[executorch][cuda] Add CudaExecutor

[Gasoonjia]

Stack from ghstack (oldest at bottom):

• [executorch][muse] Move off-graph KV to the batching path and test batching in OSS CI #23375
• [executorch][muse] Add run_solo_batching #23374
• [executorch][muse] Add export_solo_batching.py #23373
• [executorch][cuda] Let CudaExecutor honor a program's prefill lower bound #23372
• [executorch][llm] Share the batching executors' session logic under extension/llm/batching/util #23371
• -> [executorch][cuda] Add CudaExecutor #23370
• [executorch][cuda] Add step_plan for the CUDA batching executor #23369
• [executorch][llm] Add Buck targets for extension/llm/batching #23368
• [executorch][cuda] Add CudaCellCache #23367
• [executorch][cuda] Lower off-graph KV in a cell layout #23366
• [executorch][cuda] Extract CudaKVPool from CudaSequenceKVCache #23365

The CUDA counterpart of ModuleExecutor, in backends/cuda/batching: implements
the neutral batching::Executor over a cell-layout off-graph KV cache, so the
neutral Runner and DecodeFirstScheduler drive a CUDA program unchanged.

The CUDA backend compiles static graphs, so where ModuleExecutor runs one
forward, CudaExecutor runs two methods and routes each slice of a batch:

• decode: one token, static. A CUDA graph is captured for it by default.
• prefill: dynamic from 2 tokens to its exported width W, eager.
  plan_slices (step_plan.h, header-only) cuts a batch into forwards of at most
  W tokens, in order; a lone tail token runs decode. Each slice declares its
  tokens to the cache, runs, and samples its selected rows.

create():

• reads the neutral metadata (context, vocab, geometry, logits mode, which
  must be selected since decode's selector is static at one row);
• validates both methods' inputs and outputs;
• reads get_offgraph_kv_max_cells and builds the batched-cell cache with
  the program's max_cells and W, which fix the step buffers' shapes;
• requires max_sessions * max_session_tokens to fit the pool;
• sets the process-wide CUDA backend options from CudaExecutorOptions:
  weight_sharing_across_methods (default on) and
  enable_cuda_graph_for_method=decode (default on). Both are overridable.

initialize() loads both methods with the cache's registry key. Session,
clone, sampling and build_step are copied from ModuleExecutor; a follow-up
diff moves the shared parts into extension/llm/batching/util.

Builds with Buck and CMake (cuda_batching, added from the root after
extension/llm/batching, installed to ExecuTorchTargets for downstream
runners).

Differential Revision: D123051481

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23370

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 6 New Failures, 3 Unclassified Failures

As of commit 2548fb5 with merge base c16dd0a ():

NEW FAILURES - The following jobs have failed:

• pull / unittest / linux / linux-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows
• pull / unittest / macos / macos-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows
• pull / unittest / windows / windows-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows
• pull / unittest-editable / linux / linux-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows
• pull / unittest-editable / macos / macos-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows
• pull / unittest-editable / windows / windows-job (gh)
  .ci/scripts/tests/test_wheel_test_modules.py::TestDroppedTestModules::test_ci_entry_points_still_match_the_workflows

UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:

• Test CUDA Builds / unittest-cuda / linux-job (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  backends/cuda/tests/test_offgraph_kv.py::OffGraphKVCellCompileTest::test_decode_and_prefill_from_two_tokens_lower
• Test CUDA Builds / unittest-cuda-runtime / linux-job (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  [ FAILED ] CudaKVCacheTest.CacheOutlivesTheStreamItLastSteppedOn
• Test ROCm AOTI / unittest-rocm-gfx950-rocm7.2 / linux-job (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  backends/cuda/tests/test_offgraph_kv.py::OffGraphKVCellCompileTest::test_decode_and_prefill_from_two_tokens_lower

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.