Float32 fidelity: which float64 lectures need float64, and can the heaviest step drop to float32?

未关闭
#11 2 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
5/5
预计耗时
一周以上
新手友好度
25/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
python

调研方向

Start with lectures/run_lecture.py and the lecture implementations listed in the issue; compare each lecture's recorded answers and convergence behavior under --dtype float32 and --dtype float64. Then examine aiyagari_jax's get_greedy for the mixed-precision variant, measuring policy changes, runtime, and equilibrium capital across the stated grids and devices. Done means producing the per-lecture and mixed-precision result tables, saving results under lectures/results/<host>/<date>/, and recording findings in the benchmarks repository.

由索引模型根据 Issue 内容生成。

描述

Run the 13 float64 lectures in lecture-jax in float32 and record whether each still reaches its answer, then try a mixed-precision version of aiyagari_jax's greedy step. The aim is to know where QuantEcon's JAX code needs float64, and where it is a precaution.

Why

  • The lectures don't say what float64 protects. 13 of the 29 lectures in lecture-jax turn it on, each with the same sentence, "to increase the precision". None says what goes wrong without it.
  • #3 found that float32 fails aiyagari_jax in two ways, and neither is a slightly worse answer.
    • Two of its 50 solves stall at an error of about 2.5e-4, against the lecture's tolerance of 1e-4, and run to the 10,000-iteration cap. That makes the lecture about 12× slower as used.
    • On the TITAN RTX at a 5,000-point grid, its values went NaN in all three runs, from a bicgstab breakdown (QuantEcon/project-benchmarks-jax#5). The loop's test error > tol is false for NaN, so it stopped after two iterations with a policy wrong almost everywhere, and raised no error. Both CPUs stayed finite.
    • When float32 did converge, equilibrium capital matched float64 to 2e-5.
  • ifp_egm_jax converges in float32, to within 1.6e-4 of the float64 policy.
  • So float32 fails by not finishing, or by failing silently, and the failure can depend on the device. Two lectures don't say which of the other eleven share that.
  • #4 runs these lectures in float32 for time. This issue asks whether the answer survives.
  • The step that float64 makes expensive may not need it. In dense discrete choice the costly step recomputes rewards and takes the maximum over choices: 26.8× slower in float64 than in float32 on the TITAN RTX (#6). That step only has to pick the best choice. The value update, the convergence test, policy evaluation and the equilibrium search could stay in float64.

Part 1: the 13 lectures in float32

Lecture Solver and tolerance, as written
aiyagari_jax Policy iteration, 1e-4; equilibrium capital by bisection, xtol 1e-4
arellano_jax Joint iteration on the repay and default value functions, 1e-8
hopenhayn Value function iteration, 1e-6; the entry price by bisection, 1e-4
ifp_egm_jax Endogenous grid method, 1e-5
inventory_ssd Value function iteration, 1e-6
jax_intro_jax None; it turns float64 on to show the difference
job_search Value function iteration, 1e-4
markov_asset_jax A linear solve for the price-dividend ratio, after a spectral-radius check
mle_jax Newton–Raphson, 1e-3 and 1e-5
newtons_method Newton's method, 1e-5; float64 is turned on in exercise 1's solution, for tolerances of 1e-7 and 1e-15
opt_invest Value function and optimistic policy iteration, 1e-5
opt_savings_1 Value function iteration, 1e-5
opt_savings_2 Value function and optimistic policy iteration, 1e-5

Also the two float32 lectures that solve to a tolerance, overborrowing (1e-5) and lucas_model_jax (1e-6): whether they meet it, or stop at their iteration cap.

Method:

  • lectures/run_lecture.py with --dtype float32 and --dtype float64, at the lecture-jax commit #4 used. Each of the 13 has exactly one jax_enable_x64 line, as --dtype float32 requires.
  • --record each lecture's answer: the equilibrium values, prices, estimates or policies it reports. Choosing the expressions lecture by lecture is part of the work.
  • For each solve: whether it met its tolerance or stopped at its cap, and how many iterations it took. Where a solver doesn't return these, read them from its printed output.
  • Check every recorded value for NaN and Inf. Don't rely on a loop's exit: a NaN error ends while error > tol as if it had converged.
  • One run per configuration, and three where results vary between runs.

Platforms:

Device Why
M6 CPU, qe-mini1, with its agents paused Seconds per lecture; the CPU answer
TITAN RTX, qe-wharton1-gpu, card 0b:00.0 The silent NaN appeared only on the GPU
T4, RunsOn g4dn.2xlarge The lectures' CI device, if it is cheap to add to #4's job

Result: one table, lecture by device, giving whether float32 converges (yes, stops at its cap, NaN, or raises), the largest gap from float64 in the recorded answers, and whether that gap is within the lecture's own tolerance.

Part 2: a mixed-precision greedy step for aiyagari_jax

  • The variant: with float64 on, get_greedy casts to float32, computes rewards and the argmax over next-period assets, and returns the policy. The value function, get_value's bicgstab solve, the convergence test and the equilibrium bisection stay in float64.
  • Then one float64 greedy step at the converged solution, counting the policy points that change. That catches choices decided by a near-tie.
  • At the grids #3 used, 200 to 10,000 asset points, on the TITAN RTX and the M6.
  • Record: time per policy-iteration loop, against pure float64 and pure float32; the policy points that differ from float64, and by how many grid steps; the gap in equilibrium capital; iterations to converge.
  • If it is ready in time, add it to #1's Phase 2 runs. A Blackwell workstation card does float64 at 1:64 of float32, so that is where the float32 step would matter most.

What the results would change

If Then
A float64 lecture reaches its answer in float32, within its tolerance, on every device Float64 is a precaution there. The lecture can say so, or say what it guards against
A lecture stalls, goes NaN, or moves its answer past its tolerance in float32 It needs float64 as written, and the lecture can say why
The mixed step gives float64's answer at a fraction of float64's cost Research-scale discrete-choice models can run their heavy step at a card's float32 rate. That changes which card suits them, which project-compute weighs
It doesn't Dense discrete choice needs float64 throughout, and a card's float64 rate decides

Deliverables

  • the results, under lectures/results/<host>/<date>/ as #4's are;
  • the per-lecture table, and Part 2's table of time and gaps;
  • a finding in QuantEcon/project-benchmarks-jax's findings/, with its hardware consequence going to project-compute.

Related

  • #1, the precision experiment, with #3 and #6, which measured aiyagari_jax's float32 failures and float64's cost by kernel type
  • #4, the whole-lecture run, which times these lectures in float32
  • QuantEcon/project-benchmarks-jax#5, the bicgstab breakdown

🤖 Generated with Claude Code

主要语言
Python
星标
0
派生
1
平均合并
3 小时 29 分钟
30 天内合并 PR
64

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

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。

QuantEcon/benchmarks 的其他 Issue

查看 QuantEcon/benchmarks 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。