Support for AI4Bharat IndicConformer Models

Feature Description

Please add support for AI4Bharat IndicConformer models.

Currently loading the GGUF model crashes during the feature downsampling/subsampling block.

Error

GGML_ASSERT(a->ne[2] == 1) failed

Stack:
ggml_conv_2d_dw_direct()
Subsampling::build_graph()

Root Cause

The current implementation assumes FastConformer subsampling:

conv.2 depthwise
conv.3 pointwise
conv.5 depthwise
conv.6 pointwise

However AI4Bharat IndicConformer uses:

conv.0 Conv2d
ReLU
conv.2 Conv2d
ReLU

ONNX inspection:

conv.0.weight
[512,1,3,3]
group=1

conv.2.weight
[512,512,3,3]
group=1

GGUF:

encoder.pre_encode.conv.2.weight
[3,3,512,512]

This is a different computation graph, not only a tensor layout issue.

Proposed implementation

1. Detect subsampling architecture during loading.

2. Preserve existing FastConformer depthwise path.

3. Add IndicConformer path:

Conv2d(conv.0)
→ ReLU
→ Conv2d(conv.2)
→ ReLU

using ggml_conv_2d().

4. Reuse existing encoder blocks after subsampling.

Additional observation

The GGUF file itself is valid:

• parakeet-cli info works correctly.
• parakeet-cli quantize works correctly.

The failure only occurs during parakeet-cli transcribe, when the inference graph is constructed and the subsampling layer is executed.

Use Case

Standalone C++ Speech-to-Text inference for Indian languages.

Start at Subsampling::build_graph() and inspect the existing FastConformer depthwise path, then compare it with the ggml_conv_2d() operations needed by the IndicConformer graph. Use parakeet-cli transcribe with an AI4Bharat GGUF model to verify that loading and inference complete while existing FastConformer behavior remains unchanged.