Add Qwen3.5 FP4 B300 AgentX MTP - #2421
Merged
Merged
Conversation
Add the initial SGLang NEXTN AgentX recipe with golden synthetic acceptance and 256k traces.\n\n中文：新增 Qwen3.5 FP4 B300 AgentX MTP 初始配置，使用 SGLang NEXTN、黄金合成接受长度和 256k 轨迹数据集。
Contributor
|
Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase For PR verification, add the PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs 感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 如需进行 PR 验证，请为此 PR 添加 PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档 |
Append the Qwen3.5 FP4 B300 AgentX MTP benchmark entry.\n\n中文：追加 Qwen3.5 FP4 B300 AgentX MTP 基准测试触发记录。
cquil11
changed the title
Add strict two-pool HiCache accounting, cache reporting, session-sticky DEP8 routing, AgentX request/graph limits, and measured-cache guidance for the B300 frontier.\n\n中文：为 B300 前沿调优加入严格的双池 HiCache 容量核算、缓存统计、会话粘性的 DEP8 路由、AgentX 请求与 CUDA Graph 上限，并记录实测缓存容量依据。
Add isolated TP2 c72 and TP4 c80 DRAM HiCache points so each cache-cliff recovery topology can be measured independently. 中文：添加独立的 TP2 c72 与 TP4 c80 DRAM HiCache 测试点，以便分别测量各拓扑在缓存工作集拐点后的恢复效果。
Account for SGLang target KV, Mamba, and native NEXTN draft host pools as H * 31/15 per rank, with page-alignment reserve, so projected use stays within runner-generated DRAM. 中文：按每个 rank 的 H * 31/15 计入 SGLang 目标 KV、Mamba 与原生 NEXTN 草稿主机缓存，并预留页对齐空间，确保预计用量不超过 runner 生成的 DRAM 上限。
中文：将最新 main 合并到 B300 提交分支，并将本 PR 的性能变更记录重新追加到文件末尾。
Add TP4 concurrency points around the measured HiCache crossover and extend the DRAM-offload curve beyond it.\n\n中文：在实测 HiCache 切换区间细化 B300 Qwen TP4 并发点，并将 DRAM 卸载曲线扩展到拐点之后。
Remove the measured no-offload thrashing tail and add dense TP2 points across the expected HiCache crossover.\n\n中文：移除实测发生抖动的无卸载尾部配置，并在预期 HiCache 切换区间细化 B300 TP2 并发点。
Drop the attention-DP arm after live AgentX profiling showed fragmented prefix caches and throughput below the TP frontier. Remove the now-unused router path.\n\n中文：AgentX 实测表明注意力 DP 会导致前缀缓存碎片化且吞吐低于 TP 前沿，因此移除该配置及不再使用的路由路径。
Probe three post-c72 offload points to locate saturation of the measured TP2 efficiency frontier.\n\n中文：在 c72 之后增加三个卸载并发点，用于定位实测 B300 TP2 效率前沿的饱和位置。
Remove the no-offload c96 point after it was dominated by HiCache c80 in both throughput and latency, and refresh the measured crossover note.\n\n中文：无卸载 c96 在吞吐与延迟上均被 HiCache c80 支配，因此移除该点并更新实测切换区间说明。
Remove offload points that exceed the measured host-cache validity boundary and densify the final safe interval below c80.\n\n中文：移除超过实测主机缓存有效边界的卸载配置，并在 c80 以下的安全区间细化最终并发点。
Remove no-offload c64 after it stalled under the long-context working set; retain the matched HiCache points beyond c56.\n\n中文：无卸载 c64 在长上下文工作集下停滞，因此移除该点，并保留 c56 之后对应的 HiCache 配置。
中文：在缓存工作集拐点测试 B300 TP4 HiCache。
中文：裁剪 B300 TP2 无卸载配置的缓存拐点尾部。
中文：优化 B300 缓存悬崖附近的并发点，并使用 SGLang 默认单 tokenizer 启动路径以避免 HiCache 初始化期间的多 tokenizer 共享内存竞态。
中文：在 B300 HiCache 缓存悬崖附近增加 c84、c92 和 c94 测试点，以更细粒度覆盖有效的帕累托前沿。
中文：在 B300 TP2 的 c32 至 c56 区间增加 HiCache 重叠测试点，以覆盖一小时 AgentX 预热下可能提前出现的缓存悬崖。
Use six tokenizer workers for TP4 long-context replay while retaining SGLang default single-worker startup for TP2 HiCache. 中文：TP4 长上下文回放使用 6 个 tokenizer worker；TP2 HiCache 保持 SGLang 默认单 worker 启动路径，避免共享内存初始化竞态。
Add HiCache concurrency 34 and 38 around the measured c36-to-c40 transition. 中文：在实测 c36 到 c40 的转折区间补充 HiCache 并发点 34 和 38。
Densify the TP4 no-offload cliff and remove TP4 HiCache points that are dominated by the TP4 no-offload and TP2 HiCache arms. Keep TP2 HiCache dense through the measured host-KV saturation region.\n\n中文：细化 B300 缓存帕累托前沿；加密 TP4 无卸载缓存临界区间，移除被 TP4 无卸载与 TP2 HiCache 方案支配的 TP4 HiCache 测点，并保留覆盖主机 KV 饱和区间的密集 TP2 HiCache 测点。
cquil11
marked this pull request as ready for review
July 30, 2026 09:13
Remove the measured no-offload thrashing tail and add dense TP2 points across the expected HiCache crossover.\n\n中文：移除实测发生抖动的无卸载尾部配置，并在预期 HiCache 切换区间细化 B300 TP2 并发点。
Drop the attention-DP arm after live AgentX profiling showed fragmented prefix caches and throughput below the TP frontier. Remove the now-unused router path.\n\n中文：AgentX 实测表明注意力 DP 会导致前缀缓存碎片化且吞吐低于 TP 前沿，因此移除该配置及不再使用的路由路径。
Probe three post-c72 offload points to locate saturation of the measured TP2 efficiency frontier.\n\n中文：在 c72 之后增加三个卸载并发点，用于定位实测 B300 TP2 效率前沿的饱和位置。
Remove the no-offload c96 point after it was dominated by HiCache c80 in both throughput and latency, and refresh the measured crossover note.\n\n中文：无卸载 c96 在吞吐与延迟上均被 HiCache c80 支配，因此移除该点并更新实测切换区间说明。
Remove offload points that exceed the measured host-cache validity boundary and densify the final safe interval below c80.\n\n中文：移除超过实测主机缓存有效边界的卸载配置，并在 c80 以下的安全区间细化最终并发点。
Remove no-offload c64 after it stalled under the long-context working set; retain the matched HiCache points beyond c56.\n\n中文：无卸载 c64 在长上下文工作集下停滞，因此移除该点，并保留 c56 之后对应的 HiCache 配置。
中文：在缓存工作集拐点测试 B300 TP4 HiCache。
中文：裁剪 B300 TP2 无卸载配置的缓存拐点尾部。
中文：优化 B300 缓存悬崖附近的并发点，并使用 SGLang 默认单 tokenizer 启动路径以避免 HiCache 初始化期间的多 tokenizer 共享内存竞态。
中文：在 B300 HiCache 缓存悬崖附近增加 c84、c92 和 c94 测试点，以更细粒度覆盖有效的帕累托前沿。
中文：在 B300 TP2 的 c32 至 c56 区间增加 HiCache 重叠测试点，以覆盖一小时 AgentX 预热下可能提前出现的缓存悬崖。
Use six tokenizer workers for TP4 long-context replay while retaining SGLang default single-worker startup for TP2 HiCache. 中文：TP4 长上下文回放使用 6 个 tokenizer worker；TP2 HiCache 保持 SGLang 默认单 worker 启动路径，避免共享内存初始化竞态。
Add HiCache concurrency 34 and 38 around the measured c36-to-c40 transition. 中文：在实测 c36 到 c40 的转折区间补充 HiCache 并发点 34 和 38。
Densify the TP4 no-offload cliff and remove TP4 HiCache points that are dominated by the TP4 no-offload and TP2 HiCache arms. Keep TP2 HiCache dense through the measured host-KV saturation region.\n\n中文：细化 B300 缓存帕累托前沿；加密 TP4 无卸载缓存临界区间，移除被 TP4 无卸载与 TP2 HiCache 方案支配的 TP4 HiCache 测点，并保留覆盖主机 KV 饱和区间的密集 TP2 HiCache 测点。
Pin the AIPerf trace idle-gap branch and set the Qwen B300 replay cap to 300 seconds.
中文：Qwen B300 AgentX 扫描使用 GSM8K 评估，避免运行耗时的 SWE-bench。
cquil11
force-pushed
the
agent/qwen35-fp4-b300-agentx-mtp
branch
from
August 3, 2026 01:29
22265bb to
0687e15
Compare
Collaborator
Author
|
As a PR reviewer and CODEOWNER, I have reviewed this and have:
Additional detail section:
Signed: |
Collaborator
❌❌❌ REJECTED ❌❌❌@cquil11 — the sweep/eval run this sign-off relies on (30730541420) executed on commit ✅ Check 0 (CODEOWNER): PASS — |
cquil11
changed the title
中文：将已验证的 Qwen B300 扫描提交重新纳入分支历史，不更改当前文件内容。
Collaborator
Author
|
/reuse-sweep-run 30730541420 中文：复用已验证的扫描运行 30730541420。 |
Collaborator
Author
|
As a PR reviewer and CODEOWNER, I have reviewed this and have:
Additional detail section:
Signed: |
Collaborator
✅✅✅ Verdict: PASS ✅✅✅✅ Check 0 (CODEOWNER): PASS — |
xiaohuguo2023
pushed a commit
to xiaohuguo2023/InferenceX
that referenced
this pull request
Aug 4, 2026
* feat: add Qwen3.5 FP4 B300 AgentX MTP Add the initial SGLang NEXTN AgentX recipe with golden synthetic acceptance and 256k traces.\n\n中文：新增 Qwen3.5 FP4 B300 AgentX MTP 初始配置，使用 SGLang NEXTN、黄金合成接受长度和 256k 轨迹数据集。 * chore: trigger B300 AgentX sweep Append the Qwen3.5 FP4 B300 AgentX MTP benchmark entry.\n\n中文：追加 Qwen3.5 FP4 B300 AgentX MTP 基准测试触发记录。 * perf: expand B300 AgentX preflight range * chore: use SGLang v0.5.16 for Qwen B300 * chore: collect SGLang cache metrics on B300 * fix: remove unsupported SGLang tool-choice flag on B300 * perf: prepare Qwen B300 cache and DEP tuning Add strict two-pool HiCache accounting, cache reporting, session-sticky DEP8 routing, AgentX request/graph limits, and measured-cache guidance for the B300 frontier.\n\n中文：为 B300 前沿调优加入严格的双池 HiCache 容量核算、缓存统计、会话粘性的 DEP8 路由、AgentX 请求与 CUDA Graph 上限，并记录实测缓存容量依据。 * perf: add B300 HiCache cliff probes Add isolated TP2 c72 and TP4 c80 DRAM HiCache points so each cache-cliff recovery topology can be measured independently. 中文：添加独立的 TP2 c72 与 TP4 c80 DRAM HiCache 测试点，以便分别测量各拓扑在缓存工作集拐点后的恢复效果。 * fix: enforce Qwen HiCache DRAM total Account for SGLang target KV, Mamba, and native NEXTN draft host pools as H * 31/15 per rank, with page-alignment reserve, so projected use stays within runner-generated DRAM. 中文：按每个 rank 的 H * 31/15 计入 SGLang 目标 KV、Mamba 与原生 NEXTN 草稿主机缓存，并预留页对齐空间，确保预计用量不超过 runner 生成的 DRAM 上限。 * perf: densify B300 Qwen cache crossover Add TP4 concurrency points around the measured HiCache crossover and extend the DRAM-offload curve beyond it.\n\n中文：在实测 HiCache 切换区间细化 B300 Qwen TP4 并发点，并将 DRAM 卸载曲线扩展到拐点之后。 * perf: bracket B300 TP2 offload frontier Remove the measured no-offload thrashing tail and add dense TP2 points across the expected HiCache crossover.\n\n中文：移除实测发生抖动的无卸载尾部配置，并在预期 HiCache 切换区间细化 B300 TP2 并发点。 * perf: remove dominated B300 Qwen DEP topology Drop the attention-DP arm after live AgentX profiling showed fragmented prefix caches and throughput below the TP frontier. Remove the now-unused router path.\n\n中文：AgentX 实测表明注意力 DP 会导致前缀缓存碎片化且吞吐低于 TP 前沿，因此移除该配置及不再使用的路由路径。 * perf: extend B300 TP2 HiCache tail Probe three post-c72 offload points to locate saturation of the measured TP2 efficiency frontier.\n\n中文：在 c72 之后增加三个卸载并发点，用于定位实测 B300 TP2 效率前沿的饱和位置。 * perf: prune dominated B300 TP4 tail Remove the no-offload c96 point after it was dominated by HiCache c80 in both throughput and latency, and refresh the measured crossover note.\n\n中文：无卸载 c96 在吞吐与延迟上均被 HiCache c80 支配，因此移除该点并更新实测切换区间说明。 * perf: cap B300 HiCache at valid working sets Remove offload points that exceed the measured host-cache validity boundary and densify the final safe interval below c80.\n\n中文：移除超过实测主机缓存有效边界的卸载配置，并在 c80 以下的安全区间细化最终并发点。 * perf: cap B300 TP2 before the cache cliff Remove no-offload c64 after it stalled under the long-context working set; retain the matched HiCache points beyond c56.\n\n中文：无卸载 c64 在长上下文工作集下停滞，因此移除该点，并保留 c56 之后对应的 HiCache 配置。 * perf: probe B300 TP4 HiCache at the cliff 中文：在缓存工作集拐点测试 B300 TP4 HiCache。 * perf: prune the B300 TP2 no-offload cliff 中文：裁剪 B300 TP2 无卸载配置的缓存拐点尾部。 * fix(agentx): refine B300 cache frontier 中文：优化 B300 缓存悬崖附近的并发点，并使用 SGLang 默认单 tokenizer 启动路径以避免 HiCache 初始化期间的多 tokenizer 共享内存竞态。 * perf(agentx): densify B300 HiCache cliff 中文：在 B300 HiCache 缓存悬崖附近增加 c84、c92 和 c94 测试点，以更细粒度覆盖有效的帕累托前沿。 * perf(agentx): densify B300 TP2 cache crossover 中文：在 B300 TP2 的 c32 至 c56 区间增加 HiCache 重叠测试点，以覆盖一小时 AgentX 预热下可能提前出现的缓存悬崖。 * fix(agentx): scale Qwen tokenization by topology Use six tokenizer workers for TP4 long-context replay while retaining SGLang default single-worker startup for TP2 HiCache. 中文：TP4 长上下文回放使用 6 个 tokenizer worker；TP2 HiCache 保持 SGLang 默认单 worker 启动路径，避免共享内存初始化竞态。 * perf(agentx): densify B300 TP2 cache boundary Add HiCache concurrency 34 and 38 around the measured c36-to-c40 transition. 中文：在实测 c36 到 c40 的转折区间补充 HiCache 并发点 34 和 38。 * perf(agentx): refine B300 cache frontier Densify the TP4 no-offload cliff and remove TP4 HiCache points that are dominated by the TP4 no-offload and TP2 HiCache arms. Keep TP2 HiCache dense through the measured host-KV saturation region.\n\n中文：细化 B300 缓存帕累托前沿；加密 TP4 无卸载缓存临界区间，移除被 TP4 无卸载与 TP2 HiCache 方案支配的 TP4 HiCache 测点，并保留覆盖主机 KV 饱和区间的密集 TP2 HiCache 测点。 * fix(agentx): cap Qwen trace idle gaps Pin the AIPerf trace idle-gap branch and set the Qwen B300 replay cap to 300 seconds. * chore(aiperf): pin merged trace idle cap branch * chore(agentx): bump AIPerf trace idle cap fix Signed-off-by: Cam Quilici <cjquilici@gmail.com> * fix(agentx): update AIPerf trace-cap cleanup Signed-off-by: Cam Quilici <cjquilici@gmail.com> * chore(agentx): pin reconstruction-only idle cap Update AIPerf after removing runtime trace idle enforcement. Signed-off-by: Cam Quilici <cjquilici@gmail.com> * fix(agentx): update AIPerf warmup handoff Signed-off-by: Cam Quilici <cjquilici@gmail.com> * fix(agentx): retain baseline parents across warmup * fix(agentx): extend SGLang keep-alive * fix(agentx): pin validated AIPerf join timing 中文：将 AIPerf 子模块固定到已验证的 AgentX 交接、请求顺序与父子会合计时修复版本。 * fix(agentx): pin idle-cap scheduler fix Pin AIPerf to the validated revision that preserves join dependencies while enforcing the whole-system idle cap across replay waits. 中文：将 AIPerf 固定到已验证版本，在保留子代理汇合依赖的同时，对所有回放等待正确执行全局空闲上限。 * fix(agentx): pin zero-idle warmup join fix Pin AIPerf to e765bcea446082b01bd7014e11fb5008b59ac65d so accelerated warmup compresses replay-time readiness without bypassing subagent completion gates.\n\n中文：将 AIPerf 固定到 e765bcea446082b01bd7014e11fb5008b59ac65d，使加速预热压缩回放时间条件，同时不绕过子代理完成门控。 * fix(agentx): pin runtime trace idle enforcement Pin AIPerf to the validated replay-root idle watchdog so the 300-second trace cap remains effective under independent parent/subagent runtime clocks without changing request order or join dependencies. 中文：将 AIPerf 固定到已验证的重放根轨迹空闲看门狗版本，使 300 秒轨迹上限在父级/子代理独立运行时时钟下仍然生效，同时不改变请求顺序或会合依赖。 * fix(agentx): pin corrected AIPerf handoff timing Pin the AIPerf submodule to the exact commit that preserves relative stream offsets through warmup-to-profiling handoff before applying one uniform trace-wide idle-cap shift. 中文：将 AIPerf 子模块固定到精确提交；该提交在预热到性能测试交接期间保留各请求流的相对时序，并仅对整条轨迹统一应用空闲时间上限偏移。 Signed-off-by: Cam Quilici <cjquilici@gmail.com> * fix(agentx): retain full warmup in Qwen B300 fast sweep Pin the Qwen B300 sweep to the cleaned AIPerf scheduler head and keep ten warmup requests per lane while limiting fast-mode profiling to 20 minutes.\n\n中文：将 Qwen B300 扫描固定到清理后的 AIPerf 调度器最新提交，并让快速模式在 profiling 缩短至 20 分钟的同时，保持每个 lane 发送 10 个预热请求。 Signed-off-by: Cam Quilici <cjquilici@gmail.com> * revert(agentx): remove Qwen B300 fast-mode override Restore the shared agentx-fast behavior to one warmup request per lane and a 20-minute profile. The PR will run without the modifier label for canonical 10-request warmup and one-hour profiling. 中文：恢复共享的 agentx-fast 行为，每个 lane 发送 1 个预热请求并运行 20 分钟 profiling。本 PR 将移除该修饰标签，以使用标准的每 lane 10 个预热请求和 1 小时 profiling。 * test(agentx): refresh Qwen B300 fast validation Pin AIPerf to the cache-coherent warmup fix and retain ten warmup requests per lane while limiting fast-mode profiling to 20 minutes. 中文：将 AIPerf 固定到预热缓存一致性修复，并让快速模式保持每个 lane 发送 10 个预热请求，同时将性能测试时长限制为 20 分钟。 Signed-off-by: Cam Quilici <cjquilici@gmail.com> * test(agentx): shorten Qwen B300 fast profile Pin the latest AIPerf PR SemiAnalysisAI#31 head and reduce the bespoke Qwen B300 fast profile to 10 minutes while retaining ten warmup requests per lane. 中文：将子模块固定到 AIPerf PR SemiAnalysisAI#31 的最新提交，并将 Qwen B300 专用快速性能测试缩短至 10 分钟，同时保持每个 lane 发送 10 个预热请求。 Signed-off-by: Cam Quilici <cjquilici@gmail.com> * test(agentx): pin globally anchored AIPerf handoff Update Qwen B300 validation to the tested AIPerf revision that preserves profiling phase spacing across the warmup handoff. 中文：将 Qwen B300 验证固定到已测试的 AIPerf 版本，确保 warmup 交接后保留 profiling 阶段的请求间隔。 Signed-off-by: Cam Quilici <cjquilici@gmail.com> * revert(agentx): remove Qwen B300 fast override Restore the shared AgentX fast-mode behavior so the Qwen B300 PR uses canonical benchmark settings when the modifier label is absent.\n\n中文：移除 Qwen B300 的一次性快速模式覆盖；去掉修饰标签后，该 PR 将恢复使用标准 AgentX 基准测试配置。 * test(agentx): pin flattened warmup handoff timing Pin AIPerf to f9438058 so non-burst profiling restores all warmup survivors to one shared trajectory clock and preserves cross-stream timing. 中文：将 AIPerf 固定到 f9438058，使非突发 profiling 将所有 warmup 后存活的请求流恢复到同一轨迹时钟，并保留跨请求流的时序。 Signed-off-by: Cam Quilici <cjquilici@gmail.com> * test(agentx): pin barrier-aware idle watchdogs Pin the AIPerf revision that preserves flattened warmup handoff timing and keeps global and per-tree idle watchdogs active across dependency-retained timers. 中文：固定包含 AgentX 扁平化预热交接时序修复的 AIPerf 版本，并确保全局及单轨迹空闲看门狗在计时器被依赖屏障保留后仍持续生效。 Signed-off-by: Cam Quilici <cjquilici@gmail.com> * test(agentx): use GSM8K for Qwen B300 eval 中文：Qwen B300 AgentX 扫描使用 GSM8K 评估，避免运行耗时的 SWE-bench。 * feat: add Qwen3.5 FP4 B300 AgentX MTP Add the initial SGLang NEXTN AgentX recipe with golden synthetic acceptance and 256k traces.\n\n中文：新增 Qwen3.5 FP4 B300 AgentX MTP 初始配置，使用 SGLang NEXTN、黄金合成接受长度和 256k 轨迹数据集。 * chore: trigger B300 AgentX sweep Append the Qwen3.5 FP4 B300 AgentX MTP benchmark entry.\n\n中文：追加 Qwen3.5 FP4 B300 AgentX MTP 基准测试触发记录。 * perf: expand B300 AgentX preflight range * chore: use SGLang v0.5.16 for Qwen B300 * chore: collect SGLang cache metrics on B300 * fix: remove unsupported SGLang tool-choice flag on B300 * perf: prepare Qwen B300 cache and DEP tuning Add strict two-pool HiCache accounting, cache reporting, session-sticky DEP8 routing, AgentX request/graph limits, and measured-cache guidance for the B300 frontier.\n\n中文：为 B300 前沿调优加入严格的双池 HiCache 容量核算、缓存统计、会话粘性的 DEP8 路由、AgentX 请求与 CUDA Graph 上限，并记录实测缓存容量依据。 * perf: add B300 HiCache cliff probes Add isolated TP2 c72 and TP4 c80 DRAM HiCache points so each cache-cliff recovery topology can be measured independently. 中文：添加独立的 TP2 c72 与 TP4 c80 DRAM HiCache 测试点，以便分别测量各拓扑在缓存工作集拐点后的恢复效果。 * fix: enforce Qwen HiCache DRAM total Account for SGLang target KV, Mamba, and native NEXTN draft host pools as H * 31/15 per rank, with page-alignment reserve, so projected use stays within runner-generated DRAM. 中文：按每个 rank 的 H * 31/15 计入 SGLang 目标 KV、Mamba 与原生 NEXTN 草稿主机缓存，并预留页对齐空间，确保预计用量不超过 runner 生成的 DRAM 上限。 * perf: densify B300 Qwen cache crossover Add TP4 concurrency points around the measured HiCache crossover and extend the DRAM-offload curve beyond it.\n\n中文：在实测 HiCache 切换区间细化 B300 Qwen TP4 并发点，并将 DRAM 卸载曲线扩展到拐点之后。 * perf: bracket B300 TP2 offload frontier Remove the measured no-offload thrashing tail and add dense TP2 points across the expected HiCache crossover.\n\n中文：移除实测发生抖动的无卸载尾部配置，并在预期 HiCache 切换区间细化 B300 TP2 并发点。 * perf: remove dominated B300 Qwen DEP topology Drop the attention-DP arm after live AgentX profiling showed fragmented prefix caches and throughput below the TP frontier. Remove the now-unused router path.\n\n中文：AgentX 实测表明注意力 DP 会导致前缀缓存碎片化且吞吐低于 TP 前沿，因此移除该配置及不再使用的路由路径。 * perf: extend B300 TP2 HiCache tail Probe three post-c72 offload points to locate saturation of the measured TP2 efficiency frontier.\n\n中文：在 c72 之后增加三个卸载并发点，用于定位实测 B300 TP2 效率前沿的饱和位置。 * perf: prune dominated B300 TP4 tail Remove the no-offload c96 point after it was dominated by HiCache c80 in both throughput and latency, and refresh the measured crossover note.\n\n中文：无卸载 c96 在吞吐与延迟上均被 HiCache c80 支配，因此移除该点并更新实测切换区间说明。 * perf: cap B300 HiCache at valid working sets Remove offload points that exceed the measured host-cache validity boundary and densify the final safe interval below c80.\n\n中文：移除超过实测主机缓存有效边界的卸载配置，并在 c80 以下的安全区间细化最终并发点。 * perf: cap B300 TP2 before the cache cliff Remove no-offload c64 after it stalled under the long-context working set; retain the matched HiCache points beyond c56.\n\n中文：无卸载 c64 在长上下文工作集下停滞，因此移除该点，并保留 c56 之后对应的 HiCache 配置。 * perf: probe B300 TP4 HiCache at the cliff 中文：在缓存工作集拐点测试 B300 TP4 HiCache。 * perf: prune the B300 TP2 no-offload cliff 中文：裁剪 B300 TP2 无卸载配置的缓存拐点尾部。 * fix(agentx): refine B300 cache frontier 中文：优化 B300 缓存悬崖附近的并发点，并使用 SGLang 默认单 tokenizer 启动路径以避免 HiCache 初始化期间的多 tokenizer 共享内存竞态。 * perf(agentx): densify B300 HiCache cliff 中文：在 B300 HiCache 缓存悬崖附近增加 c84、c92 和 c94 测试点，以更细粒度覆盖有效的帕累托前沿。 * perf(agentx): densify B300 TP2 cache crossover 中文：在 B300 TP2 的 c32 至 c56 区间增加 HiCache 重叠测试点，以覆盖一小时 AgentX 预热下可能提前出现的缓存悬崖。 * fix(agentx): scale Qwen tokenization by topology Use six tokenizer workers for TP4 long-context replay while retaining SGLang default single-worker startup for TP2 HiCache. 中文：TP4 长上下文回放使用 6 个 tokenizer worker；TP2 HiCache 保持 SGLang 默认单 worker 启动路径，避免共享内存初始化竞态。 * perf(agentx): densify B300 TP2 cache boundary Add HiCache concurrency 34 and 38 around the measured c36-to-c40 transition. 中文：在实测 c36 到 c40 的转折区间补充 HiCache 并发点 34 和 38。 * perf(agentx): refine B300 cache frontier Densify the TP4 no-offload cliff and remove TP4 HiCache points that are dominated by the TP4 no-offload and TP2 HiCache arms. Keep TP2 HiCache dense through the measured host-KV saturation region.\n\n中文：细化 B300 缓存帕累托前沿；加密 TP4 无卸载缓存临界区间，移除被 TP4 无卸载与 TP2 HiCache 方案支配的 TP4 HiCache 测点，并保留覆盖主机 KV 饱和区间的密集 TP2 HiCache 测点。 * fix(agentx): cap Qwen trace idle gaps Pin the AIPerf trace idle-gap branch and set the Qwen B300 replay cap to 300 seconds. * fix(agentx): extend SGLang keep-alive * test(agentx): use GSM8K for Qwen B300 eval 中文：Qwen B300 AgentX 扫描使用 GSM8K 评估，避免运行耗时的 SWE-bench。 * Update nvidia-master.yaml --------- Signed-off-by: Cam Quilici <cjquilici@gmail.com>
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.
Summary
Add Qwen3.5-397B-A17B NVFP4 AgentX coverage on B300 with SGLang native NEXTN MTP, golden synthetic acceptance, prefix caching, and the pinned 256k trace dataset.
Pin AIPerf to the exact validated revision with:
Dataset timestamps and derived end-to-start delays are not rewritten.