Skip to content

perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token - #5043

Merged
adibarra merged 14 commits into
mainfrom
hyukjlee/kimik3.1-mi355x-vllm-2ded1ca
Nov 14, 2026

Conversation

@hyukjlee

@hyukjlee hyukjlee commented Nov 12, 2026 •

Copy link
Copy Markdown
Collaborator

MI355X Kimi-K3.1 vLLM AgentX (kimik3.1-fp4-mi355x-vllm-agentic-dspark):

  • Image kimi-k31-d73dd529 → kimi-k31-6d74a15 (vLLM mum-tracking 6d74a152f: KV null-block fix seungrokj/mum#17, prefill-cadence hold seungrokj/mum#21, DCP V-up-first combine, multi-stream DSA decode, memory headroom, one-launch DSA decode prologue, and Kimi-K3.1 decode fusions).
  • c4 moves out of the DRAM-offload row into its own TP8 x DCP8 row without KV offload; c1 and c16-c48 keep their layouts.
  • All points: VLLM_ROCM_DCP_VUP_RS=1, VLLM_ROCM_DSA_MULTISTREAM=1, VLLM_ROCM_MEM_HEADROOM_GIB=8.
  • c1: VLLM_KIMI_SD_DSA_PROLOGUE=1.
  • c4: DCP8 + latency_specialized decode, 5 draft tokens, no KV offload, breakable graphs off.
  • c48: 3 draft tokens; drop prefill-schedule-interval, long-prefill-token-threshold and long-prefill-token-threshold-adaptive.
  • c16 / c24 / c32: prefill-schedule-interval: 32 + VLLM_PREFILL_CADENCE_NO_ESCAPE=1 (seungrokj/mum#21).
  • Remove unused override_dram_c10.

Draft data types: no change. Inferact/Kimi-K3.1-Pro-DSpark-preview resolves to Hugging Face main 6a2ef20a, whose weights are all BF16 with no quantization_config. vLLM 6d74a152f loads it unquantized with an FP8 KV cache, as d73dd529 did. Evidence: #5043 (comment)

Accuracy-sensitive optimizations: none used. Quick reduce is NONE; the new VLLM_ROCM_DCP_VUP_RS reduce-scatter and the VLLM_ROCM_DSA_MULTISTREAM all-gather run in BF16/FP32 without quantization; there is no online or shared-expert quantization and no cross-indexer sharing; the KV cache is FP8 (the FP4 indexer cache is allowed by the K3.1 week-1 policy); and the KDA state cache is FP32. Same comment as above.

Validation: run-sweep 43399050596 on 34099327, whose Kimi-K3.1 recipe and master config match the head. All six benchmark points (c1, c4, c16, c24, c32, c48) and the c1, c4, c32 and c48 evals are green. Results were staged with /reuse-sweep-run 43399050596.

中文

MI355X Kimi-K3.1 vLLM AgentX(kimik3.1-fp4-mi355x-vllm-agentic-dspark):

  • 镜像 kimi-k31-d73dd529 → kimi-k31-6d74a15(vLLM mum-tracking 6d74a152f:KV null block 修复 seungrokj/mum#17、prefill 节奏保持 seungrokj/mum#21、DCP V-up-first 合并、多流 DSA decode、显存余量、单次 launch 的 DSA decode prologue,以及 Kimi-K3.1 decode 融合)。
  • c4 从 DRAM offload 行拆出,单独成为不使用 KV offload 的 TP8 x DCP8 行;c1 和 c16-c48 的布局不变。
  • 所有点:VLLM_ROCM_DCP_VUP_RS=1、VLLM_ROCM_DSA_MULTISTREAM=1、VLLM_ROCM_MEM_HEADROOM_GIB=8。
  • c1:设置 VLLM_KIMI_SD_DSA_PROLOGUE=1。
  • c4:DCP8 + latency_specialized decode,5 个草稿 token,无 KV offload,关闭 breakable graphs。
  • c48:3 个草稿 token;移除 prefill-schedule-interval、long-prefill-token-threshold 和 long-prefill-token-threshold-adaptive。
  • c16 / c24 / c32:prefill-schedule-interval: 32 + VLLM_PREFILL_CADENCE_NO_ESCAPE=1(seungrokj/mum#21)。
  • 删除未使用的 override_dram_c10。

草稿数据类型:无变化。 Inferact/Kimi-K3.1-Pro-DSpark-preview 解析为 Hugging Face main 6a2ef20a,其权重全部为 BF16,没有 quantization_config。vLLM 6d74a152f 不做量化地加载它,KV cache 为 FP8,与 d73dd529 相同。证据:#5043 (comment)

精度敏感优化:均未使用。 quick reduce 为 NONE;新增的 VLLM_ROCM_DCP_VUP_RS reduce-scatter 和 VLLM_ROCM_DSA_MULTISTREAM all-gather 均以 BF16/FP32 运行,不做量化;没有在线量化或 shared expert 量化,也没有跨 indexer 共享;KV cache 为 FP8(K3.1 第一周精度策略允许 FP4 indexer cache);KDA 状态缓存为 FP32。证据同上。

验证:run-sweep 43399050596,运行于 34099327,其 Kimi-K3.1 recipe 和 master config 与 head 相同。六个 benchmark 点(c1、c4、c16、c24、c32、c48)以及 c1、c4、c32、c48 的 eval 全部通过。结果通过 /reuse-sweep-run 43399050596 暂存。

AI model disclosure
  • claude-opus-5-5 (Claude Code): synced the perf-changelog block and this description with the diff, and checked the draft data types and accuracy-sensitive optimizations in the pinned source.

🤖 Generated with Claude Code

@github-actions

Copy link
Copy Markdown
Contributor

Thanks for the contribution!

  • Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
  • PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
  • After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.
中文

感谢你的贡献!

  • **审阅:**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER,请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md,再通过 Slack 联系核心维护者。必须严格遵循模板,并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have,才能触发签核验证。
  • **PR 验证:**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目,并且恰好带有一个主标签时,才会运行扫描:full-sweep-fail-fast(强烈推荐;canary 加逐矩阵 fail-fast)、full-sweep-enabled(有 canary;矩阵任务在失败后继续运行)或 non-canary-full-sweep-enabled(无 canary,也无 fail-fast)。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
  • **合并后:**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复;参见重新运行失败任务的说明。

@hyukjlee
hyukjlee force-pushed the hyukjlee/kimik3.1-mi355x-vllm-2ded1ca branch 2 times, most recently from f6651b4 to 73c4344 Compare November 12, 2026 16:38

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Beyond the inline finding, I checked how the bare agentic.yaml c4 row binds. The single-node binder matches variants on CONC and KV_OFFLOADING, so only override_gpu_kv_c4 fits. It keeps the base decode-context-parallel-size: 8, which matches the row's dcp-size: 8. The two CONC=1/none variants would be ambiguous under a bare reference, which is why the c1 rows now name their overrides.

Extended reasoning...

This is a ruled-out note added to the inline finding. The single-node binder in infx/srt_slurm/single_node.py matches each recipe variant on its CONC, KV_OFFLOADING and parallelism fields. That leaves exactly one valid binding for the c4 row and one for each named c1 row.

@hyukjlee
hyukjlee force-pushed the hyukjlee/kimik3.1-mi355x-vllm-2ded1ca branch from 73c4344 to 64e8bfc Compare November 12, 2026 16:47
@hyukjlee hyukjlee changed the title perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-2ded1ca, env flags, c1/c4 specialized decode, prefill cadence at c16-c32 / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-2ded1ca、环境变量、c1/c4 specialized decode、c16-c32 prefill cadence perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-2ded1ca, env flags, c1/c4 specialized decode, c48 three draft tokens / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-2ded1ca、环境变量、c1/c4 specialized decode、c48 三个草稿 token Nov 12, 2026
@hyukjlee hyukjlee added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Nov 12, 2026
@github-actions

github-actions Bot commented Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

@hyukjlee hyukjlee added non-canary-full-sweep-enabled Run the full sweep without the canary gate (full search space, no trim) and removed full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) labels Nov 12, 2026
…flags, c1 nst 7/5, c4 GPU KV DCP8 latency_specialized, c48 nst 3

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@hyukjlee
hyukjlee force-pushed the hyukjlee/kimik3.1-mi355x-vllm-2ded1ca branch from 64e8bfc to 3bec727 Compare November 12, 2026 16:57
@seungrokj seungrokj removed the non-canary-full-sweep-enabled Run the full sweep without the canary gate (full search space, no trim) label Nov 13, 2026
seungrokj and others added 3 commits November 13, 2026 09:34
Both concurrency-1 overrides (override_gpu_kv_c1, override_gpu_kv_c1_nst5)
run each DSA layer's specialized-decode prologue as one launch (mum PR 3,
D1a). It needs DCP1, so the DCP8 c4 override is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Remove override_gpu_kv_c1_nst5 and its master-config row; c1 runs once
with override_gpu_kv_c1 (7 draft tokens), selected by CONC as on main.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… cadence

Image kimi-k31-6d74a15 (mum-tracking 6d74a152f) adds
VLLM_PREFILL_CADENCE_NO_ESCAPE (mum PR 21). c16, c24 and c32 set
prefill-schedule-interval 32 with the flag on, since their max-num-seqs is
below the concurrency and the capacity escape would drop the hold.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@seungrokj seungrokj changed the title perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-2ded1ca, env flags, c1/c4 specialized decode, c48 three draft tokens / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-2ded1ca、环境变量、c1/c4 specialized decode、c48 三个草稿 token perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token Nov 13, 2026
seungrokj and others added 2 commits November 13, 2026 17:04
…refill threshold

c48 no longer sets prefill-schedule-interval 8, long-prefill-token-threshold
10240 or long-prefill-token-threshold-adaptive; it uses vLLM's defaults.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
override_dram_c52 and override_dram_c56 copy c48 (TP8 x DCP8, DRAM offload,
3 draft tokens, max-num-batched-tokens 10240, gpu-memory-utilization 0.91)
with max-num-seqs equal to the concurrency and decode capture sizes up to
max-num-seqs x 4 (208, 224); the DRAM row lists 52 and 56.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@seungrokj seungrokj changed the title perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens, add c52/c56 / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token、新增 c52/c56 Nov 13, 2026
override_dram_c60 copies c48 with max-num-seqs 60 and decode capture
sizes up to 240 (max-num-seqs x 4); the DRAM row lists 60.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@seungrokj seungrokj changed the title perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens, add c52/c56 / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token、新增 c52/c56 perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens, add c52/c56/c60 / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token、新增 c52/c56/c60 Nov 13, 2026
c32 no longer sets long-prefill-token-threshold 10240 or
long-prefill-token-threshold-adaptive; prefill-schedule-interval 32 and
VLLM_PREFILL_CADENCE_NO_ESCAPE=1 stay.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@seungrokj seungrokj added the non-canary-full-sweep-enabled Run the full sweep without the canary gate (full search space, no trim) label Nov 13, 2026
@seungrokj seungrokj changed the title perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens, add c52/c56/c60 / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token、新增 c52/c56/c60 perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-6d74a15, env flags, c1/c4 specialized decode, c16-c32 prefill cadence, c48 three draft tokens / perf(amd):Kimi-K3.1 MI355X vLLM AgentX:镜像 kimi-k31-6d74a15、环境变量、c1/c4 specialized decode、c16-c32 prefill 节奏、c48 三个草稿 token Nov 13, 2026
@seungrokj

Copy link
Copy Markdown
Collaborator

/use 43399050596

@seungrokj

Copy link
Copy Markdown
Collaborator

/reuse-sweep-run 43399050596

@github-actions

github-actions Bot commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

@seungrokj staged run 43399050596.

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

@seungrokj

Copy link
Copy Markdown
Collaborator

recipe: vllm-project/vllm#67487

- Update the image from vllm/vllm-openai-rocm:kimi-k31-d73dd529 (sha256:5b54f223) to vllm/vllm-openai-rocm:kimi-k31-6d74a15 (sha256:c470d0de), vLLM mum-tracking 6d74a152f.
- Split concurrency 4 out of the DRAM-offload row into its own TP8 x DCP8 row with kv-offloading none (recipe override override_dram_c4 renamed override_gpu_kv_c4). Concurrency 1 (TP8 x DCP1, GPU KV) and 16, 24, 32 and 48 (TP8 x DCP8, Simple CPU DRAM offload) keep their layouts. Remove override_dram_c10, unused since concurrency 10 left the sweep.
- Every point sets VLLM_ROCM_DCP_VUP_RS=1 (DCP V-up reduce-scatter, on top of the existing VLLM_ROCM_DCP_AITER_ALLGATHER=1), VLLM_ROCM_DSA_MULTISTREAM=1 and VLLM_ROCM_MEM_HEADROOM_GIB=8.
- Concurrency 1 (latency_specialized decode backend) sets VLLM_KIMI_SD_DSA_PROLOGUE=1, which runs each DSA layer's decode prologue in one launch.
- Concurrency 4 keeps max-num-seqs 2 and switches to the latency_specialized decode backend, 5 draft tokens (golden AL 3.41, from 7), decode-only capture sizes 6 and 12 and breakable CUDA graphs off.
- Concurrency 16, 24 and 32 set prefill-schedule-interval 32 (from 1, 1 and 8) and VLLM_PREFILL_CADENCE_NO_ESCAPE=1. Their max-num-seqs is below the concurrency, so requests always wait; without the flag vLLM drops the prefill hold (capacity escape). Concurrency 32 drops long-prefill-token-threshold 10240 and long-prefill-token-threshold-adaptive.
- Concurrency 48 uses 3 draft tokens (golden AL 2.88, from 6) with decode capture sizes 4-192, and drops prefill-schedule-interval 8, long-prefill-token-threshold 10240 and long-prefill-token-threshold-adaptive (back to vLLM defaults).
- The new vLLM (mum-tracking d73dd529 to 6d74a152f) zeroes the KV null block after CUDA graph capture in the V2 runner (seungrokj/mum#17) and adds the opt-in VLLM_PREFILL_CADENCE_NO_ESCAPE hold (seungrokj/mum#21), the DCP V-up-first combine and multi-stream DSA decode, VLLM_ROCM_MEM_HEADROOM_GIB and the one-launch DSA decode prologue. It also fuses the MXFP4 indexer decode prep, the q_a/kv_a RMSNorm with the sigmoid gate, the KDA f_b_proj GEMM, the AttnRes register tile and the latent-MoE norm, compiles the FlyDSL sparse MLA at load and sizes the packed DSA KV blocks tightly.
- No change to the DSpark draft's data types: speculative-config does not pin a revision, so Inferact/Kimi-K3.1-Pro-DSpark-preview resolves to Hugging Face main 6a2ef20a (current since 2026-11-05), whose weights are all BF16 with no quantization_config. vLLM 6d74a152f loads it unquantized with an FP8 KV cache, as d73dd529 did; VLLM_ROCM_DCP_VUP_RS also combines the draft's DCP attention output, in BF16 with an FP32 LSE.

中文:
- 将镜像从 vllm/vllm-openai-rocm:kimi-k31-d73dd529(sha256:5b54f223)更新为 vllm/vllm-openai-rocm:kimi-k31-6d74a15(sha256:c470d0de),即 vLLM mum-tracking 6d74a152f。
- 将并发 4 从 DRAM offload 行拆出,单独成为 kv-offloading none 的 TP8 x DCP8 行(recipe override override_dram_c4 改名为 override_gpu_kv_c4)。并发 1(TP8 x DCP1,KV 在 GPU 上)以及 16、24、32、48(TP8 x DCP8,Simple CPU DRAM offload)的布局不变。删除 override_dram_c10,并发 10 退出扫描后已不再使用。
- 所有点设置 VLLM_ROCM_DCP_VUP_RS=1(DCP V-up reduce-scatter,叠加在已有的 VLLM_ROCM_DCP_AITER_ALLGATHER=1 之上)、VLLM_ROCM_DSA_MULTISTREAM=1 和 VLLM_ROCM_MEM_HEADROOM_GIB=8。
- 并发 1(latency_specialized decode 后端)设置 VLLM_KIMI_SD_DSA_PROLOGUE=1,将每个 DSA 层的 decode prologue 合并为一次 launch。
- 并发 4 保持 max-num-seqs 2,改用 latency_specialized decode 后端、5 个草稿 token(golden AL 3.41,原为 7)、仅 decode 的 capture sizes 6 和 12,并关闭 breakable CUDA graph。
- 并发 16、24、32 设置 prefill-schedule-interval 32(原为 1、1、8)和 VLLM_PREFILL_CADENCE_NO_ESCAPE=1。它们的 max-num-seqs 低于并发数,请求始终在排队;不设此标志时 vLLM 会放弃 prefill 保持(capacity escape)。并发 32 去掉 long-prefill-token-threshold 10240 和 long-prefill-token-threshold-adaptive。
- 并发 48 使用 3 个草稿 token(golden AL 2.88,原为 6),decode capture sizes 为 4-192,并去掉 prefill-schedule-interval 8、long-prefill-token-threshold 10240 和 long-prefill-token-threshold-adaptive(恢复 vLLM 默认值)。
- 新 vLLM(mum-tracking d73dd529 到 6d74a152f)在 V2 runner 中于 CUDA graph capture 之后将 KV null block 清零(seungrokj/mum#17),并新增可选的 VLLM_PREFILL_CADENCE_NO_ESCAPE 保持(seungrokj/mum#21)、DCP V-up-first 合并与多流 DSA decode、VLLM_ROCM_MEM_HEADROOM_GIB 以及单次 launch 的 DSA decode prologue。它还融合了 MXFP4 indexer decode 准备、q_a/kv_a RMSNorm 与 sigmoid gate、KDA f_b_proj GEMM、AttnRes register tile 和 latent-MoE norm,在加载时编译 FlyDSL sparse MLA,并收紧 packed DSA KV block 的大小。
- DSpark 草稿的数据类型无变化:speculative-config 未固定 revision,因此 Inferact/Kimi-K3.1-Pro-DSpark-preview 解析为 Hugging Face main 6a2ef20a(自 2026-11-05 起为当前版本),其权重全部为 BF16,没有 quantization_config。vLLM 6d74a152f 不做量化地加载它,KV cache 为 FP8,与 d73dd529 相同;VLLM_ROCM_DCP_VUP_RS 也用于合并草稿的 DCP attention 输出,使用 BF16 和 FP32 LSE。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@seungrokj

Copy link
Copy Markdown
Collaborator

Draft data types: no change. The DSpark draft Inferact/Kimi-K3.1-Pro-DSpark-preview is not pinned to a revision in speculative-config, so it resolves to Hugging Face main, revision 6a2ef20a (current since 2026-11-05 and the revision of the golden AL curve). It is a standalone 5-layer K3DSparkModel checkpoint with no quantization_config, and all 67 tensors in its safetensors header are BF16:

Part Data type
Attention projections (q_a_proj, q_b_proj, kv_a_proj_with_mqa, kv_b_proj, o_proj) BF16
Dense MLP (gate_proj, up_proj, down_proj) BF16
context_proj, Markov head, confidence head BF16
Norms BF16
Embedding and LM head not in the draft checkpoint; shared from the target, whose lm_head is on the compressed-tensors ignore list and stays BF16
KV cache FP8, from speculative-config kv_cache_dtype: fp8, the same as the target's kv-cache-dtype: fp8

The image moves from vLLM d73dd529f2d5ccfc9735f0c1126b2331ca1cc58f to 6d74a152f79c14ff366f546476b9d542adc290a2 (kimi-k31-6d74a15, mum-tracking). These commits live in a non-public repository, not vllm-project/vllm, so there are no public permalinks. Paths are at 6d74a152f, and none of the files below changes between the two commits:

  • vllm/config/speculative.py:1308-1345: the draft inherits the target's quantization only when speculative-config gives no model. Every point names the draft repo, so quantization stays unset. :1495 builds the draft ModelConfig with it, and :1620 keeps K3DSparkModel out of the DeepSeek DSpark branch that copies the target's quantization (:1636).
  • vllm/config/vllm.py:937: with no quantization, the quantization config is None, so get_draft_quant_config() (vllm/model_executor/models/utils.py:1062-1084) returns None.
  • vllm/models/kimi_k3/nvidia/dspark_mla.py:72,145: every draft linear layer takes that None config and loads in BF16. :434-443 drops embed_tokens and lm_head from the draft and shares the target's. vllm/model_executor/models/registry.py:702-705 maps K3DSparkModel to this class on ROCm too.
  • vllm/config/speculative.py:421-430: only speculative-config's own attention_backend (ROCM_AITER_MLA) and kv_cache_dtype (fp8) override the draft's attention backend and KV dtype.

The recipe sets no quantization-config, --quantization or draft quantization at any point, before or after this PR, so the draft runs as shipped: BF16 weights and activations with an FP8 KV cache matching the target's. The new draft lengths (5 at c4, 3 at c48) do not change any data type. The only new code on the draft's path is VLLM_ROCM_DCP_VUP_RS, which also combines the draft MLA's DCP partial outputs (the c48 server log shows v_up-first combine site=drafter(kimi_k3 mla) used=True); that combine is BF16 with an FP32 LSE (below).

Accuracy-sensitive optimizations: none used. Checked against the accuracy-sensitive optimization policy (#4705, not yet merged) and the Kimi K3.1 Pro week-1 precision policy. The setup is aggregated TP8 (DCP1 at c1, DCP8 elsewhere) with the DSpark draft worker in the same engine. I checked image defaults at vLLM 6d74a152f:

Optimization Effective setting (base → head) Evidence
Reduced-precision all-reduce / reduce-scatter Off → off. VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: NONE in the recipe (also the engine default), and quick reduce returns before initializing on NONE. The new VLLM_ROCM_DCP_VUP_RS=1 path all-gathers an FP32 LSE and reduce-scatters the v-projected output in the model dtype (BF16) through the AITER IPC comm, with no quantization. The new VLLM_ROCM_DSA_MULTISTREAM=1 aux comm only all-gathers FP32 indexer scores. AITER_FLYDSL_STAGE2_FP8=0, AITER_BF16_FP8_MOE_BOUND=0. vllm/envs.py:1520-1524; vllm/distributed/device_communicators/quick_all_reduce.py:186-191; vllm/v1/attention/ops/dcp.py:1204-1205,1226,1243; vllm/models/kimi_k3/amd/dsa.py:538-540,710,887; recipe agentic.yaml:84-86
Online linear-layer quantization Not used → not used. No quantization-config or --quantization; the target runs its shipped compressed-tensors MXFP4 (routed experts only). recipe; checkpoint quantization_config
Shared-expert quantization Not used → not used. re:.*shared_experts.* is on the checkpoint's ignore list (BF16 as shipped), and nothing overrides it. config.json text_config.quantization_config.ignore
Non-native cross-indexer sharing Not used → not used. No hf-overrides; the dsa_config index settings are the checkpoint's own. recipe; config.json text_config.dsa_config
Non-native FP4 KV cache Not used → not used. kv-cache-dtype: fp8 (target and draft). The indexer_kv_dtype: mxfp4 indexer cache is unchanged and is allowed by the K3.1 policy ("Indexer cache: FP4 allowed"). recipe agentic.yaml:43,47; MODELS.md K3.1 policy
SSM/KDA/Mamba state cache below FP32 Not used → not used. mamba-ssm-cache-dtype: float32; the server config in run 43399050596 shows mamba_ssm_cache_dtype: 'float32'. recipe agentic.yaml:45; c1/c48 server logs

No optimization on the list is enabled, so the three-arm KVV qualification does not apply. I found no model-trainer veto in the PR or the linked issues. Validation: run 43399050596 on 34099327 (same Kimi-K3.1 recipe and master config as the head), with all six benchmark points and the c1/c4/c32/c48 evals green.

中文

草稿数据类型:无变化。 DSpark 草稿 Inferact/Kimi-K3.1-Pro-DSpark-preview 在 speculative-config 中没有固定 revision,因此解析为 Hugging Face main,即 revision 6a2ef20a(自 2026-11-05 起为当前版本,也是 golden AL 曲线所用的 revision)。它是独立的 5 层 K3DSparkModel checkpoint,没有 quantization_config,其 safetensors header 中的 67 个张量全部为 BF16:

部分 数据类型
注意力投影(q_a_proj、q_b_proj、kv_a_proj_with_mqa、kv_b_proj、o_proj) BF16
Dense MLP(gate_proj、up_proj、down_proj) BF16
context_proj、Markov head、confidence head BF16
Norm BF16
Embedding 和 LM head 不在草稿 checkpoint 中;与目标模型共享,目标模型的 lm_head 在 compressed-tensors 的 ignore 列表中,保持 BF16
KV cache FP8,来自 speculative-config 的 kv_cache_dtype: fp8,与目标模型的 kv-cache-dtype: fp8 相同

镜像从 vLLM d73dd529f2d5ccfc9735f0c1126b2331ca1cc58f 更新为 6d74a152f79c14ff366f546476b9d542adc290a2(kimi-k31-6d74a15,mum-tracking)。这些 commit 位于非公开仓库,不在 vllm-project/vllm 中,因此没有公开的永久链接。以下路径均基于 6d74a152f,且下列文件在两个 commit 之间均无改动:

  • vllm/config/speculative.py:1308-1345:只有在 speculative-config 未指定 model 时,草稿才会继承目标模型的量化。每个点都指定了草稿仓库,因此 quantization 保持未设置。:1495 用它构建草稿的 ModelConfig,:1620 将 K3DSparkModel 排除在会复制目标模型量化的 DeepSeek DSpark 分支(:1636)之外。
  • vllm/config/vllm.py:937:没有量化时量化配置为 None,因此 get_draft_quant_config()(vllm/model_executor/models/utils.py:1062-1084)返回 None。
  • vllm/models/kimi_k3/nvidia/dspark_mla.py:72,145:草稿的每个线性层都使用这个 None 配置,以 BF16 加载。:434-443 从草稿中丢弃 embed_tokens 和 lm_head,改为共享目标模型的。vllm/model_executor/models/registry.py:702-705 在 ROCm 上同样将 K3DSparkModel 映射到这个类。
  • vllm/config/speculative.py:421-430:只有 speculative-config 自身的 attention_backend(ROCM_AITER_MLA)和 kv_cache_dtype(fp8)会覆盖草稿的注意力后端和 KV dtype。

本 recipe 在本 PR 前后的任何点都没有设置 quantization-config、--quantization 或草稿量化,因此草稿按发布形式运行:BF16 权重和激活,FP8 KV cache 与目标模型一致。新的草稿长度(c4 为 5,c48 为 3)不改变任何数据类型。草稿路径上唯一的新代码是 VLLM_ROCM_DCP_VUP_RS,它也用于合并草稿 MLA 的 DCP 部分输出(c48 服务器日志显示 v_up-first combine site=drafter(kimi_k3 mla) used=True);该合并使用 BF16 和 FP32 LSE(见下)。

精度敏感优化:均未使用。 依据精度敏感优化策略(#4705,尚未合并)和 Kimi K3.1 Pro 第一周精度策略进行检查。部署为聚合式 TP8(c1 为 DCP1,其余为 DCP8),DSpark 草稿 worker 位于同一引擎中。检查范围包括 vLLM 6d74a152f 的镜像默认值:

优化 实际设置(base → head) 证据
降精度 all-reduce / reduce-scatter 关闭 → 关闭。recipe 设置 VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: NONE(也是引擎默认值),quick reduce 在 NONE 时不初始化即返回。新的 VLLM_ROCM_DCP_VUP_RS=1 路径通过 AITER IPC comm 对 FP32 LSE 做 all-gather,并以模型 dtype(BF16)对 v 投影后的输出做 reduce-scatter,不做量化。新的 VLLM_ROCM_DSA_MULTISTREAM=1 辅助 comm 只对 FP32 indexer 分数做 all-gather。AITER_FLYDSL_STAGE2_FP8=0,AITER_BF16_FP8_MOE_BOUND=0。 vllm/envs.py:1520-1524;vllm/distributed/device_communicators/quick_all_reduce.py:186-191;vllm/v1/attention/ops/dcp.py:1204-1205,1226,1243;vllm/models/kimi_k3/amd/dsa.py:538-540,710,887;recipe agentic.yaml:84-86
线性层在线量化 未使用 → 未使用。没有 quantization-config 或 --quantization;目标模型使用其发布的 compressed-tensors MXFP4(仅 routed experts)。 recipe;checkpoint quantization_config
Shared expert 量化 未使用 → 未使用。re:.*shared_experts.* 在 checkpoint 的 ignore 列表中(按发布形式为 BF16),且没有任何覆盖。 config.json text_config.quantization_config.ignore
非原生跨 indexer 共享 未使用 → 未使用。没有 hf-overrides;dsa_config 的 index 设置均为 checkpoint 自带。 recipe;config.json text_config.dsa_config
非原生 FP4 KV cache 未使用 → 未使用。kv-cache-dtype: fp8(目标与草稿)。indexer_kv_dtype: mxfp4 的 indexer cache 未改动,K3.1 策略允许(“Indexer cache: FP4 allowed”)。 recipe agentic.yaml:43,47;MODELS.md K3.1 策略
SSM/KDA/Mamba 状态缓存低于 FP32 未使用 → 未使用。mamba-ssm-cache-dtype: float32;run 43399050596 的服务器配置显示 mamba_ssm_cache_dtype: 'float32'。 recipe agentic.yaml:45;c1/c48 服务器日志

列表中的优化均未启用,因此不需要三组 KVV 资格验证。在 PR 及其关联 issue 中未发现模型训练方否决。验证:run 43399050596,运行于 34099327(Kimi-K3.1 recipe 和 master config 与 head 相同),六个 benchmark 点以及 c1/c4/c32/c48 的 eval 全部通过。

@seungrokj seungrokj left a comment •

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

As a PR reviewer and CODEOWNER, I have reviewed this and have:

Checklist items
  • Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
  • Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
  • Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this.
  • Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
  • Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
  • Verified that every draft model and draft head is served as it ships: the draft that ships with the served checkpoint, at its stored precision, through the pinned upstream image's default handling, with the shipped and effective draft precision recorded in the additional detail section. No submission-side quantization, dtype override, checkpoint substitution, or patch may lower draft precision below that default, regardless of eval results or AL. Explicitly verified that SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not enabled in the effective recipe, including inherited settings; enabling it is prohibited going forward, and historical runs do not grant an exception. See Draft-model precision for what counts as the default and the MLPerf comparison.
  • For agentic workloads: verified that speculative-decoding configs (EAGLE / MTP / draft models) run with simulated synthetic acceptance, with the acceptance-length value taken from the committed golden AL curve in infx/golden_al_distribution/ for that model, thinking mode, and draft length. A submission may choose any supported draft length, but it may not substitute a different acceptance target.
  • Verified against the current MODELS.md that this PR does not submit a deprecated model, scenario, or model-scenario combination.
  • Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; target/verifier FLOPs at lower precisions is fine, given that the config passes private evals, but this does not permit lowering draft-model or draft-head precision below what ships. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
  • Verified that new models disable integer all-reduce, with SGLang ROCm explicitly setting ROCM_QUICK_REDUCE_QUANTIZATION=NONE. Existing models are grandfathered per MODELS.md.
  • For affected Dynamo TRT-LLM benchmarks, verified that iteration/KV-cache metrics are collected and published under the pinned Dynamo implementation and effective engine settings. Resolve the installed Dynamo version separately from the container tag; use supported metrics controls without requiring KV events. Record the version and configuration evidence in the additional detail section. See verifier Check 17. N/A for unaffected serving paths or frameworks without these Dynamo settings.
  • If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
  • For official InferenceX (non-preview), verified that established models use upstream engine images built from master (or the equivalent default development branch), with new models transitioning within a reasonable time after release. See MODELS.md.
  • If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
  • Verified that every affected vLLM/SGLang production serving configuration in this PR, including Dynamo deployments, is documented in the official vLLM recipes and/or the SGLang cookbook, for both single-node and multi-node aggregated and disaggregated serving. The documentation covers production-useful flags, environment variables, and launch commands (including prefill, decode, and router/frontend commands for disaggregation); InferenceX-only benchmark/harness settings are excluded with reasoning in the additional detail section:
    • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang cookbook source repo and verified that it is MERGED before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
  • Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at docs/waiver/<PR_NUMBER>.md — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked below in the additional detail section.
  • If this PR uses append-only: true, verified that it only adds generated points or recipe variants inside a selected existing config/scenario and existing same-image visual curve: every previously generated point remains present with the same recipe, no prior point is removed or rerun, and every benchmark-affecting change in the complete diff can affect only the corresponding newly appended points (never an existing point), regardless of which file contains it.
  • If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.
  • Reported measured throughput/E2EL Pareto counts and evidence per affected curve (≥5 points strongly recommended). Below 5 or unverifiable: tag a core maintainer for review; recorded admin bypass required before merge. N/A if no curves are affected. Details.

Additional detail section:

Signed: seungrokj

@github-actions

github-actions Bot commented Nov 13, 2026 •

Copy link
Copy Markdown
Contributor

⚠️ Verdict: WARN ⚠️

⚠️ Check 18 (Recipe consistency): WARN — Re-tuned points on kimik3.1-fp4-mi355x-vllm-agentic-dspark regress >10% vs the d73dd529 run 42638222682 attempt 3 (new: 43399050596 on 34099327, P90): c16 E2EL 25.90 -> 28.60 s (+10.4%), TTFT 6.29 -> 10.37 s (+64.9%), E2E norm. interactivity 23.33 -> 15.17 (-35.0%); c24 E2EL 37.39 -> 41.19 s (+10.2%), TTFT 10.93 -> 16.53 s (+51.2%), E2E norm. interactivity 15.28 -> 9.71 (-36.5%); c32 throughput/GPU 17108 -> 15331 (-10.4%), E2EL 43.75 -> 50.45 s (+15.3%), TTFT 10.28 -> 21.10 s (+105.2%), E2E norm. interactivity 15.00 -> 7.62 (-49.2%); c1/c4/c48 improve and no removed point is undominated; (a)-(e), (g), (i) pass (configs generate, master rows match recipe overrides, image digest resolves, stream-interval 10 on every point).

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — seungrokj is a listed owner of inferencex-e2e/configs/amd-master.yaml and the **/mi355x-*/ srt-slurm recipes; perf-changelog.yaml falls to the catch-all.

✅ Check 1 (Sweep on in-PR commit): PASS — run 43399050596 on in-PR 34099327 has all six agentic / points (c1, c4, c16, c24, c32, c48) and four agentic eval / jobs successful; later commits are non-material: recipe and master entry are byte-identical at 0f287688, 6cbe3c5e is changelog-only, and the merges bring only other configs, CI plumbing, and the client-side agentx-harness 1.0.13 -> 1.0.14 memory/scheduling efficiency fix for very high concurrency (#5071).

✅ Check 2 (Evals): PASS — GSM8K em_strict on the PR image kimi-k31-6d74a15@sha256:c470d0de: c1 0.979, c4 0.981, c32 0.980, c48 0.981 (n=1319); every gap against the best sibling and the exact-scope cohort (42638222682, 0.983) is within 2-sigma; evals use the same recipe overrides (c4 eval launches latency_specialized with block verification, c1 the DSA-prologue point, c32 the prefill-cadence point).

✅ Check 3 (Upstream recipe): PASS — vllm-project/vllm#67600 is MERGED (2026-11-14, 5ec99956), and vllm-project/recipes syncs from that commit (sources.lock.json); recipes.vllm.ai Kimi-K3.1-Pro publishes the MI355X AgentX TP8 preset with all six points, the pinned image digest, VLLM_ROCM_DCP_VUP_RS, VLLM_ROCM_DSA_MULTISTREAM, VLLM_ROCM_MEM_HEADROOM_GIB=8, VLLM_KIMI_SD_DSA_PROLOGUE, VLLM_PREFILL_CADENCE_NO_ESCAPE, --prefill-schedule-interval 32, per-point draft counts, capture sizes, and offload settings; synthetic acceptance is documented as throughput-only.

✅ Check 4 (Reuse command): PASS — seungrokj (permission write, role write) posted /use 43399050596 and /reuse-sweep-run 43399050596 on their own lines; that run is the Check 1 run.

✅ Check 5 (Latest checklist): PASS — every item in the current template is present and checked, except the conditional "cannot reasonably be satisfied" item, which needs no explanation because nothing is unsatisfied.

✅ Check 6 (Upstream images): PASS — framework: vllm uses vllm/vllm-openai-rocm:kimi-k31-6d74a15@sha256:c470d0de from the upstream vLLM Docker Hub org; engine-first ordering does not apply because this is a vLLM entry; the image is a temporary model-enablement build (the non-public mum-tracking branch 6d74a152f, features not yet on vLLM main per the merged recipe) for Kimi-K3.1, which was added on 2026-11-04 and is 10 days old, so the new-model transition window allows it.

✅ Check 7 (Deprecated models): PASS — kimik3.1 agentic coding is active in MODELS.md (added 2026-11-04, no retirement).

✅ Check 8 (No architecture hacks): PASS — no hf-overrides or config edits; VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: NONE for this new model, with no other quick-reduce control set; the new DCP reduce-scatter and DSA multistream comms run in BF16/FP32; the MXFP4 indexer cache is allowed by the K3.1 week-1 policy and is covered by the evals in Check 2.

✅ Check 9 (Chat template): PASS — AgentX replays chat-format agentic traces through agentx profile (agentx-harness), not raw completions; GSM8K evals use the same chat route.

✅ Check 10 (No engine patches): PASS — no patch files, setup scripts, heredocs, or wheel installs in the recipe or diff; the pinned image runs as shipped.

✅ Check 11 (Agentic golden AL): PASS — throughput runs pin vLLM rejection_sample_method: synthetic at the golden K3.1 DSpark thinking_on values; job logs show c4 5 tokens -> 3.41 and c48 3 tokens -> 2.88, matching kimi-k3.1-pro__dspark__...probabilistic-block.yaml (c1 7 -> 3.69, c16/c24 4 -> 3.19, c32 3 -> 2.88 come from the same launcher path).

➖ Check 12 (Append-only): N/A — the new changelog entry is not append-only: true.

✅ Check 13 (Draft precision): PASS — the draft is Inferact/Kimi-K3.1-Pro-DSpark-preview @ 6a2ef20a, a 5-layer K3DSparkModel stored as bfloat16 with no quantization_config; speculative-config sets no quantization or dtype field and draft FP8 KV matches the target's kv-cache-dtype: fp8; there is no SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE (vLLM path); the sign-off's linked evidence records the pinned load path.

✅ Check 14 (Pareto coverage): PASS — one curve (kimik3.1 / agentic-coding / mi355x vLLM / fp4 / P90 E2EL, run 43399050596 results_bmk, image kimi-k31-6d74a15) has 6/6 frontier points: (22.59 s, 2033), (25.41, 3076), (28.60, 8901), (41.19, 13198), (50.45, 15331), (63.01, 19201) tok/s/GPU.

✅ Check 15 (PR description): PASS — the title, body, and changelog match the pinned recipe: image, c4 row split with no offload, the new env flags, c1 prologue, c4 5 tokens / latency_specialized / capture sizes 6 and 12, c16-c32 interval 32 + NO_ESCAPE, c48 3 tokens with the long-prefill knobs dropped, override_dram_c10 removed, and validation on 34099327.

➖ Check 16 (NVIDIA overview): N/A — the only affected hardware is AMD MI355X.

➖ Check 17 (Metrics publication): N/A — this is a direct vLLM deployment with no Dynamo TRT-LLM worker.

Assessed commit: 0f28768825026ee075b9fe0e493208116880f7b4.

@adibarra
adibarra merged commit fabf7ba into main Nov 14, 2026
51 of 52 checks passed
@adibarra
adibarra deleted the hyukjlee/kimik3.1-mi355x-vllm-2ded1ca branch November 14, 2026 18:12
@github-actions

Copy link
Copy Markdown
Contributor

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category new duplicate skipped unmapped failed
benchmark 6 0 6 0 0
eval 4 4 4 0 0
trace 6 0 0 0 0
power 6 0 0 0 0

benchmarkFiles 7, rawBenchmarkRows 12, coveredByPerJobFiles 6, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 5276, gpuMetricSamplesNew 257456, powerPublicationPoints 6, changelogEntriesWritten 1

Counts cover only this run. Rows: benchmark_results, eval_results, agentic rows linked to a trace, gpu_metrics series. This comment updates if the ingest is rerun.

中文

Ingest 报告 — ✅ 结果已写入 production 数据库。

上表仅统计本次运行写入的数据,各行单位依次为 benchmark_results 行、eval_results 行、已关联 trace 的 agentic 行和 gpu_metrics 序列。重新运行 ingest 时本评论会原地更新。

Fridge003 added a commit that referenced this pull request Nov 16, 2026
…ernel-c1

Resolve the perf-changelog.yaml tail conflict: keep main's new entries
(#5043, #5081, #5084, #5082, #5089) byte for byte and append this PR's
Muse Spark GB300 megakernel entry after them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

non-canary-full-sweep-enabled Run the full sweep without the canary gate (full search space, no trim)

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

3 participants