-
In training, I used swiglu, TP=4, PP=2. I use deepspeed_to_deepspeed.py to convert the checkpoint to a TP=1, PP=1 one. When evaluating the obtained checkpoint, it is found that the accuracy is inconsi…
-
Trying to compile a train.py file with lightning-thunder file (https://github.com/Lightning-AI/lightning-thunder/tree/main/examples/llama2.c).
Running this all on WSL 2 (Ubuntu). Laptop has 64GB r…
-
**Describe the issue**
I have an SSD enclosure(m.2 to USB converter) that uses an RTL9210B chip, when attaching this to a Linux VM (I tested it on the Debian VM from the UTM gallery) it doesn't work …
-
https://arxiv.org/abs/1901.10008
-
# A Tensor Shape DSL For PyTorch
### What do we want ?
* Propagate arbitrary information about tensors shapes on JIT IR, such as input `C, H, W = 3, 32, 32` but that the batch sizes is dynamic
…
-
## 🚀 Feature
Intermediate tensors are currently re-computed instead of re-used in fragmented computation. This could cause regression comparing to eager, which reuses tensor and avoids wasted re-co…
-
At present, Zygote will use forward mode AD (outsourced to FowardDiff) under 2 circumstances:
1. Based on a heuristic for broadcasting
2. Upon an explicit call to `Zygote.forwarddiff`
As shown by…
-
The inference speed of the int 8 quantization version of SDXL is much slower than that of fp16. I am runing trt9.3 sdxl demo and here is the result. (I changed shape to 768x1344 manually)
fp16 : pyt…
-
thanks for the demo of your team on-going research, from the issue of other classmate, I find the two papers you refer is same with me, I have some confused question in reproducing the code in the pae…
-
##### Expected Behavior
System theme being set on QT5 programs
##### Current Behavior
QT5 programs use the default fusion theme instead
##### Possible Solution
##### Steps to Reproduce (…