TensorRT-LLM v1.4.0rc0 Labels Two API Changes BREAKING, Including Removal of the KVCacheManagerV2 Python Backend, and Lists Five Known Issues

Overview

NVIDIA’s TensorRT-LLM repository carries a release tagged v1.4.0rc0, published on GitHub on October 14, 2026 at 16:35 UTC. In the release notes, two entries in the API list are labelled “(BREAKING)”, and the Highlights section opens with five known issues. This report covers what the notes and a linked pull request say, as read on October 15, 2026.

What We Know

The two entries labelled BREAKING

The release notes list these under API: “Remove the Python backend of KVCacheManagerV2 (BREAKING)” and “Migrate LlmRequest fields to Python (BREAKING)”.

For the first, the pull request that made the change describes the background in its description: KVCacheManagerV2 “shipped two implementations of the same subsystem behind TLLM_KV_CACHE_MANAGER_V2_BACKEND”. It states that TLLM_KV_CACHE_MANAGER_V2_BACKEND=python “no longer exists” and that build_wheel.py --mypyc and TRTLLM_ENABLE_MYPYC “are gone”. The same description says “Streaming KV events (kv_cache_config.kv_events_config) are dropped” and that validate_streaming_support now rejects that configuration and points at the buffered path via event_buffer_max_size. The description also says the tensorrt_llm/runtime/kv_cache_manager_v2/ directory went from 36 tracked files down to 4.

For the second, the title of the linked pull request reads “[None][refactor] BREAKING: Migrate LlmRequest fields to Python”. The release notes give no further detail on what the migration changes for users, and this report does not characterize it beyond that title.

The five known issues

The release notes list five known issues, each scoped by hardware or configuration:

• Multi-GPU (TP2) streaming or RPC executor workloads may hang during worker initialization when MPI worker pools are reused across LLM instances, with the note “(observed on H100)”.
• DeepSeek MLA models on SM120/SM121 GPUs (the notes give RTX PRO 6000 Blackwell as an example) may crash with an illegal memory access when CUDA graphs are captured with BF16 BMM.
• GPT-OSS models on SM120 GPUs may fail with attention data parallelism when automatic communication selects NCCL EP for unsupported hidden sizes.
• Qwen3 models with tied embeddings (the notes give Qwen3-0.6B as an example) may fail at model construction when attention data parallelism is enabled with TP greater than 1.
• OpenTelemetry tracing may export additional FastAPI spans when FastAPI 0.142 or newer is installed.

Other items in the notes

Under API, the release notes also list command-line configuration overrides, per-request speculative-decoding acceptance in responses, OpenAI-style errors when a request exceeds the context length, a per-model serving-extension registry for the OpenAI server, and OpenEngine gRPC bindings shipped in the wheel “so no separate package install is needed”.

The Model Support section includes GLM-5.3-Flash (glm5_next) support and K-EXAONE-2 in the PyTorch backend. The Feature section includes a Mooncake store pool, CLI and V2 scheduler preemption (part 1), beam search with the C++ KVCacheManagerV2, and streaming of C++ KV events with multimodal payloads. The Documentation section says docs were updated for “NGC PyTorch 26.08 and PyTorch 2.14.0”. The Test & Infra section includes removing Conan from the C++ build and shipping a BOLT-optimized SBSA release wheel.

What We Don’t Know

• Terms the notes do not define. The notes use “OpenEngine”, “EVS”, “EPD” and “Mooncake store” without explanation. This report quotes them as written and does not define them.
• How two items relate. The pull request description says streaming KV events configured through kv_cache_config.kv_events_config are dropped, while the release notes separately list “Stream C++ KV events with multimodal payloads”. The notes do not say how these two items relate.
• Impact of the LlmRequest migration. The release notes do not say which users or interfaces the “Migrate LlmRequest fields to Python” change affects.
• Fixes for the known issues. The notes list the issues without workarounds or fix timelines.
• Release timing. The notes do not say when a final v1.4.0 will follow the v1.4.0rc0 tag.
• Single-source reporting. The release content rests on NVIDIA’s own GitHub release page and pull request. No independent coverage or benchmark of the release was reviewed.