{"payload":{"feedbackUrl":"https://github.com/orgs/community/discussions/53140","repo":{"id":523007292,"defaultBranch":"main","name":"xla","ownerLogin":"openxla","currentUserCanPush":false,"isFork":false,"isEmpty":false,"createdAt":"2022-08-09T15:30:01.000Z","ownerAvatar":"https://avatars.githubusercontent.com/u/107584881?v=4","public":true,"private":false,"isOrgOwned":true},"refInfo":{"name":"","listCacheKey":"v0:1721744129.0","currentOid":""},"activityList":{"items":[{"before":"aec1282d9491b7a2b4b42ce6553294de0d254fe4","after":"36e5fd00f62a50537ef3f8e44a811da4183282fe","ref":"refs/heads/test_654697195","pushedAt":"2024-07-23T14:18:05.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:CPU] Add runtime check if `stochastic-convert` was decomposed.\n\nIn the current runtime, `stochastic-convert` op emitting is not supported by design, normally it is rewritten by `StochasticConvertDecomposer` HLO pass and never reaches the emit phase. This CL adds a runtime check if this op was actually rewritten and returns an explicit message if it wasn't (instead of cryptic message that `kStochasticConvert` opcode was not handled in elemental IR emitter).\n\nAlso added fundamental tests for stochastic convert op for both runtimes - current and thunks.\n\nPiperOrigin-RevId: 654697195","shortMessageHtmlLink":"[XLA:CPU] Add runtime check if <code>stochastic-convert</code> was decomposed."}},{"before":"a02433bdc5e6dbe68446d6c95c4959b5e1715cf9","after":null,"ref":"refs/heads/test_654795065","pushedAt":"2024-07-23T14:15:29.000Z","pushType":"branch_deletion","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"}},{"before":"5ef10e53fdfc0a7069fb4c5532051df0086e0fd3","after":"a02433bdc5e6dbe68446d6c95c4959b5e1715cf9","ref":"refs/heads/main","pushedAt":"2024-07-23T14:15:28.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"Integrate LLVM at llvm/llvm-project@acc159aea1e6\n\nUpdates LLVM usage to match\n[acc159aea1e6](https://github.com/llvm/llvm-project/commit/acc159aea1e6)\n\nPiperOrigin-RevId: 655146366","shortMessageHtmlLink":"Integrate LLVM at <a class=\"commit-link\" data-hovercard-type=\"commit\" data-hovercard-url=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6/hovercard\" href=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6\">llvm/llvm-project@<tt>acc159aea1e6</tt></a>"}},{"before":"b85e935c2fb6fe9aae688ab868f16b4c140cdb6d","after":"a02433bdc5e6dbe68446d6c95c4959b5e1715cf9","ref":"refs/heads/test_654795065","pushedAt":"2024-07-23T14:15:23.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"Integrate LLVM at llvm/llvm-project@acc159aea1e6\n\nUpdates LLVM usage to match\n[acc159aea1e6](https://github.com/llvm/llvm-project/commit/acc159aea1e6)\n\nPiperOrigin-RevId: 655146366","shortMessageHtmlLink":"Integrate LLVM at <a class=\"commit-link\" data-hovercard-type=\"commit\" data-hovercard-url=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6/hovercard\" href=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6\">llvm/llvm-project@<tt>acc159aea1e6</tt></a>"}},{"before":"f85603b61712cb144c0c6ba14139a64ec5ea2f0a","after":"afb0fe68eb66e4cd8f93223e9d00013be189763d","ref":"refs/heads/test_654865806","pushedAt":"2024-07-23T14:09:05.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[xla:cpu] Support for up to 16 sorted inputs\n\n+ enable more jax/lax tests for XLA CPU thunks\n\nPiperOrigin-RevId: 654865806","shortMessageHtmlLink":"[xla:cpu] Support for up to 16 sorted inputs"}},{"before":"359ea704c4ad0cbfc5ee59a4487f1f0cc2db4601","after":"5ef10e53fdfc0a7069fb4c5532051df0086e0fd3","ref":"refs/heads/main","pushedAt":"2024-07-23T13:58:40.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"PR #15222: [GPU] Fix cuDNN workspace test condition.\n\nImported from GitHub PR https://github.com/openxla/xla/pull/15222\n\nCopybara import of the project:\n\n--\ne57258f605e6df6bf61ac55632e939b5badd6c22 by Ilia Sergachev <isergachev@nvidia.com>:\n\n[GPU] Fix cuDNN workspace test condition.\n\nMerging this change closes #15222\n\nCOPYBARA_INTEGRATE_REVIEW=https://github.com/openxla/xla/pull/15222 from openxla:fix_cudnn_test e57258f605e6df6bf61ac55632e939b5badd6c22\nPiperOrigin-RevId: 655145324","shortMessageHtmlLink":"PR <a class=\"issue-link js-issue-link\" data-error-text=\"Failed to load title\" data-id=\"2425053990\" data-permission-text=\"Title is private\" data-url=\"https://github.com/openxla/xla/issues/15222\" data-hovercard-type=\"pull_request\" data-hovercard-url=\"/openxla/xla/pull/15222/hovercard\" href=\"https://github.com/openxla/xla/pull/15222\">#15222</a>: [GPU] Fix cuDNN workspace test condition."}},{"before":"05a4ca3f57fa6029eeb64c55a372e1c7cb431b9d","after":"747f7e460451761f0f624399ff8198f15dafdbf1","ref":"refs/heads/test_655128242","pushedAt":"2024-07-23T13:45:20.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:CPU] Add runtime check for whether `batch-norm-training` is rewritten\n\nIn the current runtime, emitting for op `batch-norm-training` is not supported by design.\nThe op is expected to be rewritten by another HLO pass before ever reaching the emit phase.\nThis CL adds a runtime check  for whether this op was actually rewritten and returns an explicit message if it wasn't.\nAlso includes a new unit test covering existing and new functionality: batch_norm_training_test.cc.\n\nPiperOrigin-RevId: 655128242","shortMessageHtmlLink":"[XLA:CPU] Add runtime check for whether <code>batch-norm-training</code> is rewr…"}},{"before":"359ea704c4ad0cbfc5ee59a4487f1f0cc2db4601","after":null,"ref":"refs/heads/test_650980466","pushedAt":"2024-07-23T13:42:55.000Z","pushType":"branch_deletion","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"}},{"before":"4e56ae4cafd5a8a2849325f8e2ee0607554b54a2","after":"359ea704c4ad0cbfc5ee59a4487f1f0cc2db4601","ref":"refs/heads/main","pushedAt":"2024-07-23T13:42:53.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[IFRT] Introduce pass to merge reshards with the same src and dst.\n\nThis optimizes the reshards to use a single RPC call instead of multiple RPC calls.\nIt currently merges only when the src is func BlockArg or an ifrt Op with `outputs` of type IfrtArrayType, and for any destination.\n\nPiperOrigin-RevId: 655145015","shortMessageHtmlLink":"[IFRT] Introduce pass to merge reshards with the same src and dst."}},{"before":"55c58021a46a4f472846bed408a58d40e7ddc1ff","after":"359ea704c4ad0cbfc5ee59a4487f1f0cc2db4601","ref":"refs/heads/test_650980466","pushedAt":"2024-07-23T13:42:52.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[IFRT] Introduce pass to merge reshards with the same src and dst.\n\nThis optimizes the reshards to use a single RPC call instead of multiple RPC calls.\nIt currently merges only when the src is func BlockArg or an ifrt Op with `outputs` of type IfrtArrayType, and for any destination.\n\nPiperOrigin-RevId: 655145015","shortMessageHtmlLink":"[IFRT] Introduce pass to merge reshards with the same src and dst."}},{"before":null,"after":"05a4ca3f57fa6029eeb64c55a372e1c7cb431b9d","ref":"refs/heads/test_655128242","pushedAt":"2024-07-23T13:35:23.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:CPU] Add runtime check for whether `batch-norm-training` is rewritten\n\nIn the current runtime, emitting for op `batch-norm-training` is not supported by design.\nThe op is expected to be rewritten by another HLO pass before ever reaching the emit phase.\nThis CL adds a runtime check  for whether this op was actually rewritten and returns an explicit message if it wasn't.\nAlso includes a new unit test covering existing and new functionality: batch_norm_training_test.cc.\n\nPiperOrigin-RevId: 655128242","shortMessageHtmlLink":"[XLA:CPU] Add runtime check for whether <code>batch-norm-training</code> is rewr…"}},{"before":"ad0d6e47547cda0cd3c90783f1961ce4658ec0a0","after":"55c58021a46a4f472846bed408a58d40e7ddc1ff","ref":"refs/heads/test_650980466","pushedAt":"2024-07-23T13:30:52.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[IFRT] Introduce pass to merge reshards with the same src and dst.\n\nThis optimizes the reshards to use a single RPC call instead of multiple RPC calls.\nIt currently merges only when the src is func BlockArg or an ifrt Op with `outputs` of type IfrtArrayType, and for any destination.\n\nPiperOrigin-RevId: 650980466","shortMessageHtmlLink":"[IFRT] Introduce pass to merge reshards with the same src and dst."}},{"before":"1b3cca09d6e9773c99018211ace517b5f5b7ebc6","after":"b85e935c2fb6fe9aae688ab868f16b4c140cdb6d","ref":"refs/heads/test_654795065","pushedAt":"2024-07-23T13:22:42.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"Integrate LLVM at llvm/llvm-project@acc159aea1e6\n\nUpdates LLVM usage to match\n[acc159aea1e6](https://github.com/llvm/llvm-project/commit/acc159aea1e6)\n\nPiperOrigin-RevId: 654795065","shortMessageHtmlLink":"Integrate LLVM at <a class=\"commit-link\" data-hovercard-type=\"commit\" data-hovercard-url=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6/hovercard\" href=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6\">llvm/llvm-project@<tt>acc159aea1e6</tt></a>"}},{"before":"c9af12461ea11e02f54704b50bab2ac2eb099879","after":"ad0d6e47547cda0cd3c90783f1961ce4658ec0a0","ref":"refs/heads/test_650980466","pushedAt":"2024-07-23T12:59:13.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[IFRT] Introduce pass to merge reshards with the same src and dst.\n\nThis optimizes the reshards to use a single RPC call instead of multiple RPC calls.\nIt currently merges only when the src is func BlockArg or an ifrt Op with `outputs` of type IfrtArrayType, and for any destination.\n\nPiperOrigin-RevId: 650980466","shortMessageHtmlLink":"[IFRT] Introduce pass to merge reshards with the same src and dst."}},{"before":"4e56ae4cafd5a8a2849325f8e2ee0607554b54a2","after":null,"ref":"refs/heads/test_648711460","pushedAt":"2024-07-23T12:48:56.000Z","pushType":"branch_deletion","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"}},{"before":"ddba514b5bdfbc25e73e37071f8ad80930ca2c5d","after":"4e56ae4cafd5a8a2849325f8e2ee0607554b54a2","ref":"refs/heads/main","pushedAt":"2024-07-23T12:48:55.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"#sdy Initial set of changes to allow for lowering to the Shardy dialect.\n\nThe OpenXLA project is working on an open source, MLIR, named-axis based propagation (and in the future SP<D partitioning) system that will be dialect agnostic (would work for any dialect - MHLO, StableHLO, YourDialect). We plan on having frontends like JAX and PyTorch target this when using XLA and wanting SPMD propagation/partitioning. See www.github.com/openxla/shardy for more info.\n\nCurrently Shardy is implemented inside the XLA compiler, requiring us to round-trip between StableHLO and HLO with `mhlo.sharding`s. But we will eventually make Shardy the first pass in the XLA pipeline while it's still working on StableHLO. Partitioning (the system that adds the collectives like all-gathers/all-reduces) will still be the GSPMD Partitioner, but next year the Shardy partitioner will be developed, allowing for propagation and partitioning to be completely in MLIR and the first pass in the pipeline. So then we'd have:\n1. Traced jaxpr\n2. Jaxpr -> StableHLO\n3. StableHLO with Shardy propagation\n4. StableHLO with Shardy partitioning\n5. StableHLO -> HLO\n6. XLA optimizations\n\nThe following test:\n\n```py\ndef test_sdy_lowering(self):\n  mesh = jtu.create_global_mesh((4, 2), ('x', 'y'))\n  np_inp = np.arange(16).reshape(8, 2)\n  s = jax.sharding.NamedSharding(mesh, P('x', 'y'))\n  arr = jax.device_put(np_inp, s)\n\n  @partial(jax.jit, out_shardings=s)\n  def f(x):\n    return x * 2\n\n  print(f.lower(arr).as_text())\n```\n\noutputs:\n\n```\nmodule @jit_f attributes {mhlo.num_partitions = 8 : i32, mhlo.num_replicas = 1 : i32} {\n  sdy.mesh @mesh = <\"x\"=4, \"y\"=2>\n  func.func public @main(%arg0: tensor<8x2xi64> {mhlo.layout_mode = \"{1,0}\", sdy.sharding = #sdy.sharding<@mesh, [{\"x\"}, {\"y\"}]>}) -> (tensor<8x2xi64> {jax.result_info = \"\", mhlo.layout_mode = \"default\", sdy.sharding = #sdy.sharding<@mesh, [{\"x\"}, {\"y\"}]>}) {\n    %c = stablehlo.constant dense<2> : tensor<i64>\n    %0 = stablehlo.broadcast_in_dim %c, dims = [] : (tensor<i64>) -> tensor<8x2xi64>\n    %1 = stablehlo.multiply %arg0, %0 : tensor<8x2xi64>\n    return %1 : tensor<8x2xi64>\n  }\n}\n```\n\nShardy will be hidden behind the `jax_use_shardy_partitioner` flag initially before becoming enabled by default in the future.\n\nPiperOrigin-RevId: 655127611","shortMessageHtmlLink":"#sdy Initial set of changes to allow for lowering to the Shardy dialect."}},{"before":"e74d98a0a19179b1024cff3de79b1e123b5af601","after":"4e56ae4cafd5a8a2849325f8e2ee0607554b54a2","ref":"refs/heads/test_648711460","pushedAt":"2024-07-23T12:48:52.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"#sdy Initial set of changes to allow for lowering to the Shardy dialect.\n\nThe OpenXLA project is working on an open source, MLIR, named-axis based propagation (and in the future SP<D partitioning) system that will be dialect agnostic (would work for any dialect - MHLO, StableHLO, YourDialect). We plan on having frontends like JAX and PyTorch target this when using XLA and wanting SPMD propagation/partitioning. See www.github.com/openxla/shardy for more info.\n\nCurrently Shardy is implemented inside the XLA compiler, requiring us to round-trip between StableHLO and HLO with `mhlo.sharding`s. But we will eventually make Shardy the first pass in the XLA pipeline while it's still working on StableHLO. Partitioning (the system that adds the collectives like all-gathers/all-reduces) will still be the GSPMD Partitioner, but next year the Shardy partitioner will be developed, allowing for propagation and partitioning to be completely in MLIR and the first pass in the pipeline. So then we'd have:\n1. Traced jaxpr\n2. Jaxpr -> StableHLO\n3. StableHLO with Shardy propagation\n4. StableHLO with Shardy partitioning\n5. StableHLO -> HLO\n6. XLA optimizations\n\nThe following test:\n\n```py\ndef test_sdy_lowering(self):\n  mesh = jtu.create_global_mesh((4, 2), ('x', 'y'))\n  np_inp = np.arange(16).reshape(8, 2)\n  s = jax.sharding.NamedSharding(mesh, P('x', 'y'))\n  arr = jax.device_put(np_inp, s)\n\n  @partial(jax.jit, out_shardings=s)\n  def f(x):\n    return x * 2\n\n  print(f.lower(arr).as_text())\n```\n\noutputs:\n\n```\nmodule @jit_f attributes {mhlo.num_partitions = 8 : i32, mhlo.num_replicas = 1 : i32} {\n  sdy.mesh @mesh = <\"x\"=4, \"y\"=2>\n  func.func public @main(%arg0: tensor<8x2xi64> {mhlo.layout_mode = \"{1,0}\", sdy.sharding = #sdy.sharding<@mesh, [{\"x\"}, {\"y\"}]>}) -> (tensor<8x2xi64> {jax.result_info = \"\", mhlo.layout_mode = \"default\", sdy.sharding = #sdy.sharding<@mesh, [{\"x\"}, {\"y\"}]>}) {\n    %c = stablehlo.constant dense<2> : tensor<i64>\n    %0 = stablehlo.broadcast_in_dim %c, dims = [] : (tensor<i64>) -> tensor<8x2xi64>\n    %1 = stablehlo.multiply %arg0, %0 : tensor<8x2xi64>\n    return %1 : tensor<8x2xi64>\n  }\n}\n```\n\nShardy will be hidden behind the `jax_use_shardy_partitioner` flag initially before becoming enabled by default in the future.\n\nPiperOrigin-RevId: 655127611","shortMessageHtmlLink":"#sdy Initial set of changes to allow for lowering to the Shardy dialect."}},{"before":null,"after":"8f59392388bde67332f1f0cb9d2d176122c6124f","ref":"refs/heads/test_651748318","pushedAt":"2024-07-23T12:40:04.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:GPU] Don't upcast supported fp8 dot operands when is are inside Triton fusion.\n\nKeep normalizing fp8 outside of Triton, but in the Triton fused computations, certain operand type combinations are fine.\n\nPiperOrigin-RevId: 651748318","shortMessageHtmlLink":"[XLA:GPU] Don't upcast supported fp8 dot operands when is are inside …"}},{"before":"af9882015f8668d370ff230e86897709cf9d2aee","after":"ddba514b5bdfbc25e73e37071f8ad80930ca2c5d","ref":"refs/heads/main","pushedAt":"2024-07-23T12:33:17.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"PR #15210: Require Matching Types in Layer Norm Fusion\n\nImported from GitHub PR https://github.com/openxla/xla/pull/15210\n\nDisables the fusion of layer norm patterns when the input and output types are not the same.\nCopybara import of the project:\n\n--\n4872f4413abbbc5e330a36f1c2ce5a5264a61fd7 by Philipp Hack <phack@nvidia.com>:\n\nDisables layer norm fusion when the input and output types differ.\n\nMerging this change closes #15210\n\nCOPYBARA_INTEGRATE_REVIEW=https://github.com/openxla/xla/pull/15210 from philipphack:u_layer_type_mismatch_xla 4872f4413abbbc5e330a36f1c2ce5a5264a61fd7\nPiperOrigin-RevId: 655126293","shortMessageHtmlLink":"PR <a class=\"issue-link js-issue-link\" data-error-text=\"Failed to load title\" data-id=\"2423994145\" data-permission-text=\"Title is private\" data-url=\"https://github.com/openxla/xla/issues/15210\" data-hovercard-type=\"pull_request\" data-hovercard-url=\"/openxla/xla/pull/15210/hovercard\" href=\"https://github.com/openxla/xla/pull/15210\">#15210</a>: Require Matching Types in Layer Norm Fusion"}},{"before":null,"after":"e57258f605e6df6bf61ac55632e939b5badd6c22","ref":"refs/heads/fix_cudnn_test","pushedAt":"2024-07-23T12:10:31.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"sergachev","name":"Ilia Sergachev","path":"/sergachev","primaryAvatarUrl":"https://avatars.githubusercontent.com/u/1894984?s=80&v=4"},"commit":{"message":"[GPU] Fix cuDNN workspace test condition.","shortMessageHtmlLink":"[GPU] Fix cuDNN workspace test condition."}},{"before":"d1afe17c5b9843e9695888b0722b7537f5c0f34b","after":"e74d98a0a19179b1024cff3de79b1e123b5af601","ref":"refs/heads/test_648711460","pushedAt":"2024-07-23T12:09:36.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"#sdy Initial set of changes to allow for lowering to the Shardy dialect.\n\nThe OpenXLA project is working on an open source, MLIR, named-axis based propagation (and in the future SP<D partitioning) system that will be dialect agnostic (would work for any dialect - MHLO, StableHLO, YourDialect). We plan on having frontends like JAX and PyTorch target this when using XLA and wanting SPMD propagation/partitioning. See www.github.com/openxla/shardy for more info.\n\nCurrently Shardy is implemented inside the XLA compiler, requiring us to round-trip between StableHLO and HLO with `mhlo.sharding`s. But we will eventually make Shardy the first pass in the XLA pipeline while it's still working on StableHLO. Partitioning (the system that adds the collectives like all-gathers/all-reduces) will still be the GSPMD Partitioner, but next year the Shardy partitioner will be developed, allowing for propagation and partitioning to be completely in MLIR and the first pass in the pipeline. So then we'd have:\n1. Traced jaxpr\n2. Jaxpr -> StableHLO\n3. StableHLO with Shardy propagation\n4. StableHLO with Shardy partitioning\n5. StableHLO -> HLO\n6. XLA optimizations\n\nThe following test:\n\n```py\ndef test_sdy_lowering(self):\n  mesh = jtu.create_global_mesh((4, 2), ('x', 'y'))\n  np_inp = np.arange(16).reshape(8, 2)\n  s = jax.sharding.NamedSharding(mesh, P('x', 'y'))\n  arr = jax.device_put(np_inp, s)\n\n  @partial(jax.jit, out_shardings=s)\n  def f(x):\n    return x * 2\n\n  print(f.lower(arr).as_text())\n```\n\noutputs:\n\n```\nmodule @jit_f attributes {mhlo.num_partitions = 8 : i32, mhlo.num_replicas = 1 : i32} {\n  sdy.mesh @mesh = <\"x\"=4, \"y\"=2>\n  func.func public @main(%arg0: tensor<8x2xi64> {mhlo.layout_mode = \"{1,0}\", sdy.sharding = #sdy.sharding<@mesh, [{\"x\"}, {\"y\"}]>}) -> (tensor<8x2xi64> {jax.result_info = \"\", mhlo.layout_mode = \"default\", sdy.sharding = #sdy.sharding<@mesh, [{\"x\"}, {\"y\"}]>}) {\n    %c = stablehlo.constant dense<2> : tensor<i64>\n    %0 = stablehlo.broadcast_in_dim %c, dims = [] : (tensor<i64>) -> tensor<8x2xi64>\n    %1 = stablehlo.multiply %arg0, %0 : tensor<8x2xi64>\n    return %1 : tensor<8x2xi64>\n  }\n}\n```\n\nShardy will be hidden behind the `jax_use_shardy_partitioner` flag initially before becoming enabled by default in the future.\n\nPiperOrigin-RevId: 648711460","shortMessageHtmlLink":"#sdy Initial set of changes to allow for lowering to the Shardy dialect."}},{"before":"0dc1d2c636d505975f44de04d83a3a1f913d1adc","after":"1b3cca09d6e9773c99018211ace517b5f5b7ebc6","ref":"refs/heads/test_654795065","pushedAt":"2024-07-23T11:57:13.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"Integrate LLVM at llvm/llvm-project@acc159aea1e6\n\nUpdates LLVM usage to match\n[acc159aea1e6](https://github.com/llvm/llvm-project/commit/acc159aea1e6)\n\nPiperOrigin-RevId: 654795065","shortMessageHtmlLink":"Integrate LLVM at <a class=\"commit-link\" data-hovercard-type=\"commit\" data-hovercard-url=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6/hovercard\" href=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6\">llvm/llvm-project@<tt>acc159aea1e6</tt></a>"}},{"before":"46f6c4764bd774a14d5cfe7f36e40d44f9147c88","after":"af9882015f8668d370ff230e86897709cf9d2aee","ref":"refs/heads/main","pushedAt":"2024-07-23T11:35:00.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"PR #15149: [XLA:GPU] print cudnn frontend check_support error message\n\nImported from GitHub PR https://github.com/openxla/xla/pull/15149\n\nCopybara import of the project:\n\n--\n0e5c2fa7e539fe97f7c5b8aadfc3fedd38ad4232 by cjkkkk <ske@nvidia.com>:\n\ninit\n\n--\n8565f356e82549e05b9d46e8a92f3e49a4370f3b by cjkkkk <ske@nvidia.com>:\n\nupdate function signature\n\nMerging this change closes #15149\n\nCOPYBARA_INTEGRATE_REVIEW=https://github.com/openxla/xla/pull/15149 from Cjkkkk:print_cudnn_check_support_message 8565f356e82549e05b9d46e8a92f3e49a4370f3b\nPiperOrigin-RevId: 655113338","shortMessageHtmlLink":"PR <a class=\"issue-link js-issue-link\" data-error-text=\"Failed to load title\" data-id=\"2420064197\" data-permission-text=\"Title is private\" data-url=\"https://github.com/openxla/xla/issues/15149\" data-hovercard-type=\"pull_request\" data-hovercard-url=\"/openxla/xla/pull/15149/hovercard\" href=\"https://github.com/openxla/xla/pull/15149\">#15149</a>: [XLA:GPU] print cudnn frontend check_support error message"}},{"before":"4089aadec1b6186c3d537c5c3d4eba4ba89dd8d5","after":"46f6c4764bd774a14d5cfe7f36e40d44f9147c88","ref":"refs/heads/main","pushedAt":"2024-07-23T11:32:48.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"PR #15163: [NVIDIA GPU] Skip processing for trip count 1 in loop double buffer unrolling\n\nImported from GitHub PR https://github.com/openxla/xla/pull/15163\n\nA minor improvement to loop double buffer unrolling pass to skip processing for loops with trip count =1\nCopybara import of the project:\n\n--\nf6bccd4612dde10cd020141f804523a75d9c84a2 by TJ Xu <tjx@nvidia.com>:\n\nSkip processing for trip count 1\n\n--\n19514326f83ee91f712ffbd74cec90b19db77df7 by TJ Xu <tjx@nvidia.com>:\n\nadded a test case\n\nMerging this change closes #15163\n\nCOPYBARA_INTEGRATE_REVIEW=https://github.com/openxla/xla/pull/15163 from Tixxx:tixxx/double_buffer_skip-1 19514326f83ee91f712ffbd74cec90b19db77df7\nPiperOrigin-RevId: 655111494","shortMessageHtmlLink":"PR <a class=\"issue-link js-issue-link\" data-error-text=\"Failed to load title\" data-id=\"2421203988\" data-permission-text=\"Title is private\" data-url=\"https://github.com/openxla/xla/issues/15163\" data-hovercard-type=\"pull_request\" data-hovercard-url=\"/openxla/xla/pull/15163/hovercard\" href=\"https://github.com/openxla/xla/pull/15163\">#15163</a>: [NVIDIA GPU] Skip processing for trip count 1 in loop doub…"}},{"before":"61d5fa148c384b630f94795f7064aff5da164f31","after":"b51ad7f5e0d95841c3f0a01ef702711aabcf8867","ref":"refs/heads/test_655104496","pushedAt":"2024-07-23T11:21:20.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:CPU] Add runtime check for whether `batch-norm-grad` is rewritten\n\nIn the current runtime, emitting for op `batch-norm-grad` is not supported by design.\nThe op is expected to be rewritten by another HLO pass before ever reaching the emit phase.\nThis CL adds a runtime check  for whether this op was actually rewritten and returns an explicit message if it wasn't.\nAlso includes a new unit test covering existing and new functionality: batch_norm_grad_test.cc.\n\nPiperOrigin-RevId: 655104496","shortMessageHtmlLink":"[XLA:CPU] Add runtime check for whether <code>batch-norm-grad</code> is rewritten"}},{"before":null,"after":"61d5fa148c384b630f94795f7064aff5da164f31","ref":"refs/heads/test_655104496","pushedAt":"2024-07-23T11:13:21.000Z","pushType":"branch_creation","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:CPU] Add runtime check for whether `batch-norm-grad` is rewritten\n\nIn the current runtime, emitting for op `batch-norm-grad` is not supported by design.\nThe op is expected to be rewritten by another HLO pass before ever reaching the emit phase.\nThis CL adds a runtime check  for whether this op was actually rewritten and returns an explicit message if it wasn't.\nAlso includes a new unit test covering existing and new functionality: batch_norm_grad_test.cc.\n\nPiperOrigin-RevId: 655104496","shortMessageHtmlLink":"[XLA:CPU] Add runtime check for whether <code>batch-norm-grad</code> is rewritten"}},{"before":"e918698506ecd1f967d50cdf8ca69a2ee7248394","after":"0dc1d2c636d505975f44de04d83a3a1f913d1adc","ref":"refs/heads/test_654795065","pushedAt":"2024-07-23T11:04:10.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"Integrate LLVM at llvm/llvm-project@acc159aea1e6\n\nUpdates LLVM usage to match\n[acc159aea1e6](https://github.com/llvm/llvm-project/commit/acc159aea1e6)\n\nPiperOrigin-RevId: 654795065","shortMessageHtmlLink":"Integrate LLVM at <a class=\"commit-link\" data-hovercard-type=\"commit\" data-hovercard-url=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6/hovercard\" href=\"https://github.com/llvm/llvm-project/commit/acc159aea1e6\">llvm/llvm-project@<tt>acc159aea1e6</tt></a>"}},{"before":"4089aadec1b6186c3d537c5c3d4eba4ba89dd8d5","after":null,"ref":"refs/heads/test_653988963","pushedAt":"2024-07-23T11:03:29.000Z","pushType":"branch_deletion","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"}},{"before":"bab3a715d1232505ae9b1e5d63468ef0c4c63ac6","after":"4089aadec1b6186c3d537c5c3d4eba4ba89dd8d5","ref":"refs/heads/main","pushedAt":"2024-07-23T11:03:27.000Z","pushType":"push","commitsCount":1,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:GPU] Only compute tile offset indexing maps for instructions when it is needed.\n\nComputing tile offset indexing maps is very expensive and not always necessary. When we're at Cost Model/Tiling stage, we only use the indexing map to deduplicate instruction. We can predict if we'll need an indexing map for a particular instruction by computation and comparing a parts of the hash.\n\nPiperOrigin-RevId: 655103934","shortMessageHtmlLink":"[XLA:GPU] Only compute tile offset indexing maps for instructions whe…"}},{"before":"8bd7d27c2d44897aa05e978b75f4b5324e672219","after":"4089aadec1b6186c3d537c5c3d4eba4ba89dd8d5","ref":"refs/heads/test_653988963","pushedAt":"2024-07-23T11:03:26.000Z","pushType":"force_push","commitsCount":0,"pusher":{"login":"copybara-service[bot]","name":null,"path":"/apps/copybara-service","primaryAvatarUrl":"https://avatars.githubusercontent.com/in/44061?s=80&v=4"},"commit":{"message":"[XLA:GPU] Only compute tile offset indexing maps for instructions when it is needed.\n\nComputing tile offset indexing maps is very expensive and not always necessary. When we're at Cost Model/Tiling stage, we only use the indexing map to deduplicate instruction. We can predict if we'll need an indexing map for a particular instruction by computation and comparing a parts of the hash.\n\nPiperOrigin-RevId: 655103934","shortMessageHtmlLink":"[XLA:GPU] Only compute tile offset indexing maps for instructions whe…"}}],"hasNextPage":true,"hasPreviousPage":false,"activityType":"all","actor":null,"timePeriod":"all","sort":"DESC","perPage":30,"cursor":"djE6ks8AAAAEhvIkkgA","startCursor":null,"endCursor":null}},"title":"Activity · openxla/xla"}