NVIDIA/TensorRT-LLM v1.3.0rc25

on GitHub

Highlights

• Annoucements

  • KV Cache Manager
    • Enabled KV cache manager V2 by default for:
      • DeepSeek V3, DeepSeek R1, and DeepSeek V3 Lite
      • DeepSeek V3.2
      • DeepSeek V4
      • GLM-5
      • GPT-OSS
      • Mistral Large 3
      • Kimi K2, Kimi K2.5, and Kimi K3
      • MiniMax M2 and MiniMax M3
      • Nemotron H and Nemotron H Puzzle
      • Qwen3-Next, Qwen3.5, and Qwen3.8
      • Gemma 3 and Gemma 4
    • KV cache manager V2 is the recommended architecture, with improved scalability and stability. All new models will use V2 by default, existing models will be migrated to V2 gradually, and V1 will be deprecated.

• Known Issues

  • Disaggregated serving may hang during startup or shutdown and leave workers running; affected workloads include GPT-OSS startup and tensor-parallel Llama cleanup.
  • Disaggregated serving with optional execution modes may fail to start.
  • Generation may hang indefinitely across multiple models, including Gemma 4.
  • Repeated create, generate, and shutdown cycles with overlap scheduling may hang across sampler and block-reuse configurations.
  • KV cache manager V2 scheduler workloads may intermittently crash or run out of memory on B200 GPUs.
  • DeepSeek-V3-Lite disaggregated attention data parallelism with overlap may crash; configurations additionally using MTP, CUDA Graphs, and KV cache manager V2 may hang.
  • DeepSeek-R1 V2 FP4 MTP disaggregated serving may crash during generation, and two-node executor initialization may hang indefinitely.
  • GPT-OSS EAGLE serving may deadlock during distributed startup or suffer a worker crash that causes subsequent requests to fail.
  • RocketKV generation may hang with either the TensorRT-LLM or vanilla model path.
  • Disaggregated KV cache transfers may hang after a sender aborts; LoRA generation with chunked prefill may livelock.
  • Qwen3 FP8 LoRA generation may hang.
  • Distributed workload data-parallel inference may crash or hang during shutdown after successful generation.
  • Multi-GPU fused-MoE startup may lose a rank and then hang.
  • Reusing LLM workers for MegaMoE workloads such as DeepSeek V4 Pro may exhaust GPU memory.
  • GPT-OSS dummy-load and W4 configurations may run out of memory while creating the KV cache.
  • DeepSeek-V3.2 FP8 executor creation may run out of memory on H200 GPUs.
  • BF16 LTX-2 and multi-GPU Wan 2.2 visual-generation workloads may produce degraded output quality.
  • Wan text-image-to-video generation with fused normalization may lose conditioning behavior when timestep shapes change.
  • DeepSeek-V3-Lite Helix context-parallel disaggregated serving may produce accuracy loss.
  • Nano V3 Omni FP8 encoder execution with CUDA Graphs may produce accuracy loss.
  • GLM 4.5 NVFP4 disaggregated NIXL serving with KV cache manager V1 may complete with zero accuracy.
  • Qwen2 and Exaone MoE GQA/SWA inference may intermittently produce NaNs and accuracy loss.
  • Qwen3 skip-softmax configurations may produce NaNs and crash with a CUDA device assertion during sampling.
  • Speculative decoding with chunked prefill may exhibit incorrect cache behavior and reduced token acceptance.
  • Stock EXAONE 4.5 and VILA checkpoints may fail to load because their configurations are not registered.
  • Nemotron-Hybrid checkpoints whose hybrid pattern contains a hyphen may fail during configuration loading.
  • DeepSeek-R1 long-context requests with V2 draft KV caching are unsupported when the required sequence length exceeds the initial cache size.
  • Mistral Large 3 FP8 DeepGEMM and NVFP4 EAGLE MoE workloads may crash during startup or generation.
  • LTX-2 NVFP4 generation may fail during startup.
  • Wan pipeline-parallel or Ulysses execution may fail during startup on B200 GPUs.
  • MiniMax-M3 MXFP8 model construction is unsupported.
  • Kimi K3 generation-only workloads with 96-head MLA configurations are unsupported.
  • AutoDeploy standalone generated installations may fail to run when dependency resolution selects incompatible versions.

• Model Support

  • Add Cosmos3-Edge support for text-to-image, text-to-video, and image-to-video generation #16773.
  • Add HunyuanVideo 1.5 text-to-video support #15562.
  • Add GLM-Image text-to-image support #16018.
  • Add Cosmos3 Transfer control-video conditioning #16394.
  • Add multimodal Kimi K3 support #17050.
  • Enable mixed-modality support for Nemotron Nano Omni V3 #16764.
  • Expand Kimi K3 with SiTU MegaMoE and NVFP4 support #17624, #17865, and #17940.
  • Expand Kimi K3 with selectable MLA decode backends, KDA tensor parallelism, MLA decode context parallelism, and SA speculative decoding in disaggregated serving #17800, #17796, and #17939.
  • Expand Qwen3.5 and Qwen3.8 with MoE, attention data parallelism, GDN replay, and improved weight loading #17700.
  • Enable DeepSeek-V4 and DSA models, including DeepSeek-V3.2 and GLM, on SM120 through FlashInfer sparse MLA #16224.
  • Qualify dense Qwen2 models for ModelExpress #16974.
  • Add configuration-driven sqrt-softplus router scoring for Laguna models #17918.

• API

  • Load public TensorRT-LLM exports lazily while preserving top-level names #17281.
  • Deprecate the legacy _autodeploy backend #15642.
  • Remove model_path from TriAttentionKvCacheCompressionConfig and derive the model configuration from the executor #17501.
  • Complete TorchSampler beam-search support and add occurrence penalties #16620 and #17189.
  • Retire two-model speculative-decoding execution in favor of one-model fallback #17366 and #17562.
  • Add prefill CUDA Graph backend and capture-bucket APIs, including experimental breakable CUDA Graphs, while deprecating corresponding TorchCompileConfig fields #16609.
  • Honor SamplingParams.seed and occurrence penalties during one-model speculative decoding #17599 and #17701.
  • Remove conversation IDs from disaggregated request parameters #17618.
  • Add race-free dynamic port allocation through --port 0 and address reporting through --report_addr #17460.
  • Add the /v1/images/edits endpoint for visual-generation models #16987.
  • Expose model-startup and weight-loading metrics through the LLM API and serving stack #15993.
  • Add RuntimeLoRAConfig and startup-time LoRA fusion for VisualGen #17435.
  • Add KV cache manager V2 cold-page codecs while changing pool_ratio ordering and cold-tier statistics #17512.
  • Remove all WIDEEP files #17797.
  • Remove Star Attention and its configuration surface #18025.
  • Revise the VisualGen serving API with path responses, new job states, and /v1/videos/sync #17490.
  • Honor named functions in tool_choice for non-Harmony models #14290.

• Feature

  • Enable KV cache manager V2 by default for GPT-OSS, Gemma 3, and Gemma 4 #16942 and #17396.
  • Extend KV cache manager V2 with DSA support, zero-copy token passing, distributed pool rebalancing, and Helix decode context parallelism #16060, #17308, #17391, and #17866.
  • Add prefix-attention and request-level metrics to KV cache manager V2 #17448 and #17815.
  • Expand disaggregated serving with Python V2 transceivers, Ray-based NIXL transfers, masked DSA cache transfers, decode context-parallel transfers, and enhanced Mamba page tables #16787, #17295, #17283, #17794, and #17691.
  • Add end-to-end dense FP8 LoRA, Blackwell FP8 support, base-model computation overlap, and specialized CUDA Graphs #16810, #17521, #16951, and #17412.
  • Batch TorchSampler beam-search finish-reason reduction #17494.
  • Eliminate redundant one-model sampling-parameter refills #17544.
  • Extend and accelerate DFlash and DSpark with RoPE, sliding-window attention, TensorRT-LLM Gen attention, fused attention/RMSNorm/RoPE, and optimized draft forwarding #16914, #17307, and #17935.
  • Support structured output with DSpark speculative drafting #16965.
  • Support loading MTP weights from standalone checkpoints #17378.
  • Hard-kill all ranks when any executor loop crashes #16592.
  • Enforce multimodal encoder runtime budgets with budgeted output storage #16051.
  • Support isolated out-of-tree builds and incremental artifact copy-back #17524, #17525, and #17538.
  • Optimize Kimi K3 by fusing KDA projections and routing quantization, preallocating residual snapshots, removing MLA padding, and eliminating recurrent-state copies #17311, #17053, #17059, #17684, and #17887.
  • Accelerate VisualGen with fused Wan upsampling, optimized parallel-VAE halo exchange, fused MLP and AdaLN operations, and native Wan VAE reuse for Cosmos3 #17003, #17001, #17693, #17695, and #17555.
  • Enable FlashAttention 4 and parallel VAE in LTX-2 examples #17610.
  • Accelerate DeepSeek sparse attention with a fused MLA prologue and overlapped DSA heuristic write-back #17273 and #16666.
  • Optimize MiniMax-M3 sparse attention with paged KV access, fewer transfers and synchronizations, FP8 indexer QK, improved block selection, decode kernels, and reduced idle time #16857, #17840, #17841, #17236, #17318, #17986, and #17842.
  • Enable INT8 weight-only MoE for non-gated activations #15550.
  • Optimize MoE execution with fused SM90 FP8 block-scale quantization and split routing kernels #16849 and #17893.
  • Add a CFT counted-write path for MoE all-to-all communication #17647.
  • Add tiered and self-sampling CuTe DSL GVR top-K decode kernels #16877 and #17821.
  • Move AllReduce autotuner dispatch into C++ #16902.
  • Eliminate quadratic copying during KV cache block reuse #17848.
  • Add low-M BF16 GEMM dispatch for SM10x decoding #17889.
  • Improve small-BF16-context performance with targeted fallback selection #18196.
  • Reduce runtime overhead from request-context construction, response timing, hang detection, and profiling bookkeeping #17582, #18053, and #17531.
  • Cache Whisper suppress-token indices on the GPU #17583.

• Fix

  • Enforce Responses conversation-history capacity limits #15043.
  • Resolve reasoning mode from rendered prompts #17305.
  • Preserve streaming SSE event boundaries #18090.
  • Align log probabilities with token IDs after stop-word trimming #18190.
  • Correct DeepSeek streaming tool parsers to emit withheld text, preserve pre-tool-call text, and complete zero-argument calls #17573, #17903, and #17575.
  • Harden Kimi K3 tool-call parsing #17980.
  • Surface malformed Harmony tool-call messages instead of silently discarding them #17464.
  • Delegate guided decoding through multimodal model wrappers #18162.
  • Skip Hugging Face tokenizer loading when no tokenizer directory is configured #17568.
  • Serialize Hugging Face remote-code loading while constructing input processors #17708.
  • Stabilize tokenizer representations by removing object-address-dependent output #16839.
  • Correct Kimi K3 MLA output propagation and KDA bias initialization #17445 and #17793.
  • Stabilize Kimi K3 CUDA Graph buffers and fused-verification constants #17421.
  • Make Kimi K3 FP8 checkpoint reads opt-in #17446.
  • Correct Kimi K3 KV cache manager V2 convolution-state layout and SSM statistics #17447.
  • Release Kimi K3 checkpoint mappings after loading and bound MegaMoE expert-weight memory at EP8 #17999 and #18059.
  • Select the compatible Marlin backend for Qwen3.5 MoE on Hopper #17293.
  • Correct Qwen3.5 FP8 row-wise checkpoint mapping #17433.
  • Normalize Qwen3.8 27B FP8 VLM quantization configuration #17786.
  • Select FlashInfer FA2 for Gemma4 on SM120 and SM121 #17557.
  • Stabilize Gemma4 FA2 CUDA Graph decoding on Hopper #18002.
  • Restore Gemma4 shared-KV draft loading and FlashInfer shared-KV speculative decoding #17837 and #17264.
  • Restore DSpark disaggregated decoding accuracy #17619.
  • Correct DeepSeek-V4 one-model MTP draft KV cache handling #16887.
  • Correct MiniMax-M3 routing inputs and end-to-end execution #17565.
  • Fix SageAttention for KV sequences not divisible by 128 #17648.
  • Stabilize CUDA Graph pool and handle lifetimes across captures #16952 and #17313.
  • Release eager outputs before CUDA Graph capture #17824.
  • Correct attention runtime-workspace and unfused-context workspace sizing #16432 and #17026.
  • Size decode buffers and KV cache admission by beam width #17014 and #16802.
  • Prevent arithmetic overflow in the FMHAv2 launcher #11263.
  • Densify sparse-FMHA warmup coverage to prevent missing kernel configurations #17961.
  • Skip no-op MXFP4 weight padding #17259.
  • Prevent LoRA host/device cache dtype reconfiguration races #17678.
  • Fence MoE shared writes before asynchronous bulk copies #17621.
  • Correct MoE and MTP execution for large hybrid-attention FP8 serving #17635.
  • Pad TensorRT-LLM Gen MoE route maps to cover kernel over-reads #17951.
  • Align Blackwell TMA-OOB MoE workspace buffers to 128 KiB #18133.
  • Constrain SM120 FP8 block-scale MoE autotuning correctly #18140.
  • Gate TritonFusedMoE by the SwiGLU activation family #18256.
  • Repair non-converged GVR top-K threshold searches in CUDA and CuTe DSL implementations #17550 and #18094.
  • Correct CuTe DSL MLA decode AutoTuner fallback buckets #17792.
  • Restore convolution-state ordering barriers in Triton convolution kernels #17539.
  • Reduce DeepSeek-V4 EPLB loading memory #18185.
  • Fit token budgets safely at request-preparation boundaries #15187.
  • Pad empty attention data-parallel batches so all ranks continue making progress #17379.
  • Generalize attention data-parallel padding-dummy lifecycle management #16921.
  • Keep attention data-parallel ranks in collective lockstep on request failures #16687.
  • Isolate single-rank disaggregated failures and peer-layout mismatches to affected requests #16834 and #17482.
  • Prevent gen-only asynchronous KV-transfer hangs #17107.
  • Reap idle disaggregated context transfers without invalid progress guards #17564, #17324, and #18263.
  • Recover disaggregated worker heartbeat registration #17461.
  • Preserve KV ownership during disaggregated transfer prechecks #17223.
  • Reject unsupported transceiver routes for Kimi K3 disaggregated serving #17480.
  • Detect worker death during initialization and abort wedged worker worlds #17578 and #17817.
  • Isolate FlashInfer JIT workspaces across MPI workers #17233.
  • Compensate Helix position IDs under overlap scheduling #17811.
  • Restore the Helix KV cache CpType import #18143.
  • Clamp conversation-affinity routing to each attention data-parallel rank's slot #17534.
  • Suspend CUDA Graph padding dummies during KV-pool rebalance #16157.
  • Skip draft-KV mirroring when KV cache manager V2 index mapping is saturated #17347.
  • Synchronize KV cache manager V2 host-tier quotas and fallback decisions across ranks #17717 and #18092.
  • Prevent KV-cache estimation probes from pinning the complete host tier #17850.
  • Restore KV cache manager V2 recompute handling for paused requests #15252.
  • Reattach radix-tree blocks detached under live requests #18095.
  • Correct C++ Mamba hybrid-cache management and propagate maximum-token limits #17492 and #18116.
  • Synchronize one-model MTP scheduler draft tokens #17836.
  • Preserve advanced and single-step-greedy sampling state through graph capture #17671 and #17828.
  • Refresh group-greedy speculative state on every iteration #17710.
  • Preserve dynamic-tree speculative-decoding state across forwards #16733.
  • Reject sampling parameters unsupported by one-model speculative decoding #17955.
  • Avoid MoE multi-GPU rendezvous port races #17831.
  • Swap NVFP4 batched-GEMV grid axes to support large output dimensions #17660.
  • Correct CuTe DSL packed-QKV copy lowering in VisualGen #17425.
  • Stabilize VisualGen multi-node Ulysses synchronization and remove conflicting NCCL flags #16486 and #16751.
  • Prevent Cosmos3 CUDA-event crashes when guardrails reject prompts #17510.
  • Reword Cosmos3 negative prompts that incorrectly trigger guardrails #17523.
  • Correct Cache-DiT statistics logging and Wan per-patch timestep handling #17697 and #17696.
  • Warn when VisualGen attention backends fall back silently #17698.
  • Reject static VisualGen quantization recipes for unquantized checkpoints #17699.
  • Preserve multimodal graph state while toggling torch.compile #17949.
  • Distinguish NVLink encryption from confidential-computing mode #16850.
  • Prevent streaming executor shutdown deadlocks #17426.
  • Clean up leftover disaggregated servers, reject dead prefetched MPI pools, and reap wedged session workers #17632, #17642, and #17644.
  • Prevent out-of-tree wheel copy-back from skipping freshly written files #17919.

• Documentation

  • Add the Qwen3.8 deployment guide and configurations #17594.
  • Publish guidance for evaluating agentic serving with trace replay and job-level metrics #16230.
  • Update disaggregated-serving guides for transceiver V2 #18248.
  • Clarify Kimi K3 container runtime dependencies and UCX configuration #17456 and #17479.
  • Document Helix source-test prerequisites #17628.
  • Correct LMCache example output guidance #17629.
  • Correct the attention data-parallel Balance per-GPU throughput figure #17905.

• Test & Infra

  • Migrate CI container images to Artifactory and harden Docker-in-Docker networking, wheel context, and image retagging #16970, #17829, and #18208.
  • Move AutoDeploy and other CPU-only tests plus documentation builds onto CPU stages and reduce their resource requirements #17314, #17339, #17639, and #18125.
  • Run mypy directly in the build stage #17560.
  • Improve test diagnostics with native pytest S3 capture, periodic result uploads, rank-zero log draining, hang tracebacks, and inner-test attribution #16739, #14138, #18240, #16374, and #16358.
  • Introduce coverage-based test selection and refine its VisualGen rules, pilot allowlist, compact coverage data, architecture aggregation, and skip-rate reporting #16776, #17251, #17253, #17996, #17803, #18058, and #16467.
  • Harden CI failure handling by deferring infrastructure-scoped K8s and Slurm aborts, avoiding retries for deterministic failures, and logging retry-classification declines #17615, #17938, #17645, and #17432.
  • Strengthen test-list parameter validation, collection parity, live-stage mapping, sysinfo detection, and empty-render checks #17975, #17651, #18004, and #17993.
  • Automate stale pull-request cleanup and scan all result pages #17383 and #17602.
  • Add a simplified infrastructure dry-run pipeline #17813.
  • Integrate PLC source scanning and improve PLC and nSpect reporting, authentication, and entitlement handling #15552, #17419, #17679, #18119, and #18217.
  • Harden Slurm test execution with a relocated finalizer, reliable artifact installation, cleanup and submission error classification, and multi-rank launch fencing #17263, #17634, #17704, #17709, and #17372.
  • Add execution, test-runner, modeling-bringup, and performance-optimization tooling for repository agents #17404, #16543, and #18069.
  • Add the LLVM BOLT profile-generation engine and enable its scheduled producer and promotion pipeline #16204 and #18212.
  • Expand KV cache manager V2 coverage for C++ pool rebalancing, multi-rank host-tier synchronization, and scheduler stability #17387, #17790, and #17823.
  • Expand Kimi K3 validation with KDA transfer, MoE, multi-node disaggregated, DFlash, DSpark, and FP8 prefill coverage #17484, #17711, #17802, #17846, and #17483.
  • Add the ModelExpress donor-receiver qualification harness and validation cleanup #17222 and #17834.
  • Add speculative-decoding acceptance-length regression baselines #17474.
  • Consolidate disaggregated and SSM end-to-end tests, enable Hopper IB transport, and replace disaggregated DWDP accuracy checks with aggregated coverage #16614, #17680, #17300, and #17546.
  • Restore pipeline-parallel regression coverage and add LoRA-manager H100 tests #17117 and #18114.
  • Add VisualGen Cosmos3 LPIPS baselines plus Nemotron-3.5-Lightning and Qwen3.8 accuracy guards #17780, #18180, and #18071.
  • Reduce GVR top-K decode unit-test compilation time through compile-signature-aware parametrization #17878.
  • Prune retired Bielik, DeepSeek-R1-Distill, LLaVA-Next, VILA, Nemotron-H, StarCoder2, Phi, Mistral, Step-3.7-Flash, and Qwen test suites #17345, #17351, #17595, #17675, #17601, #17674, #17600, #17598, #18013, and #17827.
  • Remove stale VBench test infrastructure #17516.
  • Expand MoE benchmark coverage with group-aware routing and Qwen3.8/Kimi K3 presets #17791 and #17959.
  • Expand performance coverage with corrected Qwen3.5 sequence lengths, benchmark-refresh mappings, and Nemotron-Ultra-V3 GB300 cases #17422, #17317, and #17609.
  • Improve benchmark configuration by inferring fixed-dataset sequence lengths and supporting MLA model dimensions #17825 and #16707.
  • Expand layer-wise benchmarking for DeepSeek-V4 and Kimi K3 and derive module performance cases from traces #17801 and #17804.
  • Harden performance testing with bounded and configurable server waits, correct pipeline-branch selection, accurate per-iteration device timing, and restored MLA decode gates #17881, #17202, #18127, #18011, #18054, and #17596.

What's Changed

• [None][fix] Enforce Responses conversation history capacity by @fallintoplace in #15043
• [None][fix] Kimi K3 MLA: pass attn_output to MLA.forward_impl by @brnguyen2 in #17445
• [None][fix] Kimi K3: eager CUDA-graph buffer allocation and prebuilt fused-verify constants by @brnguyen2 in #17421
• [https://nvbugs/6434512][fix] Select Marlin for Qwen3.5 MoE on Hopper by @lfr-0531 in #17293
• [TRTLLM-14798][perf] Fuse Wan DupUp3D output mapping by @taianz-nv in #17003
• [TRTLLM-15151][chore] load models lazily by @QiJune in #17281
• [None][feat] Opt GPT-OSS in to KV cache manager V2 by default by @eopXD in #16942
• [https://nvbugs/6475346][fix] Avoid stale CUDA Graph memory-pool handles across captures by @liji-nv in #16952
• [None][infra] Recognize SM107 (Rubin) in build config and arch detection by @zhangcl in #17336
• [None][test] Fix qwen3.5_9b avg_seq_len and restore H100 coverage for three Qwen3.5 perf cases by @yufeiwu-nv in #17422
• [https://nvbugs/6517846][fix] Raise AGG server-ready timeout to 3600s and unwaive 4 perf-sanity cases by @chenfeiz0326 in #17459
• [TRTLLM-15017][chore] Unify one-model speculative decoding samplers into SpecSampler by @zhaoyangwang-nvidia in #17292
• [None][fix] Skip no-op MXFP4 weight padding by @jiaganc in #17259
• [None][infra] Log infra-retry classify declines instead of silent rethrow by @brnguyen2 in #17432
• [https://nvbugs/6159132][fix] Differentiate the two paths via extra_acc_spec="tp_attn" when attention_dp=False by @tensorrt-cicd in #13922
• [TRTLLMINF-213][infra] Artifactory container image migration by @tburt-nv in #16970
• [https://nvbugs/6506990][fix] Don't treat NVLE-only as CC enabled by @dhansen-nvidia in #16850
• [None][doc] Note runtime-dependency requirement for the kimi_k3 Slurm container image by @brnguyen2 in #17456
• [https://nvbugs/6566737][fix] Set the Mamba SSM cache dtype to FP32 in Qwen3.5 weight-update tests by @trtllm-agent in #17415
• [TRTLLM-15215][fix] Kimi K3: make the FP8 weight-read master switch opt-in by @brnguyen2 in #17446
• [TRTLLM-15177][chore] Kimi K3 post-merge cleanup: config/import/test hygiene + L0 wiring by @brnguyen2 in #17413
• [None][chore] Default enable PDL for benchmoe by @leslie-fang25 in #17463
• [TRTLLM-13696][test] Part2.3: Migrate CPU only tests - auto deploy by @tongyuantongyu in #17314
• [None][fix] Enable INT8 weight-only (W8A16) MoE for non-gated activations by @Dorijan10 in #15550
• [None][fix] Fix gen-only async kvtransfer hang by @reasonsolo in #17107
• [None][test] use skip_less_mpi_world_size instead of skip_less_device by @crazydemo in #17405
• [None][feat] Support dense FP8 LoRA end to end by @achartier in #16810
• [https://nvbugs/6565412][fix] Size trtllm-gen and thop decode buffers for beam search by @pranav-nvidia in #17014
• [TRTLLM-13409][feat] hard-kill all ranks when one rank's executor loop crashes by @JunyiXu-nv in #16592
• [None][perf] enable zero-copy token passing in KVCacheManagerV2 by @lowsfer in #17308
• [None][infra] Replace --gpus with --gpus-per-node in single node multi gpus by @yiqingy0 in #17359
• [None][infra] Align VisualGen CBTS rule with CODEOWNERS scope by @crazydemo in #17251
• [TRTLLM-14555][perf] Optimize parallel VAE halo exchange path by @taianz-nv in #17001
• [None][feat] Add KV cache manager V2 support for DSA by @yizhang-nv in #16060
• [None][fix] fix oom by @xinhe-nv in #17486
• [TRTLLMINF-191][infra] Use native pytest capture for S3 logs by @niukuo in #16739
• [None][fix] Update to get central version variable for trtllm-bench. by @FrankD412 in #11562
• [None][chore] Use public flashinfer APIs by @yihwang-nv in #15300
• [TRTLLM-15264][doc] Kimi K3 disagg: stop recommending a UCX_TLS pin by default by @brnguyen2 in #17479
• [TRTLLM-15178][fix] Pad an empty attention-DP scheduled batch so the fleet can make forward progress by @xwang233 in #17379
• [#17146][fix] Resolve reasoning mode from the rendered prompt by @joerowell in #17305
• [None][feat] Add Cosmos3-Edge (Nemotron-dense) support by @ishovkun in #16773
• [TRTLLM-15264][test] Wire KDA disagg transfer tests into CI (cpu_only CPU stage + l0_b200 GPU) by @brnguyen2 in #17484
• [TRTLLM-14806][feat] Prefer Python V2 transceiver for LlamaForCausalLM and Gemma3ForCausalLM by @moraxu in #16787
• [None][fix] Qwen3.5 weight mapper for FP8 per-channel checkpoints by @amukkara in #17433
• [#13318][fix] Gracefully fit token budget at prep boundary by @thorjohnsen in #15187
• [TRTLLM-14814][chore] Add SA speculative-decoding eval config for Kimi K3 by @brnguyen2 in #17455
• [TRTLLM-13948][feat] Clean up DeepSeek tests using CPP Transceiver v1 by @asfiyab-nvidia in #17090
• [None][chore] Add deprecation notice for _autodeploy backend by @nvchenghaoz in #15642
• [https://nvbugs/6262973][perf] Move AllReduce autotuner dispatch to C++ by @nv-lschneider in #16902
• [TRTLLM-14955][refactor] Declare MoE backend behaviour instead of comparing classes by @xxi-nv in #17411
• [https://nvbugs/6472256][fix] Fix disagg stress cluster flapping and DeepSeek R1 FP4 ctx OOM; add aiperf error-rate gate by @moraxu in #17427
• [https://nvbugs/6490036][test] Isolate part1 FP8 weight-update tests by @shikicloud in #17503
• [https://nvbugs/6529792][fix] Avoid GPT-OSS V2 cache estimation OOM by @jiaganc in #17470
• [TRTLLM-15076][test] Remove Bielik-11B-v2.2-Instruct functional and unit tests by @xinhe-nv in #17345
• [None][test] Cover KVCacheManagerV2 C++ pool rebalance path by @thorjohnsen in #17387
• [TRTLLM-14704][feat] Import agent-flow and modeling bringup agent into TensorRT-LLM by @WeiHaocheng in #16543
• [None][feat] support NIXL cache transceiver with Ray by @chuangz0 in #17295
• [None][fix] Structured Output with DSpark Speculative Drafter by @chungen04 in #16965
• [None][test] Dump a stack traceback when a test hangs by @BowenFu in #16374
• [None][chore] add backup owner for scaffolding by @yuxianq in #17549
• [None][infra] Stop CodeRabbit requiring copyright header on test-list files by @brnguyen2 in #17528
• [None][fix] Skip draft KV mirror when IndexMapper is saturated in KVCacheManagerV2 by @erictsai-nv in #17347
• [None][refactor] BREAKING: Remove model_path from TriAttention config and reuse the executor's pretrained config by @Hudayday in #17501
• [None][infra] Allow normal review for waiver updates by @chzblych in #17547
• [TRTLLM-15078][test] Remove all DeepSeek-R1-Distill-* tests by @xinhe-nv in #17351
• [TRTLLM-13234][feat] Complete TorchSampler beam search: length_penalty, diversity_rate, early_stopping, VBWS, and CBA performance by @zhaoyangwang-nvidia in #16620
• [TRTLLM-12714][fix] Suspend CUDA-graph padding dummies before pool rebalance adjust() by @thorjohnsen in #16157
• [TRTLLM-14388][refactor] BREAKING: Force 2 model spec dec to fall back to 1 model by @mikeiovine in #17366
• [None][feat] Support the masked DSA indexer k-cache pool in the Python cache transceiver by @Tabrizian in #17283
• [https://nvbugs/6566735][fix] recover disaggregated worker heartbeat registration by @reasonsolo in #17461
• [None][fix] Fix CuTeDSL packed QKV copy lowering by @karljang in #17425
• [None][feat] Remove stale VBench test infrastructure by @yibinl-nvidia in #17516
• [https://nvbugs/6480621][fix] Preserve KV ownership in disaggregated precheck by @chienchunhung in #17223
• [TRTLLMINF-237][infra] Re-home L0_Test SLURM finalizer by @dpitman-nvda in #17263
• [None][fix] remove duplicated test by @tburt-nv in #17579
• [https://nvbugs/6487039][fix] Generalize ADP dummy lifecycle by @chienchunhung in #16921
• [None][fix] Update llm_function_core.txt by @SimengLiu-nv in #17571
• [https://nvbugs/6550803][fix] Pin non-greedy capture state in speculative metadata tests by @trtllm-agent in #17219
• [TRTLLM-15216][fix] Kimi K3 on KVCacheManagerV2: conv-state layout and SSM iteration stats by @brnguyen2 in #17447
• [TRTLLM-13696][test] Part2.4: Migrate CPU only tests - others by @tongyuantongyu in #17339
• [None][test] Extend the compressor BF16 tie tolerance to the remaining prefill assertions by @brnguyen2 in #17563
• [https://nvbugs/6599150][fix] Initialize dt_bias in KDA verify-parity test by @brnguyen2 in #17584
• [TRTLLM-15293][perf] Add tiered GVR CuTe DSL top-k decode kernels (stacked on #16457) by @longcheng-nv in #16877
• [None][infra] CBTS code coverage date early save by @crazydemo in #17253
• [TRTLLM-15157][infra] automate stale PR cleanup by @karljang in #17383
• [None][infra] Run mypy type check in the build stage without pre-commit by @brnguyen2 in #17560
• [None][fix] Remove obsolete GPT-OSS two-model Eagle3 tests by @achartier in #17597
• [None][test] Adjust timeout test cases to avoid large log by @yufeiwu-nv in #17596
• [TRTLLM-15157][fix] scan all pages in stale PR cleanup by @karljang in #17602
• [None][refactor] Organize SMG gRPC adapter by protocol by @QiJune in #17179
• [https://nvbugs/6487038][fix] Stop single-rank disagg errors from crashing all gen ranks by @Shixiaowei02 in #16834
• [None][fix] Fix possible arithmetic overflow in FMHAv2 launcher by @tongyuantongyu in #11263
• [None][fix] Suppress mypy redundant-cast for the sampler module by @brnguyen2 in #17637
• [None][chore] Drop the skip_* sampling flags from one-model spec metadata by @zhaoyangwang-nvidia in #17554
• [TRTLLM-13215][perf] Skip redundant one-model sampling-param refills by @zhaoyangwang-nvidia in #17544
• [None][doc] Add Qwen3.8 deployment guide and configs by @Wanli-Jiang in #17594
• [TRTLLM-14865][feat] Support occurrence penalties with beam search for TorchSampler by @lori-ren in #17189
• [TRTLLM-14628][feat] Support out-of-tree build state via --build_root by @brnguyen2 in #17524
• [https://nvbugs/6566891][fix] Use FlashInfer FA2 for Gemma4 on SM120 and SM121 by @lfr-0531 in #17557
• [None][perf] fp8 block scale quant fusion in SM90 Cutlass MoE by @amukkara in #16849
• [TRTLLM-14704][feat] Support multi-modal part of K3 by @WeiHaocheng in #17050
• [None][ci] Run doc build on a CPU pod instead of an a10 GPU tester by @brnguyen2 in #17639
• [None][fix] Keep ADP ranks in collective lockstep on request errors and fail fast on desync by @roborluo in #16687
• [TRTLLM-15093][test] Prune LLaVA-Next functional and unit tests by @xinhe-nv in #17595
• [None][test] Consolidate dis-agg E2E Tests by @Shixiaowei02 in #16614
• [None][infra] Add blossom-ci authorized users by @yiqingy0 in #17676
• [TRTLLMINF-300][infra] fix slurm_install.sh use broken artifact by @weiminwang-nv in #17634
• [None][chore] Scope TRTLLM_DISABLE_KV_CACHE_TRANSFER_OVERLAP to the gen worker in disagg gen_only by @dc3671 in #17535
• [None][perf] Fuse Kimi K3 KDA projections by @jiaganc in #17311
• [None][infra] Improve retry-decision messages for failed tests by @yiqingy0 in #17682
• [TRTLLM-14628][feat] Add --out-of-tree wheel builds that never write into the checkout by @brnguyen2 in #17525
• [None][fix] Restore DSpark disaggregated decoding accuracy by @longlee0622 in #17619
• [TRTLLMINF-311][infra] Infra-scoped fail-fast: defer K8s infra aborts instead of cascading by @brnguyen2 in #17615
• [https://nvbugs/6427411][test] Re-enable PP regression tests by @mingyangHao in #17117
• [TRTLLM-15177][chore] Consolidate trtllm-gen SiTu activation slot handling by @brnguyen2 in #17414
• [TRTLLM-15102][fix] Prune VILA family models by @xinhe-nv in #17675
• [https://nvbugs/6507082][fix] remove a test case of falcon model by @QiJune in #17677
• [TRTLLMINF-40][fix] Quote SLURM no-job-ID cleanup echo so it no longer exits 127 by @brnguyen2 in #17704
• [TRTLLM-13394][feat] Support loading MTP weights from standalone checkpoint by @mikeiovine in #17378
• [TRTLLM-15402][fix] VisualGen: fix Cache-DiT stats log line (lazy %-formatting never applied) by @chang-l in #17697
• [https://nvbugs/6272397][fix] Prevent host OOM during checkpoint prefetch by @moraxu in #17430
• [None][perf] Skip request_context class construction on the non-draft path by @pranav-nvidia in #17582
• [TRTLLM-14628][perf] Incremental, streamed artifact copy-back into the checkout by @brnguyen2 in #17538
• [TRTLLM-13694][infra] Add benchmark refresh mappings by @Mgluhovskoi in #17317
• [None][feat] Add HunyuanVideo 1.5 text-to-video support to VisualGen by @jloftin-nv in #15562
• [https://nvbugs/6525008][fix] Isolate FlashInfer JIT workspaces for MPI workers by @VALLIS-NERIA in #17233
• [https://nvbugs/6476233][fix] Cap max_seq_len on H200 DeepSeek-V3.2 blockscale test by @pranav-nvidia in #17659
• [TRTLLM-13579][feat] BREAKING: Support BCG in Prefill by @GuanhuaWang2001 in #16609
• [None][fix] Widen Ulysses async-barrier timeout 10s->60s for multi-node warmup by @luyiyun1021 in #16486
• [TRTLLM-15284][feat] add Kimi K3 SiTU MegaMoE support by @longlee0622 in #17624
• [None][test] Nemotron-Ultra-V3 perf-sanity cases (GB300); de-enroll DeepSeek-V3.2, Kimi-K2.5 & Llama cases by @chenfeiz0326 in #17609
• [None][perf] Avoid paged MSA K/V materialization during prefill by @peihu-nv in #16857
• [TRTLLM-14956][refactor] make MoE implementation selection reproducible by @xxi-nv in #17532
• [TRTLLMINF-263][infra] authenticate Artifactory nSpect checks by @niukuo in #17679
• [TRTLLM-15037][fix] KVCM-V2: synchronize auto host-tier quota across ranks by @dongfengy in #17717
• [TRTLLM-15177][test] Wire remaining Kimi K3 MoE unit tests into L0 (Hopper) by @brnguyen2 in #17711
• [None][feat] Enforce multimodal encoder runtime budgets with budgeted output storage by @yechank-nvidia in #16051
• [None][fix] Fix latent mypy errors in sampler.py (no-any-return, comparison-overlap) by @brnguyen2 in #17783
• [TRTLLM-11628][perf] Batch the beam-search finish-reason reduction by @zhaoyangwang-nvidia in #17494
• [https://nvbugs/6490028][fix] Increase the cross-library cuBLAS tolerance in autotuner tests by @trtllm-agent in #16788
• [https://nvbugs/6435126][test] Attribute fatal test_unittests_v2 failures to inner tests by @BowenFu in #16358
• [None][infra] Add execution and test-runner skills for Claude Code by @kaiyux in #17404
• [TRTLLM-15264][fix] Reject non-Python transceiver routes for Kimi K3 disaggregated serving by @brnguyen2 in #17480
• [None][feat] Honor SamplingParams.seed on the one-model speculative path by @zhaoyangwang-nvidia in #17599
• [None][feat] Support DFlash RoPE, sliding-window configuration, and TRTLLM-gen attention backend by @jhaotingc in #16914
• [TRTLLM-14833][chore] Introduce executor/ray/ for Ray executor integration by @lori-ren in #17487
• [TRTLLM-15100][test] Prune Nemotron-H functional and unit tests by @xinhe-nv in #17601
• [None][fix] Preserve advanced sampling state through graph capture by @jhaotingc in #17671
• [TRTLLM-15115][test] Prune Starcoder2-3B functional and unit tests by @xinhe-nv in #17674
• [None][fix] Declare attention runtime-workspace bytes/token as a backend contract by @eopXD in #16432
• [https://nvbugs/6435121][fix] Eliminate the trtllm-serve port reservation race with --port 0 + --report_addr by @JunyiXu-nv in #17460
• [https://nvbugs/6566772][fix] Prevent streaming executor shutdown from blocking on an empty queue by @trtllm-agent in #17426
• [None][refactor] BREAKING remove conversation ID from disaggregated params by @jiaganc in #17618
• [None][fix] Fence MoE shared writes before async bulk copy by @liji-nv in #17621
• [None][fix] Drop the stale choices list on --moe-backend-for-prefill by @ruodil in #17612
• [https://nvbugs/6572835][fix] Use HashStore for single-rank MegaMoE tests by @Barry-Delaney in #17497
• [https://nvbugs/6478723][fix] Preserve extra dynamic-tree state across forward calls by @trtllm-agent in #16733
• [TRTLLMINF-40][chore] Dedupe COMMON_SSH_OPTIONS to reference bloom's DEFAULT_CUSTOM_SSH_OPTIONS by @brnguyen2 in #17706
• [None][perf] fuse DSpark attention and RMSNorm RoPE by @longlee0622 in #17307
• [https://nvbugs/6445494][infra] Upgrade Triton to 3.7.0 for Torch 2.12.0 compatibility by @HandongLi-01 in #16303
• [None][fix] PostprocWorker: skip load_hf_tokenizer when tokenizer_dir is None by @aswinvisva in #17568
• [TRTLLM-14727][test] Create MX donor-receiver qualification test harness by @chienchunhung in #17222
• [None][refactor] Mixed Modality Support for Nemotron Nano Omni V3 by @aswinvisva in #16764
• [TRTLLM-13308][feat] Make the KVCacheManagerV2 KV pool rebalance safe under TP, CP, PP and attention DP by @thorjohnsen in #17391
• [TRTLLM-15218][chore] KVCacheManagerV2: report the prefix attention alone supports by @brnguyen2 in #17448
• [TRTLLM-15204][fix] Initialize KDA dt_bias and re-enable the prefill parity suite on B200 by @brnguyen2 in #17793
• [TRTLLM-15264][fix] Fail only the affected requests on disagg peer-layout mismatch by @brnguyen2 in #17482
• [https://nvbugs/6601574][fix] Clean up leftover disagg server processes between Ray tests by @chuangz0 in #17632
• [TRTLLM-15304][fix] MoE and MTP fixes for large hybrid-attention FP8 serving by @Wanli-Jiang in #17635
• [None][test] Allow IB transport in disaggregated tests on Hopper by @erictsai-nv in #17300
• [None][feat] Enable KVCacheManagerV2 by default for Gemma3 and Gemma4 by @erictsai-nv in #17396
• [None][fix] Update ModelExpress dependency version by @chienchunhung in #17716
• [https://nvbugs/6198760][fix] Refresh FMHA cubins to fix SageAttention when KV-sequence is not a multiple of 128 by @xrq-phys in #17648
• [None][refactor] Refactor Kimi K3 MLP by @jiaganc in #17312
• [https://nvbugs/6437410][fix] fix nemotron weight update test by @shuyixiong in #16712
• [TRTLLM-14719][infra] Add spec-dec acceptance-length regression baselines and remove PARD CnnDailymail coverage by @allisonlim-nv in #17474
• [https://nvbugs/6590666][fix] Detect worker death during initialization by @chienchunhung in #17578
• [https://nvbugs/6571220][fix] Correct unfused attention context workspace sizing by @pranav-nvidia in #17026
• [TRTLLM-14597][perf] Fuse the DSv4 MLA prologue: kv_a_layernorm, q_nope FP8 quant and Q RoPE by @dc3671 in #17273
• [None][fix] Synchronize one-model MTP scheduler draft tokens by @2ez4bz in #17836
• [None][test] Add kimi k3 cases for multi-node disagg by @fredricz-20070104 in #17802
• [TRTLLM-14957][refactor] split the MoE base class by responsibility and converge the loader owner gate by @xxi-nv in #17777
• [https://nvbugs/6627248][fix] Initialize multimodal flag in PP loop fixture by @yechank-nvidia in #17874
• [None][infra] Fix UpdateTestDurations.groovy concurrent-update conflict by @EmmaQiaoCh in #17689
• [#17572][fix] Emit the withheld buffer in DeepSeek streaming tool parsers by @Yigtwxx in #17573
• [https://nvbugs/6581049][ci] Mark flaky test as xfail instead of skipping by @2ez4bz in #17719
• [None][test] Add back kimi k25 and deepseek v32 cases from qa side by @fredricz-20070104 in #17861
• [https://nvbugs/6604925][fix] Swap NVFP4 batched-GEMV grid axes for large output dimensions by @trtllm-agent in #17660
• [TRTLLM-14730][feat] Add image edit serving endpoint for visual generation models by @yibinl-nvidia in #16987
• [https://nvbugs/5547275][test] Scope stage mapping checks to live stages by @BowenFu in #17651
• [TRTLLM-14775][doc] Document Helix source test prerequisites by @BowenFu in https://github.com/NVIDIA/TensorRT-LLM/pull/17628
• [https://nvbugs/6566765][fix] Release Qwen MoE CUDA memory between tests by @BowenFu in https://github.com/NVIDIA/TensorRT-LLM/pull/17641
• [https://nvbugs/6607487][fix] fix LoRA host/device cache dtype reconfiguration race by @achartier in https://github.com/NVIDIA/TensorRT-LLM/pull/17678
• [None][perf] Preallocate Kimi attention residual snapshots by @longlee0622 in https://github.com/NVIDIA/TensorRT-LLM/pull/17053
• [TRTLLMINF-40][fix] Fold sbatch submit stderr into the SLURM failure classifier (+ standby backoff) by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17709
• [https://nvbugs/5573856][fix] Re-enable NCCL symmetric shape-growth test by @nv-lschneider in https://github.com/NVIDIA/TensorRT-LLM/pull/17507
• [None][infra] Fix DIND networking/runtime compatibility and release wheel context by @chzblych in https://github.com/NVIDIA/TensorRT-LLM/pull/17829
• [https://nvbugs/6618106][fix] Carry single_step_greedy on SampleStateTensorsHostTorch by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17828
• [TRTLLM-15103][test] clean retired phi tests by @xinhe-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17600
• [None][perf] Remove spurious sync in sparse fmha forward by @brb-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17840
• [None][fix] Make disaggregated benchmark server health timeout configurable by @longlee0622 in https://github.com/NVIDIA/TensorRT-LLM/pull/17881
• [https://nvbugs/6373561][fix] Fix minimax M3 E2E test by @2ez4bz in https://github.com/NVIDIA/TensorRT-LLM/pull/17565
• [TRTLLM-14778][perf] Cache the Whisper suppress-token index as a device tensor by @pranav-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/17583
• [None][chore] Sync vendored agent-flow with upstream by @kaiyux in https://github.com/NVIDIA/TensorRT-LLM/pull/17869
• [TRTLLM-14839][chore] Relocate root PEFT modules into _torch/peft/ by @lori-ren in https://github.com/NVIDIA/TensorRT-LLM/pull/17806
• [TRTLLM-15179][test] Add multi-rank test for KVCM v2 host-tier quota sync by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17790
• [TRTLLM-14727][test] Clean up MX qualification follow-ups by @chienchunhung in https://github.com/NVIDIA/TensorRT-LLM/pull/17834
• [None][fix] helix: compensate position_id for the overlap scheduler by @lancelly in https://github.com/NVIDIA/TensorRT-LLM/pull/17811
• [https://nvbugs/6507080][fix] Make TokenizerBase.__repr__ deterministic by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/16839
• [https://nvbugs/6607481][fix] Isolate stateful KV-cache comparison by @BowenFu in https://github.com/NVIDIA/TensorRT-LLM/pull/17673
• [None][feat] bench_moe: group-aware routing support by @guqiqi in https://github.com/NVIDIA/TensorRT-LLM/pull/17791
• [None][fix] Restore Gemma4 shared-KV draft loading by @2ez4bz in https://github.com/NVIDIA/TensorRT-LLM/pull/17837
• [None][test] GVR top-K decode UT: compile-signature-aware parametrization (27m31s -> 13m34s) by @longcheng-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17878
• [TRTLLM-14575][perf] Avoid quadratic copy in KV cache block reuse by @xwang233 in https://github.com/NVIDIA/TensorRT-LLM/pull/17848
• [https://nvbugs/6602928][fix] Serialize HF remote-code loading when building input processors by @tianyuxbear in https://github.com/NVIDIA/TensorRT-LLM/pull/17708
• [None][doc] Correct per-GPU throughput in ADP Balance blog figure by @yunruis in https://github.com/NVIDIA/TensorRT-LLM/pull/17905
• [None][fix] Add KV cache V2 recompute pause path by @yizhang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/15252
• [TRTLLM-14777][doc] Correct LMCache example output guidance by @BowenFu in https://github.com/NVIDIA/TensorRT-LLM/pull/17629
• [https://nvbugs/6581121][fix] Keep the parallel-factor CSV row aligned with eight workers by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17506
• [None][perf] fuse Kimi K3 routing and MXFP8 quantization by @longlee0622 in https://github.com/NVIDIA/TensorRT-LLM/pull/17059
• [https://nvbugs/6626655][test] Align QA test-list timeout for multimodal Kimi-K2.5 dep8 with pre-merge by @tianyuxbear in https://github.com/NVIDIA/TensorRT-LLM/pull/17945
• [TRTLLM-15406][feat] Support occurrence penalties in one-model speculative decoding by @zhaoyangwang-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/17701
• [TRTLLM-15433][chore] BREAKING: Remove all WIDEEP files by @xxi-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17797
• [#17580][fix] Stream the text that precedes a tool call in DeepSeek parsers by @Yigtwxx in https://github.com/NVIDIA/TensorRT-LLM/pull/17903
• [TRTLLM-15465][feat] Support SA speculative decoding under disaggregated serving for Kimi K3 by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17939
• [TRTLLM-14839][chore] Add shim files for root PEFT module relocation by @lori-ren in https://github.com/NVIDIA/TensorRT-LLM/pull/17857
• [TRTLLM-12838][infra] CBTS: coverage-based test selection by @crazydemo in https://github.com/NVIDIA/TensorRT-LLM/pull/16776
• [https://nvbugs/6567554][fix] Make DeepSeek-V4 layer-wise benchmarks run, and derive module perf cases from the trace by @fredricz-20070104 in https://github.com/NVIDIA/TensorRT-LLM/pull/17801
• [None][fix] Don't infra-retry deterministic SLURM test failures by @dpitman-nvda in https://github.com/NVIDIA/TensorRT-LLM/pull/17645
• [TRTLLM-11901][feat] Add trtllm weight loading metrics by @yijingl-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/15993
• [None][fix] Fix Cosmos3 CUDA event crash when text guardrail blocks prompt by @ishovkun in https://github.com/NVIDIA/TensorRT-LLM/pull/17510
• [TRTLLM-14628][fix] Do not trust just-written mtimes in sync_tree copy-back by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17919
• [TRTLLMINF-320][infra] Extend infra-scoped fail-fast deferral to SLURM-scoped aborts by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17938
• [None][feat] Remove padding in Kimi K3 MLA module by @pengbowang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17684
• [TRTLLM-15403][fix] VisualGen: warn when a requested attention backend silently falls back to VANILLA by @chang-l in https://github.com/NVIDIA/TensorRT-LLM/pull/17698
• [TRTLLM-14879][feat] qualify Qwen2 dense for MX by @chienchunhung in https://github.com/NVIDIA/TensorRT-LLM/pull/16974
• [None][chore] Extend OpenEngine ownership to dynamo dev by @tanmayv25 in https://github.com/NVIDIA/TensorRT-LLM/pull/17657
• [None][perf] Address D2D copies in mixed batches by @brb-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17841
• [None][fix] Reword Cosmos3 negative prompt terms that trip the guardrail blocklist by @ishovkun in https://github.com/NVIDIA/TensorRT-LLM/pull/17523
• [None][feat] Add GlmImage text-to-image pipeline support to VisualGen by @jloftin-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/16018
• [https://nvbugs/6463967][fix] DeepSeek-V4 one-model MTP separate draft kv cache (TEP) by @nvxuanyuc in https://github.com/NVIDIA/TensorRT-LLM/pull/16887
• [TRTLLM-13767][chore] upgrade CUTLASS DSL stack to 4.6.1 by @taianz-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17274
• [None][perf] Overlap DSA heuristic prev_topk write-back on the aux stream by @longcheng-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/16666
• [None][test] Replace disaggregated DWDP accuracy tests with aggregated coverage by @tianyuz-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17546
• [None][feat] Enable FA4 + parallel VAE in LTX-2 examples by @luyiyun1021 in https://github.com/NVIDIA/TensorRT-LLM/pull/17610
• [TRTLLM-15404][fix] VisualGen: refuse static quant recipes against unquantized checkpoints (silent weight corruption) by @chang-l in https://github.com/NVIDIA/TensorRT-LLM/pull/17699
• [None][fix] Pin distro for CI sysinfo detection by @chzblych in https://github.com/NVIDIA/TensorRT-LLM/pull/18004
• [https://nvbugs/6541356][fix] Align the transceiver precheck with the serving KV cache setup by @chuangz0 in https://github.com/NVIDIA/TensorRT-LLM/pull/17121
• [https://nvbugs/6600098][test] Stabilize KV cache V2 scheduler tests by @yizhang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17823
• [None][test] Increase KV transfer timeout for PP4-to-TP4 disaggregated test by @erictsai-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17873
• [https://nvbugs/6596064][fix] Release eager outputs before CUDA graph capture by @yuxianq in https://github.com/NVIDIA/TensorRT-LLM/pull/17824
• [None][feat] Add Kimi K3 to layer-wise benchmarks by @dc3671 in https://github.com/NVIDIA/TensorRT-LLM/pull/17804
• [None][doc] Add tech blog: Evaluating Agentic Serving with Trace Replay and Job-Level Metrics by @KleinBlueC in https://github.com/NVIDIA/TensorRT-LLM/pull/16230
• [None][fix] Keep sysinfo distro probe working without the distro module; fail empty test-list renders loudly by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17993
• [None][infra] Trigger PLC source code scannning pipeline in pre-merge by @yuanjingx87 in https://github.com/NVIDIA/TensorRT-LLM/pull/15552
• [https://nvbugs/6467684][fix] Bump the golang image tag to 1.23 by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/16531
• [None][perf] Qwen3.5/3.8 wave-2: MoE, attention-DP, GDN replay, weight loading by @Wanli-Jiang in https://github.com/NVIDIA/TensorRT-LLM/pull/17700
• [None][feat] bring up Kimi K3 NVFP4 with CUTLASS and cuteDSL MegaMoE SiTU by @longlee0622 in https://github.com/NVIDIA/TensorRT-LLM/pull/17865
• [None][fix] Use class-qualified access for clarity by @Fan-Yunfan in https://github.com/NVIDIA/TensorRT-LLM/pull/8855
• [None][fix] normalize Qwen3.8 27B FP8 VLM quantization config by @Wanli-Jiang in https://github.com/NVIDIA/TensorRT-LLM/pull/17786
• [None][infra] Cap diffusers<0.40 to avoid deps conflicts by @weiminwang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18047
• [https://nvbugs/6631848][fix] Size Mistral encoder budget for MMMU by @2ez4bz in https://github.com/NVIDIA/TensorRT-LLM/pull/18028
• [None][infra] Add blossom-ci authorized users by @yiqingy0 in https://github.com/NVIDIA/TensorRT-LLM/pull/18057
• [None][fix] Clamp conversation-affinity ADP routing to per-rank slot by @lancelly in https://github.com/NVIDIA/TensorRT-LLM/pull/17534
• [None][infra] cbts-v2 coverage pilot allowlist by @crazydemo in https://github.com/NVIDIA/TensorRT-LLM/pull/17996
• [TRTLLM-15304][perf] add low-M BF16 GEMM dispatcher for SM10x decode by @Wanli-Jiang in https://github.com/NVIDIA/TensorRT-LLM/pull/17889
• [TRTLLMINF-136][infra] Strengthen check_test_list.py param-ID validation and gate on validate<->collection parity by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17975
• [https://nvbugs/6329155][fix] Raise glm5 tep8 8k1k max_num_tokens to 8192 to fit isl=8192 prefill by @chenfeiz0326 in https://github.com/NVIDIA/TensorRT-LLM/pull/18006
• [https://nvbugs/6418815][test] Pin fp32-matmul precision in VisualGen LPIPS tests and re-baseline Cosmos3 goldens by @ishovkun in https://github.com/NVIDIA/TensorRT-LLM/pull/17780
• [TRTLLM-14818][test] Port Kimi K3 DFlash/DSpark eval helpers and KDA FP8 prefill test to main by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17846
• [None][fix] test_cbts_coverage_pilot: use capfd to avoid autouse-fixture clash by @chang-l in https://github.com/NVIDIA/TensorRT-LLM/pull/18076
• [None][fix] Fix FlashInfer shared-KV speculative decode by @2ez4bz in https://github.com/NVIDIA/TensorRT-LLM/pull/17264
• [None][feat] Add cold-page codec support to KVCM2 by @lowsfer in https://github.com/NVIDIA/TensorRT-LLM/pull/17512
• [None][refactor] Modularize sparse Top-K selection by @lfr-0531 in https://github.com/NVIDIA/TensorRT-LLM/pull/17622
• [None][fix] PYTHON transceiver: decode-CP transfer support and concurrency corruption fixes by @lancelly in https://github.com/NVIDIA/TensorRT-LLM/pull/17794
• [TRTLLM-14268][feat] Cosmos3 Transfer (control-video conditioning) by @ishovkun in https://github.com/NVIDIA/TensorRT-LLM/pull/16394
• [None][fix] release Kimi K3 checkpoint mappings after loading by @jiaganc in https://github.com/NVIDIA/TensorRT-LLM/pull/17999
• [None][fix] bound CLI dependency minimums by @Mgluhovskoi in https://github.com/NVIDIA/TensorRT-LLM/pull/18045
• [TRTLLM-15264][test] Kimi K3 disagg review fixups: KDA test geometry, gate docs, example cleanup by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17483
• [TRTLLM-15033][feat] Upstream Kimi K3 MLA decode backend selection to main by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17800
• [TRTLLM-15398][perf] VisualGen MLP: cublasLt GELU-tanh epilogue for the unquantized bf16 path by @chang-l in https://github.com/NVIDIA/TensorRT-LLM/pull/17693
• [None][fix] Stop KV cache estimation probes from pinning the full host tier by @brb-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17850
• [https://nvbugs/6652876][fix] Fix release checks and waive Laguna XS FP8 by @chzblych in https://github.com/NVIDIA/TensorRT-LLM/pull/18099
• [None][perf] Optimize MiniMax-M3 MSA block selection by @peihu-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17236
• [https://nvbugs/6590418][perf] Infer avg sequence length for fixed perf datasets by @VALLIS-NERIA in https://github.com/NVIDIA/TensorRT-LLM/pull/17825
• [TRTLLM-15498][refactor] consolidate Kimi KDA production frontend by @jiaganc in https://github.com/NVIDIA/TensorRT-LLM/pull/17822
• [TRTLLM-15314][feat] Add FP8 LoRA support for B200 by @achartier in https://github.com/NVIDIA/TensorRT-LLM/pull/17521
• [None][perf] Use FP8 MiniMax-M3 MSA indexer QK by @peihu-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17318
• [https://nvbugs/6601578][fix] Avoid MoE multi-GPU rendezvous port race by @Barry-Delaney in https://github.com/NVIDIA/TensorRT-LLM/pull/17831
• [TRTLLM-15099][test] Prune Mistral functional and unit tests by @xinhe-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17598
• [None][refactor] Move disagg transfer helpers from py_executor into disaggregation/executor by @nv-xtf in https://github.com/NVIDIA/TensorRT-LLM/pull/17966
• [https://nvbugs/6541343][fix] Add a shared-filesystem barrier for all Slurm ranks by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17367
• [None][doc] Fix some typos of modeling agent by @WeiHaocheng in https://github.com/NVIDIA/TensorRT-LLM/pull/18108
• [https://nvbugs/6456085][fix] Harmony: stop discarding malformed tool-call messages silently by @JunyiXu-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17464
• [None][feat] KVCacheManagerV2: helix decode-CP via a global super-block ledger by @lancelly in https://github.com/NVIDIA/TensorRT-LLM/pull/17866
• [https://nvbugs/6337226][fix] Count profiled calls with min() of positive line hits by @tensorrt-cicd in https://github.com/NVIDIA/TensorRT-LLM/pull/15487
• [https://nvbugs/6633928][fix] Limit Gemma3 FP8 accuracy test sequence length by @erictsai-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18056
• [None][test] Set Llama4 QA max sequence length by @erictsai-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18062
• [TRTLLM-14622][feat] Add VisualGen dynamic LoRA support by @yibinl-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/17435
• [TRTLLM-15400][perf] fuse per-token AdaLN for VisualGen Wan 2.2 5B by @chang-l in https://github.com/NVIDIA/TensorRT-LLM/pull/17695
• [https://nvbugs/6581065][fix] Reap wedged workers during session drain by @BowenFu in https://github.com/NVIDIA/TensorRT-LLM/pull/17644
• [None][infra] Disable RTXPro6000D multi gpu stages due to some nodes are offline by @yiqingy0 in https://github.com/NVIDIA/TensorRT-LLM/pull/17543
• [TRTLLM-15120][test] Prune Step-3.7-Flash functional tests by @xinhe-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18013
• [https://nvbugs/6550099][fix] Raise no-top-k equivalence tolerance by @lori-ren in https://github.com/NVIDIA/TensorRT-LLM/pull/17998
• [None][fix] preserve streaming SSE event boundaries by @reasonsolo in https://github.com/NVIDIA/TensorRT-LLM/pull/18090
• [TRTLLM-13409][fix] give the benchmark-disagg fill gate's retry loop a deadline by @JunyiXu-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17202
• [TRTLLM-14692][feat] Overlap LoRA and base model computations by @AlessioNetti in https://github.com/NVIDIA/TensorRT-LLM/pull/16951
• [None][fix] fix CppMambaHybridCacheManager by @bo-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17492
• [None][fix] Refresh group greedy state per iteration by @lishicheng1996-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17710
• [None][chore] Split TorchSampler feature code out of sampler.py by @zhaoyangwang-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/17890
• [None][feat] Optimize human trigger replan for modeling agent by @WeiHaocheng in https://github.com/NVIDIA/TensorRT-LLM/pull/18069
• [TRTLLM-14851][chore] Relocate VisualGen kernel tests by @lori-ren in https://github.com/NVIDIA/TensorRT-LLM/pull/18063
• [None][infra] Reduce resource requested by CPU stages by @tongyuantongyu in https://github.com/NVIDIA/TensorRT-LLM/pull/18125
• [None][feat] Enable DeepSeek-V4 and DSA (DeepSeek-V3.2/GLM) serving on SM120 via FlashInfer sparse-MLA by @lucifer1004 in https://github.com/NVIDIA/TensorRT-LLM/pull/16224
• [TRTLLM-15302][fix] Reject dead prefetched MPI pools by @BowenFu in https://github.com/NVIDIA/TensorRT-LLM/pull/17642
• [None][fix] propagate max token limits to hybrid cache managers by @Wanli-Jiang in https://github.com/NVIDIA/TensorRT-LLM/pull/18116
• [TRTLLM-15655][chore] BREAKING: remove star attention by @lori-ren in https://github.com/NVIDIA/TensorRT-LLM/pull/18025
• [https://nvbugs/6632606][fix] Pass server_start_timeout to Ray disagg by @chuangz0 in https://github.com/NVIDIA/TensorRT-LLM/pull/18012
• [None][fix] Restore CpType import for HELIX KV cache by @chzblych in https://github.com/NVIDIA/TensorRT-LLM/pull/18143
• [TRTLLM-15314][test] Add LoRA manager tests to H100 L0 by @achartier in https://github.com/NVIDIA/TensorRT-LLM/pull/18114
• [None][fix] Report request-level KV cache metrics with KVCM V2 by @yizhang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17815
• [None][feat] Add CFT counted-write path to MoE all-to-all by @zhangcl in https://github.com/NVIDIA/TensorRT-LLM/pull/17647
• [https://nvbugs/6625710][fix] Re-attach radix-tree blocks detached under a live request by @lowsfer in https://github.com/NVIDIA/TensorRT-LLM/pull/18095
• [https://nvbugs/6428091][fix] Move use_host_stop_criteria into shared sampling metadata by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/16132
• [TRTLLM-15401][fix] VisualGen Wan: collapse uniform per-patch timesteps — unlock TeaCache on Wan2.2-TI2V-5B by @chang-l in https://github.com/NVIDIA/TensorRT-LLM/pull/17696
• [None][infra] PLC pipeline display fix by @yuanjingx87 in https://github.com/NVIDIA/TensorRT-LLM/pull/17419
• [https://nvbugs/6631019][fix] Add a torch_compiling context manager for multimodal graph state by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17949
• [TRTLLM-15176][fix] Harden Kimi K3 tool-call parsing by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17980
• [https://nvbugs/6546909][fix] Match CUDA Graph handle lifetime to captured graphs by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17313
• [https://nvbugs/6464169][fix] Pad trtllm-gen MoE route map to cover kernel over-read by @rosong11 in https://github.com/NVIDIA/TensorRT-LLM/pull/17951
• [TRTLLMINF-263][infra] unify nSpect version organization by @niukuo in https://github.com/NVIDIA/TensorRT-LLM/pull/18119
• [None][fix] Kimi K3: bound MegaMoE expert-weight memory at EP8; drop the MoE TP/EP env overrides by @longlee0622 in https://github.com/NVIDIA/TensorRT-LLM/pull/18059
• [None][infra] CBTS compact coverage data by @crazydemo in https://github.com/NVIDIA/TensorRT-LLM/pull/17803
• [None][fix] GVR indexer top-K: repair the non-converged threshold search by @longcheng-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17550
• [None][fix] CuTe DSL GVR top-K decode: repair the non-converged threshold search by @longcheng-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18094
• [None][feat] Kimi K3: KDA-TP + MLA-DCP (helix) wiring by @lancelly in https://github.com/NVIDIA/TensorRT-LLM/pull/17796
• [None][feat] Add qwen3_8 / kimi_k3 bench_moe presets and activation plumbing by @guqiqi in https://github.com/NVIDIA/TensorRT-LLM/pull/17959
• [None][refactor] Compose phased FMHA implementations by @yuxianq in https://github.com/NVIDIA/TensorRT-LLM/pull/18026
• [https://nvbugs/6581066][fix] Abort the wedged worker world when a rank dies during init by @sunnyqgg in https://github.com/NVIDIA/TensorRT-LLM/pull/17817
• [https://nvbugs/6384747][fix] Accept and forward input_ids in MiniMax-M3 MoE routing by @tensorrt-cicd in https://github.com/NVIDIA/TensorRT-LLM/pull/15712
• [https://nvbugs/6561777][fix] Add job-and-step-keyed Slurm rank barriers by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17370
• [https://nvbugs/6525059][fix] 128KiB-align TMA-OOB MoE workspace buffers on Blackwell by @sunnyqgg in https://github.com/NVIDIA/TensorRT-LLM/pull/18133
• [https://nvbugs/6617948][fix] Restore trtllm-gen MLA decode perf gate dropped by #15300 by @chenfeiz0326 in https://github.com/NVIDIA/TensorRT-LLM/pull/18054
• [TRTLLMINF-161][infra] Add simplified infrastructure dry-run pipeline by @mzweilz in https://github.com/NVIDIA/TensorRT-LLM/pull/17813
• [None][test] Add Nemotron-3.5-Lightning NVFP4 Marlin MTP3 accuracy test by @Wanli-Jiang in https://github.com/NVIDIA/TensorRT-LLM/pull/18180
• [https://nvbugs/6566707][fix] Isolate LoRA peft-cache-override test from MPI session reuse by @sunnyqgg in https://github.com/NVIDIA/TensorRT-LLM/pull/18184
• [TRTLLM-15192][feat] Specialize CUDA graphs for LoRA by @AlessioNetti in https://github.com/NVIDIA/TensorRT-LLM/pull/17412
• [#17522][fix] Restore conv_state ordering barriers in the Triton conv kernels by @wilyan09007 in https://github.com/NVIDIA/TensorRT-LLM/pull/17539
• [https://nvbugs/6627979][fix] Apply chat template in test_eagle3_output_repetition_4gpus by @zhaoyangwang-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/18170
• [TRTLLM-15349][feat] Mamba page table enhancements by @bo-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17691
• [TRTLLM-14958][refactor] separate MoE execution units from complete layers by @xxi-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18018
• [TRTLLM-15719][test] add Qwen3.8-2.4T-A95B accuracy guards by @Wanli-Jiang in https://github.com/NVIDIA/TensorRT-LLM/pull/18071
• [TRTLLM-15621][fix] Reject sampling params the one-model speculative path cannot honor by @zhaoyangwang-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/17955
• [TRTLLM-13143][feat] BREAKING: VisualGen serving API: response_format support path and modify async job status by @luyiyun1021 in https://github.com/NVIDIA/TensorRT-LLM/pull/17490
• [None][infra] Set perf-sanity s_branch from the pipeline branch, not the Jenkins folder by @chenfeiz0326 in https://github.com/NVIDIA/TensorRT-LLM/pull/18127
• [None][test] Revert the gen_only warmup probe and fix the per-iter device step time metric by @chenfeiz0326 in https://github.com/NVIDIA/TensorRT-LLM/pull/18011
• [None][perf] split routing kernels to reduce compile time by @jiahanc in https://github.com/NVIDIA/TensorRT-LLM/pull/17893
• [TRTLLM-14388][refactor] Remove 2 model spec dec drafting loops by @mikeiovine in https://github.com/NVIDIA/TensorRT-LLM/pull/17562
• [None][refactor] add Vanilla sparse attention primitives by @yihwang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18044
• [https://nvbugs/6458213][test] use TRTLLM MoE for Blackwell skip-softmax tests by @bobboli in https://github.com/NVIDIA/TensorRT-LLM/pull/17623
• [https://nvbugs/6425321][fix] remove unnecessary VSA + Ulysses NCCL flag by @o-stoner in https://github.com/NVIDIA/TensorRT-LLM/pull/16751
• [None][perf] Optimize DFlash draft forward by @amukkara in https://github.com/NVIDIA/TensorRT-LLM/pull/17935
• [https://nvbugs/5977180][fix] size KV cache and admit requests per beam width by @athena-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/16802
• [https://nvbugs/6590664][fix] Reap idle single-rank CTX transfers by @chienchunhung in https://github.com/NVIDIA/TensorRT-LLM/pull/17564
• [TRTLLMINF-95][infra] Add LLVM BOLT profile-generation engine and pipeline by @mlefeb01 in https://github.com/NVIDIA/TensorRT-LLM/pull/16204
• [#17574][fix] Complete zero-argument tool calls in the streaming tool parser by @Yigtwxx in https://github.com/NVIDIA/TensorRT-LLM/pull/17575
• [None][perf] Address inter-idle times and decode-first assumption in MSA by @brb-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17986
• [https://nvbugs/6459792][fix] Add MLA dimensions to benchmark ModelConfig by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/16707
• [None][feat] Add the ported MiniMax-M3 decode kernels ahead of their wiring by @brb-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17842
• [None][perf] use native Wan VAE for Cosmos3 by @taianz-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17555
• [None][fix] Synchronize KV cache V2 host fallback across ranks by @yizhang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18092
• [None][test] Remove deepseek v32 test cases on the qa side for disagg multinode perf testing by @fredricz-20070104 in https://github.com/NVIDIA/TensorRT-LLM/pull/18225
• [https://nvbugs/6561778][fix] Fence all ranks before pytest launch in multi-node slurm_run.sh by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17372
• [None][infra] Fix CBTS skip-rate calculation method by @crazydemo in https://github.com/NVIDIA/TensorRT-LLM/pull/16467
• [https://nvbugs/6647349][fix] Replace FuzzyWuzzy with RapidFuzz by @lfr-0531 in https://github.com/NVIDIA/TensorRT-LLM/pull/18201
• [None][fix] Stabilize Gemma4 FA2 CUDA Graph decode on Hopper by @lfr-0531 in https://github.com/NVIDIA/TensorRT-LLM/pull/18002
• [https://nvbugs/6596590][fix] Densify warmup mesh for sparse fmha kernel by @pengbowang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17961
• [TRTLLM-15748][chore] Remove dead code and de-reflect lazy attributes for ModelEngine by @lori-ren in https://github.com/NVIDIA/TensorRT-LLM/pull/18181
• [None][feat] Add nvfp4 situ moe cubins by @rosong11 in https://github.com/NVIDIA/TensorRT-LLM/pull/17940
• [None][chore] Add Top-K code owners by @yuxianq in https://github.com/NVIDIA/TensorRT-LLM/pull/18246
• [None][ci] disable autodeploy test stages by @QiJune in https://github.com/NVIDIA/TensorRT-LLM/pull/18107
• [TRTLLM-15293][perf] Add self-sampling (GVR V2) top-K decode kernels by @longcheng-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17821
• [https://nvbugs/6525010][fix] Isolate torch compile test sessions by @liji-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18065
• [None][perf] Compute response GPU timings once per batch by @liji-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18053
• [https://nvbugs/6579626][perf] Fallback small BF16 context batches by @yihwang-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18196
• [None][fix] Multimodal Mixin Guided Decoder Delegation by @aswinvisva in https://github.com/NVIDIA/TensorRT-LLM/pull/18162
• [https://nvbugs/6327718][fix] Drain EPD thread pool before proxy shutdown by @aswinvisva in https://github.com/NVIDIA/TensorRT-LLM/pull/17988
• [None][fix] Simplify idle disagg KV transfer progress check by @Tabrizian in https://github.com/NVIDIA/TensorRT-LLM/pull/17324
• [https://nvbugs/6652876][fix] Add correct autotuner constraint for sm120 fp8 block scale moe by @DomBrown in https://github.com/NVIDIA/TensorRT-LLM/pull/18140
• [None][infra] Set --platform when retagging CI image by @yuanjingx87 in https://github.com/NVIDIA/TensorRT-LLM/pull/18208
• [None][fix] Drop the unbound is_idle guard left in the idle disagg CTX reap by @pranav-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/18263
• [TRTLLMINF-339][infra] Enable BOLT profile-gen producer (job + cadence + promote) by @mlefeb01 in https://github.com/NVIDIA/TensorRT-LLM/pull/18212
• [None][perf] Cut per-iteration executor bookkeeping in the hang detector and profiler by @pranav-nvidia in https://github.com/NVIDIA/TensorRT-LLM/pull/17531
• [TRTLLM-15030][fix] CuteDSL MLA decode follow-ups: bucket AutoTuner fallback, autotune + disagg tests by @brnguyen2 in https://github.com/NVIDIA/TensorRT-LLM/pull/17792
• [https://nvbugs/6644226][fix] Reduce DeepSeek V4 EPLB loading memory by @lfr-0531 in https://github.com/NVIDIA/TensorRT-LLM/pull/18185
• [None][test] Consolidate ssm dis-agg E2E Tests by @bo-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17680
• [TRTLLM-9905][infra] Upload results periodically by @yiqingy0 in https://github.com/NVIDIA/TensorRT-LLM/pull/14138
• [TRTLLMINF-187][fix] Drain S3 test logs from rank zero by @niukuo in https://github.com/NVIDIA/TensorRT-LLM/pull/18240
• [None][infra] CBTS coverage arch combine v2 by @crazydemo in https://github.com/NVIDIA/TensorRT-LLM/pull/18058
• [None][doc] Point the dis-agg docs at transceiver v2 by @Shixiaowei02 in https://github.com/NVIDIA/TensorRT-LLM/pull/18248
• [TRTLLMINF-263][infra] pass nSpect release entitlement by @niukuo in https://github.com/NVIDIA/TensorRT-LLM/pull/18217
• [TRTLLM-15498][perf] eliminate KDA recurrent-state and metadata copies by @jiaganc in https://github.com/NVIDIA/TensorRT-LLM/pull/17887
• [TRTLLM-15108][test] clean obsolete qwen tests by @xinhe-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/17827
• [None][infra] Upload build info during pipeline setup by @crazydemo in https://github.com/NVIDIA/TensorRT-LLM/pull/18237
• [None][infra] Upgrade NIXL to v1.4.0 and UCX to v1.22.x by @zackyoray in https://github.com/NVIDIA/TensorRT-LLM/pull/17891
• [None][test] Remove unused performance test utilities and clean up test list configurations in llm_perf_core.yml. by @yufeiwu-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18287
• [None][feat] Add locality domain runtime and bindings for Rubin by @zhangcl in https://github.com/NVIDIA/TensorRT-LLM/pull/17662
• [TRTLLM-12758][feat] honor named function in tool_choice for non-harmony models by @JunyiXu-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/14290
• [https://nvbugs/6581067][fix] Compute the unbiased GEMM and bias in FP32 before casting by @trtllm-agent in https://github.com/NVIDIA/TensorRT-LLM/pull/17505
• [https://nvbugs/6663062][fix] remove Mistral moe tests by @xinhe-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18288
• [None][feat] laguna: config-driven sqrt-softplus router scoring by @joerowell in https://github.com/NVIDIA/TensorRT-LLM/pull/17918
• [https://nvbugs/6660905][fix] Gate TritonFusedMoE on the SwiGLU family, not the gpt-oss flavour by @xxi-nv in https://github.com/NVIDIA/TensorRT-LLM/pull/18256
• [None][chore] Make disagg test paths sole-owned by disagg-devs by @Shixiaowei02 in https://github.com/NVIDIA/TensorRT-LLM/pull/18235
• [https://nvbugs/6645731][fix] align logprobs with token_ids when stopwords are trimmed by @Wanli-Jiang in https://github.com/NVIDIA/TensorRT-LLM/pull/18190

New Contributors

• @fallintoplace made their first contribution in #15043
• @zhangcl made their first contribution in #17336
• @Dorijan10 made their first contribution in #15550
• @BowenFu made their first contribution in #16374
• @roborluo made their first contribution in #16687
• @jloftin-nv made their first contribution in #15562
• @GuanhuaWang2001 made their first contribution in #16609
• @HandongLi-01 made their first contribution in #16303
• @tanmayv25 made their first contribution in https://github.com/NVIDIA/TensorRT-LLM/pull/17657
• @wilyan09007 made their first contribution in https://github.com/NVIDIA/TensorRT-LLM/pull/17539

Full Changelog: v1.3.0rc24...v1.3.0rc25