"""Test native reduction matching for full-width RMSNorm + AdaLN fusion.""" import argparse import copy import json import types from pathlib import Path import mlx.core as mx from experiments.r3.common import paired from experiments.r3.fusions import norm_native_reduce from h3_mlx.benchmark import error from h3_mlx.config import SIGMAS from h3_mlx.dit import DiT from h3_mlx.layout import image_positions, rope_cache from h3_mlx.ops import gated_add, modulate from h3_mlx.runtime import configure_memory, environment def main(a): configure_memory(wired_gb=14) data = mx.load(a.trace) model = DiT(a.model, kernel="metal-exact", tune_matmul=True) model.prepare_modulation(SIGMAS) text = model.refine(data["conditioning"]) x = mx.concatenate( ( text, model._linear("video_patch_proj")(data["initial_rows"]).astype(mx.bfloat16), ), axis=0, ) rope = rope_cache( image_positions(len(text), 128, 128), model._constant("rope.inv_freq") ) indices = mx.concatenate((data["text_tags"], mx.zeros(4096, mx.int32))) params = tuple(model.modulation["block.0"][0, :, j] for j in range(6)) block = model._block("blocks.0") report = {"environment": environment(), "trials": []} ref = modulate(x, block.n1, params[0], params[1], indices, block.config.norm_eps) mx.eval(ref) for threads in (256, 512, 1024): def fn(threads=threads): return norm_native_reduce( x, block.n1, params[0], params[1], indices, block.config.norm_eps, threads, ) def native(): return modulate( x, block.n1, params[0], params[1], indices, block.config.norm_eps ) row = { "group": "norm", "threads": threads, **error(ref, fn()), **paired(native, fn, a.repeats), } report["trials"].append(row) print(json.dumps(row), flush=True) native = mx.compile( lambda x, params, indices, rope: block( x, params=params, indices=indices, rope=rope, mlp_chunk=0, kernel="metal-exact", ) ) reference = native(x, params, indices, rope) mx.eval(reference) def forward(self, xx, *, params, indices, rope, **kwargs): sa, ca, ga, sm, cm, gm = params h = norm_native_reduce(xx, self.n1, sa, ca, indices, self.config.norm_eps) xx = gated_add(xx, self.attn(h, rope, 0, "metal-exact"), ga, indices) h = norm_native_reduce(xx, self.n2, sm, cm, indices, self.config.norm_eps) return gated_add(xx, self.mlp(h, 0), gm, indices) # __call__ is looked up on the class, so use an explicit bound method. fused = copy.copy(block) fused.forward = types.MethodType(forward, fused) candidate = mx.compile( lambda x, params, indices, rope: fused.forward( x, params=params, indices=indices, rope=rope ) ) row = { "group": "block", **error(reference, candidate(x, params, indices, rope)), **paired( lambda: native(x, params, indices, rope), lambda: candidate(x, params, indices, rope), a.repeats, ), } report["trials"].append(row) print(json.dumps(row), flush=True) Path(a.output).write_text(json.dumps(report, indent=2) + "\n") model.clear() if __name__ == "__main__": p = argparse.ArgumentParser(description=__doc__) p.add_argument("--model", required=True) p.add_argument( "--trace", default="outputs/suite/product/native-3.trace.safetensors" ) p.add_argument("--repeats", type=int, default=10) p.add_argument("--output", default="reports/optimization-r3/norm.json") main(p.parse_args())