Add Strix Halo (gfx1151) Atlas Inference Hub-kernel path for Qwen3.5/3.6/3.8 Gated DeltaNet

[AzeezIsh]

On AMD Strix Halo (Ryzen AI Max, gfx1151, RDNA 3.5) the fla and causal_conv1d fast paths have no ROCm build, so Qwen3.5-family models (Qwen3.5, Qwen3.6-27B, Qwen3.8-27B) fall back to the slow pure-torch Gated DeltaNet (is_fast_path_available=False). This adds a rocm device entry, gated to that GPU, to the existing Qwen3_5GatedDeltaNet mapping from #46423, pointing at the ROCm build of the same Hub kernel (Atlas-Inference/gdn). Strix Halo users get the fast path automatically; every other GPU, including other ROCm GPUs, keeps its existing path.

• One Device(type="rocm", properties=ROCMProperties(min_capability=115, max_capability=115)) entry next to the GB10 one in integrations/hub_kernels.py, same layer name and repo, revision pinned.
• A usage note with measured numbers in docs/source/en/model_doc/qwen3_5.md.

The kernel is the same stateless kernels layer as the GB10 build. it reuses the host module's projections, gated RMSNorm and out_proj, and replaces only the conv1d, q/k L2-norm and delta-rule cores. On gfx1151 it runs the kernels the Atlas engine serves Qwen3.6/3.8-27B with on Strix Halo.

• prefill: causal_conv1d_update_prefill → l2_norm_bf16 → FLA-chunked delta rule (recompute_wu → chunk_delta_h → chunk_fwd_o)
• decode: causal_conv1d_update_l2norm_f32 → gated_delta_rule_decode_f32

What does this PR do?

The gated rocm entry in integrations/hub_kernels.py plus the docs note. @use_kernel_forward_from_hub("Qwen3_5GatedDeltaNet") decorators from #46423 already cover the dense and MoE classes.

Verified on Strix Halo (gfx1151, torch.cuda.get_device_capability() == (11, 5)), bf16, torch 2.14.0 ROCm 7.2 and ROCm 7.14 wheels, transformers main + kernels 0.17:

• Parity against a float64 torch reference for every op, including ragged prefill lengths (1, 33, 65, 100, 257, 1024): 23/23 tests on both wheels.
• kernelize() swaps the forward of all 24/24 GDN layers of Qwen/Qwen3.5-9B, and the profiler shows the kernel ops running.
• End-user path: transformers main + this PR, from_pretrained(..., use_kernels=True) with an empty kernel cache pulls Atlas-Inference/gdn at the pinned revision. All 24/24 GDN layers of Qwen/Qwen3.5-9B run it, and greedy generate() produces identical token IDs

Measured through the end-user path (from_pretrained(..., use_kernels=True), kernel pulled from the Hub) on Qwen/Qwen3.5-9B (bf16, gfx1151, torch 2.14 + ROCm 7.2, 1024-token prompt, greedy decode of 256 tokens; mean of two runs):

use_kernels	TTFT (prefill)	Decode
False (PyTorch fallback)	1.33 s	5.77 tok/s
True (Atlas-Inference/gdn)	1.16 s (1.15x faster)	6.23 tok/s

About 74% of this prefill is the dense projection GEMMs, which the kernel doesn't touch. Per Gated DeltaNet layer of Qwen/Qwen3.8-27B (real weights), the kernel is 1.65x faster on a 1024-token prefill and 1.26x faster per decode step. The published build targets torch 2.14 with ROCm 7.2.

Follow-up to #46423 (the GB10/SM121 entry). The ROCm build needs kernel-builder with gfx1151 support, which landed in huggingface/kernels#839 (tracked in huggingface/kernels#834). No overlapping PR exists.

The same trust note applies as for the GB10 entry: the mapped kernel needs trust_remote_code=True until Atlas-Inference is on the kernels trusted-publisher allowlist, would love to enable this for continued future contributions without any blocks!

Code Agent Policy

• I confirm that this is not a pure code agent PR.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Follow-up to Add GB10/SM121 Hub-kernel path for Qwen3.6 Gated DeltaNet #46423; gfx1151 builder support via Kernel-builder drops for Strix Halo architecture kernels#834 / How to restore a training? #839.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests? (Parity and generate tests were run on gfx1151 hardware; in-repo unit tests aren't included since the path needs a real gfx1151 GPU and a Hub fetch.)

Who can review?

@vasqu original PR merger, @danieldk first issue
Potentially @drbh (kernels integration) and @Cyrilvallez (text models / Qwen3.5) as well :)
If you could enable myself on the kernels trusted-publisher allowlist that would help for the future!

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 36276919751:2
Result: failure | Jobs: 2 | Tests: 11 | Failures: 1 | Duration: 59s

Code quality check failed: test jobs were skipped. Fix the code quality issues and push again to run tests.

[Rocketknight1]

cc @Abdennacer-Badaoui