Enable wight-free export for VLM - #1311
Conversation
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
Signed-off-by: Amit Raj <amitraj@qti.qualcomm.com>
|
Legacy output: ['Based on the image provided, the following colors are visible:\n\n- Black: This is the dominant color, seen in the fur of the puppy.\n- Brown: The wooden surface has various shades of brown, from light tan to darker, weathered brown.\n- Gray: There are significant gray tones in the wood, particularly in the shadows and the weathered parts of the planks.\n- White: The highlights in the puppy's eyes appear as bright white or"'] Weight-free output: ['Based on the image provided, the following colors are visible:\n\n- Black: This is the dominant color, seen in the fur of the puppy.\n- Brown: The wooden surface has various shades of brown, from light tan to darker, weathered brown, creating a rustic appearance.\n- Gray: There are noticeable gray tones in the wood, particularly in the shadows and the weathered parts of the planks.\n- White: The highlights in the puppy's eyes"'] |
Summary
This PR enables the VLM weight-free export path for
Qwen/Qwen3-VL-30B-A3B-Instruct, a Qwen3-VL-MoE image-text-to-text model.Model validated:
Validation script:
Performance Comparison
Measured with the full-layer model using:
Legacy means no weight-free export. Weight-free uses the Dynamo/
torch.exportpath because weight-free export is implemented on top of Dynamo export.Export Time
Compile Time
Artifact Size Comparison
For ONNX size, this table uses the complete wrapper export directory excluding QPC, because legacy export stores large external tensor files beside the small
.onnxprotobuf.ONNX Export Artifact Size
QPC Size
The QPC size is effectively unchanged, as expected. The major storage improvement is in the exported ONNX artifact before compile.
Runtime Comparison
Runtime is effectively at parity while export, compile, and pre-compile artifact size improve substantially.