[eble-amd]

Purpose

Merge c01368f from upstream and adapt local Hybrid W4A16 MoE to the new N-first weight layout.

With the upstream change alone, compressed-tensors W4A16 MoE models crashed on gfx11 at the first forward pass with HSA_STATUS_ERROR_MEMORY_FAULT. The merge from upstream brought in vllm-project#52798, which makes CompressedTensorsWNA16MoEMethod always load weights N-first ([E, N, K/8], scales [E, N, K/G]), but the gfx11 hybrid W4A16 path still converted them as K-first.

The hybrid path now unpacks along K without transposing before the ExLlama shuffle, and no longer transposes the scales, which already have the layout HybridW4A16MoEExperts expects.

Assisted-by: LLM

Test Plan & Results (agentic)

gfx1151 (Strix Halo), ROCm 7.17.26431, PyTorch 2.13.0+rocm10.3.0a20261027, Triton 3.8.0.
Base: 49c3a2e2ae05 (before the merge). PR: merge 55a5667e9aa4 with this fix.

Without the fix, the merge crashes with HSA_STATUS_ERROR_MEMORY_FAULT on the first forward pass of compressed-tensors W4A16 MoE models (e.g. cyankiwi/Qwen3-30B-A3B-Instruct-2507-AWQ-4bit). With it, all cases below pass, with identical token counts.

Case	TTFT ms base	TTFT ms PR	TTFT Δ	TPOT ms base	TPOT ms PR	TPOT Δ
Qwen3-30B-A3B-Instruct-2507-AWQ-4bit_128	184.0	185.3	+0.7%	11.82	11.82	0.0%
Qwen3-30B-A3B-Instruct-2507-AWQ-4bit_specdec-eagle3-aime	222.2	221.2	−0.4%	13.89	13.90	+0.1%
Qwen3.6-35B-A3B-W4A16-llmcompressor_VLM_4ksys_1k_320x240_100out_dyn-lm-int8g32	502.8	504.9	+0.4%	12.08	12.15	+0.6%
Qwen3.6-35B-A3B-W4A16-llmcompressor-mtpfix_VLM_dyn-lm-int8g32_MTP-mmstar ¹	188.1	188.0	−0.1%	9.81	9.92	+1.1%

¹ Median of 5 runs per side. This case samples at temperature 1.0, so TPOT varies with MTP acceptance length (base 1.861–1.890, PR 1.861–1.888).

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[roberteg16]

LGTM