linearhit

linearhit

Member Since 7 years ago

Alibaba-PAI, Beijing

Experience Points
11
follower
Lessons Completed
6
follow
Lessons Completed
2
stars
Best Reply Awards
4
repos

156 contributions in the last year

Pinned
⚡ The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. Note: the repository does not accept github pull requests at this moment. Please submit your patches at http://reviews.llvm.org.
⚡ vim translate plugin (zh to en)
Activity
Apr
26
3 weeks ago
Activity icon
created branch

linearhit in linearhit/vimrc create branch main

createdAt 3 weeks ago
Activity icon
created repository
createdAt 3 weeks ago
Activity icon
fork

linearhit forked tlhc/vim-yd-trans

⚡ vim translate plugin (zh to en)
linearhit Updated
fork time in 3 weeks ago
started
started time in 3 weeks ago
Apr
22
4 weeks ago
push

linearhit push alibaba/BladeDISC

linearhit
linearhit

Add UnrealizedConversionCast for codegen of si/ui const (#272)

commit sha: b39e9a8ed3902b40e95c16a2b23a81c3704b4520

push time in 4 weeks ago
pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Add UnrealizedConversionCast for codegen of si/ui const

pull request

linearhit merge to alibaba/BladeDISC

linearhit
linearhit

Not print large const to stdout/stderr

Apr
20
1 month ago
push

linearhit push linearhit/BladeDISC

linearhit
linearhit

Add UnrealizedConversionCast for codegen of si/ui const

commit sha: 7abc266fa45b3c221e50f4645dee4743319045f4

push time in 1 month ago
pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Add UnrealizedConversionCast for codegen of si/ui const

push

linearhit push linearhit/BladeDISC

linearhit
linearhit

Fix a bug in disc-dot-merger pass (#258)

linearhit
linearhit

Fix missing source command in dash shell (#256)

linearhit
linearhit

Add UnrealizedConversionCast for codegen of si/ui const

commit sha: a52d9c3ce3ff9883659543f09219092c7585f0e2

push time in 1 month ago
push

linearhit push pai-disc/tensorflow

linearhit
linearhit

Add a patch not to fold UnrealizedConversionCastOp with ui/si type

commit sha: 0c08b201c8d26a459b44616262b5e6394383e95e

push time in 1 month ago
Apr
15
1 month ago
push

linearhit push alibaba/BladeDISC

linearhit
linearhit

Fix a bug in disc-dot-merger pass (#258)

commit sha: db241f41cbbfe53e66f21158206edb96d7be9ffa

push time in 1 month ago
pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Fix a bug in disc-dot-merger pass

open pull request

linearhit wants to merge alibaba/BladeDISC

linearhit
linearhit

Fix a bug in disc-dot-merger pass

linearhit
linearhit

the init value of concat_dim_sum is zero, resulting in unexpected zero dim size in case the concat_dim_size is -1

pull request

linearhit merge to alibaba/BladeDISC

linearhit
linearhit

Fix a bug in disc-dot-merger pass

pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Fix a bug in disc-dot-merger pass

push

linearhit push linearhit/BladeDISC

linearhit
linearhit

Not clustering tf.SqueezeOp with empty squeeze_dims attr to avoid dynamic rank issue (#251)

linearhit
linearhit

Fix cpu demo crash on aarch64 platform (#252)

We try to fix #244 in this PR. The root cause of the crash is still not clear. After doing some experiments, we think that it isn't likely to be the bug of BladeDISC itself.

The crash could be re-produced on a graph with a single GEMM op using BladeDISC TF bridge. Same GEMM op test is ok if we do the test in UT environments (not using TF bridge) or use BladeDISC torch bridge. Codegen kernel runs successfully using BladeDISC TF bridge. if we use the wheel file from ARM community instead of pip install tensorflow-aarch64, GEMM op test is ok using BladeDISC TF bridge. Thus, we provide a workaround to fix #244 by replacing the wheel package.

linearhit
linearhit

[design doc] accelerate PyTorch training via Disc and LTC (#237)

  • add ltc+disc design

  • fix typo

  • update as comment

linearhit
linearhit

Fix a bug in disc-dot-merger pass

commit sha: 3c63c51bb75fb2b9f3986d7ab338ac9342483abc

push time in 1 month ago
pull request

linearhit merge to alibaba/BladeDISC

linearhit
linearhit

Fix cpu demo crash on aarch64 platform

We try to fix #244 in this PR. The root cause of the crash is still not clear. After doing some experiments, we think that it isn't likely to be the bug of BladeDISC itself.

  • The crash could be re-produced on a graph with a single GEMM op using BladeDISC TF bridge.
  • Same GEMM op test is ok if we do the test in UT environments (not using TF bridge) or use BladeDISC torch bridge.
  • Codegen kernel runs successfully using BladeDISC TF bridge.
  • if we use the wheel file from ARM community instead of pip install tensorflow-aarch64, GEMM op test is ok using BladeDISC TF bridge.

Thus, we provide a workaround to fix #244 by replacing the wheel package.

Apr
14
1 month ago
push

linearhit push alibaba/BladeDISC

linearhit
linearhit

Not clustering tf.SqueezeOp with empty squeeze_dims attr to avoid dynamic rank issue (#251)

commit sha: 5602c9f994332937f2bde5eaceea63fdeb7c1134

push time in 1 month ago
pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Not clustering tf.SqueezeOp with empty squeeze_dims attr to avoid dynamic rank issue

pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Not clustering tf.SqueezeOp with empty squeeze_dims attr to avoid dynamic rank issue

push

linearhit push linearhit/BladeDISC

linearhit
linearhit

Fix a bug in speculation pass in processing dead fusions (#250)

linearhit
linearhit

Not clustering tf.SqueezeOp with empty squeeze_dims attr to avoid dynamic rank issue

commit sha: bb41998fc8b06bd9247452eb70264ef6618cad22

push time in 1 month ago
push

linearhit push alibaba/BladeDISC

linearhit
linearhit

Fix a bug in speculation pass in processing dead fusions (#250)

commit sha: 4bdebe038c9821d60f9f32d92f201c9843bc3095

push time in 1 month ago
pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Fix a bug in speculation pass in processing dead fusions

push

linearhit push linearhit/BladeDISC

linearhit
linearhit

Not clustering tf.SqueezeOp with empty squeeze_dims attr to avoid dynamic rank issue

commit sha: d88a53eefb52649be9ff266f6716194a82d3771e

push time in 1 month ago
pull request

linearhit pull request alibaba/BladeDISC

linearhit
linearhit

Fix a bug in speculation pass in processing dead fusions

push

linearhit push linearhit/BladeDISC

linearhit
linearhit

Add lowering of shape.const_shape in disc-shape-to-std pass (#248)

linearhit
linearhit

Fix a bug in speculation pass in processing dead fusions

commit sha: f51b908c808e47f7c84271f07ba0bd9a29f71f41

push time in 1 month ago
Previous