Issues

• A MoE combine fused with the next RMSNorm gets a slow one-block-per-token kernel (bf16)

  pytorch/TensorRT#5042 · Filter by author shoumikhin opened on Nov 13, 2026

• compile_and_save needs twice the engine size on disk while it saves to ExecuTorch

  pytorch/TensorRT#5041 · Filter by author shoumikhin opened on Nov 13, 2026

• torch_tensorrt.dynamo.compile silently drops settings it does not know, so a typo changes the result with no warning

  pytorch/TensorRT#5040 · Filter by author shoumikhin opened on Nov 13, 2026

• Add a converter for PyTorch's packed variable-length attention (varlen_attn)

  pytorch/TensorRT#5039 · Filter by author shoumikhin opened on Nov 13, 2026

• aten._grouped_mm converter raises TypeError when the offsets are a constant

  pytorch/TensorRT#5038 · Filter by author shoumikhin opened on Nov 13, 2026

• Dynamic sequence length: attention and split fail to build in 2.15 when the length is a symbol

  pytorch/TensorRT#5037 · Filter by author shoumikhin opened on Nov 13, 2026

• A weight read by several TensorRT engines is stored once per engine

  pytorch/TensorRT#5036 · Filter by author shoumikhin opened on Nov 13, 2026

• The nightly runtime wheel pins an ExecuTorch nightly from Nov 5, so the nightly stack cannot follow ExecuTorch

  pytorch/TensorRT#5035 · Filter by author shoumikhin opened on Nov 13, 2026

• Torch-TensorRT 2.15.0 has only a manylinux_2_39 wheel on aarch64, so pip on older glibc silently installs an old release

  pytorch/TensorRT#5034 · Filter by author shoumikhin opened on Nov 13, 2026

• 🐛 [Bug] dynamo.compile silently ignores kwarg_inputs entries the program does not take

  pytorch/TensorRT#5031 · Filter by author shoumikhin opened on Nov 13, 2026
  

• 🐛 [Bug] enabled_precisions={torch.float16} is silently ignored in 2.15, and the engine stays FP32

  pytorch/TensorRT#5030 · Filter by author shoumikhin opened on Nov 13, 2026
  4

  

• TensorRT masked attention silently ignores the mask in float16 with a single query and 513+ keys

  pytorch/TensorRT#5011 · Filter by author shoumikhin opened on Nov 11, 2026

• tensorrt.plugin AOT plugins fail at enqueue when the kernel needs more than 48 KiB of dynamic shared memory

  pytorch/TensorRT#5010 · Filter by author shoumikhin opened on Nov 11, 2026
  1

  

• TensorRT masked attention is 2 to 3 times slower when the key length is odd

  story: Performance & Benchmarking
  pytorch/TensorRT#5009 · Filter by author shoumikhin opened on Nov 11, 2026

• TensorRT reduce-sum over a short axis of a bf16 tensor is 8x slower than torch.sum (MoE top-k combine)

  backend: TensorRTstory: Performance & Benchmarking
  Bug

  pytorch/TensorRT#5008 · Filter by author shoumikhin opened on Nov 11, 2026

• TensorRT IMoELayer on Jetson Thor: the top-k combine step is 18x to 1600x slower than PyTorch, and bf16/fp16 scores fail to build

  backend: TensorRTstory: Performance & Benchmarking
  Bug

  pytorch/TensorRT#5007 · Filter by author shoumikhin opened on Nov 11, 2026

• ExecuTorch: review the design of CUDA graph replay across multiple green contexts (#4983), and benchmark with VLA models

  pytorch/TensorRT#5003 · Filter by author micwill755 opened on Nov 10, 2026
  1

• ✨[Feature] Resource-partitioning recompute: redesign choices for what to recompute across engines (follow-up to #4961)

  component: partitioningfeature requestRFC: Under Review
  pytorch/TensorRT#5002 · Filter by author cehongwang opened on Nov 10, 2026

• Low-memory compile API (map_weights, low_memory): 2.16 second pass and follow-ups

  component: api [Python]component: dynamofeature request
  pytorch/TensorRT#5001 · Filter by author cehongwang opened on Nov 10, 2026 · v2.16.0

• 🐛 [Bug] Multi-device safe mode runs the engine on the caller's GPU instead of the GPU it was loaded on

  pytorch/TensorRT#4997 · Filter by author 0z5a opened on Nov 9, 2026
  

• 🐛 [Bug] The engine cache returns an engine built by another TensorRT version or on another host, and compile() fails in the refitter

  pytorch/TensorRT#4996 · Filter by author 0z5a opened on Nov 9, 2026
  

• ExecuTorch CUDA graph replay stops replaying when buffers move and shapes change often

  pytorch/TensorRT#4991 · Filter by author shoumikhin opened on Nov 6, 2026

• 🐛 [Bug] With enable_autocast, an in-place update is lost through a view taken before it

  pytorch/TensorRT#4990 · Filter by author shoumikhin opened on Nov 6, 2026

• 🐛 [Bug] int8 math with an int8 constant does not wrap around on overflow

  pytorch/TensorRT#4989 · Filter by author shoumikhin opened on Nov 6, 2026

• 🐛 [Bug] FP16 Dynamo compile produces incorrect results for batched attention when query originates from an expanded tensor

  bug
  pytorch/TensorRT#4968 · Filter by author unilin-pietergobin opened on Nov 3, 2026
  3