pyre-code-ml-practice

name	pyre-code-ml-practice
description	Self-hosted ML coding practice platform with 68 problems covering Transformers, diffusion, RLHF, and more — instant browser feedback, no GPU required.
triggers	["set up pyre code practice platform","add a new ml coding problem to pyre","how do I run the grading service","configure AI help for pyre code","pyre code problem set and learning paths","implement a coding challenge in pyre","troubleshoot pyre code submission grading","deploy pyre code with docker"]

Pyre Code ML Practice Platform

Skill by ara.so — Daily 2026 Skills collection.

Pyre Code is a self-hosted ML coding practice platform with 68 problems ranging from ReLU to flow matching. Users implement internals of modern AI systems (Transformers, vLLM, TRL, diffusion models) in a browser editor with instant pass/fail feedback, no GPU required.

---

Installation

Option A — One-liner (recommended)

git clone https://github.com/whwangovo/pyre-code.git
cd pyre-code
./setup.sh
npm run dev

setup.sh creates a .venv (prefers uv, falls back to python3 -m venv), installs all Python deps, then prints the start command.

Option B — Conda

git clone https://github.com/whwangovo/pyre-code.git
cd pyre-code
conda create -n pyre python=3.11 -y && conda activate pyre
pip install -e ".[dev]"
npm install
npm run dev

Option C — Docker

git clone https://github.com/whwangovo/pyre-code.git
cd pyre-code
docker compose up --build

Progress persists in a Docker volume. Reset with docker compose down -v.

After installation

• Grading service: http://localhost:8000
• Web app: http://localhost:3000

---

Project Structure

pyre/
├── web/                        # Next.js frontend
│   ├── src/app/                # Pages and API routes
│   ├── src/components/         # UI components
│   └── src/lib/                # Utilities, problem data
├── grading_service/            # FastAPI backend (grading API)
├── torch_judge/                # Judge engine — problem definitions + test runner
│   ├── problems/               # Individual problem modules
│   └── runner.py               # Test execution logic
├── setup.sh                    # Environment bootstrap script
├── package.json                # Dev scripts (runs frontend + backend concurrently)
└── pyproject.toml              # Python package config

---

Key Commands

# Start both frontend and backend concurrently
npm run dev

# Start only the grading service (FastAPI)
cd grading_service && uvicorn main:app --reload --port 8000

# Start only the frontend (Next.js)
cd web && npm run dev

# Run Python tests
pytest torch_judge/

# Install Python package in editable mode with dev deps
pip install -e ".[dev]"

# Docker: build and start
docker compose up --build

# Docker: stop and remove volumes (reset progress)
docker compose down -v

---

Configuration

Environment Variables

Create web/.env.local to override defaults:

# URL of the FastAPI grading service
GRADING_SERVICE_URL=http://localhost:8000

# SQLite database path for progress tracking
DB_PATH=./data/pyre.db

AI Help (Optional)

Copy web/.env.example to web/.env and configure:

AI_HELP_BASE_URL=https://api.openai.com/v1
AI_HELP_API_KEY=$OPENAI_API_KEY
AI_HELP_MODEL=gpt-4o-mini

Any OpenAI-compatible endpoint works: OpenAI, Anthropic via proxy, Ollama, etc. Users can also set their own key in the UI if no server-side config is present.

---

Problem Categories

Category	Examples
Fundamentals	ReLU, Softmax, GELU, SwiGLU, Dropout, Embedding, Linear, Kaiming Init
Normalization	LayerNorm, BatchNorm, RMSNorm
Attention	Scaled Dot-Product, Multi-Head, Causal, GQA, Flash, Differential, MLA
Position Encoding	Sinusoidal PE, RoPE, ALiBi, NTK-aware RoPE
Architecture	GPT-2 Block, ViT Block, Conv2D, MoE, Depthwise Conv
Training	Adam, Cosine LR, Gradient Clipping, Mixed Precision, Activation Checkpointing
Distributed	Tensor Parallel, FSDP, Ring Attention
Inference	KV Cache, Top-k Sampling, Beam Search, Speculative Decoding, Paged Attention
Alignment	DPO, GRPO, PPO, Reward Model
Diffusion	Noise Schedule, DDIM Step, Flow Matching, adaLN-Zero
Adaptation	LoRA, QLoRA
Reasoning	MCTS, Multi-Token Prediction
SSM	Mamba SSM

---

Adding a New Problem

Problems live in torch_judge/problems/. Each problem is a Python module with a standard structure:

# torch_judge/problems/my_new_problem.py

import torch
import torch.nn as nn
from typing import Any

PROBLEM_ID = "my_new_problem"
TITLE = "My New Problem: Implement Foo"
DIFFICULTY = "medium"  # "easy" | "medium" | "hard"
CATEGORY = "Fundamentals"

DESCRIPTION = """
## My New Problem

Implement the `foo` function that does XYZ.

### Input
- `x` (Tensor): shape `(batch, dim)`

### Output
- Tensor of shape `(batch, dim)`

### Formula
$$\\text{foo}(x) = x^2 + 1$$
"""

STARTER_CODE = """
import torch

def foo(x: torch.Tensor) -> torch.Tensor:
    # Your implementation here
    pass
"""

REFERENCE_SOLUTION = """
import torch

def foo(x: torch.Tensor) -> torch.Tensor:
    return x ** 2 + 1
"""

def make_test_cases() -> list[dict[str, Any]]:
    \"\"\"Return a list of test cases, each with inputs and expected outputs.\"\"\"
    cases = []
    
    # Basic case
    x = torch.tensor([[1.0, 2.0, 3.0]])
    cases.append({
        "input": {"x": x},
        "expected": x ** 2 + 1,
        "description": "Basic 1x3 tensor",
    })
    
    # Batch case
    x = torch.randn(4, 16)
    cases.append({
        "input": {"x": x},
        "expected": x ** 2 + 1,
        "description": "Batch of 4, dim 16",
    })
    
    # Edge case: zeros
    x = torch.zeros(2, 8)
    cases.append({
        "input": {"x": x},
        "expecteddescriptionZero tensofoopassedtotalerroFunction   foundfooexpecteddescriptiondescriptionpassederroOutput mismatch: got {output}, expected {[]}descriptiondescriptionpassederropassedpassedtotalresults

Register the problem

After creating the module, register it in the problem registry (typically torch_judge/registry.py or equivalent):

from torch_judge.problems import my_new_problem

PROBLEMS = [
    # ... existing problems ...
    my_new_problem,
]

---

Grading Service API

The FastAPI grading service at http://localhost:8000 exposes:

# Health check
GET /health

# List all problems
GET /problems

# Get a specific problem
GET /problems/{problem_id}

# Submit a solution
POST /submit
Content-Type: application/json

{
  "problem_id": "relu",
  "code": "import torch\n\ndef relu(x):\n    return torch.clamp(x, min=0)"
}

# Response
{
  "problem_id": "relu",
  "passed": 3,
  "total": 3,
  "results": [
    {"case": 1, "description": "Basic positive values", "passed": true, "error": null},
    {"case": 2, "description": "Negative values", "passed": true, "error": null},
    {"case": 3, "description": "Mixed values", "passed": true, "error": null}
  ]
}

Calling the grading API from Python

import requests

response = requests.post(
    "http://localhost:8000/submit",
    json={
        "problem_id": "softmax",
        "code": """
import torch

def softmax(x: torch.Tensor, dim: int = -1) -> torch.Tensor:
    x_max = x.max(dim=dim, keepdim=True).values
    x_exp = torch.exp(x - x_max)
    return x_exp / x_exp.sum(dim=dim, keepdim=True)
"""
    }
)

result = response.json()
print(f"Passed {result['passed']}/{result['total']} test cases")
for r in result["results"]:
    status = "✓" if r["passed"] else "✗"
    print(f"  {status} Case {r['case']}: {r['description']}")
    if r["error"]:
        print(f"      Error: {r['error']}")

---

Example Implementations

Scaled Dot-Product Attention

import torch
import torch.nn.functional as F
import math

def scaled_dot_product_attention(
    q: torch.Tensor,  # (batch, heads, seq, d_k)
    k: torch.Tensor,
    v: torch.Tensor,
    mask: torch.Tensor | None = None,
) -> torch.Tensor:
    d_k = q.size(-1)
    scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(d_k)
    if mask is not None:
        scores = scores.masked_fill(mask == 0, float('-inf'))
    weights = F.softmax(scores, dim=-1)
    return torch.matmul(weights, v)

RMSNorm

import torch

def rms_norm(x: torch.Tensor, weight: torch.Tensor, eps: float = 1e-6) -> torch.Tensor:
    rms = x.pow(2).mean(dim=-1, keepdim=True).add(eps).sqrt()
    return x / rms * weight

LoRA Linear Layer

import torch
import torch.nn as nn

class LoRALinear(nn.Module):
    def __init__(self, in_features: int, out_features: int, rank: int = 4, alpha: float = 1.0):
        super().__init__()
        self.weight = nn.Parameter(torch.randn(out_features, in_features) * 0.02)
        self.lora_A = nn.Parameter(torch.randn(rank, in_features) * 0.02)
        self.lora_B = nn.Parameter(torch.zeros(out_features, rank))
        self.scale = alpha / rank

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        base = x @ self.weight.T
        lora = x @ self.lora_A.T @ self.lora_B.T
        return base + self.scale * lora

Cosine Learning Rate Schedule

import math

def cosine_lr(step: int, max_steps: int, lr_max: float, lr_min: float = 0.0) -> float:
    if step >= max_steps:
        return lr_min
    progress = step / max_steps
    return lr_min + 0.5 * (lr_max - lr_min) * (1 + math.cos(math.pi * progress))

KV Cache (Inference)

import torch
from dataclasses import dataclass, field

@dataclass
class KVCache:
    keys: list[torch.Tensor] = field(default_factory=list)
    values: list[torch.Tensor] = field(default_factory=list)

    def update(self, new_k: torch.Tensor, new_v: torch.Tensor):
        self.keys.append(new_k)
        self.values.append(new_v)

    def get(self) -> tuple[torch.Tensor, torch.Tensor]:
        return torch.cat(self.keys, dim=-2), torch.cat(self.values, dim=-2)

    def __len__(self) -> int:
        return len(self.keys)

---

Learning Paths

Choose a path based on your goal:

Path	Focus
Transformer Internals	Activations → Normalization → Attention → GPT-2 Block
Attention & Position Encoding	Every attention variant + RoPE, ALiBi, NTK-RoPE
Train a GPT from Scratch	Embeddings → architecture → loss → optimizer → tricks
Inference & Distributed	KV cache, quantization, sampling, tensor parallel, FSDP
Alignment & Reasoning	Reward model → DPO → GRPO → PPO → MCTS
Vision Transformer	Conv → patch embedding → ViT block
Diffusion & DiT	Noise schedule → DDIM → flow matching → adaLN-Zero
LLM Frontier Architectures	GQA, Differential Attention, MLA, MoE, MTP

Recommended progression:

Fundamentals → Transformer Internals → Train a GPT from Scratch
                      │                         │
                      ▼                         ▼
             Attention & PE            Inference & Distributed
                      │                         │
                      ▼                         ▼
             LLM Frontier Archs        Alignment & Reasoning

---

Troubleshooting

Grading service not reachable

# Check if the service is running
curl http://localhost:8000/health

# If not, start it manually
cd grading_service
source ../.venv/bin/activate
uvicorn main:app --reload --port 8000

Python environment issues

# Verify correct Python is active
which python && python --version  # should be 3.11+

# Reinstall deps
pip install -e ".[dev]"

# With uv
uv pip install -e ".[dev]"

Frontend can't connect to grading service

Check web/.env.local:

GRADING_SERVICE_URL=http://localhost:8000

Restart Next.js after changing .env.local.

Docker: port conflicts

# Check what's on port 3000 or 8000
lsof -i :3000
lsof -i :8000

# Stop conflicting processes, then retry
docker compose up --build

Submission always fails with import errors

Ensure the submission code only uses packages available in the environment. Core deps include torch, numpy, math. Check pyproject.toml for the full list.

Progress not persisting

The SQLite DB lives at ./data/pyre.db by default. For Docker, ensure the volume is mounted:

# docker-compose.yml
volumes:
  - pyre_data:/app/data

---

Contributing a Problem

1. Create torch_judge/problems/{problem_id}.py using the structure above
2. Include PROBLEM_ID, TITLE, DIFFICULTY, CATEGORY, DESCRIPTION, STARTER_CODE, REFERENCE_SOLUTION, make_test_cases(), and grade()
3. Register in the problem registry
4. Write at least 3 test cases: basic, edge case, and a larger/random tensor case
5. Verify with pytest torch_judge/ before opening a PR
6. Open an issue first for new categories or structural changes