Extract optimized causal conv1d prefill algorithm from Qwen3.5 into standalone function + tests

[Copilot]

The optimized prefill-path depth-wise causal conv1d in qwen3_5.py (lines 377–412) was inlined inside the model's forward() with no reusable interface or test coverage.

Changes

• vllm/model_executor/models/qwen3_5_conv1d.py — new standalone function optimized_causal_conv1d_prefill

  • Accepts x: [bs, seq_len, dim], preprocessed weight: [kernel_size, dim], optional bias: [dim]
  • Implements the loop-based in-place accumulation strategy (avoids large intermediate tensors from F.conv1d), then applies SiLU
  • Input validation with descriptive errors

• tests/kernels/test_qwen3_5_conv1d.py — test suite covering:

  • Numerical correctness vs. nn.Conv1d reference across shapes, kernel sizes, dtypes, and bias combinations
  • Causality: future token perturbations must not affect past outputs
  • Degenerate cases: kernel_size=1 (pointwise), zero weights, single-token sequence
  • Shape invariance
  • Validation error paths

Usage

from vllm.model_executor.models.qwen3_5_conv1d import optimized_causal_conv1d_prefill

# weight in preprocessed [kernel_size, dim] form:
# conv1d.weight.squeeze(1).transpose(0,1).flatten().reshape(kernel_size, dim)
out = optimized_causal_conv1d_prefill(x, weight, bias=bias, activation="silu")
# out: [bs, seq_len, dim]