Eikan Wang

Fix the performance issue that the for-loop before ExternallCall could not be parallelized.

4

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #85056 Currently, NNC only parallelizes the loop statement of the graph outputs. The logic could bypass some loop statements that could be...

Use high precision accmulate buffer for bf16 accmulation

4

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #85140 * __->__ #84402 * #84041 Accumulation operation is not friendly to BFloat16 because its mantissa part is only 7bits while the operand...

Support BF16ImmPtr

7

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #85140 * #84402 * __->__ #84041 - To support BF16 Immediate value by converting it to uint16. The behavior is as same as...

Optimize to if the datatyep of the source tensor is as same as the dest datatype

3

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #85140 * #84402 * #84041 The AMP inserts `_autocast_to_reduced_precision` and `_autocast_to_full_precision` automatically. The aten implementation provides a fast path to bypass the...

[Inductor] Eliminate redundant to_dtype node

4

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #96650 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @soumith @voznesenskym @penguinwu @anijain2305 @Guobing-Chen @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @peterbell10 @desertfire

Modularize aten parameter parser and checker

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #126883 * #126517 * #125897 * #125831 * #125819 * __->__ #125308 * #124926 In this PR, we abstracted the different types of...

Support aten operations with out tensor

2

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #126883 * #126517 * #125897 * #125831 * #125819 * #125308 * __->__ #124926 This PR intends to support the aten operations with...

[2/N] Non-Tensor: Scalar Support: Add scalar to the cache for eager-through-torch.compile

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #125308 * #124926 * __->__ #124070 * #124177 * #116368 * #124836 Add scalar information to the kernel configuration. #### Additional Context Currently,...

[1/N] Non-Tensor: Scalar Support: Enable aot compile to support aten operations with scalar input like alpha

3

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #125308 * #124926 * #124070 * __->__ #124177 * #116368 * #124836 Some operations have a scalar input parameter, like `torch.add(a, b, alpha=2.0)`....

Add a cache mechanism to accelerate torch.compile-for-eager

17

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #125308 * #124926 * #124070 * #124177 * __->__ #116368 * #124836 This PR is a follow-up of RFC https://github.com/pytorch/pytorch/issues/115545. In this PR,...