MLX mode never reaches the mx.addmm dispatches because BlasOpt is excluded

未关闭
#2,525 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
1-2 天
新手友好度
57/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
python

调研方向

Start by reading MLXLinker and the BlasOpt rewrite rules, then inspect the Gemv and Ger dispatches in link/mlx/dispatch/blas.py. Reproduce the two examples in the issue and inspect the MLX graph to confirm which rewrites run. Done when both expressions reach the intended mx.addmm dispatches and the outer product is handled in MLX mode.

由索引模型根据 Issue 内容生成。

描述

mlx performance

MLXLinker lists BlasOpt among its incompatible rewrites, so Gemv and Ger never appear in an MLX graph, and their mx.addmm dispatches in link/mlx/dispatch/blas.py are unreachable. A @ v + b runs as a Matmul plus a separate Add, and an outer product col @ row runs as a Matmul. local_dot_to_mul leaves core outer products alone so GER can handle them, but MLX mode never produces a GER. #2514 adds local_outer_dot_to_mul for Numba, which has no GER either, and tagging it mlx would cover the outer product. #2406 turns Gemv and Ger into plain Ops, but MLX still excludes BlasOpt there.

import re

import mlx.core as mx
import numpy as np
import pytensor
import pytensor.tensor as pt

A = pt.matrix("A", dtype="float32")
v = pt.vector("v", dtype="float32")
b = pt.vector("b", dtype="float32")
col = pt.matrix("col", shape=(None, 1), dtype="float32")
row = pt.matrix("row", shape=(1, None), dtype="float32")

rng = np.random.default_rng(0)
cases = {
    "A @ v + b": ([A, v, b], A @ v + b, [rng.normal(size=(5, 3)), rng.normal(size=3), rng.normal(size=5)]),
    "col @ row": ([col, row], col @ row, [rng.normal(size=(5, 1)), rng.normal(size=(1, 3))]),
}
for name, (inputs, output, values) in cases.items():
    fn = pytensor.function(inputs, output, mode="MLX")
    mx.export_to_dot("graph.dot", fn.vm.jit_fn(*[mx.array(value.astype("float32")) for value in values]))
    print(name, [type(node.op).__name__ for node in fn.maker.fgraph.toposort()], re.findall(r'label ="([^"]+)"', open("graph.dot").read()))
# A @ v + b ['DimShuffle', 'Dot', 'DimShuffle', 'Elemwise'] ['Transpose', 'Reshape', 'Matmul', 'Transpose', 'Reshape', 'Add']
# col @ row ['Dot'] ['Matmul']
主要语言
Python
星标
654
派生
258
平均合并
2 天 7 小时
30 天内合并 PR
39

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

pymc-devs/pytensor 的其他 Issue

查看 pymc-devs/pytensor 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。