"""
Multi-Function Module Example for PyPTO
This example demonstrates how to use multiple `@pypto.frontend.jit` functions together
to build complex computation pipelines. It shows:
- Multiple JIT-compiled functions
- Data flow between functions
- Function composition patterns
- Reusing compiled functions
- Switching between different functions at runtime
This pattern is useful for building modular neural network components.
"""
import os
import sys
import argparse
import pypto
import torch
import numpy as np
from numpy.testing import assert_allclose
from dataclasses import dataclass
from typing import Optional
def _peek_run_mode_from_argv(default: str = "npu") -> str:
"""Read run_mode early so module-level decorators can use it."""
for idx, arg in enumerate(sys.argv):
if arg == "--run_mode" and idx + 1 < len(sys.argv):
value = sys.argv[idx + 1]
if value in ("npu", "sim"):
return value
if arg.startswith("--run_mode="):
value = arg.split("=", 1)[1]
if value in ("npu", "sim"):
return value
return default
global_run_mode = pypto.RunMode.NPU
if _peek_run_mode_from_argv("npu") == "sim":
global_run_mode = pypto.RunMode.SIM
def get_device_id():
"""
Get and validate TILE_FWK_DEVICE_ID from environment variable.
Returns:
int: The device ID if valid, None otherwise.
"""
if 'TILE_FWK_DEVICE_ID' not in os.environ:
print("Please set the environment variable TILE_FWK_DEVICE_ID before running:")
print(" export TILE_FWK_DEVICE_ID=0")
return None
try:
device_id = int(os.environ['TILE_FWK_DEVICE_ID'])
return device_id
except ValueError:
print(f"ERROR: TILE_FWK_DEVICE_ID must be an integer, got: {os.environ['TILE_FWK_DEVICE_ID']}")
return None
@dataclass
class ModuleConfig:
"""Configuration for multi-function module."""
hidden_size: int = 128
intermediate_size: int = 256
dtype: pypto.DataType = pypto.DT_BF16
use_dynamic_shape: bool = False
def layer_norm_golden(x: torch.Tensor, gamma: torch.Tensor, beta: torch.Tensor, eps: float) -> torch.Tensor:
"""PyTorch reference for layer norm."""
mean = x.mean(dim=-1, keepdim=True)
var = x.var(dim=-1, keepdim=True, unbiased=False)
normalized = (x - mean) / torch.sqrt(var + eps)
return normalized * gamma + beta
def gelu_golden(x: torch.Tensor) -> torch.Tensor:
"""PyTorch reference for GELU."""
return torch.nn.functional.gelu(x)
def layernorm_core(x: pypto.Tensor, gamma: pypto.Tensor, beta: pypto.Tensor, eps: float = 1e-6) -> pypto.Tensor:
hidden_size = x.shape[-1]
mean = pypto.sum(x, dim=-1, keepdim=True)
mean = mean / hidden_size
centered = x - mean
squared = centered * centered
var = pypto.sum(squared, dim=-1, keepdim=True)
var = var / hidden_size
var_eps = var + eps
std = pypto.sqrt(var_eps)
normalized = centered / std
scaled = normalized * gamma
return scaled + beta
@pypto.frontend.jit(runtime_options={"run_mode": global_run_mode})
def layer_norm_kernel(
x: pypto.Tensor(),
gamma: pypto.Tensor(),
beta: pypto.Tensor(),
out: pypto.Tensor()):
"""Layer Normalization."""
pypto.set_vec_tile_shapes(64, 128)
out[:] = layernorm_core(x, gamma, beta)
@pypto.frontend.jit(runtime_options={"run_mode": global_run_mode})
def linear_projection_kernel(
x: pypto.Tensor(),
weight: pypto.Tensor(),
out: pypto.Tensor()):
bias = None
pypto.set_cube_tile_shapes([64, 64], [64, 64], [64, 64])
if bias is not None:
out[:] = pypto.add(pypto.matmul(x, weight, out_dtype=x.dtype), bias)
else:
out[:] = pypto.matmul(x, weight, out_dtype=x.dtype)
@pypto.frontend.jit(runtime_options={"run_mode": global_run_mode})
def gelu_activation_kernel(
x: pypto.Tensor(),
out: pypto.Tensor()):
tile_shapes = [32 for _ in range(len(x.shape))]
pypto.set_vec_tile_shapes(*tile_shapes)
coeff = 1.702
x_scaled = x * coeff
out[:] = x * pypto.sigmoid(x_scaled)
@pypto.frontend.jit(runtime_options={"run_mode": global_run_mode})
def residual_add_kernel(
x: pypto.tensor(),
residual: pypto.tensor(),
out: pypto.tensor()):
pypto.set_vec_tile_shapes(64, 128)
out[:] = pypto.add(x, residual)
@pypto.frontend.jit(runtime_options={"run_mode": global_run_mode})
def attention_kernel(
q: pypto.tensor(),
k: pypto.tensor(),
v: pypto.tensor(),
out: pypto.tensor()):
pypto.set_cube_tile_shapes([64, 64], [64, 64], [64, 64])
k_t = pypto.transpose(k, [0, 1, 3, 2])
scores = pypto.matmul(q, k_t, out_dtype=out.dtype)
scores_scaled = pypto.mul(scores, scale)
attn_weights = pypto.softmax(scores_scaled, dim=-1)
out[:] = pypto.matmul(attn_weights, v, out_dtype=out.dtype)
def test_sequential_functions(device_id: int = None, dynamic: bool = False) -> None:
"""Test multiple functions in sequence."""
print("=" * 60)
print("Test: Sequential Functions")
print("=" * 60)
device = f'npu:{device_id}' if global_run_mode == pypto.RunMode.NPU and device_id is not None else 'cpu'
atol_val = 1e-1
batch_size, hidden_size = 32, 128
x = torch.randn(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
gamma = torch.ones(hidden_size, dtype=torch.bfloat16, device=device)
beta = torch.zeros(hidden_size, dtype=torch.bfloat16, device=device)
normed = torch.empty(x.shape, dtype=torch.bfloat16, device=device)
layer_norm_kernel(x, gamma, beta, normed)
activated = torch.empty(normed.shape, dtype=torch.bfloat16, device=device)
gelu_activation_kernel(normed, activated)
expected_normed = layer_norm_golden(x, gamma, beta, 1e-6)
expected_activated = gelu_golden(expected_normed)
max_diff_norm = (normed - expected_normed).abs().max().item()
max_diff_act = (activated - expected_activated).abs().max().item()
print(f"Input shape: {x.shape}")
if global_run_mode == pypto.RunMode.NPU:
print(f"Normalized max diff: {max_diff_norm:.6f}")
print(f"Activated max diff: {max_diff_act:.6f}")
assert max_diff_norm < atol_val, "Layer norm mismatch!"
assert max_diff_act < atol_val, "GELU mismatch!"
print("✓ Sequential functions passed")
print()
def test_residual_connection(device_id: int = None, dynamic: bool = False) -> None:
"""Test residual connection pattern."""
print("=" * 60)
print("Test: Residual Connection")
print("=" * 60)
device = f'npu:{device_id}' if global_run_mode == pypto.RunMode.NPU and device_id is not None else 'cpu'
batch_size, hidden_size = 32, 128
x = torch.randn(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
residual = torch.randn(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
out = torch.empty(x.shape, dtype=torch.bfloat16, device=device)
residual_add_kernel(x, residual, out)
expected = x + residual
max_diff = (out - expected).abs().max().item()
print(f"Input shape: {x.shape}")
print(f"Residual shape: {residual.shape}")
print(f"Output shape: {out.shape}")
if global_run_mode == pypto.RunMode.NPU:
print(f"Max difference: {max_diff:.6f}")
assert max_diff < 1e-2, "Residual connection mismatch!"
print("✓ Residual connection passed")
print()
def test_transformer_block(device_id: int = None, dynamic: bool = False) -> None:
"""Test a complete transformer block using multiple functions."""
print("=" * 60)
print("Test: Transformer Block (Multi-Function)")
print("=" * 60)
device = f'npu:{device_id}' if global_run_mode == pypto.RunMode.NPU and device_id is not None else 'cpu'
batch_size, hidden_size, intermediate_size = 32, 128, 256
x = torch.randn(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
gamma = torch.ones(hidden_size, dtype=torch.bfloat16, device=device)
beta = torch.zeros(hidden_size, dtype=torch.bfloat16, device=device)
gate_weight = torch.randn(hidden_size, intermediate_size, dtype=torch.bfloat16, device=device)
up_weight = torch.randn(hidden_size, intermediate_size, dtype=torch.bfloat16, device=device)
down_weight = torch.randn(intermediate_size, hidden_size, dtype=torch.bfloat16, device=device)
normed = torch.zeros(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
gate = torch.zeros(batch_size, intermediate_size, dtype=torch.bfloat16, device=device)
up = torch.zeros(batch_size, intermediate_size, dtype=torch.bfloat16, device=device)
activated = torch.zeros(batch_size, intermediate_size, dtype=torch.bfloat16, device=device)
ffn_out = torch.zeros(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
output = torch.zeros(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
normed = torch.empty(x.shape, dtype=torch.bfloat16, device=device)
layer_norm_kernel(x, gamma, beta, normed)
if global_run_mode == pypto.RunMode.NPU:
torch.npu.synchronize()
linear_projection_kernel(normed, gate_weight, gate)
if global_run_mode == pypto.RunMode.NPU:
torch.npu.synchronize()
linear_projection_kernel(normed, up_weight, up)
if global_run_mode == pypto.RunMode.NPU:
torch.npu.synchronize()
activated = torch.empty(gate.shape, dtype=torch.bfloat16, device=device)
gelu_activation_kernel(gate, activated)
if global_run_mode == pypto.RunMode.NPU:
torch.npu.synchronize()
activated = activated * up
linear_projection_kernel(activated, down_weight, ffn_out)
output = torch.empty(x.shape, dtype=torch.bfloat16, device=device)
residual_add_kernel(x, ffn_out, output)
print(f"Input shape: {x.shape}")
print(f"Output shape: {output.shape}")
print(f"Output range: [{output.min():.4f}, {output.max():.4f}]")
print("✓ Transformer block (multi-function) completed")
print()
def test_function_reuse(device_id: int = None, dynamic: bool = True) -> None:
"""Test reusing the same function multiple times."""
print("=" * 60)
print("Test: Function Reuse")
print("=" * 60)
device = f'npu:{device_id}' if global_run_mode == pypto.RunMode.NPU and device_id is not None else 'cpu'
batch_size, hidden_size = 32, 128
x1 = torch.randn(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
x2 = torch.randn(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
x3 = torch.randn(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
gamma = torch.ones(hidden_size, dtype=torch.bfloat16, device=device)
beta = torch.zeros(hidden_size, dtype=torch.bfloat16, device=device)
out1 = torch.zeros(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
out2 = torch.zeros(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
out3 = torch.zeros(batch_size, hidden_size, dtype=torch.bfloat16, device=device)
layer_norm_kernel(x1, gamma, beta, out1)
if global_run_mode == pypto.RunMode.NPU:
torch.npu.synchronize()
layer_norm_kernel(x2, gamma, beta, out2)
if global_run_mode == pypto.RunMode.NPU:
torch.npu.synchronize()
layer_norm_kernel(x3, gamma, beta, out3)
if global_run_mode == pypto.RunMode.NPU:
torch.npu.synchronize()
expected1 = layer_norm_golden(x1, gamma, beta, 1e-6)
expected2 = layer_norm_golden(x2, gamma, beta, 1e-6)
expected3 = layer_norm_golden(x3, gamma, beta, 1e-6)
max_diff1 = (out1 - expected1).abs().max().item()
max_diff2 = (out2 - expected2).abs().max().item()
max_diff3 = (out3 - expected3).abs().max().item()
print(f"Function reused 3 times with different inputs")
if global_run_mode == pypto.RunMode.NPU:
print(f"Max diff 1: {max_diff1:.6f}")
print(f"Max diff 2: {max_diff2:.6f}")
print(f"Max diff 3: {max_diff3:.6f}")
assert max_diff1 < 1e-1 and max_diff2 < 1e-1 and max_diff3 < 1e-1, "Function reuse mismatch!"
print("✓ Function reuse passed")
print()
def main():
"""Run multi-function module examples.
Usage:
python multi_function_module.py # Run all examples
python multi_function_module.py 1 # Run example 1 only
python multi_function_module.py --list # List all available examples
"""
parser = argparse.ArgumentParser(
description="PyPTO Multi-Function Module Examples",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
Examples:
%(prog)s Run all examples
%(prog)s function_reuse::test_function_reuse
Run example function_reuse::test_function_reuse
%(prog)s --list List all available examples
"""
)
parser.add_argument(
'example_id',
type=str,
nargs='?',
help='Example ID to run (1-4). If not specified, all examples will run.'
)
parser.add_argument(
'--list',
action='store_true',
help='List all available examples and exit'
)
parser.add_argument(
'--run_mode',
type=str,
nargs='?',
default='npu',
choices=["npu", "sim"],
help='Run mode, supports npu and sim.'
)
args = parser.parse_args()
examples = {
'sequential_functions::test_sequential_functions': {
'name': 'Sequential Functions',
'description': 'Using multiple functions in sequence',
'function': test_sequential_functions,
'requires_npu': True
},
'residual_connection::test_residual_connection': {
'name': 'Residual Connection',
'description': 'Residual connection pattern',
'function': test_residual_connection,
'requires_npu': True
},
'transformer_block::test_transformer_block': {
'name': 'Transformer Block',
'description': 'Complete transformer block with multiple functions',
'function': test_transformer_block,
'requires_npu': True
},
'function_reuse::test_function_reuse': {
'name': 'Function Reuse',
'description': 'Reusing the same function with different inputs',
'function': test_function_reuse,
'requires_npu': True
}
}
if args.list:
print("\n" + "=" * 60)
print("Available Examples")
print("=" * 60 + "\n")
for ex_id, ex_info in sorted(examples.items()):
print(f" ID: {ex_id}")
print(f" name: {ex_info['name']}")
print(f" description: {ex_info['description']}\n")
return
if args.example_id is not None:
if args.example_id not in examples:
print(f"ERROR: Invalid example ID: {args.example_id}")
print(f"Valid example IDs are: {', '.join(map(str, sorted(examples.keys())))}")
print("\nUse --list to see all available examples.")
sys.exit(1)
print("\n" + "=" * 60)
print("PyPTO Multi-Function Module Examples")
print("=" * 60 + "\n")
device_id = None
examples_to_run = []
if args.example_id is not None:
examples_to_run = [(args.example_id, examples[args.example_id])]
else:
examples_to_run = list(examples.items())
if args.run_mode == "npu":
device_id = get_device_id()
if device_id is None:
return
import torch_npu
torch.npu.set_device(device_id)
print("Running examples that require NPU hardware...")
print("(Make sure CANN environment is configured and NPU is available)\n")
try:
for ex_id, ex_info in examples_to_run:
print(f"Running Example {ex_id}: {ex_info['name']}")
ex_info['function'](device_id)
if len(examples_to_run) > 1:
print("=" * 60)
print("All multi-function module tests passed!")
print("=" * 60)
except Exception as e:
print(f"\nError: {e}")
raise
if __name__ == "__main__":
main()