How do torch.no_grad() and tensor.detach() differ, and when do you use each?

import torch

model_param = torch.tensor(2.0, requires_grad=True)

# ── torch.no_grad(): context manager — disables ALL grad tracking
# Use for inference and validation loops
with torch.no_grad():
    out = model_param * 3       # no graph built
    print(out.requires_grad)    # False
    print(out.grad_fn)          # None
# Faster + less memory — standard pattern for eval

# ── .detach(): detaches a SPECIFIC tensor from the graph
# The tensor still knows about grad, but is cut off from history
a = model_param * 4
print(a.requires_grad)          # True  (still tracking)
b = a.detach()                  # b shares data with a
print(b.requires_grad)          # False (disconnected)
print(b.data_ptr() == a.data_ptr())  # True — SAME memory!

# Common use case: compute a "stop gradient" target
# in actor-critic / target networks
target = a.detach()             # stop gradient through target
loss = (a - target) ** 2       # gradient only flows through a, not target

# ── @torch.no_grad() decorator variant
@torch.no_grad()
def predict(x):
    return model_param * x      # no grad even without with block

# Validation loop pattern
def validate(model, loader):
    model.eval()                # turns off dropout, batchnorm train mode
    with torch.no_grad():       # no gradient computation
        for x, y in loader:
            pred = model(x)
            # compute metrics...