32/EP32 across 4 nodes, so the page shipped a recipe nobody serves.Modifications
docs_new/src/snippets/configs/moonshotai/kimi-k3.jsx— the H200 Unified High-Throughput cell:nnodes: 2 → 4,--tp-size/--ep-size16 → 32--dist-timeout 3600and the cross-node fabric env the H100 4×8 cells already carry:SGLANG_HOST_IP,NCCL_SOCKET_IFNAME,GLOO_SOCKET_IFNAME(via the existing{{LOCAL_IP}}/{{NETWORK_IFACE}}placeholders),PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,SGLANG_K3_ATTN_RES_MODE=jit,SGLANG_MOE_FUSED_GATE_RADIX=1extra_buffer_lazy--nnodes/--node-rank/--dist-init-addrstay out of the cell (engine-injected), and--mamba-full-memory-ratiostays out too so the ratio calculator keeps driving itAs run — DSPARK and HiCache L1+L2 layered on, ratio
0.058— the static allocation leaves 12.54 GB free per GPU and 1940352 KV tokens. Recorded in the cell comment.Node count is a cell property, so the three prose spots that state H200 = 2×8 are updated in
Kimi-K3.mdx(hardware-recipe line, Strategy list, per-platform table), along with theh200multi-node hint that said "BOTH ranks".Checklist
mint validatepassesmint dev): the H200 High-Throughput command emits--nnodes 4 --node-rank <rank> --dist-init-addr <node0-ip>:20000with TP32/EP32CI States
Latest PR Test (Base): ✅ Run #30501811286
Latest PR Test (Extra): ❌ Run #30501811122