Allocate fused MoE bias parameters in the locality-domain device context #18336

Summary

FusedMoEMethodBase.create_weights selects device=torch.device("cuda") when a locality-domain weight context is active. It applies this device to w3_w1_weight and w2_weight, but not to w3_w1_bias and w2_bias.

Required change

Update tensorrt_llm/_torch/modules/fused_moe/quantization.py so both bias torch.empty(...) allocations in FusedMoEMethodBase.create_weights receive device=device.

Rationale

When module.bias is enabled in partition_weight_context(), the current code can allocate weights on CUDA while bias parameters use the ambient device. This can place related parameters on different devices and prevents the biases from using the locality-domain memory pool.

Acceptance criteria

• w3_w1_bias and w2_bias use the same device variable as w3_w1_weight and w2_weight.
• The behavior is unchanged when module.bias is disabled.
• The non-locality-domain path continues to use the existing default allocation behavior.

Backlinks

• PR: [None][feat] Add Rubin SM107 CuTe DSL foundation and BF16 kernels #18311
• Review discussion: [None][feat] Add Rubin SM107 CuTe DSL foundation and BF16 kernels #18311 (comment)

Requested by: @farazkh80