[6078291][OMNIML-3716] Add ViT FP8 + Torch-TRT example, wire softmax_quantizer in _QuantAttention - #1569

Merged

ajrasane merged 17 commits into

mainfrom

ajrasane/nvbug_6078291-vit-torch-trt-fp8

Jul 1, 2026

Commits on Jul 1, 2026

[6078291] Add ViT FP8/NVFP4 recipes + Torch-TRT example

ajrasane
and
claude
committed
bfaeaf2
Rename example script to torch_tensorrt_ptq.py + strip benchmarking comparison
ajrasane
committed
0869610
Add e2e integration test for torch_trt example
ajrasane
committed
b2b3f90
Remove quantize-only smoke variant from torch_trt e2e test
ajrasane
committed
814c970
[6078291] Simplify ViT FP8/NVFP4 recipes via shared config units

ajrasane
and
claude
committed
280a9ae
[6078291] Add Torch-TRT ViT ImageNet accuracy example

ajrasane
and
claude
committed
a6c5707
[6078291] Torch-TRT-compile the accuracy baseline + document the example

ajrasane
and
claude
committed
712f13b
[6078291] Comment the local torch_tensorrt import in the example

ajrasane
and
claude
committed
934e2c0
[6078291] Address PR review: wire torch_trt into example CI + deps

ajrasane
and
claude
committed
5d47e12
[6078291] Document torch_trt vs torch_onnx/onnx_ptq in example README

ajrasane
and
claude
committed
cfa8444
[6078291] torch_trt example: default to dynamic FP8 engine, simplify CLI

ajrasane
and
claude
committed
c3eda36
[6078291] torch_trt example: FP16, dynamic-only engine, simpler loader

ajrasane
and
claude
committed
93bf5a4
[6078291] torch_trt example: load ViT with eager attention for softmax PTQ

ajrasane
and
claude
committed
80134eb
[6078291] torch_trt test: skip nvfp4 case on non-Blackwell GPUs

ajrasane
and
claude
committed
406ca1a
[6078291] torch_trt test: use tiny offline ViT, drop NVFP4 case

ajrasane
and
claude
committed
59fa568
[6078291] torch_trt: remove the ViT NVFP4 recipe and its references

ajrasane
and
claude
committed
3e4d4ec
[6078291] torch_trt: quantize non-causal (ViT) attention softmax-P via eager p_bmm_quantizer wrapper

ajrasane
and
claude
committed
3b0a302

Navigation Menu

[6078291][OMNIML-3716] Add ViT FP8 + Torch-TRT example, wire softmax_quantizer in _QuantAttention - #1569

[6078291][OMNIML-3716] Add ViT FP8 + Torch-TRT example, wire softmax_quantizer in _QuantAttention#1569
ajrasane merged 17 commits into
mainfrom
ajrasane/nvbug_6078291-vit-torch-trt-fp8

Commits on Jul 1, 2026

[6078291] Add ViT FP8/NVFP4 recipes + Torch-TRT example

Rename example script to torch_tensorrt_ptq.py + strip benchmarking comparison

Add e2e integration test for torch_trt example

Remove quantize-only smoke variant from torch_trt e2e test

[6078291] Simplify ViT FP8/NVFP4 recipes via shared config units

[6078291] Add Torch-TRT ViT ImageNet accuracy example

[6078291] Torch-TRT-compile the accuracy baseline + document the example

[6078291] Comment the local torch_tensorrt import in the example

[6078291] Address PR review: wire torch_trt into example CI + deps

[6078291] Document torch_trt vs torch_onnx/onnx_ptq in example README

[6078291] torch_trt example: default to dynamic FP8 engine, simplify CLI

[6078291] torch_trt example: FP16, dynamic-only engine, simpler loader

[6078291] torch_trt example: load ViT with eager attention for softmax PTQ

[6078291] torch_trt test: skip nvfp4 case on non-Blackwell GPUs

[6078291] torch_trt test: use tiny offline ViT, drop NVFP4 case

[6078291] torch_trt: remove the ViT NVFP4 recipe and its references

[6078291] torch_trt: quantize non-causal (ViT) attention softmax-P via eager p_bmm_quantizer wrapper