All pull requests

Pull requests

• fix #254: Pass object tool-call arguments to the chat template as mappings

  #257·mihai-chiorean opened Oct 19, 2026·5 of 7 tasks·

  ·
  1

  1

• fix #253: Untie the lm_head before quantizing it for tied-embedding models

  #256·mihai-chiorean opened Oct 19, 2026·6 of 7 tasks·

  ·
  1

  1

• fix #150: Load CNN/DailyMail by its namespaced Hub id in reduce-vocab and zh_en_mixed

  #255·mihai-chiorean opened Oct 19, 2026·4 of 7 tasks·

  ·
  1

• fix: feed channel-first patches to the Cosmos3-Edge visual engine (#232)

  #251·0z5a opened Oct 16, 2026·3 of 7 tasks·

  ·
  1

  1

• feat: Add an INT8 Gemma 4 PLE runtime sidecar

  #249·mihai-chiorean opened Oct 15, 2026·7 tasks done·

  ·
  1

• perf(tokenizer): O(n log n) byte-pair merge for long pieces

  #248·drakarah opened Oct 13, 2026·6 of 7 tasks·

  ·
  2

• feat #240: Expose the encoder embedding cache budget through pybind and tensorrt-edgellm-serve

  #246·mihai-chiorean opened Oct 10, 2026·7 tasks done·

  ·
  1

  1

• fix #239: Name the tensor, its size and the free device memory when an allocation fails

  #245·mihai-chiorean opened Oct 10, 2026·6 of 7 tasks·

  ·
  1

  1

• fix #237: Attach the LoRA branch to the unsmoothed activation of AWQ INT4 GEMMs

  #238·mihai-chiorean opened Oct 10, 2026·5 of 7 tasks·

  ·
  1

  1

• fix #183: Define cudaLibrary_t for the CuTe DSL F16 MoE, LayerNorm and RMSNorm runners on CUDA 12.0-12.7

  #236·mihai-chiorean opened Oct 9, 2026·4 of 7 tasks·

  ·
  1

  1

• feat: Add an INT8 token-embedding runtime sidecar

  #235·mihai-chiorean opened Oct 9, 2026·7 tasks done·

  ·
  4

• fix(server): keep JSON-schema property order when canonicalising guides

  #234·drakarah opened Oct 8, 2026·7 tasks done·

  ·
  1

• test: Verify copy-pr-bot and CPU runner access

  #231·jhalabi-nv opened Oct 6, 2026Collaborator·

  ·
  1

  

• fix: avoid unused prefill memory during speculative engine builds

  #229·0z5a opened Sep 30, 2026·

  ·

• fix: avoid native NVFP4 projection compiler crash with rank-two GEMM

  #225·0z5a opened Sep 28, 2026·

  ·

• fix: lower native log-softmax through FP32 reductions

  #224·0z5a opened Sep 28, 2026·

  ·

• feat: Add blockwise FP8 GEMM support for Qwen3-VL

  #220·Sunk011 opened Sep 23, 2026·7 of 8 tasks·

  ·
  1

  3

• feat: add native continuous batching and chunked prefill for LLM runtime

  #215·ajmalrasi opened Sep 15, 2026·

  ·

• Fix Qwen3 ONNX export bug related to 0/1 specialization

  #213·victoroliv2 opened Sep 14, 2026·7 tasks·

  ·

• fix https://github.com/NVIDIA/TensorRT-Edge-LLM/issues/211: advance the Philox offset per sampled token

  #212·filipemartinsubrobotics opened Sep 12, 2026·7 tasks done·

  ·
  1

• fix: say why Int4GroupwiseGemmPluginV2 cannot execute when its kernels are absent

  #209·filipemartinsubrobotics opened Sep 11, 2026·

  ·
  1

• feat: enable SM103 (B300 / Blackwell Ultra) across the attention SM allowlists

  #207·filipemartinsubrobotics opened Sep 11, 2026·

  ·
  1

• fix https://github.com/NVIDIA/TensorRT-Edge-LLM/issues/204: pass the CUDA include path to the XQA NVRTC JIT

  #205·filipemartinsubrobotics opened Sep 11, 2026·

  ·
  1

• fix: Share the target lm_head for DSpark drafts without their own head

  #200·rain-sicoreai opened Sep 9, 2026·7 tasks done·

  ·

• feat: Add request batching and spec tree-size options to the experimental server, fix FP8-KV and unquantized MTP draft builds

  #199·rain-sicoreai opened Sep 8, 2026·7 tasks done·

  ·
  1