[Metax]Fix metax backend bugs #432
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
1. update multi-backend code 2. fix argmin op might test failed under int types
mx-flaggems-user
changed the title
Galaxy1458
approved these changes
Jan 24, 2025
DuanYaQi
pushed a commit
that referenced
this pull request
Feb 11, 2025
1. update multi-backend code 2. fix argmin op might test failed under int types Co-authored-by: mx-flaggems-user <[email protected]>
machuanjiang
pushed a commit
that referenced
this pull request
Feb 25, 2025
1. update multi-backend code 2. fix argmin op might test failed under int types Co-authored-by: mx-flaggems-user <[email protected]>
machuanjiang
pushed a commit
that referenced
this pull request
Feb 25, 2025
1. update multi-backend code 2. fix argmin op might test failed under int types Co-authored-by: mx-flaggems-user <[email protected]>
Galaxy1458
added a commit
that referenced
this pull request
Feb 27, 2025
* add triton_musa submodule * Modify testcase from cuda to musa. * Workaround for musa testcase. * modify test_unary_pointwise_ops from cuda to musa * modify test_reduction_ops from cuda to musa * Fix bug of reduceOp and shared memory. * fix dropout bug. * fix softmax exceeds shared memory error * Promote cpu reference accuracy to float32 * Modify benchmark performance test script to musa * Add torch_musa unsupported op test case. * Support bert model. * Update submodule url. * Comment v3/v4 test case. * fix: vectornorm upcast to fp64 * fix: group_norm modify case because of out-of-shared-memory * Promote golden accuracy from fp32 to fp64. * Rebase on master commit 1e49d6. * Update triton_musa submodule. * align perf utils to profiling pack 0717 * rebase on master commit 9000685 Signed-off-by: Jian Li <[email protected]> * Support op cumsum. config: {BLOCK_M: 8, num_warps: 8} will cause the number of registers within a single thread to be exceeded when the tensor shape is 4096 * 2304, so reduce BLOCK_M to 4 to supprot cumsum. * fix embedding tensor usage Signed-off-by: Jian Li <[email protected]> * uncomment supported op test Signed-off-by: Jian Li <[email protected]> * Support isclose() and allclose() - Torch_musa does not support fp64 input type, so CPU is used as a reference * Open up some tests that have already passed - Does not support test_accuracy_groupnorm - Some use cases have accuracy issues in test_embedding * rebase on master commit 801377f Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2e55d66 Signed-off-by: Jian Li <[email protected]> * fix distribution ops warps num Signed-off-by: Jian Li <[email protected]> * rebase on master commit a156268 Signed-off-by: Jian Li <[email protected]> * Support op argmax. * adapt masked_fill and dropout resolution Signed-off-by: Jian Li <[email protected]> * fix _log2 in topk op Signed-off-by: Jian Li <[email protected]> * rebase on master commit a1138bf Signed-off-by: Jian Li <[email protected]> * Open BF16 * rebase on master commit edb09f0 Signed-off-by: Jian Li <[email protected]> * fix div_trunc and div_floor Signed-off-by: Jian Li <[email protected]> * fix div_trunc cont. Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2c9ae67 Signed-off-by: Jian Li <[email protected]> * rebase on master commit 204f3d4 Signed-off-by: Jian Li <[email protected]> * Support op vstack * rebase on master commit 4e5081 rebase on master commit 8f669d * SW-46066: remove submodule(triton_musa) in flaggems Signed-off-by: chuanjiang.ma <[email protected]> * [Operator] Fix vstack (#237) Modify the function parameter type declaration so that it can run in python 3.8 --------- Co-authored-by: zhengyang <[email protected]> * [Operator] Add upsample_nearest2d op [MooreThreads] (#193) * SW-46093: rebase on master commit 98924c Signed-off-by: jiaqi.wang <[email protected]> * succeed for norm (#233) Add _weight_norm op, while the original _weight_norm op changed to _weight_norm_interface op. * [Operator] slice&select scatter (#143) * add Ops & UT & Bench * add full zero ones Ops & UT & Bench * split normal op * [Operator] init slice&select scatter * code format * PR comment * split test_special_ops * add K-S test * split special perf * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * resolve conflict * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * test for op * test for op * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * add argparse * fix desc * fix num * Update test_specific_ops.py * split UT files * fix * fix * [Operator] Optimize CrossEntropyLoss (#131) reimplement cross_entropy_loss forward and backward support; indices/probabilities/weight/reduction/ignore_index/label_smoothing; perform better than torch eager on large scale tensors * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * [Test] Test for op (#151) * [chore] solve slice&select scatter's test cases * [fix] fix slice&select scatter's test cases * [chore] remove out-of-range indices in select_scatter's test cases * [chore] simplify slice_scatter's test cases * [fix] Added range that is deleted by mistake * Merge branch 'master' into slice&select_scatter * [chore] reformat * [fix] typo * [chore] Considering perf, pause the replacement of some aTen operators * slice_scatter * select_scatter * index_select * [fix] Add libentry in op.cumsum * [fix] Del slice&select scatter's perf tests * [Chore] Add pytest mark for slice&select scatter's test * [Fix] Correct slice_scatter test * [Fix] Replace CPU Tensor --------- Co-authored-by: Bowen12992 <[email protected]> Co-authored-by: Tongxin Bai <[email protected]> Co-authored-by: Clement Chan <[email protected]> Co-authored-by: Bowen <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator] Add slice&select_scatter's benchmark (#262) * benchmark fix (#229) * benchmark fix * add seven new testing parameters * move shapes info to yaml file * Added the BenchmarkMetrics & BenchmarkResult abstraction * Specializing slice_scatter. (#270) * specializing slice_scatter. WIP. * polish and refine 2d_inner cases. * fix slice_scatter error on 1d inputs. * test slice_scatter fallback * Enhance Benchmarking for repeat_interleave Operation (#274) * Relocate select and slice benchmarks to test_select_and_slice_perf.py * sort keys for summary result * clean cuda cache after benchmark * fix repeat_interleave * modify format for summary info * SW-46093: rebase on master commit 2bd92c Signed-off-by: jiaqi.wang <[email protected]> * add test entry & use skip to avoid error Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: update mt test script by waiting frontend Signed-off-by: chuanjiang.ma <[email protected]> * SW-46093: adjust the skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: currently test accuracy ref to cpu Signed-off-by: chuanjiang.ma <[email protected]> * SW-47833: skip one case for half type op not supported by torch_musa Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: update test suite setup Signed-off-by: machuanjiang <[email protected]> * mthreads: fix review comments Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: skip weight_nor testing for driver complaining 1. one test in special_op test change the device type from cuda to musa Signed-off-by: chuanjiang.ma <[email protected]> * fix #304 (#308) fix max: input tensor with big shape may cause "numel (2097152) exceeds triton maximum tensor numel" * add program_id & num_programs that returns tensor in tl.int64 (#327) * add program_id & num_programs that returns tensor in tl.int64 to avoid integer overflow * fix bmm * fix vector norm, treat empty dim list as full-reduce; use tle in fused ops, too * block pointer requires offsets/block shape to be int32 * [bugfix] replace tle with tl for operators using philox (#337) * [bugfix] replace tle with tl for operators using philox * [bugfix] tle in rand * [bugfix] uniform * Remove INT_DTYPES from isfinite operation in benchmark and tests (#336) * [Operator] Fix full & full_like op (#338) Co-authored-by: zhengyang <[email protected]> Co-authored-by: Bowen <[email protected]> * [Operator] Add diag_embed (#288) * diag_embed operator * add diag_embed ops * resolve conflict * SW-46093: add test python script for 50 ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: support output compared_speedup for benchmark plot Signed-off-by: root <chuanjiang.ma> * SW-46093: add an optin to control not output compared speedup Signed-off-by: root <chuanjiang.ma> * SW-46093: cherry-pick updates from commit 54a471 to ca13b7 Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Adapt Multiple backends code for MUSA Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: Adjust skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Update _mthreads backend config Signed-off-by: jiaqi.wang <[email protected]> * add pytest alluredir options for generating report * [Operator] Add sort op (#322) * [Operator] Add sort op --------- Co-authored-by: zhengyang <[email protected]> Co-authored-by: MARD1NO <[email protected]> * [Muti_backend]part_2_device (#344) * new feature, muti_backend * update auto_tune_module * update auto_tune_module * update auto_tune_module * update __init__ * rebase * fix bug * modifiy auto_tune_config * fix bug * fix bug * update * update * update scatter&gather * fix auto_tune * add gen_torch_device_fn * fix codestyle * fix codestyle * Modify code based on comments * Modify gen_impl with loops instead of recursion * Update code structure * Polish code * update * Polish code * Modify code based on comments * modify based on comment * Modify code based on comments * update * final fix * modify_device_unify * fix bug * update * remove '-> list' * modify * modify * modify * fix bug * muti_backend_part_3 enhancement of ability (#361) * muti_backend_part_3 enhancement of ability * fix bug * modify * modify * fix bug * modify * SW-49945: adjust merged code Signed-off-by: jiaqi.wang <[email protected]> * Reimplement slice_scatter/select_scatter using pointwise_dynamic (#358) * merged with origin. * updated * renames copy_ to copy, fixes benchmark code. * fixes benchmark. * handle self-overlapping input in select_scatter and slice_scatter --------- Co-authored-by: Clement Chan <[email protected]> * modify code (#371) * modify code * fix bug * fix bug * modify * fix * fix * [Polish code] Adjust some directory structures and variable names (#373) * modify code * fix bug * fix bug * modify * fix * fix * polish code * polish code * polish code * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [Muti backend] add heuristics_config for muti_bacnkend (#377) * muti_backend * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [bugfix] disable more shapes from blas benchmark (#375) * add heuristics_config * modify * modify * modify * modify --------- Co-authored-by: StrongSpoon <[email protected]> * SW-49945: update configuration for mthreads Signed-off-by: jiaqi.wang <[email protected]> * [benchmark] skip perf test of cummin when triton < 3.0 (#385) * SW-46093: fix test configuration Signed-off-by: jiaqi.wang <[email protected]> * SW-50265: optimize the perf of gelu, tanh op and most other pointwise ops * Gather bwd (#382) * add gather_backward op * add debug log in * perf gather backward * rebased with master * scatter rewrite done. * scatter handling internally overlapping input. * Scatter reduce now uses atomics. * remove fp16 from scatter reduce UT. * sets threadblock size to 128 for scatter. * Change atomic memory order to relaxed in scatter. --------- Co-authored-by: awayzjj <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator]Fix metax backend bugs (#432) 1. update multi-backend code 2. fix argmin op might test failed under int types Co-authored-by: mx-flaggems-user <[email protected]> * [TEST] fix error in argmin UT when dtype=int16 (#431) Co-authored-by: junjian.zhan <[email protected]> * SW-49945: Update configurations for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-47833: Fix case resolve_conj and fill. 1. resolve_conj: ref to this link: https://jira.mthreads.com/browse/MTAI-1530 2. fill: torch_musa does not support case torch.fill(dtype=cpu, dtype=musa). * add backward of conv2d (#365) * add backward of conv2d * delete useless code * format code of tests * modify configs for tuning * modify autotune config * delete test flag * delete useless type convert --------- Co-authored-by: Jiang Bin <[email protected]> * SW-52470: Cherry-pick updates from commit f1ba20c to 5f31f35. * SW-49945: rebase for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-52470: adjust skipped ops for merge Signed-off-by: jiaqi.wang <[email protected]> * Mthreads: adjust ops for multi-backends merge Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * add triton_musa submodule * Modify testcase from cuda to musa. * Workaround for musa testcase. * modify test_unary_pointwise_ops from cuda to musa * modify test_reduction_ops from cuda to musa * Fix bug of reduceOp and shared memory. * fix dropout bug. * fix softmax exceeds shared memory error * Promote cpu reference accuracy to float32 * Modify benchmark performance test script to musa * Add torch_musa unsupported op test case. * Support bert model. * Update submodule url. * Comment v3/v4 test case. * fix: vectornorm upcast to fp64 * fix: group_norm modify case because of out-of-shared-memory * Promote golden accuracy from fp32 to fp64. * Rebase on master commit 1e49d6. * Update triton_musa submodule. * align perf utils to profiling pack 0717 * rebase on master commit 9000685 Signed-off-by: Jian Li <[email protected]> * Support op cumsum. config: {BLOCK_M: 8, num_warps: 8} will cause the number of registers within a single thread to be exceeded when the tensor shape is 4096 * 2304, so reduce BLOCK_M to 4 to supprot cumsum. * fix embedding tensor usage Signed-off-by: Jian Li <[email protected]> * uncomment supported op test Signed-off-by: Jian Li <[email protected]> * Support isclose() and allclose() - Torch_musa does not support fp64 input type, so CPU is used as a reference * Open up some tests that have already passed - Does not support test_accuracy_groupnorm - Some use cases have accuracy issues in test_embedding * rebase on master commit 801377f Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2e55d66 Signed-off-by: Jian Li <[email protected]> * fix distribution ops warps num Signed-off-by: Jian Li <[email protected]> * rebase on master commit a156268 Signed-off-by: Jian Li <[email protected]> * Support op argmax. * adapt masked_fill and dropout resolution Signed-off-by: Jian Li <[email protected]> * fix _log2 in topk op Signed-off-by: Jian Li <[email protected]> * rebase on master commit a1138bf Signed-off-by: Jian Li <[email protected]> * Open BF16 * rebase on master commit edb09f0 Signed-off-by: Jian Li <[email protected]> * fix div_trunc and div_floor Signed-off-by: Jian Li <[email protected]> * fix div_trunc cont. Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2c9ae67 Signed-off-by: Jian Li <[email protected]> * rebase on master commit 204f3d4 Signed-off-by: Jian Li <[email protected]> * Support op vstack * rebase on master commit 4e5081 rebase on master commit 8f669d * SW-46066: remove submodule(triton_musa) in flaggems Signed-off-by: chuanjiang.ma <[email protected]> * [Operator] Fix vstack (#237) Modify the function parameter type declaration so that it can run in python 3.8 --------- Co-authored-by: zhengyang <[email protected]> * [Operator] Add upsample_nearest2d op [MooreThreads] (#193) * SW-46093: rebase on master commit 98924c Signed-off-by: jiaqi.wang <[email protected]> * succeed for norm (#233) Add _weight_norm op, while the original _weight_norm op changed to _weight_norm_interface op. * [Operator] slice&select scatter (#143) * add Ops & UT & Bench * add full zero ones Ops & UT & Bench * split normal op * [Operator] init slice&select scatter * code format * PR comment * split test_special_ops * add K-S test * split special perf * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * resolve conflict * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * test for op * test for op * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * add argparse * fix desc * fix num * Update test_specific_ops.py * split UT files * fix * fix * [Operator] Optimize CrossEntropyLoss (#131) reimplement cross_entropy_loss forward and backward support; indices/probabilities/weight/reduction/ignore_index/label_smoothing; perform better than torch eager on large scale tensors * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * [Test] Test for op (#151) * [chore] solve slice&select scatter's test cases * [fix] fix slice&select scatter's test cases * [chore] remove out-of-range indices in select_scatter's test cases * [chore] simplify slice_scatter's test cases * [fix] Added range that is deleted by mistake * Merge branch 'master' into slice&select_scatter * [chore] reformat * [fix] typo * [chore] Considering perf, pause the replacement of some aTen operators * slice_scatter * select_scatter * index_select * [fix] Add libentry in op.cumsum * [fix] Del slice&select scatter's perf tests * [Chore] Add pytest mark for slice&select scatter's test * [Fix] Correct slice_scatter test * [Fix] Replace CPU Tensor --------- Co-authored-by: Bowen12992 <[email protected]> Co-authored-by: Tongxin Bai <[email protected]> Co-authored-by: Clement Chan <[email protected]> Co-authored-by: Bowen <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator] Add slice&select_scatter's benchmark (#262) * benchmark fix (#229) * benchmark fix * add seven new testing parameters * move shapes info to yaml file * Added the BenchmarkMetrics & BenchmarkResult abstraction * Specializing slice_scatter. (#270) * specializing slice_scatter. WIP. * polish and refine 2d_inner cases. * fix slice_scatter error on 1d inputs. * test slice_scatter fallback * Enhance Benchmarking for repeat_interleave Operation (#274) * Relocate select and slice benchmarks to test_select_and_slice_perf.py * sort keys for summary result * clean cuda cache after benchmark * fix repeat_interleave * modify format for summary info * SW-46093: rebase on master commit 2bd92c Signed-off-by: jiaqi.wang <[email protected]> * add test entry & use skip to avoid error Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: update mt test script by waiting frontend Signed-off-by: chuanjiang.ma <[email protected]> * SW-46093: adjust the skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: currently test accuracy ref to cpu Signed-off-by: chuanjiang.ma <[email protected]> * SW-47833: skip one case for half type op not supported by torch_musa Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: update test suite setup Signed-off-by: machuanjiang <[email protected]> * mthreads: fix review comments Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: skip weight_nor testing for driver complaining 1. one test in special_op test change the device type from cuda to musa Signed-off-by: chuanjiang.ma <[email protected]> * fix #304 (#308) fix max: input tensor with big shape may cause "numel (2097152) exceeds triton maximum tensor numel" * add program_id & num_programs that returns tensor in tl.int64 (#327) * add program_id & num_programs that returns tensor in tl.int64 to avoid integer overflow * fix bmm * fix vector norm, treat empty dim list as full-reduce; use tle in fused ops, too * block pointer requires offsets/block shape to be int32 * [bugfix] replace tle with tl for operators using philox (#337) * [bugfix] replace tle with tl for operators using philox * [bugfix] tle in rand * [bugfix] uniform * Remove INT_DTYPES from isfinite operation in benchmark and tests (#336) * [Operator] Fix full & full_like op (#338) Co-authored-by: zhengyang <[email protected]> Co-authored-by: Bowen <[email protected]> * [Operator] Add diag_embed (#288) * diag_embed operator * add diag_embed ops * resolve conflict * SW-46093: add test python script for 50 ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: support output compared_speedup for benchmark plot Signed-off-by: root <chuanjiang.ma> * SW-46093: add an optin to control not output compared speedup Signed-off-by: root <chuanjiang.ma> * SW-46093: cherry-pick updates from commit 54a471 to ca13b7 Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Adapt Multiple backends code for MUSA Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: Adjust skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Update _mthreads backend config Signed-off-by: jiaqi.wang <[email protected]> * add pytest alluredir options for generating report * [Operator] Add sort op (#322) * [Operator] Add sort op --------- Co-authored-by: zhengyang <[email protected]> Co-authored-by: MARD1NO <[email protected]> * [Muti_backend]part_2_device (#344) * new feature, muti_backend * update auto_tune_module * update auto_tune_module * update auto_tune_module * update __init__ * rebase * fix bug * modifiy auto_tune_config * fix bug * fix bug * update * update * update scatter&gather * fix auto_tune * add gen_torch_device_fn * fix codestyle * fix codestyle * Modify code based on comments * Modify gen_impl with loops instead of recursion * Update code structure * Polish code * update * Polish code * Modify code based on comments * modify based on comment * Modify code based on comments * update * final fix * modify_device_unify * fix bug * update * remove '-> list' * modify * modify * modify * fix bug * muti_backend_part_3 enhancement of ability (#361) * muti_backend_part_3 enhancement of ability * fix bug * modify * modify * fix bug * modify * SW-49945: adjust merged code Signed-off-by: jiaqi.wang <[email protected]> * Reimplement slice_scatter/select_scatter using pointwise_dynamic (#358) * merged with origin. * updated * renames copy_ to copy, fixes benchmark code. * fixes benchmark. * handle self-overlapping input in select_scatter and slice_scatter --------- Co-authored-by: Clement Chan <[email protected]> * modify code (#371) * modify code * fix bug * fix bug * modify * fix * fix * [Polish code] Adjust some directory structures and variable names (#373) * modify code * fix bug * fix bug * modify * fix * fix * polish code * polish code * polish code * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [Muti backend] add heuristics_config for muti_bacnkend (#377) * muti_backend * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [bugfix] disable more shapes from blas benchmark (#375) * add heuristics_config * modify * modify * modify * modify --------- Co-authored-by: StrongSpoon <[email protected]> * SW-49945: update configuration for mthreads Signed-off-by: jiaqi.wang <[email protected]> * [benchmark] skip perf test of cummin when triton < 3.0 (#385) * SW-46093: fix test configuration Signed-off-by: jiaqi.wang <[email protected]> * SW-50265: optimize the perf of gelu, tanh op and most other pointwise ops * Gather bwd (#382) * add gather_backward op * add debug log in * perf gather backward * rebased with master * scatter rewrite done. * scatter handling internally overlapping input. * Scatter reduce now uses atomics. * remove fp16 from scatter reduce UT. * sets threadblock size to 128 for scatter. * Change atomic memory order to relaxed in scatter. --------- Co-authored-by: awayzjj <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator]Fix metax backend bugs (#432) 1. update multi-backend code 2. fix argmin op might test failed under int types Co-authored-by: mx-flaggems-user <[email protected]> * [TEST] fix error in argmin UT when dtype=int16 (#431) Co-authored-by: junjian.zhan <[email protected]> * SW-49945: Update configurations for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-47833: Fix case resolve_conj and fill. 1. resolve_conj: ref to this link: https://jira.mthreads.com/browse/MTAI-1530 2. fill: torch_musa does not support case torch.fill(dtype=cpu, dtype=musa). * add backward of conv2d (#365) * add backward of conv2d * delete useless code * format code of tests * modify configs for tuning * modify autotune config * delete test flag * delete useless type convert --------- Co-authored-by: Jiang Bin <[email protected]> * SW-52470: Cherry-pick updates from commit f1ba20c to 5f31f35. * SW-49945: rebase for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-52470: adjust skipped ops for merge Signed-off-by: jiaqi.wang <[email protected]> * Mthreads: adjust ops for multi-backends merge Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * MTHREADS: Checking Code Format Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * MTHREADS: fix bugs for multi-backends ops Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * MTHREADS: fix according to review issue Signed-off-by: chuanjiang.ma <[email protected]> * MTHREADS: add annotation to explain the log-file-cross calc Signed-off-by: chuanjiang.ma <[email protected]> --------- Signed-off-by: Jian Li <[email protected]> Signed-off-by: chuanjiang.ma <[email protected]> Signed-off-by: jiaqi.wang <[email protected]> Signed-off-by: machuanjiang <[email protected]> Signed-off-by: root <chuanjiang.ma> Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> Co-authored-by: yuzhe-wu <[email protected]> Co-authored-by: Buzz <[email protected]> Co-authored-by: hang.zhang <[email protected]> Co-authored-by: lingfeng.qiu <[email protected]> Co-authored-by: jialuo.bai <[email protected]> Co-authored-by: Jian Li <[email protected]> Co-authored-by: Tianjie Ling <[email protected]> Co-authored-by: jiaqi.wang <[email protected]> Co-authored-by: zhzhcookie <[email protected]> Co-authored-by: zhengyang <[email protected]> Co-authored-by: zaccur <[email protected]> Co-authored-by: TZWX-0 <[email protected]> Co-authored-by: Hiujin Gwok <[email protected]> Co-authored-by: Bowen12992 <[email protected]> Co-authored-by: Tongxin Bai <[email protected]> Co-authored-by: Clement Chan <[email protected]> Co-authored-by: Bowen <[email protected]> Co-authored-by: StrongSpoon <[email protected]> Co-authored-by: kiddyjinjin <[email protected]> Co-authored-by: haowen-han <[email protected]> Co-authored-by: niu_he <[email protected]> Co-authored-by: root <chuanjiang.ma> Co-authored-by: yu.huang <[email protected]> Co-authored-by: MARD1NO <[email protected]> Co-authored-by: Galaxy1458 <[email protected]> Co-authored-by: Sheng Wang <[email protected]> Co-authored-by: awayzjj <[email protected]> Co-authored-by: StrongSpoon <[email protected]> Co-authored-by: mx-flaggems-user <[email protected]> Co-authored-by: mx-flaggems-user <[email protected]> Co-authored-by: junjian.zhan <[email protected]> Co-authored-by: FatJhon <[email protected]> Co-authored-by: Jiang Bin <[email protected]> Co-authored-by: jiaqi.wang <jiaqi.wang @mthreads.com>
StrongSpoon
added a commit
that referenced
this pull request
Mar 17, 2025
* add triton_musa submodule * Modify testcase from cuda to musa. * Workaround for musa testcase. * modify test_unary_pointwise_ops from cuda to musa * modify test_reduction_ops from cuda to musa * Fix bug of reduceOp and shared memory. * fix dropout bug. * fix softmax exceeds shared memory error * Promote cpu reference accuracy to float32 * Modify benchmark performance test script to musa * Add torch_musa unsupported op test case. * Support bert model. * Update submodule url. * Comment v3/v4 test case. * fix: vectornorm upcast to fp64 * fix: group_norm modify case because of out-of-shared-memory * Promote golden accuracy from fp32 to fp64. * Rebase on master commit 1e49d6. * Update triton_musa submodule. * align perf utils to profiling pack 0717 * rebase on master commit 9000685 Signed-off-by: Jian Li <[email protected]> * Support op cumsum. config: {BLOCK_M: 8, num_warps: 8} will cause the number of registers within a single thread to be exceeded when the tensor shape is 4096 * 2304, so reduce BLOCK_M to 4 to supprot cumsum. * fix embedding tensor usage Signed-off-by: Jian Li <[email protected]> * uncomment supported op test Signed-off-by: Jian Li <[email protected]> * Support isclose() and allclose() - Torch_musa does not support fp64 input type, so CPU is used as a reference * Open up some tests that have already passed - Does not support test_accuracy_groupnorm - Some use cases have accuracy issues in test_embedding * rebase on master commit 801377f Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2e55d66 Signed-off-by: Jian Li <[email protected]> * fix distribution ops warps num Signed-off-by: Jian Li <[email protected]> * rebase on master commit a156268 Signed-off-by: Jian Li <[email protected]> * Support op argmax. * adapt masked_fill and dropout resolution Signed-off-by: Jian Li <[email protected]> * fix _log2 in topk op Signed-off-by: Jian Li <[email protected]> * rebase on master commit a1138bf Signed-off-by: Jian Li <[email protected]> * Open BF16 * rebase on master commit edb09f0 Signed-off-by: Jian Li <[email protected]> * fix div_trunc and div_floor Signed-off-by: Jian Li <[email protected]> * fix div_trunc cont. Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2c9ae67 Signed-off-by: Jian Li <[email protected]> * rebase on master commit 204f3d4 Signed-off-by: Jian Li <[email protected]> * Support op vstack * rebase on master commit 4e5081 rebase on master commit 8f669d * SW-46066: remove submodule(triton_musa) in flaggems Signed-off-by: chuanjiang.ma <[email protected]> * [Operator] Fix vstack (#237) Modify the function parameter type declaration so that it can run in python 3.8 --------- Co-authored-by: zhengyang <[email protected]> * [Operator] Add upsample_nearest2d op [MooreThreads] (#193) * SW-46093: rebase on master commit 98924c Signed-off-by: jiaqi.wang <[email protected]> * succeed for norm (#233) Add _weight_norm op, while the original _weight_norm op changed to _weight_norm_interface op. * [Operator] slice&select scatter (#143) * add Ops & UT & Bench * add full zero ones Ops & UT & Bench * split normal op * [Operator] init slice&select scatter * code format * PR comment * split test_special_ops * add K-S test * split special perf * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * resolve conflict * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * test for op * test for op * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * add argparse * fix desc * fix num * Update test_specific_ops.py * split UT files * fix * fix * [Operator] Optimize CrossEntropyLoss (#131) reimplement cross_entropy_loss forward and backward support; indices/probabilities/weight/reduction/ignore_index/label_smoothing; perform better than torch eager on large scale tensors * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * [Test] Test for op (#151) * [chore] solve slice&select scatter's test cases * [fix] fix slice&select scatter's test cases * [chore] remove out-of-range indices in select_scatter's test cases * [chore] simplify slice_scatter's test cases * [fix] Added range that is deleted by mistake * Merge branch 'master' into slice&select_scatter * [chore] reformat * [fix] typo * [chore] Considering perf, pause the replacement of some aTen operators * slice_scatter * select_scatter * index_select * [fix] Add libentry in op.cumsum * [fix] Del slice&select scatter's perf tests * [Chore] Add pytest mark for slice&select scatter's test * [Fix] Correct slice_scatter test * [Fix] Replace CPU Tensor --------- Co-authored-by: Bowen12992 <[email protected]> Co-authored-by: Tongxin Bai <[email protected]> Co-authored-by: Clement Chan <[email protected]> Co-authored-by: Bowen <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator] Add slice&select_scatter's benchmark (#262) * benchmark fix (#229) * benchmark fix * add seven new testing parameters * move shapes info to yaml file * Added the BenchmarkMetrics & BenchmarkResult abstraction * Specializing slice_scatter. (#270) * specializing slice_scatter. WIP. * polish and refine 2d_inner cases. * fix slice_scatter error on 1d inputs. * test slice_scatter fallback * Enhance Benchmarking for repeat_interleave Operation (#274) * Relocate select and slice benchmarks to test_select_and_slice_perf.py * sort keys for summary result * clean cuda cache after benchmark * fix repeat_interleave * modify format for summary info * SW-46093: rebase on master commit 2bd92c Signed-off-by: jiaqi.wang <[email protected]> * add test entry & use skip to avoid error Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: update mt test script by waiting frontend Signed-off-by: chuanjiang.ma <[email protected]> * SW-46093: adjust the skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: currently test accuracy ref to cpu Signed-off-by: chuanjiang.ma <[email protected]> * SW-47833: skip one case for half type op not supported by torch_musa Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: update test suite setup Signed-off-by: machuanjiang <[email protected]> * mthreads: fix review comments Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: skip weight_nor testing for driver complaining 1. one test in special_op test change the device type from cuda to musa Signed-off-by: chuanjiang.ma <[email protected]> * fix #304 (#308) fix max: input tensor with big shape may cause "numel (2097152) exceeds triton maximum tensor numel" * add program_id & num_programs that returns tensor in tl.int64 (#327) * add program_id & num_programs that returns tensor in tl.int64 to avoid integer overflow * fix bmm * fix vector norm, treat empty dim list as full-reduce; use tle in fused ops, too * block pointer requires offsets/block shape to be int32 * [bugfix] replace tle with tl for operators using philox (#337) * [bugfix] replace tle with tl for operators using philox * [bugfix] tle in rand * [bugfix] uniform * Remove INT_DTYPES from isfinite operation in benchmark and tests (#336) * [Operator] Fix full & full_like op (#338) Co-authored-by: zhengyang <[email protected]> Co-authored-by: Bowen <[email protected]> * [Operator] Add diag_embed (#288) * diag_embed operator * add diag_embed ops * resolve conflict * SW-46093: add test python script for 50 ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: support output compared_speedup for benchmark plot Signed-off-by: root <chuanjiang.ma> * SW-46093: add an optin to control not output compared speedup Signed-off-by: root <chuanjiang.ma> * SW-46093: cherry-pick updates from commit 54a471 to ca13b7 Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Adapt Multiple backends code for MUSA Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: Adjust skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Update _mthreads backend config Signed-off-by: jiaqi.wang <[email protected]> * add pytest alluredir options for generating report * [Operator] Add sort op (#322) * [Operator] Add sort op --------- Co-authored-by: zhengyang <[email protected]> Co-authored-by: MARD1NO <[email protected]> * [Muti_backend]part_2_device (#344) * new feature, muti_backend * update auto_tune_module * update auto_tune_module * update auto_tune_module * update __init__ * rebase * fix bug * modifiy auto_tune_config * fix bug * fix bug * update * update * update scatter&gather * fix auto_tune * add gen_torch_device_fn * fix codestyle * fix codestyle * Modify code based on comments * Modify gen_impl with loops instead of recursion * Update code structure * Polish code * update * Polish code * Modify code based on comments * modify based on comment * Modify code based on comments * update * final fix * modify_device_unify * fix bug * update * remove '-> list' * modify * modify * modify * fix bug * muti_backend_part_3 enhancement of ability (#361) * muti_backend_part_3 enhancement of ability * fix bug * modify * modify * fix bug * modify * SW-49945: adjust merged code Signed-off-by: jiaqi.wang <[email protected]> * Reimplement slice_scatter/select_scatter using pointwise_dynamic (#358) * merged with origin. * updated * renames copy_ to copy, fixes benchmark code. * fixes benchmark. * handle self-overlapping input in select_scatter and slice_scatter --------- Co-authored-by: Clement Chan <[email protected]> * modify code (#371) * modify code * fix bug * fix bug * modify * fix * fix * [Polish code] Adjust some directory structures and variable names (#373) * modify code * fix bug * fix bug * modify * fix * fix * polish code * polish code * polish code * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [Muti backend] add heuristics_config for muti_bacnkend (#377) * muti_backend * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [bugfix] disable more shapes from blas benchmark (#375) * add heuristics_config * modify * modify * modify * modify --------- Co-authored-by: StrongSpoon <[email protected]> * SW-49945: update configuration for mthreads Signed-off-by: jiaqi.wang <[email protected]> * [benchmark] skip perf test of cummin when triton < 3.0 (#385) * SW-46093: fix test configuration Signed-off-by: jiaqi.wang <[email protected]> * SW-50265: optimize the perf of gelu, tanh op and most other pointwise ops * Gather bwd (#382) * add gather_backward op * add debug log in * perf gather backward * rebased with master * scatter rewrite done. * scatter handling internally overlapping input. * Scatter reduce now uses atomics. * remove fp16 from scatter reduce UT. * sets threadblock size to 128 for scatter. * Change atomic memory order to relaxed in scatter. --------- Co-authored-by: awayzjj <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator]Fix metax backend bugs (#432) 1. update multi-backend code 2. fix argmin op might test failed under int types Co-authored-by: mx-flaggems-user <[email protected]> * [TEST] fix error in argmin UT when dtype=int16 (#431) Co-authored-by: junjian.zhan <[email protected]> * SW-49945: Update configurations for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-47833: Fix case resolve_conj and fill. 1. resolve_conj: ref to this link: https://jira.mthreads.com/browse/MTAI-1530 2. fill: torch_musa does not support case torch.fill(dtype=cpu, dtype=musa). * add backward of conv2d (#365) * add backward of conv2d * delete useless code * format code of tests * modify configs for tuning * modify autotune config * delete test flag * delete useless type convert --------- Co-authored-by: Jiang Bin <[email protected]> * SW-52470: Cherry-pick updates from commit f1ba20c to 5f31f35. * SW-49945: rebase for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-52470: adjust skipped ops for merge Signed-off-by: jiaqi.wang <[email protected]> * Mthreads: adjust ops for multi-backends merge Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * add triton_musa submodule * Modify testcase from cuda to musa. * Workaround for musa testcase. * modify test_unary_pointwise_ops from cuda to musa * modify test_reduction_ops from cuda to musa * Fix bug of reduceOp and shared memory. * fix dropout bug. * fix softmax exceeds shared memory error * Promote cpu reference accuracy to float32 * Modify benchmark performance test script to musa * Add torch_musa unsupported op test case. * Support bert model. * Update submodule url. * Comment v3/v4 test case. * fix: vectornorm upcast to fp64 * fix: group_norm modify case because of out-of-shared-memory * Promote golden accuracy from fp32 to fp64. * Rebase on master commit 1e49d6. * Update triton_musa submodule. * align perf utils to profiling pack 0717 * rebase on master commit 9000685 Signed-off-by: Jian Li <[email protected]> * Support op cumsum. config: {BLOCK_M: 8, num_warps: 8} will cause the number of registers within a single thread to be exceeded when the tensor shape is 4096 * 2304, so reduce BLOCK_M to 4 to supprot cumsum. * fix embedding tensor usage Signed-off-by: Jian Li <[email protected]> * uncomment supported op test Signed-off-by: Jian Li <[email protected]> * Support isclose() and allclose() - Torch_musa does not support fp64 input type, so CPU is used as a reference * Open up some tests that have already passed - Does not support test_accuracy_groupnorm - Some use cases have accuracy issues in test_embedding * rebase on master commit 801377f Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2e55d66 Signed-off-by: Jian Li <[email protected]> * fix distribution ops warps num Signed-off-by: Jian Li <[email protected]> * rebase on master commit a156268 Signed-off-by: Jian Li <[email protected]> * Support op argmax. * adapt masked_fill and dropout resolution Signed-off-by: Jian Li <[email protected]> * fix _log2 in topk op Signed-off-by: Jian Li <[email protected]> * rebase on master commit a1138bf Signed-off-by: Jian Li <[email protected]> * Open BF16 * rebase on master commit edb09f0 Signed-off-by: Jian Li <[email protected]> * fix div_trunc and div_floor Signed-off-by: Jian Li <[email protected]> * fix div_trunc cont. Signed-off-by: Jian Li <[email protected]> * rebase on master commit 2c9ae67 Signed-off-by: Jian Li <[email protected]> * rebase on master commit 204f3d4 Signed-off-by: Jian Li <[email protected]> * Support op vstack * rebase on master commit 4e5081 rebase on master commit 8f669d * SW-46066: remove submodule(triton_musa) in flaggems Signed-off-by: chuanjiang.ma <[email protected]> * [Operator] Fix vstack (#237) Modify the function parameter type declaration so that it can run in python 3.8 --------- Co-authored-by: zhengyang <[email protected]> * [Operator] Add upsample_nearest2d op [MooreThreads] (#193) * SW-46093: rebase on master commit 98924c Signed-off-by: jiaqi.wang <[email protected]> * succeed for norm (#233) Add _weight_norm op, while the original _weight_norm op changed to _weight_norm_interface op. * [Operator] slice&select scatter (#143) * add Ops & UT & Bench * add full zero ones Ops & UT & Bench * split normal op * [Operator] init slice&select scatter * code format * PR comment * split test_special_ops * add K-S test * split special perf * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * resolve conflict * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * test for op * test for op * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * add argparse * fix desc * fix num * Update test_specific_ops.py * split UT files * fix * fix * [Operator] Optimize CrossEntropyLoss (#131) reimplement cross_entropy_loss forward and backward support; indices/probabilities/weight/reduction/ignore_index/label_smoothing; perform better than torch eager on large scale tensors * Exponential added. (#138) * exponential added. * Added K-S tests to exponential_, fp64 corrected. * aligned with aten prototype * Exponential_ uses uint64 offsets in Triton kernel. * Update pyproject config for new test dependencies. * Use int64 indexing when needed & fix argmax (#146) 1. fix amax, armax and triu, use int64 indexing when the largest tensor's size_in_bytes exceed int32's max; 2. change the tiling scheme for argmax to loop in the reduction dimension, instead of data-size-dependent-tile-size * Making libentry thread safe (#136) * libentry now is lock protected. * Add multithreading tests for libentry. * polish code. * [Test] Test for op (#151) * [chore] solve slice&select scatter's test cases * [fix] fix slice&select scatter's test cases * [chore] remove out-of-range indices in select_scatter's test cases * [chore] simplify slice_scatter's test cases * [fix] Added range that is deleted by mistake * Merge branch 'master' into slice&select_scatter * [chore] reformat * [fix] typo * [chore] Considering perf, pause the replacement of some aTen operators * slice_scatter * select_scatter * index_select * [fix] Add libentry in op.cumsum * [fix] Del slice&select scatter's perf tests * [Chore] Add pytest mark for slice&select scatter's test * [Fix] Correct slice_scatter test * [Fix] Replace CPU Tensor --------- Co-authored-by: Bowen12992 <[email protected]> Co-authored-by: Tongxin Bai <[email protected]> Co-authored-by: Clement Chan <[email protected]> Co-authored-by: Bowen <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator] Add slice&select_scatter's benchmark (#262) * benchmark fix (#229) * benchmark fix * add seven new testing parameters * move shapes info to yaml file * Added the BenchmarkMetrics & BenchmarkResult abstraction * Specializing slice_scatter. (#270) * specializing slice_scatter. WIP. * polish and refine 2d_inner cases. * fix slice_scatter error on 1d inputs. * test slice_scatter fallback * Enhance Benchmarking for repeat_interleave Operation (#274) * Relocate select and slice benchmarks to test_select_and_slice_perf.py * sort keys for summary result * clean cuda cache after benchmark * fix repeat_interleave * modify format for summary info * SW-46093: rebase on master commit 2bd92c Signed-off-by: jiaqi.wang <[email protected]> * add test entry & use skip to avoid error Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: update mt test script by waiting frontend Signed-off-by: chuanjiang.ma <[email protected]> * SW-46093: adjust the skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: currently test accuracy ref to cpu Signed-off-by: chuanjiang.ma <[email protected]> * SW-47833: skip one case for half type op not supported by torch_musa Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: update test suite setup Signed-off-by: machuanjiang <[email protected]> * mthreads: fix review comments Signed-off-by: chuanjiang.ma <[email protected]> * mthreads: skip weight_nor testing for driver complaining 1. one test in special_op test change the device type from cuda to musa Signed-off-by: chuanjiang.ma <[email protected]> * fix #304 (#308) fix max: input tensor with big shape may cause "numel (2097152) exceeds triton maximum tensor numel" * add program_id & num_programs that returns tensor in tl.int64 (#327) * add program_id & num_programs that returns tensor in tl.int64 to avoid integer overflow * fix bmm * fix vector norm, treat empty dim list as full-reduce; use tle in fused ops, too * block pointer requires offsets/block shape to be int32 * [bugfix] replace tle with tl for operators using philox (#337) * [bugfix] replace tle with tl for operators using philox * [bugfix] tle in rand * [bugfix] uniform * Remove INT_DTYPES from isfinite operation in benchmark and tests (#336) * [Operator] Fix full & full_like op (#338) Co-authored-by: zhengyang <[email protected]> Co-authored-by: Bowen <[email protected]> * [Operator] Add diag_embed (#288) * diag_embed operator * add diag_embed ops * resolve conflict * SW-46093: add test python script for 50 ops Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: support output compared_speedup for benchmark plot Signed-off-by: root <chuanjiang.ma> * SW-46093: add an optin to control not output compared speedup Signed-off-by: root <chuanjiang.ma> * SW-46093: cherry-pick updates from commit 54a471 to ca13b7 Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Adapt Multiple backends code for MUSA Signed-off-by: jiaqi.wang <[email protected]> * SW-46093: Adjust skipped ops Signed-off-by: jiaqi.wang <[email protected]> * SW-49945: Update _mthreads backend config Signed-off-by: jiaqi.wang <[email protected]> * add pytest alluredir options for generating report * [Operator] Add sort op (#322) * [Operator] Add sort op --------- Co-authored-by: zhengyang <[email protected]> Co-authored-by: MARD1NO <[email protected]> * [Muti_backend]part_2_device (#344) * new feature, muti_backend * update auto_tune_module * update auto_tune_module * update auto_tune_module * update __init__ * rebase * fix bug * modifiy auto_tune_config * fix bug * fix bug * update * update * update scatter&gather * fix auto_tune * add gen_torch_device_fn * fix codestyle * fix codestyle * Modify code based on comments * Modify gen_impl with loops instead of recursion * Update code structure * Polish code * update * Polish code * Modify code based on comments * modify based on comment * Modify code based on comments * update * final fix * modify_device_unify * fix bug * update * remove '-> list' * modify * modify * modify * fix bug * muti_backend_part_3 enhancement of ability (#361) * muti_backend_part_3 enhancement of ability * fix bug * modify * modify * fix bug * modify * SW-49945: adjust merged code Signed-off-by: jiaqi.wang <[email protected]> * Reimplement slice_scatter/select_scatter using pointwise_dynamic (#358) * merged with origin. * updated * renames copy_ to copy, fixes benchmark code. * fixes benchmark. * handle self-overlapping input in select_scatter and slice_scatter --------- Co-authored-by: Clement Chan <[email protected]> * modify code (#371) * modify code * fix bug * fix bug * modify * fix * fix * [Polish code] Adjust some directory structures and variable names (#373) * modify code * fix bug * fix bug * modify * fix * fix * polish code * polish code * polish code * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [Muti backend] add heuristics_config for muti_bacnkend (#377) * muti_backend * [Fix bugs] Variable reference dependency error (#374) * fix benchmark bugs * fix * [bugfix] disable more shapes from blas benchmark (#375) * add heuristics_config * modify * modify * modify * modify --------- Co-authored-by: StrongSpoon <[email protected]> * SW-49945: update configuration for mthreads Signed-off-by: jiaqi.wang <[email protected]> * [benchmark] skip perf test of cummin when triton < 3.0 (#385) * SW-46093: fix test configuration Signed-off-by: jiaqi.wang <[email protected]> * SW-50265: optimize the perf of gelu, tanh op and most other pointwise ops * Gather bwd (#382) * add gather_backward op * add debug log in * perf gather backward * rebased with master * scatter rewrite done. * scatter handling internally overlapping input. * Scatter reduce now uses atomics. * remove fp16 from scatter reduce UT. * sets threadblock size to 128 for scatter. * Change atomic memory order to relaxed in scatter. --------- Co-authored-by: awayzjj <[email protected]> Co-authored-by: StrongSpoon <[email protected]> * [Operator]Fix metax backend bugs (#432) 1. update multi-backend code 2. fix argmin op might test failed under int types Co-authored-by: mx-flaggems-user <[email protected]> * [TEST] fix error in argmin UT when dtype=int16 (#431) Co-authored-by: junjian.zhan <[email protected]> * SW-49945: Update configurations for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-47833: Fix case resolve_conj and fill. 1. resolve_conj: ref to this link: https://jira.mthreads.com/browse/MTAI-1530 2. fill: torch_musa does not support case torch.fill(dtype=cpu, dtype=musa). * add backward of conv2d (#365) * add backward of conv2d * delete useless code * format code of tests * modify configs for tuning * modify autotune config * delete test flag * delete useless type convert --------- Co-authored-by: Jiang Bin <[email protected]> * SW-52470: Cherry-pick updates from commit f1ba20c to 5f31f35. * SW-49945: rebase for multi-backends Signed-off-by: jiaqi.wang <[email protected]> * SW-52470: adjust skipped ops for merge Signed-off-by: jiaqi.wang <[email protected]> * Mthreads: adjust ops for multi-backends merge Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * MTHREADS: Checking Code Format Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * MTHREADS: fix bugs for multi-backends ops Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> * MTHREADS: fix according to review issue Signed-off-by: chuanjiang.ma <[email protected]> * MTHREADS: add annotation to explain the log-file-cross calc Signed-off-by: chuanjiang.ma <[email protected]> --------- Signed-off-by: Jian Li <[email protected]> Signed-off-by: chuanjiang.ma <[email protected]> Signed-off-by: jiaqi.wang <[email protected]> Signed-off-by: machuanjiang <[email protected]> Signed-off-by: root <chuanjiang.ma> Signed-off-by: jiaqi.wang <jiaqi.wang @mthreads.com> Co-authored-by: yuzhe-wu <[email protected]> Co-authored-by: Buzz <[email protected]> Co-authored-by: hang.zhang <[email protected]> Co-authored-by: lingfeng.qiu <[email protected]> Co-authored-by: jialuo.bai <[email protected]> Co-authored-by: Jian Li <[email protected]> Co-authored-by: Tianjie Ling <[email protected]> Co-authored-by: jiaqi.wang <[email protected]> Co-authored-by: zhzhcookie <[email protected]> Co-authored-by: zhengyang <[email protected]> Co-authored-by: zaccur <[email protected]> Co-authored-by: TZWX-0 <[email protected]> Co-authored-by: Hiujin Gwok <[email protected]> Co-authored-by: Bowen12992 <[email protected]> Co-authored-by: Tongxin Bai <[email protected]> Co-authored-by: Clement Chan <[email protected]> Co-authored-by: Bowen <[email protected]> Co-authored-by: StrongSpoon <[email protected]> Co-authored-by: kiddyjinjin <[email protected]> Co-authored-by: haowen-han <[email protected]> Co-authored-by: niu_he <[email protected]> Co-authored-by: root <chuanjiang.ma> Co-authored-by: yu.huang <[email protected]> Co-authored-by: MARD1NO <[email protected]> Co-authored-by: Galaxy1458 <[email protected]> Co-authored-by: Sheng Wang <[email protected]> Co-authored-by: awayzjj <[email protected]> Co-authored-by: StrongSpoon <[email protected]> Co-authored-by: mx-flaggems-user <[email protected]> Co-authored-by: mx-flaggems-user <[email protected]> Co-authored-by: junjian.zhan <[email protected]> Co-authored-by: FatJhon <[email protected]> Co-authored-by: Jiang Bin <[email protected]> Co-authored-by: jiaqi.wang <jiaqi.wang @mthreads.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR Category
Operator
Type of Change
Bug Fix
Description
_metaxmulti-backend codedim=None