Sanjoy Das

icon indicating a website link https://sanjoy.link icon indicating an email [email protected]

icon company @NVIDIA icon location CA icon location DL Compiler lead at @NVIDIA. Ex-Cruise, Ex-Google, Ex-AzulSystems.

Results 27 comments of Sanjoy Das

InvalidArgumentError: Cannot assign a device for operation

@dgoldenberg-audiomack Does this problem reproduce in tf-nightly? NVIDIA added a GPU kernel for `ResourceSparseApplyAdagradV2` in January, so it is possible that this just works now.

Stuck after printing 'Successfully opened dynamic library libcublas.so.10.0'

For such cases, it'd be useful to get a stacktrace of where TF is stuck. You can obtain this using gdb: start gdb, [attach](https://sourceware.org/gdb/onlinedocs/gdb/Attach.html) to the hung TF process, and...

Stuck after printing 'Successfully opened dynamic library libcublas.so.10.0'

@alanzyt311 I missed that you're running TF 1.14. 1.14 is very old and does not have native support for your GPU (which I believe is Ampere based). So TensorFlow blocks...

Image Classification Example Int8 Quantization-Out of memory error

Can you try using [`set_memory_growth`](https://www.tensorflow.org/guide/gpu#limiting_gpu_memory_growth) to prevent TF from allocating all of the GPU memory on startup? Does that address this issue?

Image Classification Example Int8 Quantization-Out of memory error

@tfeher Can you PTAL?

Tf2.1.0, tensorrt7, batchSize > 0 && batchSize <= mEngine.getMaxBatchSize(). Note: Batch size was: 0, but engine max batch size was: 1

@bixia1

TensorRT is Unable to convert the FusedBatchNormV2 op

Thanks for the detailed triage @danfischetti ! @bixia1 can you PTAL?

Missing GPU op for zeros_like for RaggedTensorVariant, error occurs when Ragged Tensor fed thru tf.map_fn

@djoshea this looks like a reasonable request, will you be able to send a PR for this?

Missing GPU op for zeros_like for RaggedTensorVariant, error occurs when Ragged Tensor fed thru tf.map_fn

> I could potentially try developing it, but I unfortunately don't know where to start. Is there a guide somewhere to implementing new ops for the GPU? I don't think...

why ExtractVolumePatches uses different implements for ThreadPoolDevice and GPUDevice.

@quintinwang5, just so I understand are you saying that TF produces substantially different results on CPU vs GPU? If yes, can you demonstrate this using a Python snippet?