topk fp16 failure fixes. #270

lcskrishna · 2018-10-15T18:57:11Z

No description provided.

iotamudelta · 2018-10-15T19:06:23Z

LGTM

Commit Messages: - reset parameters for FusedDenseGeluDense similar to FusedDense to make the test_gelu pass (ROCm#269) (ROCm#270) Co-authored-by: Sriram Kumar <[email protected]> PRs: - ROCm/apex#270 Fixes: - https://ontrack-internal.amd.com/browse/SWDEV-540029

Commit Messages: - reset parameters for FusedDenseGeluDense similar to FusedDense to make the test_gelu pass (#269) (#270) Co-authored-by: Sriram Kumar <[email protected]> PRs: - ROCm/apex#270 Fixes: - https://ontrack-internal.amd.com/browse/SWDEV-540029

Commit Messages: - reset parameters for FusedDenseGeluDense similar to FusedDense to make the test_gelu pass (#269) (#270) Co-authored-by: Sriram Kumar <[email protected]> - Fix build error (#263) PRs: - ROCm/apex#270 Fixes: - https://example.com/issue-263 - https://ontrack-internal.amd.com/browse/SWDEV-540029

Commit Messages: - reset parameters for FusedDenseGeluDense similar to FusedDense to make the test_gelu pass (#269) (#270) Co-authored-by: Sriram Kumar <[email protected]> - Fix build error (#263) PRs: - ROCm/apex#270 - ROCm/apex#263 Fixes: - https://ontrack-internal.amd.com/browse/SWDEV-540029 Co-authored-by: sriram <[email protected]> Co-authored-by: Pruthvi Madugundu <[email protected]>

Commit Messages: - update the param_id calculation so that it works on both CPX and SPX modes (#271) (#272) - reset parameters for FusedDenseGeluDense similar to FusedDense to make the test_gelu pass (#269) (#270) Co-authored-by: Sriram Kumar <[email protected]> - Fix build error (#263) - Fix warp size (#256) * replace c10_warp_size in fused rope * replace c10_warp_size in fused softmax * replace c10_warp_size in group batch norm * replace c10_warp_size in multiheadattention * replace c10_warp_size in tramsducer * replace c10_warp_size in xentropy * replace c10_warp_size in sync batch normalization * replace c10_warp_size in group batch norm * replace warp_size in multihead attention - Disabling Aiter Installation in default build (#254) * made a flag to switch on/off aiter compile using --aiter when installing apex * Added information on building AITER during installation in readme - Replaced warpsize with C10_WARP_SIZE (#249) - correct the approach to get to the apex folder from the test file (#248) - Apex extensions import test (#245) * add test to extract extensions from setup.py and test if there can be imported * moved test outside tests/L0 - Fixing the C10_warpsize issue. replacing the macros with at::cuda::warp_size() (#237) - Replacing c10_warp_size with platform based warp_size values (#228) fixes :https://ontrack-internal.amd.com/browse/SWDEV-541725 - [master] Added AITER as a submodule and use in fused_rope.py (#222) * Added aiter support in fused_rope.py for all 4 variants. Updated fused rope test, reduced tolerances according to unit test in aiter repo. * Add aiter as a submodule and install it if it is rocm. Switch on aiter backend if it is rocm and aiter is installed * add pandas to the requirements so that aiter can be used without numpy error - ValueError: numpy.dtype size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject * Replace ROCM_HOME condition to IS_ROCM_PYTORCH for installing aiter and use pip install -e . instead of python setup.py develop for installing aiter. * Create apex and aiter subclasses for the four variants of FusedRoPEFunc and select apex or aiter subclass based on AITER_ROPE_BACKEND value. The user can specify the environment variable USE_ROCM_AITER_ROPE_BACKEND to select between aiter and apex backends for fused rope. * If the AITER backend is selected, use lowered precision in the unit test otherwise use the original precision 1e-3 * warn user about the lower precision when using aiter backend for fused rope * Update fused_rope.py remove spaces * simplify the switch between aiter and apex subclasses * install aiter without editable mode - Merge pull request #227 from ROCm/amd/dev/iassiour/SWDEV-541770 Do not use warpSize as a constexpr in nhwc_batch_norm_kernel.h - Do not use warpSize as a constexpr in nhwc_batch_norm_kernel.h In ROCm 7.0, the warpSize variable is no longer constexpr. This commit replaces the variable use with the correct values based on the architecture we're running on. - change epilogue parameter for hipblaslt matmul in cuda kernel for fused dense gelu dense (#223) Fixes : https://ontrack-internal.amd.com/browse/SWDEV-534531 - Reset torch default device to cpu after running the amp unit tests. (#220) - Fix unit tests for transformer, fused dense, mlp (#218) * Fix fused_dense_gelu_dense, change the names of the parameters so that they can be accessed by the test appropriately * Update the absolute tolerances in test_mlp from 0 and 1e-7 to 1e-5 * Deactivate the amp state handle for optimization level other than O0. This helps to pass the UT after this. * Update condition for deactivating amp state handle from opt level equal to 1 to opt level not equal to 0 * Update torch set default dtype method to remove warning * Update the method to create overflow buffer for amp optimizer * Update the method to create overflow buffer for amp optimizer * Update the method to create overflow buffer for amp optimizer * reset the default device to cpu so that the generator uses cuda, as run_amp tests set its to cuda - Update fused layer norm code from upstream apex repo. The intra-warp reductions code inside cuWelfordMuSigma2() function in layer norm kernel assumes a warp size of 32, so added a condition for rocm to support gpu warp size (based on earlier apex code). For rocm, adjust the threadsize, based on earlier apex code. (#215) - upgrade matplotlib to resolve setuptools_scm error. (#213) The error: File /tmp/easy_install-_pfhn8pn/matplotlib-3.5.1/.eggs/setuptools_scm-8.3.1-py3.12.egg/setuptools_scm/_integration/pyproject_reading.py, line 36, in read_pyproject section = defn.get(tool, {})[tool_name] ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^ KeyError: 'setuptools_scm' Solution : https://github.com/matplotlib/matplotlib/blob/v3.8.x/pyproject.toml#L22 matplotlib 3.8 is the first version to have pyproject.toml with this tool.setuptools_scm section. This higher version of setuptools expects this structure in the python packages it installs. Matplotlib 3.5.1 doesn't satisfy this condition. The solution is to change the condition to matplotlib>=3.8. - Update distributed fused adam - integrate Pipeline operations and support different grad (#207) * Fix `DistributedFusedAdam` for grad dtype != param dtype (#1893) * Pipeline `reduce-scatter` and `all-reduce`. (#1895) --------- Co-authored-by: Tailing Yuan <[email protected]> Co-authored-by: Wil Kong <[email protected]> - Update the condition for building the NCCL allocator, PyTorch should be greater than or equal to 2.6 (#204) - Update version.txt (#203) change the version from 1.7.0 to 1.8.0 - [ROCm] Use at::empty to manage workspace memory to avoid hip runtime calls (#197) Optimize the memory for fused_weight_gradient_mlp_cuda module - Update README.md (#198) Add release notes for release/1.5, 1.6 and 1.7 - Update README.md (#196) updated the support versions for apex 1.7.0 PRs: - https://github.com/ROCm/apex/pull/1895 Fixes: - https://example.com/issue-271 - https://example.com/issue-249 - https://example.com/issue-254 - https://example.com/issue-228 - https://example.com/issue-263 - https://example.com/issue-223 - https://example.com/issue-237 - https://example.com/issue-203 - https://example.com/issue-256 - https://example.com/issue-245 - https://example.com/issue-272 - https://example.com/issue-204 - https://ontrack-internal.amd.com/browse/SWDEV-540029 - https://example.com/issue-222 - https://example.com/issue-220 - https://example.com/issue-248 - https://example.com/issue-1893 - https://example.com/issue-198 - https://example.com/issue-215 - https://example.com/issue-213 - https://example.com/issue-1895 - https://example.com/issue-218 - https://example.com/issue-227 - https://example.com/issue-196 - https://example.com/issue-197 - https://example.com/issue-207

Commit Messages: - update the param_id calculation so that it works on both CPX and SPX modes (#271) (#272) - reset parameters for FusedDenseGeluDense similar to FusedDense to make the test_gelu pass (#269) (#270) Co-authored-by: Sriram Kumar <[email protected]> PRs: - ROCm/apex#272 Fixes: - https://ontrack-internal.amd.com/browse/SWDEV-540029 - https://example.com/issue-272 - https://example.com/issue-271

Commit Messages: - update the param_id calculation so that it works on both CPX and SPX modes (#271) (#272) - reset parameters for FusedDenseGeluDense similar to FusedDense to make the test_gelu pass (#269) (#270) PRs: - ROCm/apex#272 - ROCm/apex#269 Fixes: - https://ontrack-internal.amd.com/browse/SWDEV-540029 - https://ontrack-internal.amd.com/browse/SWDEV-548434

fixes for topk fp16

d7f6aa3

lcskrishna requested a review from ezyang as a code owner October 15, 2018 18:57

iotamudelta merged commit 3aab7c4 into ROCm:master Oct 15, 2018

iotamudelta pushed a commit to iotamudelta/pytorch that referenced this pull request Oct 22, 2018

fixes for topk fp16 (ROCm#270)

0ddd079

iotamudelta added the cherry-picked This pull request has been cherry-picked into pytorch/pytorch upstream. label Oct 22, 2018

amd-sriram mentioned this pull request Jul 29, 2025

[release/2.8] update related_commit #2431

Closed

amd-sriram mentioned this pull request Jul 29, 2025

[rocm7.0_internal_testing] update related_commit #2433

Merged

amd-sriram mentioned this pull request Aug 12, 2025

[release/2.8] update related_commit #2490

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

topk fp16 failure fixes. #270

topk fp16 failure fixes. #270

Uh oh!

lcskrishna commented Oct 15, 2018

Uh oh!

iotamudelta commented Oct 15, 2018

Uh oh!

Uh oh!

topk fp16 failure fixes. #270

topk fp16 failure fixes. #270

Uh oh!

Conversation

lcskrishna commented Oct 15, 2018

Uh oh!

iotamudelta commented Oct 15, 2018

Uh oh!

Uh oh!