Conv device wrappers: update() leaves stale geometry and fused-output pointers, two wrappers miss the 2 GiB gate, negative output extents not rejected

Description

Three validation/update gaps in the 2.x conv device wrappers, plus two smaller latent inconsistencies.

1. update() refreshes pointers but leaves geometry stale

conv/device/implicit_gemm_convolution.h:305-316 (the fusion implicit_gemm_convolution_fusion.h:209-222 and direct direct_convolution.h:192-204 wrappers have the same shape):

Status update(Arguments const &args, void *workspace = nullptr) {
  params_.ptr_A = args.ref_A.data();
  params_.ptr_B = args.ref_B.data();
  params_.ptr_C = args.ref_C.data();
  params_.ptr_D = args.ref_D.data();
  params_.output_op = args.output_op;
  params_.semaphore = static_cast<int *>(workspace);
  return Status::kSuccess;
}

problem_size, implicit_gemm_problem_size, grid_tiled_shape, swizzle_log_tile, gemm_k_iterations(_per_channel), all four iterator Params, and split_k_mode are never refreshed. Reusing the operation via update() with a different problem size or split-k mode launches the old grid and old K trip count over new tensors: silent wrong results or OOB.

Additionally, when the plain wrapper is instantiated on kernel::ImplicitGemmConvolutionWithFusedEpilogue or ...WithAbsMax (default_conv2d_fprop_with_broadcast.h), Params also carry ptr_Vector/ptr_Tensor/ldr/ldt (and ptr_Aux), which update() skips even though it is exactly the class of pointer it does refresh for C/D. After updating to a new layer's bias/residual tensor, the kernel keeps reading and writing the previous one.

2. The >=2 GiB offset-overflow gate is missing from two sibling wrappers

The plain wrapper rejects any activation/filter/output tensor of 2 GiB or more because the 2.x iterators use 32-bit offsets:

// conv/device/implicit_gemm_convolution.h:117-126
if (args.problem_size.activation_size() * sizeof(ElementA) >= (1ull << 31) || ...

That gate was added there only (3.8). implicit_gemm_convolution_fusion.h:104-133 and direct_convolution.h:104-157 run the same iterator family but their can_implement goes straight to per-iterator alignment checks without any size gate. A >=2 GiB problem passes validation and overflows iterator offsets at runtime.

3. Degenerate problems yield negative output extents that are never rejected

conv/convnd_problem_shape.hpp:554-572:

return 1 + (act_ext + pad_total - ((filter_ext - 1) * dilation + 1)) / tstride;

When the filter exceeds the padded input this goes negative (C++ truncation): e.g. fprop with activation {5x5}, filter {8x8}, no padding, stride 1 gives z = p = -2, and the linearized M of the GEMM comes back positive (-2 * -2). Nothing in ConvProblemShape construction or the sm90/sm100 collectives' can_implement rejects it (alignment checks see only the positive A/B shapes), so the 3.x path builds TMA im2col descriptors from garbage geometry instead of returning kErrorInvalidProblem. The result is stride-dependent (extent 0 at stride 2, -2 at stride 1).

Additional latent observations (no in-tree trigger today)

• conv/kernel/sm100_implicit_gemm_tma_warpspecialized.hpp: to_underlying_arguments sizes the scheduler workspace segment from the transformed problem-shape form while get_workspace_size/initialize_workspace use the linearized form; benign while every reachable conv scheduler reports zero workspace.
• Group handling differs between helpers: conv2d_problem_size.h:264-269 groups the filter extent by groups, conv3d_problem_size.h:295-300 does not; grouped conv3d kernels do not exist so the pair never disagrees in practice.

Suggested fixes

1. Refresh the full Params set in each wrapper's update() (or drop the entry points).
2. Port the 2 GiB gate into the fusion and direct wrappers' can_implement.
3. Reject non-positive output extents in calculate_xformed_act consumers (or at can_implement) with kErrorInvalidProblem.

Start with the update() and can_implement() implementations in conv/device/implicit_gemm_convolution.h, conv/device/implicit_gemm_convolution_fusion.h, and conv/device/direct_convolution.h, then trace output-shape handling in conv/convnd_problem_shape.hpp and the sm90/sm100 collective validation paths. Done means all relevant Params and tensor pointers are refreshed, both sibling wrappers reject sizes at or above 2 GiB, and non-positive output extents return kErrorInvalidProblem.