[CPU][Profiler] Group torch profiler tables by input shape when record_shapes is on

[Chinmay-Kulkarni-AMD]

Purpose

Group PyTorch profiler key_averages tables by input shape on the CPU platform when the user enables --profiler-config.torch_profiler_record_shapes=true.

Without this, CPU dumps collapse every call of an op into one row, which hides prefill vs decode and GEMM size differences. GPU dumps are unchanged: group_by_input_shape is only set when current_platform.is_cpu() and torch_profiler_record_shapes are both true. Recording shapes remains user-controlled; it is not forced on the CPU platform.

Test Plan

CPU throughput bench with torch profiler, shapes on:

export CUDA_VISIBLE_DEVICES=""
export VLLM_CPU_KVCACHE_SPACE=8

vllm bench throughput \
  --model Qwen/Qwen3.5-4B --language-model-only --dtype bfloat16 \
  --max-model-len 2048 --max-num-seqs 8 --num-prompts 16 \
  --random-input-len 16 --random-output-len 16 \
  --profile \
  --profiler-config "{\"profiler\": \"torch\", \"torch_profiler_dir\": \"torch-logs\", \"torch_profiler_with_stack\": false, \"torch_profiler_use_gzip\":false, \"torch_profiler_record_shapes\":true}"

Also confirm that "torch_profiler_record_shapes": false still prints an ungrouped table, and that a GPU run does not pass group_by_input_shape=True.

Test Result

CPU dump (torch_profiler_record_shapes=true) now has an Input Shapes column. The same op is split by GEMM / sequence shape (decode 8 vs prefill 128):

Name                                   Self CPU %   Self CPU    # of Calls   Input Shapes
vllm::cpu_gdn_attention_core             2.15%       77.060ms  720          [[8, 8192], [8, 32], [8, 32], [8, 32, 128], []]
vllm::cpu_gdn_attention_core             0.59%       21.323ms   48          [[128, 8192], [128, 32], [128, 32], [128, 32, 128], []]
_C::cpu_attention_with_kv_cache        0.25%        8.829ms  240          [[8, 16, 256], [409, 4, 640, 256], ...]

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[amukho]

Looks good to me...

[mergify]

Hi @Chinmay-Kulkarni-AMD, the pre-commit checks have failed. Please run:

uv pip install pre-commit>=4.5.1
pre-commit install
pre-commit run --all-files

Then, commit the changes and push to your branch.

For future commits, pre-commit will run automatically on changed files before each commit.

[bigPYJ1151]

Thanks @Chinmay-Kulkarni-AMD , LGTM :)
Please fix the pre-commit check.

[Chinmay-Kulkarni-AMD]

Thanks @Chinmay-Kulkarni-AMD , LGTM :) Please fix the pre-commit check.

Hello @bigPYJ1151 , the pre-commit/linter issues are now fixed.
I had missed one check. Apologies for the inconvenience.

[bigPYJ1151]

/ci run

[github-actions]

✅ Triggered Buildkite CI #88758 for commit 22bd55b9d567.