[amd-nprotaso]

Adds a FlyDSL kernel for the Qwen3.8 hyper-connection mix projection (_hc_project) on gfx950: BF16 input, decode-sized M (up to 128 rows).

Changes

• aiter.ops.flydsl.hc_mix: down projection, then split-K reduce + SiLU, then up projection with gated HC reduction. Math runs in FP32; only the SiLU intermediate and the output are rounded to BF16.
• pack_hc_weights: packs the weights once into the MFMA operand layout.

Performance

Setup:

• MI355X (gfx950), torch 2.11.0+rocm7.2, FlyDSL 0.3.4.1, BF16.
• x [M, 10240], down [320, 10240], up [10240, 320].
• 64 weight sets are rotated between calls, so the weights are cold in L2, as in a real layer stack.
• Each measurement is a CUDA-graph replay; the time is the per-call median.
• Script: benchmark/kernels/bench_hc_mix_flydsl.py in the SGLang PR.

M	FlyDSL hc_mix (µs)	SGLang persistent Triton fused_hc_mix (µs)	torch.compile (µs)	vs Triton	vs torch.compile
1	18.1	36.3	34.3	2.00x	1.90x
2	18.8	44.9	38.1	2.38x	2.02x
4	20.3	44.9	38.1	2.21x	1.88x
8	22.1	45.7	38.2	2.07x	1.73x
16	26.3	45.5	38.6	1.73x	1.47x
geomean				2.07x	1.79x

The weights alone are 2 × 320 × 10240 × 2 B = 13.1 MB. At M=1 that is about 0.72 TB/s effective, against the Triton kernel's 0.36 TB/s.

Accuracy

Against the float64 reference: atol=4e-3, rtol=1e-2, and atol=5e-3 in the gate-range test.

Tests

$ python -m pytest -q op_tests/test_flydsl_hc_mix.py
25 passed

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:gfx1250-ffm-triton	Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x	Run an additional Triton test job on MI300X in PRs (added automatically when gfx942 configs change); main branch always runs both MI35X and MI300X
ci:triton-355	Run the full Triton test suite on MI35X, not only the tests the change affects
multigpu	Aiter multi-GPU tests on the 8-GPU runner
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 5993 --add-label <label>

One backend per PR:
A PR changes one kernel backend: [Triton/Gluon] (Triton and Gluon count as one), [HIP], [ASM], [CK], [OPUS] or [FlyDSL]. If the title ends up with two backend tags, split the PR -- as stacked pull requests when one part cannot merge without the other.

PR title tags & labels:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title and as PR labels automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf], op tags like [MLA], and human labels (ci:*) are left untouched. Add the no-auto-title label to stop the title rewrites; labels stay in sync either way.