[SimengLiu-nv]

Dev Engineer Review

This forward-port combines fixes across attention, checkpoint loading, disaggregation, distributed execution, EPLB, telemetry, and documentation. It adds APIs and code paths, changes defaults and dispatch behavior, and updates test infrastructure.

Verify FMHA head-size-104 dispatch and kernel generation across supported architectures. Check that NCCL window-allocation suppression is restored after parallel autotuning, and that shared cubin reads and publication handle concurrent ranks. Verify NIXL descriptor validation and transfer ownership, LoRA adapter identity rules, Qwen3.5 INT4 packing, and Mistral image-processing behavior.

No current review findings or severity counts were supplied. The reported tests/unittest/scripts/test_perf_submit.py run passed 127 tests, but the supplied CI summary also reports a failed /LLM/main/L0_MergeRequest_PR pipeline. GPU tests were not run locally. Overall completion status is not established.

QA Engineer Review

The change set adds and updates tests for FMHA head size 104, Pixtral and sparse attention, MLA page tables, checkpoint mapping, distributed workspace caching, cubin pooling, NIXL transfers, LoRA identity, and EPLB mounts. It also updates CI and QA test lists, moves one Qwen3 case between QA lists, and removes one waiver.

The head-104 C++ test is listed in test-db/l0_b200.yml, test-db/l0_gb202.yml, and test-db/l0_rtx_pro_6000.yml. The Pixtral test is listed in the GB202 and RTX Pro 6000 lists. The cubin-pool test is listed in test-db/l0_cpu.yml. The Qwen3 case moves from qa/llm_function_core.txt to qa/llm_function_multinode.txt. The supplied list changes do not identify entries for the other new tests.

Coverage verdict: needs follow-up. The supplied evidence does not establish that every new test is included in an appropriate CI or manual-QA list. GPU tests were not run locally, and the reported merge-request pipeline failed.

Per-File QA Perspective

Source and documentation files

• cpp/kernels/fmha_v2/setup.py — Verify generated head-size-104 mask kernels across SM90, SM100, and SM120, including dtype, warp-specialization, and tiling conditions.
• cpp/tensorrt_llm/common/ncclUtils.cpp — Verify suppressed allocations return an invalid buffer before communicator and size validation.
• cpp/tensorrt_llm/common/ncclUtils.h — Verify the setter declaration matches its implementation and thread-local contract.
• cpp/tensorrt_llm/executor/cache_transmission/nixl_utils/transferAgent.cpp — Verify descriptor splitting against recorded VMM regions, including behavior when an agent has no recorded map.
• cpp/tensorrt_llm/kernels/fmhaDispatcher.cpp — Verify head-size-104 dispatch selects FMHA v2 only for the specified SM100-family types and masks.
• cpp/tensorrt_llm/runtime/moeLoadBalancer/hostAccessibleDeviceAllocator.cpp — Verify managed-memory support, allocation advice, and fallback behavior across GDRCopy and direct-host-access configurations.
• cpp/tensorrt_llm/runtime/moeLoadBalancer/hostAccessibleDeviceAllocator.h — Verify the new access-mode state and documented preference order match allocator behavior.
• cpp/tensorrt_llm/thop/allreduceOp.cpp — Verify the new operator returns the prior suppression state and behaves as intended without multi-device support.
• tensorrt_llm/_torch/attention/kernels/blackwell/mla/mla_decode_fp8.py — Verify partial page-table tiles do not read beyond the table.
• tensorrt_llm/_torch/autotuner.py — Verify suppression applies only to distributed parallel tuning and is restored after profiling errors.
• tensorrt_llm/_torch/configs/nemotron_h.py — Verify the None cache-dtype default resolves through the intended model-dtype logic.
• tensorrt_llm/_torch/disaggregation/native/transfer.py — Verify WRITE preflight failures prevent submission and retire ownership correctly; verify bounced sends validate original destinations.
• tensorrt_llm/_torch/disaggregation/nixl/_agent_py.py — Verify MemoryDescs conversion and errors match native-agent expectations.
• tensorrt_llm/_torch/distributed/ops.py — Verify process-wide workspace reuse is safe across threads and distinct mappings.
• tensorrt_llm/_torch/flashinfer_cubin_pool.py — Verify pool eligibility, checksum validation, atomic publication, and fallback behavior.
• tensorrt_llm/_torch/flashinfer_utils.py — Verify pool installation runs only when FlashInfer is available and preserves normal loader behavior.
• tensorrt_llm/_torch/models/checkpoints/hf/qwen3_5_weight_mapper.py — Verify packed-INT4 dimensions, scales, biases, and AWQ preprocessing for split and fused layouts.
• tensorrt_llm/_torch/models/checkpoints/hf/qwen4_exp_weight_mapper.py — Verify shared TP-rank concatenation preserves projection ordering.
• tensorrt_llm/_torch/models/modeling_minimaxm3.py — Verify reshaped GQA SDPA output preserves batch, KV-head, and group ordering.
• tensorrt_llm/_torch/models/modeling_mistral.py — Verify image geometry and scheduling metadata for both processors, including text-only checkpoint behavior.
• tensorrt_llm/_torch/models/modeling_utils.py — Verify TP-rank concatenation preserves rank-major ordering and rejects invalid shapes.
• tensorrt_llm/_torch/modules/linear.py — Verify optional AWQ pre-quant scales are loaded, sharded, and cleared correctly.
• tensorrt_llm/_torch/nccl_window_graph.py — Verify the context manager restores the previous suppression state, including after exceptions.
• tensorrt_llm/bench/tuning/heuristics.py — Verify missing and auto cache dtypes use the model dtype when sizing Mamba cache memory.
• tensorrt_llm/executor/base_worker.py — Verify LoRA ID bindings reject conflicting identities while allowing aliases and retries after unsuccessful lookups.
• tensorrt_llm/llmapi/mpi_session.py — Verify check_worker_error() reports a dead manager thread but not a live or stopped pool.
• tensorrt_llm/serve/disagg_coordinator.py — Verify HTTP cluster-storage scope validation rejects unsafe host and scope combinations.
• tensorrt_llm/usage/usage_lib.py — Verify deployment IDs are normalized as UUIDs and invalid values are omitted from telemetry.
• jenkins/scripts/perf/benchmark_utils.py — Verify EPLB detection handles inline and file-based configs and adds the shared-memory mount only when needed.
• jenkins/scripts/perf/local/submit.py — Verify local submission passes the loaded config into mount generation without changing other launch arguments.
• jenkins/scripts/perf/submit.py — Verify CI submission changes container mounts only and preserves other srun arguments.
• docs/source/blogs/tech_blog/blog13_Inference_Time_Compute_Implementation_in_TensorRT-LLM.md — Documentation-only model-name capitalization change.
• docs/source/developer-guide/sparse-attention-development-guide.md — Verify the documented summary contract matches kernel weighting for partial KV blocks.
• docs/source/features/sparse-attention.md — Verify the public sparse-attention contract states means, valid-token counts, and no pre-weighting.
• docs/source/llm-api/index.md — Verify shared-cubin documentation matches implementation and environment-variable behavior.
• docs/source/overview.md — Documentation-only capitalization correction.
• docs/source/release-notes.md — Documentation-only typo correction.
• examples/wide_ep/README.md — Verify the managed-memory fallback is limited to the stated coherent platforms.
• tensorrt_llm/_torch/attention/ATTENTION_DEVELOPER_GUIDE.md — Verify developer guidance matches sparse-kernel proxy weighting.
• tensorrt_llm/usage/schemas/README.md — Verify UUID normalization guidance matches telemetry behavior.

Test-code files

• cpp/tests/unit_tests/executor/transferAgentTest.cpp — Covers NIXL descriptor subranges, unknown agents, writes, and adjacent VMM chunks; no test-list entry is identified in the supplied summaries.
• cpp/tests/unit_tests/kernels/fmhaHead104Test.cpp — Covers packed-QKV head-size-104 output across masks, dtypes, and ragged lengths; its integration test is listed in GPU test lists.
• cpp/tests/unit_tests/runtime/hostAccessibleDeviceAllocatorTest.cu — Covers direct host access to managed memory without GDRCopy; no test-list entry is identified in the supplied summaries.
• tests/integration/defs/cpp/conftest.py — Adds the build fixture for the head-104 C++ test; the integration test is listed in GPU test lists.
• tests/integration/defs/cpp/test_unit_tests.py — Adds the head-104 integration test for targets 100 and 120; corresponding entries appear in the supplied GPU test-list changes.
• tests/unittest/_torch/attention/multi_gpu/test_mha_helix.py — Covers attention output lifetime inside an NCCL window tensor scope; no test-list entry is identified.
• tests/unittest/_torch/attention/multi_gpu/test_mla_helix.py — Covers MLA output lifetime inside an NCCL window tensor scope; no test-list entry is identified.
• tests/unittest/_torch/attention/sparse/test_prims_ts_block_sparse.py — Covers partial-block proxy weighting and CPU/GPU behavior; no test-list entry is identified.
• tests/unittest/_torch/attention/test_cute_dsl_mla_page_table.py — Covers unpadded page-table widths and partial final pages on SM100/SM103; no test-list entry is identified.
• tests/unittest/_torch/attention/test_fmha_v2_head_size_coverage.py — Covers generated head-size-104 kernel inclusion and exclusions; no test-list entry is identified.
• tests/unittest/_torch/attention/test_pixtral_head104_attention.py — Covers Pixtral head-size-104 attention on supported architectures; listed for GB202 and RTX Pro 6000.
• tests/unittest/_torch/attention/test_prims_ts_attention_backend.py — Covers lazy rescaling around the skip threshold, including CUDA graph replay; no test-list entry is identified.
• tests/unittest/_torch/attention/test_prims_ts_fmha.py — Updates wrapper mocks to use autospec across FMHA paths; no test-list entry is identified.
• tests/unittest/_torch/distributed/test_allreduce_workspace_cache.py — Covers cross-thread reuse, concurrent initialization, and distinct mappings; no test-list entry is identified.
• tests/unittest/_torch/flashinfer/test_flashinfer_cubin_pool.py — Covers pool eligibility, reuse, corruption, failed downloads, concurrency, and installation; listed in test-db/l0_cpu.yml.
• tests/unittest/_torch/modeling/test_modeling_mistral.py — Covers image geometry and scheduling for both Mistral processors and text-only checkpoints; no test-list entry is identified.
• tests/unittest/_torch/modeling/test_qwen4_exp_support.py — Tests the shared TP-rank concatenation helper; no test-list entry is identified.
• tests/unittest/_torch/models/checkpoints/hf/test_qwen3_5_weight_mapper.py — Covers INT4 mapping, layouts, TP sizes, scales, and invalid inputs; no test-list entry is identified.
• tests/unittest/_torch/models/test_minimax_m3.py — Covers dense SDPA decode against explicit GQA repetition with scattered cache pages; no test-list entry is identified.
• tests/unittest/_torch/models/test_qwen3_5_int4_loading.py — Covers INT4 mapping into GDN projections across TP ranks and optional scales; listed in test-db/l0_b200.yml.
• tests/unittest/_torch/modules/test_linear_uneven_tp.py — Covers optional AWQ pre-quant scales in COLUMN and ROW modes; listed in test-db/l0_b200.yml.
• tests/unittest/_torch/multi_gpu/test_allocate_output_buffer_kinds.py — Covers NCCL window allocation after distributed parallel tuning; no test-list entry is identified.
• tests/unittest/_torch/test_custom_config_registration.py — Covers Nemotron-H cache dtype resolution for absent and explicit values; no test-list entry is identified.
• tests/unittest/disaggregated/test_agent.py — Covers Python NIXL descriptor conversion, metadata results, and errors; no test-list entry is identified.
• tests/unittest/disaggregated/test_coordinator_worker.py — Covers loopback-only HTTP cluster-storage validation; no test-list entry is identified.
• tests/unittest/disaggregated/test_kv_transfer_fail_stop.py — Updates WRITE request fixtures for descriptor validation; no test-list entry is identified.
• tests/unittest/disaggregated/test_retirement_deadline.py — Updates NIXL WRITE fixtures across retirement and deadline cases; no test-list entry is identified.
• tests/unittest/disaggregated/test_transfer_late_settlement.py — Updates sender fixtures for descriptor checks; no test-list entry is identified.
• tests/unittest/disaggregated/test_transfer_ownership_regressions.py — Covers preflight rejection, submission failure, retirement state, and sender ownership; no test-list entry is identified.
• tests/unittest/executor/test_base_worker.py — Covers LoRA identity conflicts, aliases, retry behavior, and concurrency; no test-list entry is identified.
• tests/unittest/executor/test_proxy_fast_death.py — Covers dead MPI manager-thread detection; no test-list entry is identified.
• tests/unittest/llmapi/test_mpi_launcher_shutdown.py — Covers launcher startup without site initialization; no test-list entry is identified.
• tests/unittest/others/test_bench.py — Covers Nemotron cache sizing for unset, auto, and float32 dtypes; no test-list entry is identified.
• tests/unittest/scripts/test_perf_submit.py — Covers EPLB detection and CI/local mount generation; the PR objectives report 127 tests passed.
• tests/unittest/usage/test_reporter.py — Covers UUID normalization and invalid deployment IDs across telemetry events; no test-list entry is identified.
• tests/test_common/session_prefetcher.py — Changes diagnostic output handling; verify messages remain available through file descriptor 1.
• tests/unittest/_torch/peft/test_lora_ownership.py — Updates runtime setup with LoRA identity and locking fields; no test-list entry is identified.

Test-list and test configuration files

• tests/integration/test_lists/qa/llm_function_core.txt — Removes the Qwen3 NVFP4 test entry; the case moves to the multinode QA list.
• tests/integration/test_lists/qa/llm_function_multinode.txt — Adds the Qwen3 NVFP4 test entry.
• tests/integration/test_lists/test-db/l0_b200.yml — Adds the head-104 C++ test and Qwen3.5 INT4/AWQ tests to pre-merge coverage.
• tests/integration/test_lists/test-db/l0_cpu.yml — Adds the FlashInfer cubin-pool unit test.
• tests/integration/test_lists/test-db/l0_gb202.yml — Adds head-104 and Pixtral attention tests.
• tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml — Adds head-104 and Pixtral attention tests.
• tests/integration/test_lists/waives.txt — Removes the waiver for TestMistralLarge3_675B::test_nvfp4_4gpus[latency_moe_trtllm]; verify the test is suitable to run without the waiver.
• tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con1_ctx1_dep2_gen1_tep4_mtp3_ccb-NIXL.yaml — Adds NUMA-aware worker affinity for the context worker; verify the launcher accepts the setting.
• tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con4096_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml — Adds NUMA-aware worker affinity for the context worker; verify the launcher accepts the setting.
• tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_1k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml — Adds NUMA-aware worker affinity for the context worker; verify the launcher accepts the setting.
• tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1024_ctx1_dep2_gen1_dep8_eplb256_mtp1_ccb-NIXL.yaml — Adds NUMA-aware worker affinity for the context worker; verify the launcher accepts the setting.
• tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con1_ctx1_dep2_gen1_tep8_eplb0_mtp3_ccb-NIXL.yaml — Adds NUMA-aware worker affinity for the context worker; verify the launcher accepts the setting.
• tests/scripts/perf-sanity/disaggregated/gb300_glm-5-fp4_8k1k_con512_ctx1_dep2_gen1_dep32_eplb0_mtp3_ccb-NIXL.yaml — Adds NUMA-aware worker affinity for the context worker; verify the launcher accepts the setting.

Description

Second mass merge of release/1.3 fixes that are still missing from main, following #20232. It forward-ports the 23 release/1.3 PRs that, as of 2026-10-22, have no counterpart on main and no open PR to main. Each original PR is one commit, and each commit message links back to the original release PR and commit and keeps the original author and sign-off. Commits that needed changes for main carry an "Adapted to main" note in their message. The branch was rebased onto main 79d1d922dc on 2026-10-26.

Release PR	Ticket	Author	Change	Port	Commit
#19788	6665906	Pranav Shrestha	Restore Mistral Large 3 multimodal inference	Adapted: uses main's MultimodalItemMetadata contract (#20391)	1c9f751655
#19815	6562942	Pranav Shrestha	Load Qwen3.5 packed INT4 checkpoints	Clean cherry-pick	da07e21397
#20029	6877539	Chenfei Zhang (via TRT-LLM agent)	Fold GQA group into SDPA query axis in MiniMax-M3 decode attention	Adapted: also updates a stale main comment	f8117136b1
#20132	—	Ivy Zhang	Move cases to multinode test list	Clean cherry-pick	682c90e6a8
#20136	6463987	Chenfei Zhang	Pin GB300 GLM-5 ctx workers to their GPU's NUMA node in perf-sanity	Clean cherry-pick	eaeba343db
#19847	6859934	Mike Iovine (via TRT-LLM agent)	Process-wide workspace caches keyed by Mapping with a per-mapping lock…	Clean cherry-pick	ca66314afa
#20152	6792615	Sunny Gai	Keep session-prefetch output off sys.stdout	Clean cherry-pick	e6ec4c51db
#20193	6918015	Stanley Sun	Fix spelling and capitalization in release/1.3 docs	Clean cherry-pick	455c15f3d4
#20156	6379406	Sunny Gai	Use coherent managed memory for EPLB when GDRCopy is unavailable	Clean cherry-pick	50dbedc167
#20195	6911393	Mike Iovine (via TRT-LLM agent)	Run the launcher test's python3 stand-in with -S so start-up hooks cannot…	Adapted: release-only waives dropped	5c078b528f
#20264	6825983	Chenfei Zhang	Share TRTLLM-Gen FMHA cubins across isolated MPI ranks	Clean cherry-pick	cbecf90da7
#20248	6912133	Asfiya Baig	Detect dead MPI pool manager thread in MpiPoolSession	Clean cherry-pick	b7cd5ebe60
#20415	6862662	Yuxian Qiu	Bound FP8 MLA page-table loads for partial tiles	Adapted: kernel moved by #19275; test uses FP8 device-scale inputs (#19960)	ab40fa8e28
#20381	—	Yuxian Qiu	Follow up PrimTS refresh backport review tests and docs	Adapted: fits main's metadata-owned plan cache; main's CPU-only Sage tests keep their marker	ca51b44a79
#20588	6932241	Yuxian Qiu	Scope NCCL window leases in helix attention tests	Adapted: release-only waives dropped	e16a1568b2
#20344	6553508	Yibin Li	Reject conflicting LoRA adapter identities	Adapted: wraps main's _load_lora_adapter() -> None (#20043); updates main's test_lora_ownership.py fixture (#20605)	57b691f506
#20346	6857649	Yibin Li	Validate native transfer destinations before submission	Adapted: main-only disagg tests pass a valid WRITE request	a9b9c772a4
#20531	6384589	XingFei Xi	Mount a 640G /dev/shm for online-EPLB perf cases	Adapted: tests reuse main's submit fixtures	1190531881
#20657	6944374	Lori Ren	Fix rank-asymmetric NCCL window allocation hang during PARALLEL autotuning	Clean cherry-pick	5020a21ea5
#20284	6885370	Shanshan Wang	Keep Nemotron-H SSM cache dtype unset	Clean cherry-pick	11f6764a42
#20651	6887489	Yibin Li	Restrict coordinator HTTP cluster storage to loopback	Clean cherry-pick	dc6104d10d
#20794	6964933	Ruodi Lu	Use FlashInfer for Gemma4 NVFP4 perf on SM107	Adapted: updates main's SM121 comment	ab6d70ed28
#20777	—	Maxim Gluhovskoi	Validate telemetry deployment IDs as UUIDs	Adapted: test patches _HEARTBEAT_STOP (renamed on main)	034f5ddf5d

Notes:

• [https://nvbugs/6463987][fix] Cherry-pick #20103 to release/1.3: pin GB300 GLM-5 ctx workers to their GPU's NUMA node in perf-sanity #20136 and [https://nvbugs/6825983][perf] Cherry-pick #20257 to release/1.3: Share TRTLLM-Gen FMHA cubins across isolated MPI ranks #20264 were release cherry-picks of main PRs [https://nvbugs/6463987][fix] Pin GB300 GLM-5 ctx workers to their GPU's NUMA node in perf-sanity #20103 and [https://nvbugs/6825983][perf] Share TRTLLM-Gen FMHA cubins across isolated MPI ranks #20257, which were closed without merging, so the fixes are ported from the release PRs.
• [https://nvbugs/6877539][fix] Fold GQA group into SDPA query axis in MiniMax-M3 decode attention #20029, [https://nvbugs/6665906][fix] Restore Mistral Large 3 multimodal inference #19788 and [https://nvbugs/6562942][fix] Load Qwen3.5 packed INT4 checkpoints #19815 had earlier forward-port PRs ([https://nvbugs/6877539][fix] Fold GQA group into SDPA query axis in MiniMax-M3 decode attention #20115, [https://nvbugs/6665906][fix] Cherry-pick #19788: Restore Mistral Large 3 multimodal inference #20172, [https://nvbugs/6562942][fix] Cherry-pick #19815: Load Qwen3.5 packed INT4 checkpoints #20171) that were closed without merging.
• The main waive for nvbugs/6862662 (test_attention_mla) is kept; it can be removed once the [https://nvbugs/6862662][fix] Bound FP8 MLA page-table loads for partial tiles #20415 fix is confirmed on B200.
• Not included: release PRs that are only partly on main ([https://nvbugs/6783973][doc] Cherry-pick #19784 to release/1.3: stop listing supported SA speculation as a Kimi K3 limitation #19863, [https://nvbugs/6801229][fix] Fix NVFP4 gather test references and skip empty FP4 quantize launches #19761, [https://nvbugs/6818710][fix] Add helix_owned_new_tokens to the base update_helix_param hook #20018, [https://nvbugs/6892484][fix] Port of the 6791675 fix to release/1.3: RemoteTask gains sys_path, the… #20058, [TRTLLMINF-474][infra] Upgrade NGC PyTorch base image to 26.09 (#19679) #20146, [https://nvbugs/6784181][fix] Fix test_relative_path_validity: generated-doc allowlist and skip installed envs (cherry-pick #19405, #19403) #20601, [None][fix] Cherry-pick #20203, #20581, #20575 to release/1.3 #20689) and those already covered by open PRs to main.

Test Coverage

• Each commit carries its original PR's tests, including the new test_qwen3_5_int4_loading.py, test_fmha_v2_head_size_coverage.py, test_pixtral_head104_attention.py, fmhaHead104Test.cpp, test_allreduce_workspace_cache.py, test_flashinfer_cubin_pool.py, test_cute_dsl_mla_page_table.py, the LoRA identity tests in test_base_worker.py, the EPLB /dev/shm tests in test_perf_submit.py, and the coordinator, telemetry and MPI pool tests.
• Tests adapted for main: test_cute_dsl_mla_page_table.py (FP8 scale inputs), test_base_worker.py (no return value), test_retirement_deadline.py, test_kv_transfer_fail_stop.py and test_transfer_late_settlement.py (WRITE requests for the new pre-submission check), test_prims_ts_block_sparse.py (cpu_only markers, including on main's Sage tests), test_modeling_mistral.py (main's get_mm_item_metadata API and tuple fields), and main's test_lora_ownership.py (adapter identity state in its worker fixture).
• Locally: all touched Python files compile with no new ruff F/E9 findings versus main, and the repo formatters pass. Before the rebase, tests/unittest/scripts/test_perf_submit.py passed (127 tests). GPU tests were not run locally and rely on CI.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

🤖 Generated with Claude Code

[SimengLiu-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #80694 [ run ] triggered by Bot. Commit: 034f5dd Link to invocation

[tensorrt-cicd]

PR_Github #80694 [ run ] completed with state SUCCESS. Commit: 034f5dd
/LLM/main/L0_MergeRequest_PR pipeline #66443 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[sunnyqgg]

LGTM