[NPU] Add causal conv1d

[zhaozx-cn]

Motivation

apply custom causal_conv1d on npu, sgl-project/sgl-kernel-npu#555 is required

Modifications

apply custom causal_conv1d on npu.

Accuracy Tests

Model	Task	Score	Official
Qwen/Qwen3.5-35B-A3B	CEVAL	0.9057	0.902
Qwen/Qwen3.5-397B-A17B	CEVAL	0.922	0.93

Speed Tests and Profiling

Model: Qwen/Qwen3.5-35B-A3B

Stage	Input / Output	Batch Size	Metric	Before this PR	After this PR
Prefill	8k / 1	bs1	TTFT (ms)	531	446
Decode	8k / 2k	bs48	TPOT (ms)	37.81	33.49
MTP	8k / 2k	bs48	TPOT (ms)	16.30	16.08

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ❌ Run #29193910582
Latest PR Test (Extra): ❌ Run #29193910464

[gemini-code-assist]

Code Review

This pull request refactors the Ascend GDN backend to use torch.ops.npu.causal_conv1d for both decode and extend forward passes, and introduces a device_print utility in utils.py for debugging on NPU. Key feedback includes caching transposed and contiguous convolution weights on the layer object to avoid performance overhead in hot paths, correcting the side-effect registration in _mark_op_side_effectful by adding ops to torch.fx.node._side_effectful_ops instead of calling has_side_effect, and removing unused imports (dataclass and List).

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Calling layer.conv_weights.transpose(0, 1).contiguous() on every forward pass introduces unnecessary overhead. Cache the transposed and contiguous weights on the layer object.

Suggested change

	mixed_qkv = torch.ops.npu.causal_conv1d(
	mixed_qkv,
	layer.conv_weights.transpose(0, 1).contiguous(),
	conv_states=conv_states,
	if not hasattr(layer, "_conv_weights_npu_cache"):
	layer._conv_weights_npu_cache = layer.conv_weights.transpose(0, 1).contiguous()
	mixed_qkv = torch.ops.npu.causal_conv1d(
	mixed_qkv,
	layer._conv_weights_npu_cache,
	conv_states=conv_states,

[gemini-code-assist]

Calling layer.conv_weights.transpose(0, 1).contiguous() on every forward pass introduces unnecessary overhead. Cache the transposed and contiguous weights on the layer object.

Suggested change

	mixed_qkv = torch.ops.npu.causal_conv1d(
	mixed_qkv,
	layer.conv_weights,
	layer.bias,
	activation=layer.activation,
	conv_states=conv_states_tmp,
	has_initial_state=has_initial_states,
	cache_indices=cache_indices,
	layer.conv_weights.transpose(0, 1).contiguous(),
	conv_states=conv_states_for_prefill,
	if not hasattr(layer, "_conv_weights_npu_cache"):
	layer._conv_weights_npu_cache = layer.conv_weights.transpose(0, 1).contiguous()
	mixed_qkv = torch.ops.npu.causal_conv1d(
	mixed_qkv,
	layer._conv_weights_npu_cache,
	conv_states=conv_states_for_prefill,