v0.4.0

Switching the repo to public

What's Changed

• Alternative subtract_intervals function optimized for scalability by @tjkemp in #89
• Add MIT license to all meaningful files by @gabeweisz in #92
• Bugfix/conv output shape by @RElbers in #95
• dim eff analysis by @ajassani in #94
• Stagingv0.4 by @ajassani in #96
• pass arch and detail to all op by @ajassani in #97
• Add additional Jax data analyses by @gabeweisz in #98
• Add in GEMM shape analysis from XLA file by @gabeweisz in #100
• move convert import up by @gabeweisz in #101
• only include linked kernels in gpu timeline by @ajassani in #110
• Feat/event replay by @ajassani in #102
• include nn module in kernel view and refine by @ajassani in #116
• add aten::bmm perf model by @ajassani in #117
• Feat/topk ops by @ajassani in #120
• Feat/nn module parent by @ajassani in #121
• feat: TraceLens UI by @mpashkovskii in #99
• Fix get_breakdown_df_multigpu to filter out CPU events by @gabeweisz in #122
• include args in perf metrics table by @ajassani in #130
• short term fix for te linear gemms by @ajassani in #126
• add perf model for aten baddbmm by @ajassani in #131
• annotate gpu events by stream index by @ajassani in #135
• Reorg examples by @ajassani in #136
• Feat/unique args by @ajassani in #137
• feat: transformer engine ver 1 GEMM ops te_gemm_ts added by @olehtika in #125
• parse trans from args and add te gemm name to list in examples by @ajassani in #140
• Add GEMM performance model support for Jax by @gabeweisz in #139
• Integrate gemmologist for modeling gemm efficiencies. by @araina-amd in #133
• use correct dtype in cmd by @ajassani in #141
• Jax support for gemmologist integration by @gabeweisz in #143
• fix: tex_ts::te_gemm_ts missing dtype by @olehtika in #142
• allow causal attention by @ajassani in #145
• support gqa by @ajassani in #147
• tev2 native support by @ajassani in #148
• native cat 2 op names by @ajassani in #149
• megatron lm custom flow by @ajassani in #152
• fix flash true in fusedattnfunc by @ajassani in #153
• Fix/unique args kernel names by @ajassani in #154
• Fix/megatron gemm dtype by @ajassani in #156
• Fix/te dtype by @ajassani in #157
• Feat/qkv stride by @ajassani in #159
• torch_op_mapping: fix typo, add clamp_max by @lauri9 in #162
• Add script to filter trace to range of given user annotation by @lauri9 in #163
• Feat/custom report by @ajassani in #166
• Feat/nn flops by @ajassani in #169
• Feat/replay from report by @ajassani in #168
• Feat/aten sdp eff atten by @ajassani in #172
• Test/event replay by @ajassani in #173
• Fix/event replay by @ajassani in #174
• Fix/torch import by @ajassani in #175
• Test/event replay improve by @ajassani in #176
• Feat/evt replay example improve by @ajassani in #177
• add conditional strenum as suggested by @jakki-amd by @gabeweisz in #161
• Fix/lib deps by @ajassani in #178
• continue on fn call error by @ajassani in #179
• explicit openpyxl import error by @ajassani in #180
• sort false to avoid error by @ajassani in #181
• Add roofline_analyzer 0.1.1 by @tykow in #182
• regression test for perf report by @ajassani in #183
• fix is tensor type by @ajassani in #188
• Feat/perf report compare by @ajassani in #193
• NCCLAnalyzer: add missing allgather_into_tensor_coalesced collective name by @lauri9 in #194
• Enabling batch gemm through gemmologist. by @araina-amd in #189
• NcclAnalyser: add gzip support by @lauri9 in #196
• add flash_attn::_flash_attn_forward by @lauri9 in #197
• feat: extend performance report for multiple GPU ranks and trace files by @olehtika in #186
• Revert "feat: extend performance report for multiple GPU ranks and tr… by @olehtika in #198
• fix: fallback for StrEnum ImportError for Python 3.10 by @olehtika in #205
• feat: use util.DataLoader for all data loading by @olehtika in #208
• fix: JaxAnalysis FP8 GEMM kernel missing issue by @olehtika in #203
• SDPA/Flash attention changes for the perf model by @araina-amd in #191
• Partition on K and V instead of Q in SDPA backward pass by @araina-amd in #211
• option to print cpu op dispatch args by @ajassani in #213
• add aten::_scaled_dot_product_flash_attention to perf model by @ajassani in #217
• Feat/call stack analysis by @ajassani in #220
• dropout not needed for perf metrics by @ajassani in #221
• warn on dtype mismatch in gemms by @ajassani in #223
• feat: jax analysis reporting command line tool by @olehtika in #210
• Fix/jax analyses do not parse file header by @jujaykka in #218
• Fix/processing performance fixes by @jujaykka in #219
• minimize installation dependencies by @ajassani in #224
• Add GitHub Actions Workflows by @spandoesai in #222
• torch op categorization by @ajassani in #228
• fix error on dtype mismatch by @ajassani in #231
• fix edge cases for perf report by @ajassani in #235
• quick fix for execute by @ajassani in #236
• add openpyxl as requirement by @stephen-youn in #230
• overcome openpyxl dependency by @ajassani in #238
• gemms conceptual by @ajassani in #134
• notebook for autocast exploration by @ajassani in #240
• Commandline interface for roofline and code refactoring for perf model by @araina-amd in #229
• Added Perf Model for aiter::flash_attn by @spandoesai in #242
• Docs/trace2tree by @ajassani in #243
• fix perf report comparison script by @ajassani in #244
• TraceMap: interactive HTML dashboards for analyzing vLLM workloads (added to custom_workflows) by @hyukjlee in #246
• warn when compute perf metrics fails by @ajassani in #248
• Check if m, n and k are not none in gemm. by @araina-amd in #247
• add feature for extension by @ajassani in #249
• kernel detail update by @ajassani in #245
• Added the TraceDiff API to TraceLens by @spandoesai in #241
• add support for grouped gemm by @ajassani in #254
• allow different d_h_qk and d_h_v by @ajassani in #255
• fix sdpa param extract by @ajassani in #256
• some fixes for double dtype by @ajassani in #259
• Fix/event replay double by @ajassani in #260
• counts in repro by @ajassani in #261
• Refact/perf report by @ajassani in #264
• refact compare perf report by @ajassani in #265
• Feat/jax trace2 tree by @olehtika in #257
• FIX: Update te_gemms.py to be compatible with TreePerfAnalyzer. by @wenchenvincent in #266
• include collective analysis in report by @ajassani in #268
• Feat/jax trace2 tree add gpu kernel operation category by @olehtika in #267
• [NcclAnalyser]: Add option to warn instead of raise on metadata consistency failure by @lauri9 in #271
• include start time stamp for df_kernels by @ajassani in #273
• TraceDiff Graph View Quality of Life Changes by @spandoesai in #269
• profiling guide by @ajassani in #274
• add docs for profile analysis by @ajassani in #275
• use fp8 arg for perf model by @ajassani in #277
• Added New Perf Model for aiter::fmha_v3 (fwd and bwd) by @spandoesai in #278
• add script for collective analysis by @ajassani in #283
• Added Perf Model for FAv3 (fwd only) by @spandoesai in #284
• Add support for flash_attn variable sequence length by @wenchenvincent in #272
• Add logic for vLLM Llama-3.3-70B FP4 by @poznano-amd in #270
• add graph launch gpu events corectly to tree by @ajassani in #292
• fix to include unlinked runtime events in kernel launchers by @ajassani in #293
• Feat/jax tree perf perf model by @gphuang in #286
• Fixes for JAX. by @araina-amd in #282
• fix missing dictionary check by @gabeweisz in #323
• Switch installed jax report from jax_analysis_report to jax_report by @gabeweisz in #327
• 325 creating one excel file with multiple tabs rather than one file per report by @gphuang in #338
• add contributing md by @ajassani in #356
• Update CONTRIBUTING.md by @ajassani in #360
• update readme by @ajassani in #362
• Feat/jax perf analysis keyword filter by @olehtika in #332
• 308 jax tree perf analyzer parse gemm metadata value error invalid operand list by @gphuang in #347
• Add trtllm::cublas_scaled_mm to torch_op_mapping by @poznano-amd in #368
• include only non empty roofline sheets in perf report by @ajassani in #369
• include unlinked kernels by @ajassani in #370
• docs(contributing): add linting and branch naming rules by @spandoesai in #371
• New and Improved Compare Report Regression Testing by @spandoesai in #372
• add logging levels by @gphuang in #367
• add feature to optionally separate idle time into micro and macro by @ajassani in #375
• Added PR and Issue Templates by @spandoesai in #376
• docs(test-traces-disclaimer): add disclaimer for open-source test traces clarifying non-official perf usage by @ajassani in #379
• chore(license): update .py files to use AMD copyright header by @ajassani in #381
• Chore/license/update headers by @ajassani in #383
• Update issue templates by @ajassani in #387
• fix(tests): resolve non-determinism in perf report comparison by @ajassani in #385
• test(ci): add copyright header validation to regression tests by @ajassani in #384
• ci: added automation test for linting check using the black formatter by @spandoesai in #388
• feat: JAX XLA op analysis including GPU kernel launch latency and operation total input bytes calculation by @olehtika in #392
• Test/339 add unit test with minimal jax conv by @gphuang in #380

New Contributors

• @tjkemp made their first contribution in #89
• @mpashkovskii made their first contribution in #99
• @olehtika made their first contribution in #125
• @araina-amd made their first contribution in #133
• @lauri9 made their first contribution in #162
• @tykow made their first contribution in #182
• @jujaykka made their first contribution in #218
• @spandoesai made their first contribution in #222
• @stephen-youn made their first contribution in #230
• @hyukjlee made their first contribution in #246
• @wenchenvincent made their first contribution in #266
• @poznano-amd made their first contribution in #270
• @gphuang made their first contribution in #286

Full Changelog: v0.3.0...v0.4.0