Proto commits in tensorflow/profiler

These commits are when the Protocol Buffers files have changed: (only the last 100 relevant commits are shown)

Commit:9d2184c
Author:Profiler Team
Committer:Copybara-Service

Add integration and unit tests for High Confidence TPU Duty Cycle and Program Goodput. PiperOrigin-RevId: 950638914

Commit:b04bc68
Author:Profiler Team
Committer:Copybara-Service

Add integration and unit tests for High Confidence TPU Duty Cycle and Program Goodput. PiperOrigin-RevId: 950638914

Commit:02a4fc6
Author:Profiler Team
Committer:Copybara-Service

XProf: Add native C++ Per-Step Straggler Diagnostics map result extractor Extracts per-step peak wait times and boundaries directly from XPlane events in C++ to improve straggler diagnostic performance, bypassing the need to parse raw events in Python. Global step alignment and statistical anomaly detection (MAD, Z-Score) are deferred to the Python frontend for flexibility and speed. PiperOrigin-RevId: 949805062

Commit:246d612
Author:Profiler Team
Committer:Copybara-Service

XProf: Add native C++ Per-Step Straggler Diagnostics map result extractor Extracts per-step peak wait times and boundaries directly from XPlane events in C++ to improve straggler diagnostic performance, bypassing the need to parse raw events in Python. Global step alignment and statistical anomaly detection (MAD, Z-Score) are deferred to the Python frontend for flexibility and speed. PiperOrigin-RevId: 949805062

Commit:1fd1e0e
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 951546827

The documentation is generated from this commit.

Commit:d0f4939
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 947661634

The documentation is generated from this commit.

Commit:d17bf15
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 947661634

The documentation is generated from this commit.

Commit:a73adce
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 937897160

The documentation is generated from this commit.

Commit:806ab89
Author:Profiler Team
Committer:Copybara-Service

XProf: Add native C++ Per-Step Straggler Diagnostics map result extractor Extracts per-step peak wait times and boundaries directly from XPlane events in C++ to improve straggler diagnostic performance, bypassing the need to parse raw events in Python. Global step alignment and statistical anomaly detection (MAD, Z-Score) are deferred to the Python frontend for flexibility and speed. PiperOrigin-RevId: 949805062

Commit:c179bae
Author:Profiler Team
Committer:Copybara-Service

Add FlatOpMetricsDb support to OpStats and StepDetails. PiperOrigin-RevId: 950998670

Commit:df6ea48
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 947661634

Commit:db41a52
Author:Profiler Team
Committer:Copybara-Service

Add integration and unit tests for High Confidence TPU Duty Cycle and Program Goodput. PiperOrigin-RevId: 950638914

Commit:1ab8a85
Author:Profiler Team
Committer:Copybara-Service

Add open-source protobuf storage fields for High Confidence TPU Duty Cycle and Program Goodput. PiperOrigin-RevId: 945616604

Commit:9a2db39
Author:Profiler Team
Committer:Copybara-Service

Expose High Confidence TPU Duty Cycle and Program Goodput metrics on the Overview Page frontend. PiperOrigin-RevId: 950136107

Commit:b896da6
Author:Profiler Team
Committer:Copybara-Service

Add serialization and aggregation for High Confidence TPU Duty Cycle and Program Goodput metrics into TfOpStats and Overview Page. PiperOrigin-RevId: 950136108

Commit:7ddd73e
Author:Profiler Team
Committer:Copybara-Service

Add ProgramGoodputForHloOpHC to calculate high-confidence program goodput by excluding opaque custom calls lacking FLOP and ICI communication metadata from efficiency calculations. PiperOrigin-RevId: 950136110

Commit:b0ed632
Author:Profiler Team
Committer:Copybara-Service

Introduce ConstructHCDutyCycleTracker to compute High Confidence TPU Duty Cycle. PiperOrigin-RevId: 950136109

Commit:9dc3329
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 947661634

Commit:9f154a3
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 947661634

Commit:f2bd3e0
Author:Profiler Team
Committer:Copybara-Service

Add a field to store a deterministic, unique op ID. PiperOrigin-RevId: 947661634

Commit:b5d1cc0
Author:Profiler Team
Committer:Copybara-Service

Fix VMEM peak memory allocation in XProf memory viewer PiperOrigin-RevId: 944511949

Commit:271caa3
Author:Profiler Team
Committer:Copybara-Service

Input Pipeline Analysis: Flat Host Metrics Downstream Consumption PiperOrigin-RevId: 939719525

Commit:34588b6
Author:Profiler Team
Committer:Copybara-Service

Input Pipeline Analysis: Flat Host Metrics Utility Updates PiperOrigin-RevId: 939708138

Commit:94ec319
Author:Profiler Team
Committer:Copybara-Service

Input Pipeline Analysis: Flat Host Metrics Converter Implementation PiperOrigin-RevId: 939701840

Commit:f2d6008
Author:Profiler Team
Committer:Copybara-Service

Input Pipeline Analysis: Flat Host Metrics Proto Changes PiperOrigin-RevId: 939684524

Commit:b857563
Author:Profiler Team
Committer:Copybara-Service

Add ConvertDeviceTraceXPlaneToFlatOpMetricsDb to parse GPU XPlanes natively into FlatOpMetricsDb using XEventsFlatOpMetricsDbBuilder. PiperOrigin-RevId: 940620823

Commit:6e40965
Author:Profiler Team
Committer:Copybara-Service

Integrate FlatOpMetricsDb into the Roofline Model pipeline and support incomplete steps. PiperOrigin-RevId: 936077242

Commit:fbf69f6
Author:Profiler Team
Committer:Copybara-Service

Support combining FlatOpMetricsDb in OpStats combination pipeline. PiperOrigin-RevId: 938552869

Commit:e5a6d08
Author:Profiler Team
Committer:Copybara-Service

Integrate FlatOpMetricsDb into XPlane-to-StepEvents conversion pipeline. PiperOrigin-RevId: 938234267

Commit:b49e3d9
Author:Profiler Team
Committer:Copybara-Service

Add GetFlatOpKeyFromXEvent utility to extract OpKey from XEvent. PiperOrigin-RevId: 938546019

Commit:ff6b072
Author:Profiler Team
Committer:Copybara-Service

Add utilities to convert FlatOpMetrics to standard Record format for Roofline model. PiperOrigin-RevId: 936077240

Commit:cfc01f6
Author:Profiler Team
Committer:Copybara-Service

Add utilities to aggregate Device FlatOpMetrics into Framework (TF/JAX) Op Metrics. PiperOrigin-RevId: 936077239

Commit:ee7cb18
Author:Profiler Team
Committer:Copybara-Service

Add FlatOpMetricsDb support to OpStats and StepDetails. PiperOrigin-RevId: 925237403

Commit:8dd658f
Author:Profiler Team
Committer:Copybara-Service

Internal change PiperOrigin-RevId: 914722263

Commit:eaa1c6a
Author:Profiler Team
Committer:Copybara-Service

Internal change PiperOrigin-RevId: 909900419

Commit:9cfd31e
Author:Profiler Team
Committer:Copybara-Service

Internal change PiperOrigin-RevId: 911717012

Commit:b9f2537
Author:Profiler Team
Committer:Copybara-Service

Refactor FlatOpMetricsDb conversion logic to code readability. PiperOrigin-RevId: 911717012

Commit:facbc7f
Author:Profiler Team
Committer:Copybara-Service

Add ConvertDeviceTraceXPlaneToFlatOpMetricsDb to parse GPU XPlanes natively into FlatOpMetricsDb using XEventsFlatOpMetricsDbBuilder. PiperOrigin-RevId: 935588208

Commit:7f40311
Author:Profiler Team
Committer:Copybara-Service

Project import generated by Copybara PiperOrigin-RevId: 929104288

Commit:bb8eb70
Author:Sai Ganesh Muthuraman
Committer:Copybara-Service

Enable scroll and zoom out in Trace Viewer V2 when loading deep-linked traces. PiperOrigin-RevId: 919479016

Commit:4fad2cd
Author:Profiler Team
Committer:Copybara-Service

Added XSpace-to-FlatProfileMap conversion support for power data. PiperOrigin-RevId: 916029727

Commit:67e98d3
Author:Profiler Team
Committer:Copybara-Service

Add additional data to overview screenshot test. PiperOrigin-RevId: 906362969

Commit:65b8d04
Author:Profiler Team
Committer:Copybara-Service

Adding Functionality : ConvertXSpaceToFlatOpMetricsDb PiperOrigin-RevId: 913633636

Commit:b03d265
Author:Profiler Team
Committer:Copybara-Service

Addition of Flat Device Op Metric Db to OpStats PiperOrigin-RevId: 913605984

Commit:9d320d4
Author:Sai Ganesh Muthuraman
Committer:Copybara-Service

Add `details` field to TraceDataResponse proto for compressed traces. PiperOrigin-RevId: 906149238

Commit:e65e8c4
Author:Nancy Li
Committer:Copybara-Service

[XProf: trace viewer] Add trace_lite.proto PiperOrigin-RevId: 906016226

Commit:f3d49a2
Author:Profiler Team
Committer:Copybara-Service

Add Sparse Core ops to HLO Op Stats tool. PiperOrigin-RevId: 905236948

Commit:def69ce
Author:Profiler Team
Committer:Copybara-Service

Add flat_op_metrics.proto to define a flat structure for operation metrics across different core types. PiperOrigin-RevId: 904282580

Commit:6f341d9
Author:Profiler Team
Committer:Copybara-Service

Add TraceDataResponse proto definition. This would be used for transferring trace data over the wire significantly reducing the payload size from the current JSON baseline. PiperOrigin-RevId: 892755054

Commit:d45fbe2
Author:Profiler Team
Committer:Copybara-Service

Internal Change PiperOrigin-RevId: 888002603

Commit:7088a10
Author:Ruiyang Zhu
Committer:Copybara-Service

Create EventFractionAnalyzerResults Proto PiperOrigin-RevId: 886331240

Commit:acf601a
Author:Charles Alaras
Committer:Copybara-Service

Fix documentation typo in OpMetrics proto PiperOrigin-RevId: 876320696

Commit:b5257f5
Author:Charles Alaras
Committer:Copybara-Service

⚠️ Deprecate `flops` → `flops_v2` and `model_flops` → `model_flops_v2` PiperOrigin-RevId: 875515357

Commit:c6a3380
Author:Profiler Team
Committer:Copybara-Service

Display disaggregated serving avg decoding step time on TPU in the XProf overview page PiperOrigin-RevId: 874418602

Commit:9fea95a
Author:Sai Ganesh Muthuraman
Committer:Copybara-Service

Add mandatory device and resource regexes to TraceFilterConfig PiperOrigin-RevId: 870869716

Commit:ee7b24f
Author:Profiler Team
Committer:Copybara-Service

Update `GetEventTimeFractionAnalyzerResult` in 3P ToolDataProviderImpl to use `GetToolDataHelper` API PiperOrigin-RevId: 867894263

Commit:b8d26f0
Author:Profiler Team
Committer:Copybara-Service

Initial proto for Disaggregared Serving Latency PiperOrigin-RevId: 858648057

Commit:c5760f7
Author:Profiler Team
Committer:Copybara-Service

Enable Memory Profile to correlate memory usage with HLO module execution. PiperOrigin-RevId: 855749344

Commit:a06e8dc
Author:Profiler Team
Committer:Copybara-Service

Op Profile adjustments. PiperOrigin-RevId: 847910843

Commit:f6efc9c
Author:Profiler Team
Committer:Copybara-Service

Add a `suppress_others` flag in `smart_suggestion.proto` to suppress other suggestions once the flag is on. PiperOrigin-RevId: 845499158

Commit:e7edebd
Author:Charles Alaras
Committer:Copybara-Service

Support SparseCore events in OpMetricsDb PiperOrigin-RevId: 844869212

Commit:59579f0
Author:Profiler Team
Committer:Copybara-Service

Include per-host event fraction for each step in `EventTimeFractionAnalyzer` proto. PiperOrigin-RevId: 843778153

Commit:a170c8d
Author:Yin Zhang
Committer:Copybara-Service

Updates to the run environment data processing in op stats. No-op for now. PiperOrigin-RevId: 843423093

Commit:7363782
Author:Yin Zhang
Committer:Copybara-Service

update the task proto, this is non-func change now. PiperOrigin-RevId: 842579755

Commit:e50b7f2
Author:Profiler Team
Committer:Copybara-Service

Added HBM Utilization metric to overview page and restructured Memory Bandwidth Utilization section to match FLOPS Utilization section. PiperOrigin-RevId: 840057194

Commit:8328bd9
Author:Profiler Team
Committer:Copybara-Service

Refactor EventTimeFractionAnalyzer to group results by chip. PiperOrigin-RevId: 839477708

Commit:e58a1b8
Author:Yin Zhang
Committer:Copybara-Service

Add dvfs time scale multiplier processing logic for roofline model. PiperOrigin-RevId: 836277274

Commit:1922ba3
Author:Profiler Team
Committer:Copybara-Service

Move event_time_fraction_analyzer proto to OSS PiperOrigin-RevId: 834859827

Commit:fc53a4f
Author:Yin Zhang
Committer:Copybara-Service

Removed the uncapp flops utilization UI controller, and always use the raw uncapped flops by default on the UI. PiperOrigin-RevId: 834150070

Commit:72e48ec
Author:Subham Soni
Committer:Copybara-Service

Add worker ID to responses and configure gRPC channels for retries and load balancing. PiperOrigin-RevId: 834135653

Commit:3da39d8
Author:Subham Soni
Committer:Copybara-Service

Add worker ID to responses and configure gRPC channels for retries and load balancing. PiperOrigin-RevId: 827391460

Commit:0ebecf4
Author:Subham Soni
Committer:Copybara-Service

Add worker ID to responses and configure gRPC channels for retries and load balancing. PiperOrigin-RevId: 827391460

Commit:c21b6f3
Author:Yin Zhang
Committer:Copybara-Service

Removed the uncapp flops utilization UI controller, and always use the raw uncapped flops by default on the UI. PiperOrigin-RevId: 805146752

Commit:d6051c1
Author:Yin Zhang
Committer:Copybara-Service

Add dvfs time scale multiplier processing logic for roofline model. PiperOrigin-RevId: 805986876

Commit:3fa0faa
Author:Yin Zhang
Committer:Copybara-Service

Add normalized_total_op_time_ps (on default dvfs) to opstats.device_op_metrics_db, together with the total_op_time_ps, it will give us a weighted average dvfs multiplier, which can be used to calculate dvfs normalized flops utilization and so on. - The aggregated dvfs multiplier will be shown in roofline page top as part of the device info, and used for device spec normalization. - The agg dvfs multiplier will also be used a an identifier if dvfs normalization control knobs should be displayed on UI or not. (show the knob with a valid multiplier, which is >0 and <1) PiperOrigin-RevId: 824384537

Commit:85fa4ce
Author:Yin Zhang
Committer:Copybara-Service

Add normalized_total_op_time_ps (on default dvfs) to opstats.device_op_metrics_db, together with the total_op_time_ps, it will give us a weighted average dvfs multiplier, which can be used to calculate dvfs normalized flops utilization and so on. - The aggregated dvfs multiplier will be shown in roofline page top as part of the device info, and used for device spec normalization. - The agg dvfs multiplier will also be used a an identifier if dvfs normalization control knobs should be displayed on UI or not. (show the knob with a valid multiplier, which is >0 and <1) PiperOrigin-RevId: 817431887

Commit:0d1b55e
Author:Yin Zhang
Committer:Copybara-Service

Add normalized_total_op_time_ps (on default dvfs) to opstats.device_op_metrics_db, together with the total_op_time_ps, it will give us a weighted average dvfs multiplier, which can be used to calculate dvfs normalized flops utilization and so on. - The aggregated dvfs multiplier will be shown in roofline page top as part of the device info, and used for device spec normalization. - The agg dvfs multiplier will also be used a an identifier if dvfs normalization control knobs should be displayed on UI or not. (show the knob with a valid multiplier, which is >0 and <1) PiperOrigin-RevId: 817431887

Commit:b073e99
Author:Yin Zhang
Committer:Copybara-Service

Add normalized_total_op_time_ps (on default dvfs) to opstats.device_op_metrics_db, together with the total_op_time_ps, it will give us a weighted average dvfs multiplier, which can be used to calculate dvfs normalized flops utilization and so on. - The aggregated dvfs multiplier will be shown in roofline page top as part of the device info, and used for device spec normalization. - The agg dvfs multiplier will also be used a an identifier if dvfs normalization control knobs should be displayed on UI or not. (show the knob with a valid multiplier, which is >0 and <1) PiperOrigin-RevId: 817431887

Commit:a5182ac
Author:Yin Zhang
Committer:Copybara-Service

Add normalized_total_op_time_ps (on default dvfs) to opstats.device_op_metrics_db, together with the total_op_time_ps, it will give us a weighted average dvfs multiplier, which can be used to calculate dvfs normalized flops utilization and so on. - The aggregated dvfs multiplier will be shown in roofline page top as part of the device info, and used for device spec normalization. - The agg dvfs multiplier will also be used a an identifier if dvfs normalization control knobs should be displayed on UI or not. (show the knob with a valid multiplier, which is >0 and <1) PiperOrigin-RevId: 817431887

Commit:f0bc3a0
Author:Yin Zhang
Committer:Copybara-Service

Add normalized_total_op_time_ps (on default dvfs) to opstats.device_op_metrics_db, together with the total_op_time_ps, it will give us a weighted average dvfs multiplier, which can be used to calculate dvfs normalized flops utilization and so on. - The aggregated dvfs multiplier will be shown in roofline page top as part of the device info, and used for device spec normalization. - The agg dvfs multiplier will also be used a an identifier if dvfs normalization control knobs should be displayed on UI or not. (show the knob with a valid multiplier, which is >0 and <1) PiperOrigin-RevId: 817431887

Commit:a154dd9
Author:Yin Zhang
Committer:Copybara-Service

Removed the uncapp flops utilization UI controller, and always use the raw uncapped flops by default on the UI. PiperOrigin-RevId: 805146752

Commit:df0a416
Author:Yin Zhang
Committer:Copybara-Service

Add dvfs time scale multiplier processing logic for roofline model. PiperOrigin-RevId: 805986876

Commit:54704f0
Author:Yin Zhang
Committer:Copybara-Service

Removed the uncapp flops utilization UI controller, and always use the raw uncapped flops by default on the UI. PiperOrigin-RevId: 805146752

Commit:4c370f6
Author:Yin Zhang
Committer:Copybara-Service

Add aggregated dvfs time scale multiplier to opstats.device_op_metrics_db - Note we assumes only core 0 has this info, so MetricsDb combiner only write it once from a src db that has a valid multiplier. - this aggregated dvfs factor will be shown in roofline page top as part of the device info, and used for device spec normalization. - it will also be used a an identifier if dvfs normalization control knobs should be displayed on UI or not. (show the knob with a valid multiplier, which is >0 and <1) PiperOrigin-RevId: 817431887

Commit:4d291b3
Author:Yin Zhang
Committer:Copybara-Service

No-op Changes. Add a `time_scale_multiplier` metadata in xplane and corresponding `normalized_time_ps` in OpMetrics proto. PiperOrigin-RevId: 814487519

Commit:a81f63e
Author:Charles Alaras
Committer:Copybara-Service

Allow `resource_id` to be 64 bits PiperOrigin-RevId: 811486066

Commit:06b9fb5
Author:Yin Zhang
Committer:Copybara-Service

No-op Changes. Add a `time_scale_multiplier` metadata in xplane and corresponding `normalized_time_ps` in OpMetrics proto. PiperOrigin-RevId: 794174377

Commit:77e7419
Author:Profiler Team
Committer:Copybara-Service

Fix typo in comment PiperOrigin-RevId: 809107812

Commit:1382786
Author:Profiler Team
Committer:Copybara-Service

Only use root nodes in computing the current source metrics. PiperOrigin-RevId: 808825812

Commit:43092e0
Author:Profiler Team
Committer:Copybara-Service

Only use root notes in computing the current source metrics. PiperOrigin-RevId: 808241860

Commit:0a943bf
Author:Profiler Team
Committer:Copybara-Service

Fix typo in comment PiperOrigin-RevId: 808204183

Commit:2a2ff2f
Author:Clive Verghese
Committer:Copybara-Service

Add support to extract custom_call metadata from hlo_ops PiperOrigin-RevId: 807906945

Commit:8f8487a
Author:Clive Verghese
Committer:Copybara-Service

Add support to extract custom_call metadata from hlo_ops PiperOrigin-RevId: 805939855

Commit:9b49225
Author:Yin Zhang
Committer:Copybara-Service

No-op Changes. Add a `time_scale_multiplier` metadata in xplane and corresponding `normalized_time_ps` in OpMetrics proto. PiperOrigin-RevId: 794174377

Commit:8d718d4
Author:Clive Verghese
Committer:Copybara-Service

Add support to extract custom_call metadata from hlo_ops PiperOrigin-RevId: 805939855

Commit:dc1d132
Author:Clive Verghese
Committer:Copybara-Service

Add support for Viewing custom call text in Graph Viewer. PiperOrigin-RevId: 805970858

Commit:db22907
Author:Profiler Team
Committer:Copybara-Service

Show the metrics. PiperOrigin-RevId: 783937622

Commit:bf0babd
Author:Subham Soni
Committer:Copybara-Service

Introduce gRPC worker service for distributed profile processing PiperOrigin-RevId: 799038615

Commit:9c8601f
Author:Yin Zhang
Committer:Copybara-Service

Add generic input pipeline op processing support and add traceme ops from xplane to op_stats.host_op_metrics_db PiperOrigin-RevId: 798351303

Commit:12f7261
Author:Yin Zhang
Committer:Copybara-Service

Add generic input pipeline op processing support and add traceme ops from xplane to op_stats.host_op_metrics_db PiperOrigin-RevId: 795264700