[Klaud Cold] Update dsr1-fp8-b200-trt TensorRT-LLM image to 1.3.0rc27 / 将 dsr1-fp8-b200-trt 的 TensorRT-LLM 镜像更新至 1.3.0rc27

[adibarra]

Goal: Update TensorRT-LLM image from nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc14 to nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc27.
Baseline: 2026-05-18 · nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc14
Mean latency · Sources: API 1, API 2, API 3

Point	Total tok/s/GPU ↑	Output tok/s/GPU ↑	TTFT ms ↓	TPOT ms ↓
1k/1k c4 TP8 EP1 3bbdf9	109.42	54.45	189.91	8.69
1k/1k c8 TP4 EP1 f3f420	294.45	147.77	194.49	13.06
1k/1k c8 TP8 EP1 f65dd1	176.52	88.59	180.97	10.86
1k/1k c16 TP4 EP1 c85941	460.04	228.8	218.01	16.92
1k/1k c64 TP8 EP1 a14bb9	749.44	374.64	569.02	20.06
1k/1k c128 TP8 EP1 b974b1	1,228.86	613.76	591.66	24.54
8k/1k c4 TP8 EP1 0d4f86	455.28	50.65	322.69	9.29
8k/1k c4 TP8 EP1 36cb9b	N/A	N/A	N/A	N/A
8k/1k c8 TP4 EP1 b46f49	N/A	N/A	N/A	N/A
8k/1k c8 TP4 EP1 c34538	1,162.66	130.7	473.53	14.55
8k/1k c8 TP8 EP1 3d1071	733.64	82.47	338	11.56
8k/1k c8 TP8 EP1 e0d29a	N/A	N/A	N/A	N/A
8k/1k c16 TP4 EP1 a53ad4	1,696.68	188	626.53	20.07
8k/1k c16 TP4 EP1 dfc182	N/A	N/A	N/A	N/A
8k/1k c32 TP4 EP1 8259b0	2,059.89	230.35	8,292.35	24.69
8k/1k c32 TP4 EP1 fd4a5c	N/A	N/A	N/A	N/A
8k/1k c64 TP8 EP1 51b444	2,215.41	245.78	878.6	31
8k/1k c64 TP8 EP1 631469	N/A	N/A	N/A	N/A
8k/1k c128 TP8 EP1 073724	2,908.37	322.15	1,674.34	47.08
8k/1k c128 TP8 EP1 9d2b6c	N/A	N/A	N/A	N/A
8k/1k c256 TP8 EP1 61dbdd	3,492.78	388.2	3,615.08	77.21
8k/1k c256 TP8 EP1 75f40d	N/A	N/A	N/A	N/A

Note: 8k/1k c4 TP8 EP1 36cb9b, 8k/1k c8 TP4 EP1 b46f49, 8k/1k c8 TP8 EP1 e0d29a, 8k/1k c16 TP4 EP1 dfc182, 8k/1k c32 TP4 EP1 fd4a5c, 8k/1k c64 TP8 EP1 631469, 8k/1k c128 TP8 EP1 9d2b6c, 8k/1k c256 TP8 EP1 75f40d: unavailable.

Eval	Score ↑	Samples
gsm8k/em_strict · c64	95.68%	1,319
gsm8k/em_strict · c256	95.45%	1,319

中文

**目标：**将 TensorRT-LLM 镜像从 nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc14 更新为 nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc27。
**基线：**2026-05-18 · nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc14
平均延迟 · 来源： API 1, API 2, API 3；数值及异常说明见上表。

[github-actions]

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 only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• 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：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **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。

[adibarra]

Initial update 0/5 · Not dispatched · no run · 2026-09-20 UTC
nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc27 · 9748705d861acc4a84f350f78f291fdbd485c2c9 · 8k/1k · TP4/EP1, TP8/EP1 · Mean latency
Change: Bump only the dsr1-fp8-b200-trt master image from the NGC release container 1.3.0rc14 (index digest sha256:da677289043b8e41b734962bca1f52991ff6eed945a603cb99dec521ca6b25d2, pushed 2026-05-06, image env TRT_LLM_GIT_COMMIT = 93cb6518 = tag v1.3.0rc14) to 1.3.0rc27 (index digest sha256:2af290225a38d163b26af3be48581a722db98a594e839ff1e3675cc2c69f1df8, pushed 2026-09-14, TRT_LLM_GIT_COMMIT = 6e1cc953 = tag v1.3.0rc27, the newest release on GitHub (2026-09-18) and on NGC; no 1.3.0 final or 1.4.0rc0 tag exists). Coupled stack moves from CUDA 13.1.1 / PyTorch 2.11.0a0 (NGC PyTorch 26.02) / NCCL 2.29 / TensorRT 10.15 to CUDA 13.4.1 / PyTorch 2.14.0a0 (26.08) / NCCL 2.30.7 / TensorRT 11.2.1. Source review of v1.3.0rc14...v1.3.0rc27 against benchmarks/single_node/fixed_seq_len/dsr1_fp8_b200_trt.sh: every trtllm-serve flag the recipe passes (--backend, --trust_remote_code, --max_seq_len, --max_num_tokens, --tp_size, --ep_size, --extra_llm_api_options) is still defined in tensorrt_llm/commands/serve.py; cuda_graph_config.{enable_padding,max_batch_size}, enable_attention_dp, print_iter_log, kv_cache_config.{dtype,free_gpu_memory_fraction,enable_block_reuse}, stream_interval and moe_config.backend: TRTLLM remain valid in llm_args.py (CudaGraphConfig is now an alias of DecodeCudaGraphConfig; TRTLLM is still in the MoE backend enum). The recipe's torch_compile_config.enable_piecewise_cuda_graph / capture_num_tokens (used at c64–c256) are deprecated in rc27 but are mapped automatically to prefill_cuda_graph_backend='piecewise' / prefill_capture_num_tokens by normalize_prefill_cuda_graph_config, so no flag change is needed. TLLM_OVERRIDE_LAYER_NUM=61 is still honored (with a deprecation warning) in _torch/pyexecutor/model_loader.py and equals DeepSeek-R1's real layer count. Behavior change to watch: since v1.3.0rc25 kv_cache_config.use_kv_cache_manager_v2 defaults to auto, which resolves to the V2 manager for DeepSeek-R1; rc25/rc26 list B200 memory-estimate known issues for V2 and rc27 ships V2 fixes (#18824, #18885) without listing that issue, so runtime verification is required. No engine patch exists on the selected launch path (runners/launch_b200-nscale-slurm.sh single-node → dsr1_fp8_b200_trt.sh); weights resolve to the node-local /scratch/models/DeepSeek-R1-0528 path used by the baseline. The tag is not digest-pinned in the master because this launcher imports with enroot import docker://$IMAGE, and enroot 3.x does not parse tag@digest (see the note in runners/launch_gb200-nv.sh); the digests above are the pin record. Not runtime-verified: no benchmark was dispatched.

Blocker · session credential identity: the GitHub credential supplied to this candidate session authenticates as a login other than Klaud-Cold, so this PR is authored by that other login and the lifecycle helper rejects it as not owned (report --kind baseline and finish both fail with Candidate ownership mismatch; check-final, check-stop and run discovery apply the same author filter, so dispatched runs could never be verified or validated). The complete 2026-05-18 public baseline (22 throughput points incl. the pre-deprecation 1k/1k points, plus the two published gsm8k evals from producer run 26016874169, head fff53738) was frozen locally and rendered into the PR body with the canonical renderer, but the typed record could not be embedded and no verified completion receipt can exist. Capacity checks for b200-nscale passed at every gate (before edits, before branch creation, before the intended smoke dispatch). No e2e-tests.yml run was dispatched, so there is nothing to cancel. Same failure as #3279 (2026-09-19) and #3311 (2026-09-20); the last Klaud PRs authored by Klaud-Cold are from 2026-09-13.

Next: Maintainer decision: restore the Klaud-Cold identity for the candidate credential, then either adopt this PR (the change is source-verified; full-sweep-fail-fast runs the full family of 8 points and 2 default gsm8k evals under maintainer ownership) or close it. This PR stays draft with no sweep labels.

中文

初始更新 0/5 · 未调度 · 无运行 · 2026-09-20 UTC
nvcr.io#nvidia/tensorrt-llm/release:1.3.0rc27 · 9748705d861acc4a84f350f78f291fdbd485c2c9 · 8k/1k · TP4/EP1, TP8/EP1 · 平均延迟
**变更：**仅将 dsr1-fp8-b200-trt 的主镜像从 NGC 发布容器 1.3.0rc14（镜像内 TRT_LLM_GIT_COMMIT 93cb6518，对应 tag v1.3.0rc14）更新为 1.3.0rc27（TRT_LLM_GIT_COMMIT 6e1cc953，对应 tag v1.3.0rc27，为 GitHub 与 NGC 上最新的发布版本）。对照 v1.3.0rc14…v1.3.0rc27 源码逐项核对了配方使用的所有 trtllm-serve 参数与 extra_llm_api_options 键：均仍有效；torch_compile_config.enable_piecewise_cuda_graph / capture_num_tokens 已被标记为弃用，但会被自动映射到 prefill_cuda_graph_backend='piecewise' / prefill_capture_num_tokens，无需修改配方；TLLM_OVERRIDE_LAYER_NUM=61 仍被读取。需关注的行为变化：自 rc25 起 DeepSeek-R1 默认使用 KV cache manager V2，rc25/rc26 列出 B200 相关内存已知问题，rc27 包含 V2 修复，需运行验证。所选启动路径无引擎补丁；权重路径与基线一致。因该启动器通过 enroot import docker://$IMAGE 导入且 enroot 3.x 不解析 tag@digest，主配置中未按摘要固定，摘要记录见上文。未进行运行时验证：未调度任何基准测试。

**阻塞 · 会话凭据身份：**本会话获得的 GitHub 凭据登录名不是 Klaud-Cold，因此本 PR 的作者为其他登录名，生命周期助手判定其不属于本会话（report --kind baseline 与 finish 均报 Candidate ownership mismatch；check-final、check-stop 与运行发现使用相同的作者过滤，即使调度也无法验证）。完整的 2026-05-18 公开基线（22 个吞吐点，含已弃用的 1k/1k 点，以及生产运行 26016874169 发布的两个 gsm8k 评测）已在本地冻结并用规范渲染器写入 PR 正文，但无法嵌入类型化记录，也不会有经验证的完成回执。b200-nscale 容量检查在各关口均通过。未调度任何 e2e-tests.yml 运行，无需取消。与 #3279（2026-09-19）、#3311（2026-09-20）为同一故障；最后由 Klaud-Cold 创建的 Klaud PR 为 2026-09-13。

**下一步：**由维护者决定：恢复候选凭据的 Klaud-Cold 身份，然后采纳本 PR（变更已经源码核验；添加 full-sweep-fail-fast 可在维护者名下运行 8 个点与 2 个默认 gsm8k 评测的完整配置族）或关闭本 PR。本 PR 保持草稿且不带 sweep 标签。