Fix WebGPU inference crash in embedding and multi-modal feature allocation

[feich-ms]

Summary

• Fix heap corruption crash during WebGPU inference by using the correct allocator (p_device_inputs_) for tensors that serve as inputs to the embedding/decoder session
• Previously, embeddings.cpp and multi_modal_features.cpp allocated these tensors on p_device_ (WebGPU GPU memory), but the embedding session runs on CPU for WebGPU (without graph capture), causing the CPU session to write into GPU memory.
• This is a pair change for this pr [WebGPU] Support models with opset 24 ops and KV-shared decoder layers (Gemma 4) onnxruntime#28501.

Changes

src/models/embeddings.cpp (lines 24, 52):

• Changed model_.p_device_->GetAllocator() → model_.p_device_inputs_->GetAllocator() for embedding input tensor allocation and reallocation during sequence length updates

src/models/multi_modal_features.cpp (lines 66, 85):

• Changed model_.p_device_->GetAllocator() → model_.p_device_inputs_->GetAllocator() for empty feature tensors in Update() and AllocateEmptyFeatures() — these are inputs to the embedding session
• Vision/speech model output allocations (lines 42, 98) correctly remain on p_device_

Impact

• WebGPU: Fixes crash (heap corruption 0xc0000374) during inference
• CUDA/DML/RyzenAI: No-op — p_device_inputs_ == p_device_ on these providers
• CPU: No-op — p_device_inputs_ == p_device_ on CPU

Test plan

• End-to-end WebGPU inference with Gemma 4 E2B (109 tokens generated correctly)
• CPU inference still works after changes
• Output matches between CPU and WebGPU execution

[Copilot]

Pull request overview

Fixes a WebGPU-specific heap corruption/crash by ensuring tensors that are fed into the embedding/decoder path are allocated using the model’s “inputs device” allocator (p_device_inputs_) rather than the main device allocator (p_device_). This aligns embedding-related buffers with the codebase’s existing device split (notably WebGPU without graph-capture using CPU memory for inputs).

Changes:

• Allocate embedding input tensors (Embeddings::Mode::Input) using model_.p_device_inputs_->GetAllocator() (constructor + sequence-length reallocation).
• Allocate empty multi-modal feature tensors (used as embedding inputs) using model_.p_device_inputs_->GetAllocator() in both Update() and AllocateEmptyFeatures().
• Leave vision/speech output allocations on p_device_ unchanged.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/models/embeddings.cpp	Switches embedding input tensor allocation/reallocation to the inputs allocator to avoid CPU writing into WebGPU-allocated buffers.
src/models/multi_modal_features.cpp	Switches empty feature-tensor allocation (embedding inputs) to the inputs allocator to match embedding session execution/memory expectations.