Expose gdn::causal_conv1d and gdn::gated_delta_rule as standalone torch ops with unit tests

[Copilot]

gdn::causal_conv1d and gdn::gated_delta_rule were only accessible as internal implementation details of gdn_attention, with no way to call them independently from Python.

Changes

C++ – new wrapper functions (gdn_attn_interface.cpp, ops.h, torch_bindings.cpp)

• Added causal_conv1d and gated_delta_rule C++ wrapper functions that forward directly to gdn::causal_conv1d / gdn::gated_delta_rule
• Registered both ops under torch.ops._xpu_C via TORCH_LIBRARY_EXPAND, following the same pattern as gdn_attention

Python – unit tests (tests/gdn_attn/)

• test_causal_conv1d.py: reference impl (ref_causal_conv1d) + parametrised test covering prefill/decode/mix modes, bias/no-bias, reorder_input=True/False; validates all six output tensors (q, k, v, z, b, a) and updated conv_states
• test_gated_delta_rule.py: reference impl (ref_gated_delta_rule) + parametrised test covering all modes; validates core_attn_out and updated ssm_state

Usage

import vllm_xpu_kernels._xpu_C  # noqa: F401

# causal_conv1d – splits and convolves mixed_qkvz, fills q/k/v/z/b/a
torch.ops._xpu_C.causal_conv1d(
    q_out, k_out, v_out, z_out, b_out, a_out,
    mixed_qkvz, mixed_ba, conv_weights, conv_bias, conv_states,
    query_start_loc, cache_indices, has_initial_state,
    activation, num_prefills, num_decodes, reorder_input)

# gated_delta_rule – runs the per-token gated delta-rule state update
torch.ops._xpu_C.gated_delta_rule(
    core_attn_out, q, k, v, b, a, A_log, dt_bias, ssm_state,
    query_start_loc, cache_indices, has_initial_state,
    num_prefills, num_decodes)

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.