Port Qwen row-map SwiGLU expert gathers

tools/issue112-extract-metal.py

@@ -0,0 +1,19 @@
+"""Regenerate only the vendored GPU shader closure from the pinned oracle."""
+import subprocess,sys
+from pathlib import Path
+checkout=Path(sys.argv[1])
+assert subprocess.check_output(['git','-C',str(checkout),'rev-parse','HEAD'],text=True).strip()=='9882703f3105363ddc37eca9f97aa09a1d387112'
+sys.path.insert(0,str(checkout))
+import mlx.core as mx
+from mtplx.kernels import moe_sorted_gather_nax as k
+assert mx.__version__=='0.32.2'
+root=k._include_root();seen=set();out=[]
+for rel in (k._PREAMBLE_HEADER,*k._KERNEL_HEADERS):k._inline(root,rel,seen,out)
+for rel,name in [('mlx/backend/metal/kernels/unary_ops.h','Sigmoid'),('mlx/backend/metal/kernels/binary_ops.h','Multiply')]:
+ functor=k._functor(root,rel,name)
+ assert functor is not None
+ out.append(functor)
+metal=Path(__file__).resolve().parents[1]/'metal'
+preamble='// MLX 0.32.2 GPU support, Copyright Apple Inc., MIT; see MLX-LICENSE.txt.\n// Flattened only the GPU header closure used by MTPLX 9882703 row-map gather.\n'
+(metal/'mtplx-moe-sorted-header.metal').write_text(preamble+'\n'.join(out)+'\nusing namespace mlx::steel;\n')
+(metal/'mtplx-moe-sorted-body.metal').write_text('// MTPLX 9882703, MIT; see MTPLX-LICENSE.txt.\n'+k._SOURCE)