[aciddelgado]

Description:

Summary

Add  com.microsoft::GatedDeltaNetStateReplay  v1 to replay captured GatedDeltaNet transitions directly on WebGPU. This provides the ORT-side building block for partial-accept state reconstruction without downloading recurrent state to the CPU.

The operator supports one request and one descriptor, with FP32 state and capsule backing tensors. CPU metadata supplies element offsets, head geometry, original capture capacity, and the number of leading transitions to keep.

Implementation

• Use one invocation per state element and 128 threads per workgroup to apply the captured recurrence:

hk = floor(h
  * Hk / Hv)
  s = source[h,v,i]
  for t = 0 .. kept_count-1:
    s
  *= decay[t,h]
    s += key[t,hk,i]
  * delta[t,h,v]
  destination[h,v,i] = s

• Use existing  ProgramBase ,  ComputeContext , and  BufferView  machinery. Logical offsets do not require storage-binding alignment, and binding-limit validation follows actual buffer ownership and segmentation.
• Require output prebinding to the inactive destination allocation. Preserve source, capsule, and neighboring destination elements; reject writable destination aliasing with read-only inputs.
• Call the existing  FlushAndWaitChecked()  helper after dispatch so successful execution returns only after checked GPU completion. Keep enqueue and completion separate for future multi-dispatch integration.
• Register the schema and kernel, regenerate contrib documentation, and document the original incoming-state and capture-capacity invariants.

Validation

Coverage includes all kept prefixes, nonzero offsets, Qwen’s unaligned capsule-row layout, segmented/shared backing, poisoned unused tails, destination sentinels, and invalid metadata. Integration coverage replays an actual WebGPU producer capsule against independent prefix runs. A separate test checks queue completion immediately after  Session::Run , before blocking readback.

Native Release validation on Vulkan/lavapipe: 73 tests passed, one expected skip. The completion test also passed ten consecutive repetitions. Numerical comparisons use explicit FP32 tolerances.

Scope

No GenAI changes or new public C API/ABI. Convolution replay, batching, graph capture, and GenAI helper-session integration are out of scope. Hardware-GPU performance and plugin/browser execution are not established by this validation.

[Copilot]

🔵 Needs a closer look

Manual aliasing, segmented GPU bindings, and blocking queue-completion semantics warrant final human validation.

0 open findings

What changed in this PR

Adds a WebGPU-only contrib operator for replaying captured GatedDeltaNet state transitions without CPU state downloads. No actionable defects were identified.

Changes:

• Adds and registers the replay schema, kernel, and WGSL implementation.
• Enforces metadata, aliasing, buffer-limit, and checked-completion requirements.
• Adds unit, integration, segmentation, validation, and completion tests.

File	Description
onnxruntime/​test/​contrib_ops/​gated_delta_net_state_replay_op_test.cc	Adds comprehensive replay tests.
onnxruntime/​test/​contrib_ops/​gated_delta_net_op_test.cc	Tests replay against producer-generated capsules.
onnxruntime/​core/​providers/​webgpu/​ep/​README.md	Documents replay behavior and constraints.
onnxruntime/​core/​providers/​webgpu/​compute_context.h	Exposes checked flush-and-wait.
onnxruntime/​core/​graph/​contrib_ops/​ms_opset.h	Registers the schema.
onnxruntime/​core/​graph/​contrib_ops/​bert_defs.cc	Defines the operator contract and inference.
onnxruntime/​contrib_ops/​webgpu/​webgpu_contrib_kernels.cc	Registers the WebGPU kernel.
onnxruntime/​contrib_ops/​webgpu/​bert/​gated_delta_net_state_replay.wgsl.template	Implements the replay recurrence.
onnxruntime/​contrib_ops/​webgpu/​bert/​gated_delta_net_state_replay.h	Declares the kernel.
onnxruntime/​contrib_ops/​webgpu/​bert/​gated_delta_net_state_replay.cc	Implements validation, dispatch, and synchronization.
docs/​ContribOperators.md	Adds generated operator documentation.

🧠 Review effort: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Baiju Meswani (baijumeswani)]

The replay calculation and buffer checks look sound as a first WebGPU building block. The points below matter before GenAI relies on it for partial acceptance; I am not asking this PR to implement all of the follow-up work.