Feat/gw rename directories#750
Draft
gabeweisz wants to merge 8 commits into
Draft
Conversation
Heuristic Findings Ordering
## Summary Replace the heuristic `_extract_launcher_path()` with the pre-computed `entry_point` column from `unified_perf_summary`, and switch `module_chain`/`call_chain` extraction to parse the new `call_stack_full` format. This PR accompanies #743. All changes in #743 are in this PR. --- ## Problem The agent's `_extract_launcher_path()` re-derived the dispatch entry point at analysis time by walking the old `" => "`-delimited `call_stack` string from `unified_perf_callstacks.csv`. Now that `unified_perf_callstacks.csv` has been replaced by two columns `entry_point` and `call_stack_full` in `unified_perf_summary`, the agent can use `entry_point` for the `Kernel Path` column in the tables. `extract_module_chain` and `extract_call_chain` can use `call_stack_full`. ## Changes **`orchestrator_prepare.py`** — Removed the `unified_perf_callstacks.csv` join block. `entry_point` and `call_stack_full` are now columns in `unified_perf_summary.csv` and flow directly to per-category CSVs. **`analysis_utils.py`**: - `build_operation_metrics()`: reads `row["entry_point"]` as primary `launcher_path`. When `entry_point` is `"Not found"`, falls back to module chain (`"LlamaModel > LlamaDecoderLayer > LlamaMLP"`), then library name (`"Triton (vendor)"`), then `"Not found"`. Reads `row["call_stack_full"]` for `module_chain`/`call_chain`. - Added `_parse_call_stack()`: parses `call_stack_full` (Python list repr) via `ast.literal_eval`, keeping only top-level string frames and discarding nested sublists. The sublists represent diverging kernel tails in multi-kernel ops — they contain GPU kernel names that the agent doesn't use in prose context. - Deleted `_extract_launcher_path()`, `_extract_backward_context()`, and `_SKIP_PY_PATTERNS`. --------- Co-authored-by: gabeweisz <162640284+gabeweisz@users.noreply.github.com>
Add kernel name information to TraceLens Agent output tables. Kernel Name comes from `call_stack_full` column in `unified_perf_summary`. If there is only one kernel called by the CPU op, the column looks like: | Operation | Args | Kernel Path | Kernel Name | Time (ms) | %E2E | Count | FLOPS/Byte | Efficiency | Bound | |-----------|------|-------------|-------------|-----------|------|-------|------------|------------|-------| | aten::addmm | — | — | nvjet_tst_64x64_64x16_2x1_2cta_v_bz_bias_TNT | — | — | — | — | — | — | | aten::addmm | — | — | nvjet_tst_64x16_64x16_2x4_2cta_h_bz_bias_TNT | — | — | — | — | — | — | | aten::mm | — | — | nvjet_tst_192x8_64x8_4x1_v_bz_TNT | — | — | — | — | — | — | If there are multiple kernels it looks like: | Operation | Args | Kernel Path | Kernel Name | Time (ms) | %E2E | Count | FLOPS/Byte | Efficiency | Bound | |-----------|------|-------------|-------------|-----------|------|-------|------------|------------|-------| | vllm::moe_forward | — | — | Kernel 1: void moe::dev::routing::routingRenormalize::routingIndicesHistogramScoresKe...<br>Kernel 2: void moe::dev::routing::routingIndicesOffsetsKernel<moe::dev::routing::rout...<br>Kernel 3: void tensorrt_llm::kernels::quantize_with_block_size<(tensorrt_llm::BlockSc...<br>Kernel 4: void moe::dev::routing::routingIndicesHistogramKernel<moe::dev::routing::ro...<br>Kernel 5: void moe::dev::finalize::finalizeKernelVecLoad<moe::dev::finalize::KernelPa...<br>Kernel 6: bmm_Bfloat16_MxE2m1MxE4m3_Fp32_t128x16x256u2_s3_et128x16_m128x16x32_cga1x1x...<br>Kernel 7: bmm_MxE4m3_MxE2m1MxE4m3_Fp32_t128x16x256u2_s3_et128x16_m128x16x32_cga1x1x1_... | — | — | — | — | — | — | The kernel name is truncated to include the first 75 characters (better kernel name shortening strategy can be explored later).
Co-authored-by: Cursor <cursoragent@cursor.com>
Move eval-post-processing, workflow-llm-eval, and quality-llm-eval into skills/<name>/SKILL.md and reference.md, remove .cursor/skills, and update README and EVAL_RUBRICS paths. Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Follow the skills spec here
Note that we fudged a bit with the sub-agent directory, but it worked in cursor
HOLD UNTIL AFTER AI DAY. Need to run evals too.
Ran test with
agent --print --force --trust --output-format stream-json "set TL_EXTENSION=TraceLens_internal for all operations and commands. Use the full Tracelens Agent Analysis using venv_tracelens_refactor to analyze ~/torch_trace_output and put the results in ~/torch_trace_output_refactor"analysis.md