Refresh vllm_for_multi_arc.patch from qwen3_next_fix_0324

[gc-fu]

Summary

• Regenerate vllm/patches/vllm_for_multi_arc.patch from intel-sandbox/llm-scaler-vllm-xpu branch qwen3_next_fix_0324 (commit 1623c86) against upstream vllm v0.14.0.
• Patch diff vs previous: +1087 / -340 lines (total 17,510 lines / 712K).
• Verified with git apply --check against a fresh upstream v0.14.0 checkout.

What's new in this refresh

• Qwen3.5 / Qwen3-Next: BSZ>1 decode paths, moe_forward_full(_v2) single-dispatch, ESIMD RMSNormGated and batched fused_add_rms_norm, cached decode tensor views, _prepare_inputs_decode_fast for BSZ=1.
• INT4 ESIMD kernels: MoE integration for Qwen3.5 MoE models (incl. transpose=False weight layout), GEMV for decode path, norm_gemv / resadd_norm_gemv.
• GDN: rollback BSZ>1 precomputed-proj path to standard gdn(), v=48 fix, missing BSZ>1 buffer/int4-flag init in Qwen3_5DecoderLayer.
• Paged attention: GQA ratio pad support for non-4-divisible ratios; gate page_attn_decode on gqaRatio % 4 == 0 to prevent kernel assertion.
• Misc: skip sym_int4 quantization for shared_expert (accuracy), cos_sin_cache + fp16 cache conversion, unified kernel package imports for moe_ops / eagle_ops.

Test plan

• git apply --check vllm_for_multi_arc.patch against upstream v0.14.0 — passes.
• Docker image built successfully: amr-registry.caas.intel.com/intelanalytics/llm-scaler-vllm:0.14.0-b8.3-0427 (34.5GB).
• Smoke / accuracy runs on Qwen3.5 models (to follow).

🤖 Generated with Claude Code