No forward-mode differentiation (.fwds missing)

[szihs]

Summary

slangpy Function has .bwds() for backward-mode differentiation but no .fwds() for forward-mode differentiation (Jacobian-vector products).

Expected Behavior

slangtorch provides both:

• .fwd() - forward-mode differentiation (computes Jacobian-vector products)
• .bwd() - backward-mode differentiation (computes vector-Jacobian products)

Actual Behavior

slangpy only provides:

• .bwds() - backward-mode differentiation

No forward-mode equivalent exists.

Use Case

Forward-mode differentiation is useful for:

• Computing directional derivatives
• Jacobian-vector products without materializing the full Jacobian
• Cases where output dimension >> input dimension (forward-mode is more efficient)

Reproduction

square_diff.slang:

[Differentiable]
void square(no_diff int tid, DiffTensorView input, DiffTensorView output)
{
    if (tid >= input.size(0)) return;
    output.store(tid, input[tid] * input[tid]);
}

test_fwd.py:

import torch
import slangpy as spy
from slangpy.torchintegration import diff_pair

device = spy.Device(type=spy.DeviceType.cuda)
module = spy.Module.load_from_source(device, "square_diff.slang")

X = torch.tensor([3.0, 4.0], device="cuda")
Y = torch.zeros_like(X)
tid = torch.arange(2, device="cuda", dtype=torch.int32)

# Backward works
module.square.bwds(tid=tid, input=diff_pair(X, torch.zeros_like(X)), 
                   output=diff_pair(Y, torch.ones_like(Y)))

# Forward doesn't exist
module.square.fwds(...)  # AttributeError: no .fwds method

slangtorch equivalent

# slangtorch forward-mode diff
m.square.fwd(input=(X, dX), output=(Y, dY)).launchRaw(...)

Impact

Medium - Forward-mode diff is less commonly used than backward-mode, but important for certain applications (sensitivity analysis, directional derivatives).

Environment

• slangpy 0.40.1
• CUDA 12.x
• PyTorch 2.10.0