qihqi comments

Results 42 comments of


                                            qihqi

Update flatbuffer test models to match pkl models

@pytorchbot merge

Make deepcopy of GraphModule able to handle circular reference.

@pytorchbot merge

Update OpenXLA pin to enable libtpu & pjrt plugin for 02282024

afaik we still need quant_dequant_converter @lsy323

Lowering aten::where encounters mismatched types and crashed

Hi, sorry for the delay. Just tried on HEAD and your example is passing. My guess is that is is fixed after https://github.com/pytorch/xla/pull/6460 So we introduces f64 when scalars are...

[export] Support capturing of non-tensor inputs

cc. @gmagogsfm @angelayi @zhxchen17 We chatted briefly on this topic and agreed that is desirable. Although there is no timeline for it. Feel free to update about the future plan....

[Feature Request] Per request sampling params

HI @kiratp few questions: On 1. prompt_token_counts would be the same behavior as per-request max_output_len? On 2> logprobs is a boolean arg on input to signify to return the logprobs...

torch.linalg.lstsq issues on GPU/TPU

Hi, You can try this in torchax: (https://github.com/pytorch/xla/tree/master/torchax) ```python import torch as torch import torchax torchax.enable_globally() device = 'jax' diff = torch.randn(8,3,1,requires_grad=True,device=device) A = torch.randn(8,251,3,requires_grad=True,device=device) B = torch.randn(8,251,1,requires_grad=True,device=device) C =...