Regenerate example traces for autotune session split (#440) by FindHao · Pull Request #440 · meta-pytorch/tritonparse

FindHao · 2026-09-23T20:09:50Z

Summary:

Regenerate the triton and inductor example traces with tritonparse.tools.generate_examples (triton 3.8.0, torch 2.15 dev+cu130, H100) on top of the parent diff's autotune session split, so the shipped website examples and tests/example_output fixtures reflect per-key sub-sessions.

What changed in the new traces:

The triton example's call-site matmul session now splits into two per-key sub-sessions (M=N=16 vs M=32,N=32): 1 -> 2 autotune_analysis events with <site>:<key> session ids.
matmul launch count 1078 -> 1041 (autotune benchmark jitter; refreshed to match the review-fixed parent, whose sub-session ids now carry full 16-char key signatures).

Test-only update alongside the regen:

test_list_kernels_multiple: matmul launches 1078 -> 1041 (the comment already says to refresh on regen).

Differential Revision: D121418449

meta-codesync · 2026-09-23T20:10:18Z

@FindHao has exported this pull request. If you are a Meta employee, you can view the originating Diff in D121418449.

Copilot

Copilot review overview

🟡 Changes recommended

Unresolved critical session-partitioning and moderate fixture/assertion consistency issues remain.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1 · 1 · 1

Open (3)

Use autotune key when launch signature is unavailable · New
Synchronize fixture launch count with assertion · New
Fix grammatical typo in test comment · New

What changed in this PR

Regenerates autotune example traces and updates parsing to split call-site sessions by autotune key.

Changes:

Captures per-launch occurrences and autotune callbacks.
Produces per-key analysis sessions.
Updates launch-count expectations and adds regression coverage.

File	Summary
`tritonparse/parse/trace_processor.py`	Captures launch and autotune callback metadata.
`tritonparse/parse/event_diff.py`	Splits sessions by key; needs an `autotune_key` fallback when launch signatures are unavailable.
`tests/cpu/test_kernel_query.py`	Assertion records 1041 launches while the regenerated fixture is stated as 1055.
`tests/cpu/test_autotune_session_split.py`	Adds session-splitting tests; comment contains “sames scalars” typo.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

+    config_args = _find_config_args(group_infos)
+    signature_by_group = {
+        group_hash: _launch_key_signature(extracted, config_args)
+        for group_hash, _comp_hash, extracted in group_infos


        kernel_dict = {k.name: k for k in result}
        self.assertEqual(kernel_dict["fused_op_kernel"].total_launches, 4)
-        self.assertEqual(kernel_dict["matmul_kernel"].total_launches, 1078)
+        self.assertEqual(kernel_dict["matmul_kernel"].total_launches, 1041)


+        self.assertEqual(set(by_winner), {"hash_a", "hash_b"})
+
+        # Each key round keeps single-valued args: M/N are constant
+        # within the round (sames scalars), only the true config param


Copilot

Copilot review overview

🟡 Changes recommended

A critical compilation-loss issue and multiple key-partitioning correctness issues remain unresolved.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 2 · 2 · 1

Open (5)

Preserve compilations without traced benchmark launches · New
Use autotune key when launch signature is unavailable
Restrict boolean parsing to valid true/false spellings · New
Synchronize fixture launch count with assertion
Fix grammatical typo in test comment

Summary: Pull Request resolved: #440 Regenerate the `triton` and `inductor` example traces with `tritonparse.tools.generate_examples` (triton 3.8.0, torch 2.15 dev+cu130, H100) on top of the parent diff's autotune session split, so the shipped website examples and `tests/example_output` fixtures reflect per-key sub-sessions. What changed in the new traces: - The triton example's call-site matmul session now splits into two per-key sub-sessions (M=N=16 vs M=32,N=32): 1 -> 2 `autotune_analysis` events with `<site>:<key>` session ids. - matmul launch count 1078 -> 1041 (autotune benchmark jitter; refreshed to match the review-fixed parent, whose sub-session ids now carry full 16-char key signatures). Test-only update alongside the regen: - `test_list_kernels_multiple`: matmul launches 1078 -> 1041 (the comment already says to refresh on regen). Differential Revision: D121418449

Copilot

Copilot review overview

🟡 Changes recommended

Autotune sessions and callback results can be partitioned or matched incorrectly.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 2 · 1 · 1

Open (4)

Preserve key arguments when deriving config fields · New
Use autotune key when launch signature is unavailable
Synchronize fixture launch count with assertion
Fix grammatical typo in test comment

Resolved since last review (2)

Preserve compilations without traced benchmark launches
Restrict boolean parsing to valid true/false spellings

+    for arg_name, per_hash in values_by_arg.items():
+        if len(per_hash) <= 1:
+            continue
+        if not all(len(values) == 1 for values in per_hash.values()):
+            continue
+        # The values must actually disagree across hashes; a constant arg
+        # is not a config param even though it trivially has "one value
+        # per hash".
+        if len(set().union(*per_hash.values())) > 1:
+            config_args.add(arg_name)


Summary: Pull Request resolved: #440 Regenerate the `triton` and `inductor` example traces with `tritonparse.tools.generate_examples` (triton 3.8.0, torch 2.15 dev+cu130, H100) on top of the parent diff's autotune session split, so the shipped website examples and `tests/example_output` fixtures reflect per-key sub-sessions. What changed in the new traces: - The triton example's call-site matmul session now splits into two per-key sub-sessions (M=N=16 vs M=32,N=32): 1 -> 2 `autotune_analysis` events with `<site>:<key>` session ids. - matmul launch count 1078 -> 1041 (autotune benchmark jitter; refreshed to match the review-fixed parent, whose sub-session ids now carry full 16-char key signatures). Test-only update alongside the regen: - `test_list_kernels_multiple`: matmul launches 1078 -> 1041 (the comment already says to refresh on regen). Differential Revision: D121418449

Copilot

Copilot review overview

🟡 Changes recommended

Critical launch-loss and callback-attribution defects must be resolved before approval.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 4 · 2 · 1

Open (7)

Note-only launches are omitted from emitted analysis · New
Partial config matches misattribute callback results · New
Preserve key arguments when deriving config fields
Use autotune key when launch signature is unavailable
Nested descriptors lose key-dependent fields · New
Synchronize fixture launch count with assertion
Fix grammatical typo in test comment

+    occurrences = [
+        record
+        for record in occurrences
+        if record.get("launch_group_hash") not in skipped_note_groups
+    ]


+        for name, expected in pairs.items():
+            if name == "num_warps" and winner_warps_base is not None:
+                actual = winner_warps_base
+            elif name in extracted and not _is_tensor_like_value(extracted[name]):
+                actual = _extract_scalar_value(extracted[name])
+            elif name in compilation_metadata:
+                actual = compilation_metadata[name]
+            else:
+                continue
+            compared += 1
+            if _config_value_matches(expected, actual):
+                matched += 1
+        if compared > 0 and matched == compared:
+            return index


+        elif isinstance(arg_val, dict):
+            # Descriptors and foreign wrappers: keep identity metadata
+            # but drop volatile nested data_ptrs for stability.
+            payload[arg_name] = _drop_data_ptrs(arg_val)


Summary: Autotune session ids hash only the user call stack, so repeated Autotuner.run() invocations at the same call site with different keys (e.g. a loop over M=N=16 then M=N=32, as in the regenerated example trace) collapse into one session. Per-config args then become multi-valued distributions that the viewer renders as an ellipsis. Expand each call-site session into per-key sub-sessions at analysis time: launches are grouped by a signature over runtime args (config params excluded, tensor data_ptrs excluded), compilations attach by referenced hash, and each sub-session resolves its own winner and AutotuneListener result (matched via best_config). Unsplit sessions keep their plain session id and behave exactly as before. Ingest records per-launch occurrence mapping and keeps autotune results as a list (previously last-wins); the coarse last-wins winner map is replaced by per-sub-session resolution from occurrence records. Differential Revision: D121412109

Summary: Pull Request resolved: #440 Regenerate the `triton` and `inductor` example traces with `tritonparse.tools.generate_examples` (triton 3.8.0, torch 2.15 dev+cu130, H100) on top of the parent diff's autotune session split, so the shipped website examples and `tests/example_output` fixtures reflect per-key sub-sessions. What changed in the new traces: - The triton example's call-site matmul session now splits into two per-key sub-sessions (M=N=16 vs M=32,N=32): 1 -> 2 `autotune_analysis` events with `<site>:<key>` session ids. - matmul launch count 1078 -> 1041 (autotune benchmark jitter; refreshed to match the review-fixed parent, whose sub-session ids now carry full 16-char key signatures). Test-only update alongside the regen: - `test_list_kernels_multiple`: matmul launches 1078 -> 1041 (the comment already says to refresh on regen). Differential Revision: D121418449

meta-codesync · 2026-09-24T20:13:20Z

This pull request has been merged in 5edb065.

Copilot AI lite review requested due to automatic review settings September 23, 2026 20:09

meta-cla Bot added the CLA Signed This label is managed by the Meta Open Source bot. label Sep 23, 2026

meta-codesync Bot added the meta-exported label Sep 23, 2026

Copilot started reviewing on behalf of FindHao September 23, 2026 20:10 View session

Copilot AI reviewed Sep 23, 2026

View reviewed changes

meta-codesync Bot force-pushed the export-D121418449 branch from 999478d to 0a2bb0a Compare September 23, 2026 23:53

FindHao requested a balanced review from Copilot September 23, 2026 23:53

Copilot started reviewing on behalf of FindHao September 23, 2026 23:54 View session

Copilot AI reviewed Sep 23, 2026

View reviewed changes

Comment thread tritonparse/parse/event_diff.py

Comment thread tritonparse/parse/event_diff.py Outdated

FindHao requested a balanced review from Copilot September 24, 2026 02:34

Copilot started reviewing on behalf of FindHao September 24, 2026 02:34 View session

meta-codesync Bot force-pushed the export-D121418449 branch from 0a2bb0a to 21a37cc Compare September 24, 2026 02:36

meta-codesync Bot changed the title ~~Regenerate example traces for autotune session split~~ Regenerate example traces for autotune session split (#440) Sep 24, 2026

Copilot AI reviewed Sep 24, 2026

View reviewed changes

meta-codesync Bot force-pushed the export-D121418449 branch from 21a37cc to 4b7cd50 Compare September 24, 2026 03:31

FindHao requested a balanced review from Copilot September 24, 2026 03:33

Copilot started reviewing on behalf of FindHao September 24, 2026 03:33 View session

Copilot AI reviewed Sep 24, 2026

View reviewed changes

FindHao added 2 commits September 24, 2026 08:56

meta-codesync Bot force-pushed the export-D121418449 branch from 4b7cd50 to 57d8dba Compare September 24, 2026 16:00

meta-codesync Bot closed this in 5edb065 Sep 24, 2026

meta-codesync Bot added the Merged label Sep 24, 2026

Navigation Menu

Repository navigation

Regenerate example traces for autotune session split (#440) - #440

Regenerate example traces for autotune session split (#440)#440
FindHao wants to merge 2 commits into
mainfrom
export-D121418449

FindHao commented Sep 23, 2026 •

edited by meta-codesync Bot

Loading

Uh oh!

meta-codesync Bot commented Sep 23, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

meta-codesync Bot commented Sep 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

FindHao commented Sep 23, 2026 • edited by meta-codesync Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

meta-codesync Bot commented Sep 23, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Copilot review overview

🟡 Changes recommended

Uh oh!

meta-codesync Bot commented Sep 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

FindHao commented Sep 23, 2026 •

edited by meta-codesync Bot

Loading