add gdn custom conv1d

[zhaozx-cn]

Description

This PR adds GDN (Grouped Deformable Normalization) custom conv1d op, which could be used in:

• Prefill
• Decode
• MTP (Multi-Token Prediction)

---

Accuracy Test

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

---

Performance Test

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

Model: Qwen/Qwen3.5-397B-A17B

Stage	Input / Output	Batch Size	Metric	Before this PR	After this PR
MTP	3k / 1.5k	bs54	output throughput (tokens)	5015	5579

---

[gemini-code-assist]

Code Review

This pull request refactors the causal_conv1d operator to support CANN graph mode, template kernels, and speculative decoding, introducing extensive host-side tiling, validation, and kernel dispatching logic. Feedback on the changes highlights several critical issues: a data type mismatch for num_accepted_tokens between the host-side implementation and the CANN/tiling validation code, thread-safety concerns regarding unprotected global static variables, potential race conditions and cross-device compatibility errors from using static global tensors (globalNumAcceptBuffer and globalTilingBuffer), and the omission of device compatibility checks for input tensors.

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.