Skip to content

T
Tensorflow

曾经的那一瞬间 / Tensorflow
10 个月 前同步成功

通知 10
Star 0
Fork 0
T
Tensorflow

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

切换分支/标签
  1. 23 2月, 2021 9 次提交
    • A
      Add some internal documentation on the difference between def_function.py and function.py · c7a91d5b
      由 Allen Lavoie 提交于 
      This is a common source of confusion; the public docstring for tf.function is pretty good, but the internal documentation has been quite sparse.

PiperOrigin-RevId: 358837475
Change-Id: I961dfa70ca8f73c7641a376b09015f5a938050f1
      c7a91d5b
    • R
      Fixed OpenGL FullyConnected. · 6927b25a
      由 Raman Sarokin 提交于 
      PiperOrigin-RevId: 358833393
Change-Id: I9d788fd74503d43f458aa212ec81c3fc222e05dd
      6927b25a
    • S
      Fix logic to capture ancestor frame in `module_wrapper._call_location()` · a9a24e0e
      由 Scott Wegner 提交于 
      While looping through the parent frames, the loop variable was never updated. As a result, each loop iteration did the exact same check, and in the end we still return the current frame.

This fix correctly captures the `parent` frame on each loop iteration.

PiperOrigin-RevId: 358832374
Change-Id: I4ccf4cf634b64c56fb008cca6f60a009c57ada86
      a9a24e0e
    • R
      [XLA:GPU] Fix Custom call implementation to correctly handle token inputs. · 35e41514
      由 Rahul Joshi 提交于 
      - Change custom call thunk to accept an optional slice for each argument and result.
- The slice will not be present when the corresponding arg or result of the custom call
  is a token type.
- When the slice is not present, pass in a nullptr for that position in the buffers passed
  into the custom call.

PiperOrigin-RevId: 358831157
Change-Id: Ib9051b52a45095394349ac68af5b61fa16bc0447
      35e41514
    • A
      Use cc_library.defines to apply most vulkan.hpp #defines · 97a1411a
      由 A. Unique TensorFlower 提交于 
      Also update vulkan_headers to ec2db85225ab410bc6829251bef6c578aaed5868, which
fixes some of the previous unguarded references to assert().

Tested:
  bazel build @vulkan_headers//...
PiperOrigin-RevId: 358830383
Change-Id: I25546d4525c5b384cd27a6897cee99870997c934
      97a1411a
    • R
      [XLA:GPU] Fix XLA:HLO -> LMHLO conversion of custom call to handle token types. · 66e0db07
      由 Rahul Joshi 提交于 
      - Handle custom calls with token type by explicitly dropping the token inputs or
  results and remembering the mapping of the LMHLO arg/results to the original
  XLA:HL arg and results. This is required because the custom call ultimately expects
  arguments as defined in XLA:HLO.

PiperOrigin-RevId: 358828332
Change-Id: Ie1230f2384aff98b1ab5814292d7b047fd0dcbff
      66e0db07
    • E
      Fix build nondeterminism. · 41ff2474
      由 Eric Spishak-Thomas 提交于 
      The zip file entries had real timestamps, which were inconsistent across builds. Passing a ZipInfo uses fixed timestamps.

And, the report.html file in the zip file contained nondeterministic tmp file paths, so this replaces those paths with a placeholder so the output is deterministic.

PiperOrigin-RevId: 358828255
Change-Id: Id6866922505164f1561ceb4c035cf9f6c17a3f60
      41ff2474
    • R
      [MLIR:LHLO] Add optional call target arg mapping to LMHLO CustomCall operations. · 9f8d5d0e
      由 Rahul Joshi 提交于 
      - XLA:HLO -> LMHLO conversion drops all token arguments and return values, however
  custom calls that users write still expect to get buffer pointers for these token types.
- To be able to support this, add an optional call target argument mapping attribute to
  LMHLO custom calls. When this attribute is present, it indicates the number of
  arguments and returns that the custom call expects and also indicates which LMHLO
  arg() or output() maps to which arg or result number of the custom call.

PiperOrigin-RevId: 358826664
Change-Id: I36e839e9ff5b73890715b71717a4c13631955fba
      9f8d5d0e
    • A
      Internal clean up · a2939994
      由 A. Unique TensorFlower 提交于 
      PiperOrigin-RevId: 358826589
Change-Id: I5acabb9db1bd8ed9ba88d393e25491a5c6ff0b97
      a2939994
  3. 22 2月, 2021 14 次提交
  5. 21 2月, 2021 17 次提交