[inureyes]

Summary

• Resolve Qwen3 and Qwen3-MoE rope_scaling once per model and propagate linear position scales or Llama 3 frequency tables to every attention layer.
• Keep linear scaling on the fused Q/K-normalization decode path, route frequency-table schemes through the graph path, and apply the same plan to dense batched decoding.
• Preserve the plan through direct block construction, pipeline stages, and tensor-parallel rank-local models.
• Make malformed or unsupported shared RoPE schemes observable through a checkpoint-keyed, once-per-model-and-scheme warning instead of a silent unscaled decode.
• Retain the existing test boundary: large checkpoint qualification uses weights already present on local accelerator hardware, while hosted CI remains limited to ordinary and small-fixture coverage.

Validation

• cargo fmt --all -- --check
• cargo test -p mlxcel --profile test-fast --lib qwen3 -- --nocapture on integrated latest main (110 passed, 20 existing hardware/large-model tests ignored)
• Focused fused/graph RoPE scale, malformed-scheme warning, batched, direct-loader, pipeline-parallel, and tensor-parallel tests
• cargo clippy -p mlxcel --all-targets -- -D warnings
• CUDA release build on NVIDIA GB10: MLX_CUDA_ARCHITECTURES=121 cargo build --release --features cuda --bin mlxcel
• Existing local models/mlx/qwen3-4b-4bit: PR and latest main matched for 128 greedy tokens
• Existing local models/mlx/qwen3-30b-a3b-4bit: PR and latest main matched for 128 greedy tokens
• Injected linear factor 8 into a temporary hard-linked view of the local 4B checkpoint; a 3,009-token prompt matched mlx-lm 0.31.3 on the graph RoPE path (The quick. The quick brown fox.), while latest main diverged because it ignored the block
• Temporary large-checkpoint Actions run cancelled and workflow removed; no 4B/30B download is part of the final CI change

Notes

The dense implementation follows upstream mlx-lm's scaling_config path. Qwen3-MoE intentionally applies the same config even though current mlx-lm still constructs a plain RoPE module there, matching the issue's documented compatibility decision.

The default fused decode path shares the graph path's scale and passes direct numerical parity tests. Cross-runtime generated text can diverge after a shared prefix through the issue's documented f16 reduction-order jitter, so the real-model oracle comparison also selected the graph path used by upstream.

Bilingual qualification details are recorded in TECHNICAL_REPORTS/1398-qwen3-rope-scaling-20260824.{en,ko}.md.

Closes #1388