Vikram Sharma Mailthody

Query: GPipe graph node assignment

7

Hi @bignamehyp and other developers, I am trying to understand the GPipe's composite layer assignment heuristic for each GPU and how the GPUs interact and I am not sure if...

Kubeflow is another set of tools widely used in the enterprise segment for deploying ML models. More details of kubeflow can be found here - https://www.kubeflow.org/ (eventually, we need to...

Name suggests.

There are different degrees to which we can enable Jupyter Notebook. The simplistic version is to use an MLflow-based mechanism as we did in https://github.com/paletteml/mlsync/issues/2. However, what I have in...

Most developer and tracking is currently done with the help of JIRA tools but having a slackbot to automatically generate summaries of the changes would be helpful with MLSync. Basically,...

Enable fp16 while running transformers with gpipe

2

Looking at the log, I noted that the Lingvo-Gpipe LM implementation used fp32 data type by default (for GPUs). Is fp16 supported for this model? If so, how should i...

Kernel panic - after drop caches

2

Hi @pakmarkthub When I run the vectorAdd program repeatedly (manually and not using a run script), I end up getting a kernel panic error. I upgraded the kernel to 5.6.3...

Issue in running first example

4

I am trying to run this repo and see that roi_pooling_param is not found. Where can I download this file from? ``` python test.py --gpu 0 --S 800 --L 2...

CUDA support enhancement

11

**Is your feature request related to a problem? Please describe.** Given the massive throughput requirement of search engines, especially for vector search, are there plans to enable CUDA support with...

Query: Maximum number of outstanding read request

1

I am wondering what is the maximum number of outstanding read request supported in the mmap interface of AXI in the axi_crossbar implementation. Xilinx IPs have a limit of 32...