Conversation
Pin the upstream image containing #40637 by digest. Add nine TP4/EP4 and TP2/EP2 points with load-specific Engram placement, the GB300 launch route, and Enroot digest normalization. Preserve native DSpark precision and use real verification for evals. 新增 GB300 SGLang DeepSeek-V4.1 AgentX 配方，使用 digest 固定包含 #40637 的上游镜像。覆盖九个 TP4/EP4 与 TP2/EP2 测试点，按负载选择 Engram 放置方式，并补充启动路由和 Enroot digest 转换。保留原生 DSpark 精度，准确率评测使用真实验证。 Validated Bash syntax, 31 local command/routing checks, nine-point matrix, three-mode eval matrix, and append-only changelog. GPU qualification remains pending; replace the changelog PR-link placeholder after PR creation.
Move TP/concurrency-specific Engram selection out of the shared GB300 launcher and into the SGLang benchmark script. Preserve the nine benchmark points and keep only routing and digest import changes in the launcher. 将按 TP 和并发选择 Engram 放置方式的逻辑移至 SGLang benchmark 脚本，保留九个测试点，减少共享 GB300 启动器的修改，仅保留路由和 digest 导入支持。 Validation: Bash syntax, git diff --check, and 31 local command-construction/routing checks passed. No GPU execution.
合并最新 main，保留已发布分支历史；采用已合并的 GB300 配方并保留 Enroot digest 转换。
更新 GB300 SGLang 镜像并调整 TP4 AgentX 预填充配置；保留 TP2 启动设置，等待全量性能与准确率验收。
Use EP1 for C1/C2 with the low-latency recipe, preserving TP2 and TP4 C4+ settings. Record measured C1 evidence and mark C2 as untested. Update English/Chinese guidance and append performance changelog entries. 中文：C1/C2 使用纯 TP4/EP1 低延迟配置，保留 TP2 与 TP4 C4+ 设置。记录 C1 实测依据并注明 C2 尚未验证，同步中英文说明并追加性能变更记录。
Merge origin/main without rewriting history. Preserve the main changelog byte-for-byte and append the three unchanged GB300 submission entries. No GB300 serving settings changed. 中文：合并 origin/main，不改写历史。逐字节保留主分支变更记录，将三条未修改的 GB300 提交记录追加至末尾，不改变 GB300 服务配置。
|
Thanks for the contribution!
中文感谢你的贡献！
|
Consolidate this branch's changelog additions and shorten the English and Chinese configuration notes. Benchmark settings are unchanged. 中文：合并本分支的曲线更新记录，精简中英文配置说明，基准测试设置不变。
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=36085901363 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=36085901363 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Not approving — the inline finding on the SGLANG_RAGGED_VERIFY_MODE=static export sitting outside the CONC > 2 guard (benchmarks/single_node/agentic/dsv41flash_fp4_gb300_sglang_mtp.sh:137) affects the new TP4/EP1 C1/C2 runs and should be looked at.
Beyond that, I also checked the Engram host-table layout export at line 51: SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT=per_rank stays exported for the new TP4 C1/C2 case even though SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE is forced to 0 right after — since the layout variable is only meaningful when the host-table feature flag is on, this looks like a harmless leftover rather than a functional bug.
Extended reasoning...
The diff requalifies a GB300 SGLang benchmark on a new nightly image, splits/narrows a config search-space entry, and adds a TP==4 override block in the benchmark shell script plus digest-to-tag translation in the launcher; no auth/crypto/permission surface is touched. A CONFIRMED finding (SGLANG_RAGGED_VERIFY_MODE=static exported outside the CONC>2 gate, affecting the new untested EP1 C1/C2 points) is already queued as an inline comment, which is enough by itself to withhold approval; I additionally traced the Engram host-table layout export for the same new C1/C2 path and found it inert (feature flag forced off), which is worth noting since it was one of the ruled-out candidates.
| if (( TP == 4 )); then | ||
| export SGLANG_RAGGED_VERIFY_MODE=static |
There was a problem hiding this comment.
🔴 TP4/EP1 C1 and C2 runs now silently pick up SGLANG_RAGGED_VERIFY_MODE=static, a setting the docs and changelog say is exclusive to TP4/EP4 C4+. export SGLANG_RAGGED_VERIFY_MODE=static at line 137 sits outside the if (( CONC > 2 )) block (line 138), so it applies to every TP==4 run, including the new EP1 C1 (CONC<=2) path added at line 52. docs/configuration-procedures.md lists only chunk 4096, interval 16, request cap 2, decode graph cap 64 and automatic SWA for TP4 C1, with static ragged verification scoped to "TP4/EP4 at C4 and above"; the C1/C2 measurements and the 450 vs 433 tok/s comparison in perf-changelog.yaml never varied this knob. …
Why this was flagged
…Fix: move the export SGLANG_RAGGED_VERIFY_MODE=static line inside the if (( CONC > 2 )) block so it only applies where it was actually validated, keeping C1/C2 on whatever ragged-verify mode was previously in effect.
Trigger: any TP==4 job (both the new EP1 C1/C2 entries in configs/nvidia-master.yaml and existing EP4 runs) executes lines 136-141. The code unconditionally exports SGLANG_RAGGED_VERIFY_MODE=static at line 137 before the CONC>2 gate at line 138, so C1/C2 EP1 runs get an untested serving mode never listed in docs/configuration-procedures.md's C1/C2 settings or in perf-changelog.yaml's description of the qualified config. On base, TP4 C1/C2 ran without this override. No safeguard catches the mismatch since the env var isn't included in SGLANG_CMD or written to sglang_command.txt/server.log filters that would flag it against the documented recipe.
Verification: normal. The claim is factually verified. In benchmarks/single_node/agentic/dsv41flash_fp4_gb300_sglang_mtp.sh, export SGLANG_RAGGED_VERIFY_MODE=static (line 137) is placed inside if (( TP == 4 )) (line 136) but ABOVE the if (( CONC > 2 )) gate (line 138), so it executes for every TP4 run — including the new EP1 C1/C2 path (line 52: if (( TP == 4 && CONC <= 2 ))). All the other C4+…
This branch has not been deployed
1 participant
Update on gb300 curve