perf(asr): Whisper conv1d uses CPU fallback instead of GPU kernel

[m96-chan]

Description

The Whisper ASR encoder uses a CPU fallback implementation for 1D convolution instead of a native GPU kernel.

Location

src/pygpukit/asr/whisper/encoder.py:78

# CPU fallback implementation using im2col
# TODO: Implement native GPU conv1d kernel
x_np = x.to_numpy()
w_np = weight.to_numpy()

Impact

• Performance: Conv1d is called for every audio frame in the encoder
• Memory: Unnecessary GPU -> CPU -> GPU transfers
• Latency: Adds significant overhead to ASR inference

Current Implementation

Uses im2col + numpy matmul pattern:

1. x.to_numpy() - GPU to CPU transfer
2. np.pad() - CPU padding
3. im2col loop (Python for-loop)
4. from_numpy() - CPU to GPU transfer

Required Work

1. Implement native CUDA conv1d kernel in native/ops/conv1d.cu
2. Support stride, padding, dilation parameters
3. Optimize for typical ASR dimensions (80 mel bins, kernel_size=3)

Related

• RFC: Image Generation Support (Stable Diffusion, Flux, DiT) #177 - Image generation also needs Conv2d
• Conv1d can share design patterns with Conv2d

Workaround

None - the feature works but is slow.